49102.2005.017-2006.128 - ICT IMPACTS - FINAL REPORT V2

FINAL REPORT

Information Society: ICT impact assessment by

linking data from different sources

Grant Agreement Number – 49102.2005.017-2006.128

August 2008

Authors

Aarno Airaksinen Statistics Finland Andrea de Panizza ISTAT, Italy and European Commission Eric Bartelsman (Professor) Vrije Universiteit Amsterdam, Tinbergen Institute and IZA Eva Hagsten Statistics Sweden George van Leeuwen Statistics Netherlands Mark Franklin Office for National Statistics, UK Mika Maliranta Statistics Finland & The Research Institute of the Finnish Economy (ETLA) Patricia Kotnik University of Ljubljana, Faculty of Economics Peter Stam Office for National Statistics, UK Petri Rouvinen The Research Institute of the Finnish Economy (ETLA) Shikeb Farooqui Office for National Statistics, UK Simon Quantin INSEE, France Stefan Svanberg Statistics Sweden Tony Clayton Office for National Statistics, UK Yoann Barbesol INSEE, France

With thanks and acknowledgement to

Allessandra Nurra ISTAT, Italy Brian Ring CSO Ireland Chiara Criscuolo London School of Economics, Centre for Economic Performance Eugene van der Pijll Statistics Netherlands Gerolamo Giungato ISTAT, Italy Hans-Olof Hagen Statistics Sweden Jonathan Haskel (Professor) Queen Mary University of London Joseph Robjohns Office for National Statistics, UK Martin Lundo Statistics Denmark Martin Mana Czech Statistical Office Natalia Cherevichenko Statistics Denmark Nina Djahangiri Statistics Austria Ole-Petter Kordahl Statistics Norway Oliver Bauer Federal Statistics Office, Germany Ritchie McMahon CSO Ireland Stefan Bender Institute for Employment Research Vaclav Kosina Czech Statistical Office We would like to thank all those who have supported and contributed to this project, not least the Information Society team at Eurostat. However, any errors or omissions are the responsibility of the authors. Disclaimer All data in this report comply with statistical disclosure measures and standards throughout all project member countries.

Furthermore, data refer to experimental and provisional research datasets. They should not be treated as - or compared to - any official national statistics.

Eurostat Agreement No. 49102.2005.017-2006.128 Contents

Chapter Page 1. Recommendations, Summary Results and Conclusions 1

Tony Clayton

2. Introduction and Background 20 Tony Clayton

3. Methods and Data Sources 29 Mark Franklin

4. Describing the Data 40 Peter Stam

5. Properties of Linked Data Evidence from the ICT Impacts Project 68 Eric Bartelsman

6. Productivity and Core ICT Metrics at Firm Level 94 Mark Franklin Shikeb Farooqui

7. ICT Characteristics of Fast Growing Firms 120 Simon Quantin Yoann Barbesol

8. Employment, Skills and Information Technology 134 Eva Hagsten Patricia Kotnik

9. ICT Business Integration 149 Mark Franklin Tony Clayton

10. ICT Investment and Productivity 163 George van Leeuwen Shikeb Farooqui

11. Offshoring 190 Andrea de Panizza Eva Hagsten Patricia Kotnik Simon Quantin Stefan Svanberg Yoann Barbesol

12. ICT, Innovation and Productivity 222 George van Leeuwen Shikeb Farooqui

13. IT Outsourcing in Finnish Business 240 Aarno Airaksinen Mika Maliranta Petri Rouvinen

14. From Micro to Macro 255 Eric Bartelsman

References 272 Appendix I – Project User guide 280

Peter Stam

Eurostat Agreement No. 49102.2005.017-2006.128 Chapter 1

Chapter 1

Summary Results, Conclusions and Recommendations Tony Clayton UK Office for National Statistics This chapter summarises the main analytical results of the project, uses them to draw conclusions from the new findings on Information and Communications Technology (ICT) impacts, alongside prior research, and sets out recommendations for action by Eurostat and NSIs. Where possible the full analytical evidence in subsequent chapters is referenced.

Section 1.1 Summary of Results This section brings together the analytical findings from the project. The range of work covers all the milestones set for the original contract (see Appendix A), and the development of new methodology. The results here derive from the following different types of analytical work:

• at the most basic level, one-off studies of firm level productivity impact where only one country has the data to perform a specific piece of analysis (for example work from Finland on ICT outsourcing).

• several examples of groups of countries collaborating on micro data analysis for topics where all have similar (but not necessarily identical) firm level data which enable a common analytical framework to be used and compared (eg Netherlands and UK on ICT investment, Sweden, France and Italy on offshoring, Sweden, Netherlands and UK on innovation).

• an encouraging range of firm level analysis using common metrics and common analytical code with identical data sources, either carried out by local researchers across most countries direct from local datasets, or using the data created for the project and centrally written code to run identical regression analysis, for all countries in the project except Denmark and Slovenia.

• construction of an extensive ‘metadata warehouse’, which is used to weight and aggregate ICT use, structural business and business register data from surveys in all 13 countries in as comparable a way as is possible, producing distributed microdata datasets (DMD); the aggregation process operates to produce estimates of complex indicators (constructed from more than one variable from a survey) as well as indicators which depend on intersections between surveys; this metadata system can also be used to generate datasets on a highly comparable basis for firm level regression analysis within countries.

• industry / country level analysis of ICT impacts, using the large (and still under-explored) dataset produced by the distributed microdata (DMD) analysis system, where we have a very high level of confidence in the comparability of indicators, and on the ability to draw reliable comparisons between industries /countries and over time.

The summary below presents results in terms of the key topics tackled in the study, and

draws on evidence from both firm level and DMD analysis as appropriate.

1


1.1.1 Results from firm level analysis

1.1.1.1 Coverage (Chapters 4, 5 and 6) All 13 countries participating in the study (including Denmark which had to withdraw in the late stages due to lack of personnel to undertake analysis and check results) have succeeded in producing regression and / or correlation results from firm level data, either individually or using the DMD analysis methodology developed in the project – and in most cases both.

Project participants have completed analysis to show results relating ICT use at firm level to labour productivity from 11 of the 13 participating countries, on common metrics using an exactly comparable method, and using the common metadata to define and link variables.

We know we can get results from the remaining countries with minor additional resources; the missing micro-data analysis is due to analytical resource constraints, or to limitations under which access to data was available for this particular project.

Analysis of the properties of linked datasets in the project, using methodologies developed in earlier studies, shows that sample reweighting, using metadata and methods included in the project, is capable of dealing with most issues of ‘representativeness’ of data. This breaks down in cases where overlap between datasets is inadequate, and we have not advocated modelling in such cases.

Linking of datasets in many countries, using sampling designs currently in use, leaves the overlap between ICT surveys and firm performance surveys heavily biased towards larger firms. This affects both firm level analysis and the DMD analysis. For impact conclusions adequately to reflect small firms, sampling strategies would need to change.

1.1.1.2 Common firm level analysis across all NSIs (Chapter 6) The core ICT use metrics used in the project (computer use, e-sales, e-purchases, fast internet enabled or using employees) show reasonably consistent, positive, labour productivity effects at firm level across manufacturing industries in all countries in the project, beyond the six which have been covered by earlier studies. This suggests that productivity impacts related to use of ICT in manufacturing are now relatively well established and transferable across countries within the EU.

The same core ICT use metrics have much more varied relationships with labour productivity across services at firm level in different countries; for the UK, France, Nordic countries and Netherlands, positive correlations seen in prior studies, and reported in early work from this project are confirmed, in other countries participating in the project, productivity effects are insignificant or even, in one or two cases, negative.

There seems to be at least some correlation between the countries (Nordic states, Netherlands, UK, France) where ICT use by firms is relatively more intensive and communications infrastructure is strong, or where there is greater market flexibility / dynamism, and the strength of the statistical relationship between ICT use and firm level productivity in services. These differences in impact for services could be explained by a number of factors, including:

• differences in competitive conditions in national services markets, and / or

• productivity gains requiring ‘critical mass’ in networks and ICT use, and / or

• measurement difficulties in services which are better tackled in some states.

The common analysis shows limited evidence for productivity impact of e-commerce as a variable on its own, and clear positive relationships between productivity and wages (used

2


later in the analysis programme as an imperfect indicator of skills), but little or no separate impact of firm age.

1.1.1.3 Country based firm level analysis across groups of NSIs

1.1.1.3.1 Impact of ICT use metrics compared to IT capital measures (Appendix to Chapter 6)

For Netherlands and UK, data are available on firm level IT capital - hardware and software for the UK, hardware only for Netherlands. This makes it possible to test the impact of measures of ICT use over and above those of IT capital services in productivity models. The results show that impacts are differentiated by firm type:

• in manufacturing firms intensity of e-procurement shows the strongest relationship to productivity advantage;

• in distribution service firms the largest impact on productivity is related to the intensity of use of e-commerce for selling;

• in other, mainly business and financial, service industries the strongest relationship with productivity over the first half of the 2000s comes from the proportion of workers with access to high speed internet;

• across all three industry types, IT capital (including software) is positively related to productivity levels in the UK, and with a much larger impact in differentiated (i.e. non distribution) services;

• across all three types in the Netherlands analysis IT capital (excluding software) is insignificant.

These differences suggest clear limits to an analytical approach which treats ICT as a ‘general purpose technology’. Its impacts in different industries suggest different processes at work, which need to be understood in context, through different practical effects of information technology (IT) and communications technology (CT). Impact analysis also needs to take account of complementary influences of ICT in combination with other factors such as skills and organisational change.

1.1.1.3.2 Fast growing firms (Chapter 7)

Analysis by INSEE shows that in France, the use of ICT does not affect the probability that a firm will show high growth characteristics (20% + employment growth over four successive years, not including firms with less than 30 employees).

However, French microdata does show that intensity of ICT use for business purposes (% e-purchases or % e-sales, % employees with high speed internet connection) is a positive influence on the – small - probability that a particular firm will achieve high growth.

Validating these French results using the cross-country (DMD) datasets shows that high growth manufacturing firms are more intensive ICT users in a majority (but not all) of the countries covered, with intensity of electronic transactions appearing to provide the best indicator.

Across the DMD dataset, fast growth firms in services in about half the countries are more intensive ICT users for business purposes – but not (paradoxically given results in the previous section) in the UK or France.

1.1.1.3.3 Employment, Skills and ICT Skills (Chapter 8)

Results from DMD analysis across Austria, Czech Republic, France, Germany, Great Britain, Italy, Netherlands, Norway and Sweden, for 26 sectors of manufacturing and services, show no clear relationship, at industry level, between ICT use metrics (internet and fast internet

3


use by employees) and employment growth. Taken together with the results for fast growth firms above, this suggests that more intensive ICT use may increase the chances of growth at individual firm level, but this may be at the expense of competitors if overall industry effects are insignificant.

Three countries within the project, Finland, Sweden and Norway, have ‘real skills’ data available at firm level, derived by linking employer and employee data. In all three there are strong, significant and simultaneous correlations between labour productivity and the proportion of employees with ICT skills, as well as those with other (non ICT) higher education levels. For both types of skill measures the size of productivity impact make a strong case for wider collection of this type of data across other countries.

In Finland and Sweden similarly strong relationships exist between Total Factor Productivity (TFP) and employee skills (both IT and non-IT), and these relationships are significant alongside the ‘fast internet enabled employees’ measure mentioned above. For the Norwegian analysis fast internet enabled employees appear insignificant in regression analysis together with non-ICT skills, but, paradoxically, ICT skills remain highly significant. General skills appear to have greater impact in TFP analysis, but ICT skills show up as more significant in labour productivity analysis.

In all three countries it is possible to test the complementarity of skills and ICT intensity by adding an interaction term (% skilled employees x % fast internet enabled employees) and only in Sweden does this show up as a significant contributor.

The productivity effects of skills show up more strongly in service firms than in manufacturing (but this may be because there is more heterogeneity across service firms).

From analysis across all three countries it is clear that wages have a stronger correlation with productivity than do real measures of skills (although this is partly an arithmetic effect because employee compensation is part of value added). The analysis shows that wages have strong limitations as a direct proxy for skills in productivity analysis, without risk of understating other impacts. However, in analyses where skills data are not available, a proxy based on wages may be useful as a check against overstating ICT impacts due to correlation between ICT use and skills.

1.1.1.3.4 Organisation / integration of e-business links, and ICT outsourcing (Chapters 9 and 13)

Analysis led by UK, Netherlands and Sweden has been undertaken using measures of ICT business process integration to test methods of combining the metrics in the Eurostat model ICT use survey in ways which relate effectively to productivity impact.

Swedish analysis, based on a hierarchical specification of advance in business process sophistication starting with any form of external link working up to use of e-commerce, internet selling and links with suppliers / customers, and also looking at specific types of links, shows that the range of indicators linked to productivity, and the magnitude of correlation has grown through to 2004, and the evidence in favour of productivity impacts is growing. However the exact form of correlation, and the evidence for channels through which productivity is impacted, is changing from year to year.

UK results suggest that the productivity effects of linkages depend on the business type, with manufacturing firms showing stronger correlation coefficients between TFP and the incidence of electronic links to suppliers (associated with supply chain management) and service firms showing stronger productivity effects associated with links to customers.

Graphical evidence from DMD analysis across Slovenia, Italy, Netherlands, UK, Czech Republic, Finland, Sweden, Austria, France and Norway shows, for most (but not all) countries, a positive relationship between position in national productivity ranking, by quartile

4


and %e-procurement in manufacturing, and for half the countries a similar relationship between productivity ranking and %e-sales in retail and distribution.

Regression analysis using firm-level data from UK, Netherlands, Sweden, France, Czech Republic, and Austria suggests, similarly, that the productivity relationships are ‘better behaved’ for manufacturing, and that elsewhere there are signs of positive relationships, but that a hierarchical model is not the best approach. Regressions also show that external links have more explanatory power than links between processes within firms – suggesting that impacts which occur through market dynamics are more important that efficiency gains through process coordination. The latest ICT use survey (2007/8) will provide better, more consistent data to explore this.

Data for Finland, which alone among the EU countries tackles organisational issues associated with IT mobility and IT services outsourcing, shows substantial and significant productivity benefits associated with:

• use of mobile access to ICT by workers (suggesting gains from more flexible work patterns);

• use of outsourced IT services (suggesting gains from specialisation compared to ‘Do it yourself’ approaches).

1.1.1.3.5 IT Investment and ICT use (Chapter 10)

There is evidence from in depth analysis from Netherlands and UK, the two countries which have firm level IT investment data (hardware only for Netherlands, hardware and purchased software for UK) that fast internet connected employees is (over the period 2000 - 2005) a good predictor of cumulative IT investment.

There is also evidence from both countries that the productivity effects of high speed communication used by workers are additional, over and above effects of measured IT investment, suggesting that fast internet use by employees may capture unmeasured (own account) software, knowledge management by employees, and more open, flexible methods of working.

In both countries’ data there is evidence that the relationship between ICT investment and ICT use metrics differs across industries and countries, and that additional indicators (eg e-sales, and measures of business IT integration) may be significant; however the directional effects of the available ICT use metrics as indicators of IT investment are similar.

Given the difficulty (and cost) of collecting IT investment data at firm level experienced by most countries which have done it, the continued development of ICT use surveys as proxies for investment, and as indicators of IT impacts, is justified by this analysis.

1.1.1.1.6 Offshoring (Chapter 11)

For France, Sweden and Italy analysis has been conducted to explore the interaction between IT use and offshoring, in order to test indications from earlier studies, and in the wider literature, that productivity advantages associated with globalised supply chains are reinforced by ICT use. Each of the studies has used a methodology which controls at firm level for ‘IT maturity’ by using a composite indicator from the ICT use survey. In all three the analysis is affected by the fact that offshoring firms are normally at the upper end of enterprise size, by employment.

The results for each country depend on data available, but in one study (for France, where it is possible to test the relationship between labour productivity and an interaction between proportion of offshored intermediate goods and the use of e-procurement) suggests no cumulative effect.

5


The French results suggest that offshored intermediate goods or services are associated with higher productivity levels, in some cases, and that for manufacturing firms there are separate, significant, effects from importing intermediates from high wage economies (large) compared to importing from low wage economies (small).

French data also suggests that the observed productivity effects related to ‘IT maturity’ do not operate directly through mechanisms associated with offshoring.

Swedish data show similar differentiated productivity effects of offshoring to high skill versus low skill countries, but the effects are visible for both manufacturing and service firms. However the offshoring effects are reduced in intensity and in significance by including IT maturity in the analysis.

Gains from offshoring by relatively high skilled Swedish firms (especially IT skills) seem be stronger where offshoring is to lower wage economies, suggesting gains from specialisation. The pattern of impacts observed for other combinations of industry / skill intensity / import source is complex.

For Italy productivity effects of offshoring, after allowing for IT maturity, are sector dependent. For ‘traditional’ and ‘scale intensive’ industries the gains are significant by offshoring to lower labour cost sources; for specialist or science based firms, sourcing from high income countries is associated with higher productivity.

One result from this analysis is that it is not yet possible to identify a cumulative effect of ICT use and offshoring of goods on firm productivity. If it were possible to extend the analysis in services the conclusion might differ. Difficulties in linking data are an important limitation to this analysis – because the sampling strategies of most NSIs make it unlikely that a firm will be included in trade, ICT use and structural business surveys in the same year.

1.1.1.1.7 Innovation (using data from the Community Innovation Survey) (Chapter 12)

Similar survey overlap limitations exist in analysing relationships between ICT, innovation and productivity. However sufficient progress has been made by Sweden, Netherlands and the UK to draw some relatively strong conclusions on the role of ICT in innovation, and the mechanism through which much of the productivity gain associated with ICT may be achieved.

UK analysis linking ICT use surveys to questions in the Community Innovation Survey on sources of innovation shows a strong link between the use of high speed internet connections by employees within firms (in the ICT use survey) and the ability to innovate using ideas from outside the firm, and outside the customer / supplier chain. This suggests a link between fast internet network use and the ability of firms to acquire and manage knowledge in the innovation process, to develop higher levels of new goods or service sales, or more use of new processes.

Evidence from Sweden and Netherlands, where more extensive analysis is possible, suggests that ICT use – reflected in the proportion of fast internet linked employees and levels of e-commerce – is related to the intensity with which firms produce and sell new products / services. This also is likely to reflect network effects on knowledge management, on the effectiveness with which firms are able to convert knowledge into new products and services, and on the speed with which they are able to commercialise them out into the market. The role of e-commerce is highlighted in Netherlands analysis as evidence that the marketing benefits of e-commerce for innovation (which analysts have sought unsuccessfully up to now) may finally be visible.

Analysis across all the participating countries using DMD shows that in industries which have relatively high levels of ICT use on the core metrics, there also tend to be higher absolute amounts of market share change (or ‘churn’). This is consistent with the view that

6


ICT intensive industries in Europe show the same tendency seen in the US by Brynjolfsson et al, for successful firms to be better able, and quicker, to replicate new market share winning innovations across production and distribution networks.

From Sweden and Netherlands there is initial evidence, from datasets restricted by the limits of overlap between production, ICT and innovation surveys, in regressions in which both effects are considered simultaneously, that productivity effects of ICT use are associated more strongly through the ‘indirect innovation’ effect (percent new products / services) than through ICT use measures directly. The Swedish analysis tests the relative strength of direct and indirect productivity effects and concludes that the ICT => innovation => productivity channel is significantly stronger than the direct ICT => productivity channel for the individual firm. The Swedish evidence is particularly concentrated on larger firms due to sampling effects.

Evidence from Netherlands suggests that ICT use can substitute in productivity equations for the CIS process innovation indicator, which suggests that ICT use may be a good proxy for process innovation in certain types of firms. This provides statistical evidence for a position that has been argued by researchers, that in service industries particularly ICT introduction is often the embodiment of process change.

As noted above, this analysis has stretched the statistical limits of overlap datasets – showing that the intersection sets of two surveys are often good enough for firm level analysis, but it is much more difficult to achieve significant analysis from matching three or more surveys. This has limited the ability of other member states to contribute to the ICT / innovation analysis. Given the fundamental nature of the findings above, it would be valuable to expand the available data for analysis.

1.1.2 Results of Country / Industry analysis, using aggregated data

1.1.2.1 Coverage Thanks to work in each NSI, and by the project’s academic consultants:

• the project group succeeded in covering all participating countries from the exercise in a new metadata framework, which can be used to produce the most comparable indicators;

• the project has shown that the statistical properties of overlap datasets between ICT use surveys and production surveys is sufficiently representative for the moments of the intersections set, for the datasets used here, to be used in analysis, for almost all industries / countries;

• the project has also shown that it would be possible to use the metadata to deliver useful indicators, including complex indicators drawing on and combining multiple responses within or between surveys.

A major limit on using this approach for wider analytical work, which needs to be tackled by Eurostat, is the lack of an agreed approach to publication of aggregated data which, while not generally disclosive, is at a lower level than NSIs would normally publish. This is a potential problem for those NSIs which are reluctant to see results placed in the public domain which may show unacceptable limits of error for individual aggregate estimates, and possibly conflict with official published data.

This problem explains why the report quotes relatively little data from the DMD database built during the project, and that such data are quoted at a high level of aggregation. The industry / country level DMD data have been made available under conditions of confidentiality to analysts working within the project (effectively within the European Statistical System) but under agreements made with NSIs we are not yet able to make the data generally available for research.

7


These issues of publication for more disaggregated data than NSIs are normally prepared to release are less relevant for cross sectional or growth accounting analysis (which uses disaggregated data that NSIs may be unwilling to see in the public domain, but whose results to not pose either disclosure or consistency problems to official statistics). The economists in the project team are keen to see this type of data used to broaden the scope of productivity analysis if the possible publication conflicts can be overcome.

1.1.2.2 Analysis on ICT indicators at industry / country level Analysis across all countries in the study using standard National Accounts based growth accounting treatment of productivity (as developed by EU KLEMS) shows worthwhile improvement in explanatory power, when carried out with ICT use indicators, constructed using the metadata to ensure comparability. This improvement is partly due to methodological differences in the treatment of ICT (especially software) in National Accounts, whereas the ICT metrics delivered by this project are much more comparable both in source and in compilation.

High speed internet use by workers shows up in this cross country analysis as a more powerful indicator related to productivity than e-commerce measures, and as the most effective ICT explanatory input over the period 2000 – 2004/5 over which most of the international data are available. This relationship is stronger at industry level than at firm level, due to reallocation effects within industries as more successful firms grow.

Figure 1.1 Labour productivity and broadband in selected EU countires (2001 - 05)

0

20

40

60

80

100

120

0 10 20 30 40 50 60 70 80 90 1

% of employees linked by broadband

Labo

ur p

rodu

ctiv

ity (€

per

hou

r)

00

However, within country analysis of high speed internet suggests (as in the firm level analysis discussed earlier) that high speed internet use by employees is insignificant or negatively related to firm level labour productivity in Germany, Austria and Italy for the early part of the 2000s.

After taking account of factors influencing adoption, both ICT adoption itself and the observed country / industry productivity effects are also positively associated with ‘dynamism’ of the market – i.e. the ability of ICT users to grow within their markets (and perhaps contribute to market expansion) and to take market share off less successful firms.

8


This analysis takes account not only of the ‘within firm’ productivity effects on which the firm level analysis has focused, but also of the competitive dynamics and reallocation of resources which takes place within industries due to differential growth, and to entry and exit. The analysis is able to combine indicators built using the metadata approach (mainly ICT) with measures available at industry level from National Accounts and labour market statistics including productivity, growth, and – for most countries – ICT capital.

The results of ICT use adoption models using this dataset seem reasonably consistent between countries, and the adoption rate is usually strongly associated with worker skills (measured by wages). The productivity conclusions are also robust to inclusion in regression equations of wages as a proxy for skills.

Initial analysis combining this work with measures of labour market flexibility suggests that more intensive ICT using industries make the fastest progress in catching up to the best practice ‘productivity frontier’ in economies where there is more labour flexibility (measured using OECD’s international framework).

The empirical process to arrive at these results is a good example of how to collect firm-level indicators and merge them into aggregate data able to track firm-level decisions, market selection and aggregate outcomes. While this style of theoretical and empirical work is new, it points in a clear direction towards types of indicators that NSIs should consider. First, by linking surveys on firm level behaviour into the business register, one can create industry-level indicators of ‘what firms do’. Next, researchers can generate industry-level indicators that capture the dynamics and reallocation of resources across firms in an industry. The exact specification of these will depend on the model setup of the researcher, but the indicators should be easy to compute if the data infrastructure (metadata) is in place.

Section 1.2 Conclusions This section sets the main project results against policy requirements and prior studies in order to derive conclusions for indicators, and for future work. The range of results from the study is very wide, and the indicator dataset that it has created provides opportunities for further research by NSIs with access to it. The discussion here focuses on the conclusions derived on:

• the way in which ICT impacts the performance of firms in countries contributing to the study;

• implications for measurement.

1.2.1 The Distributed Microdata method The project has demonstrated, beyond proof of concept through to application in 13 different NSIs, that:

• linked micro-data can be used at the firm level for analysis of productivity effects associated with ICT and with complementary inputs;

• in all these countries it is possible to use a metadata framework to derive, for each national dataset, indicators based on linked data at industry level, and the results of this process can be used either as comparable indicators in their own right, or as inputs to cross-country productivity analysis along with other National Accounts based aggregates. In the recommendations below this process is referred to as ‘distributed microdata (DMD) analysis’.

The process for using linked surveys to produce industry level indicators can be operated without compromising disclosure practices in each country, and this should allow a useful extension of micro-data analysis across the EU (using ICT and other surveys), subject to appropriate institutional arrangements.

9


Earlier studies have shown that both IT investment and ICT use are linked to high productivity (in both level and change models). In this project, using the indicators directly derived from ICT use surveys we show a positive correlation between productivity and the percentage of workers in an industry with access to (high speed) internet. These results continue to hold, after controlling for ICT adoption. The evidence shows that countries or industries with a high proportion of skilled workers are more intense users of new technology. Further, the amount of variability in firm-level output growth across firms in an industry is seen to boost the intensity of firm use of internet.

A key ingredient missing from the statistical system, for policy analysts interested in productivity mechanisms, is direct evidence on how transactions are conducted in a market place. The underlying ‘Keynesian’ design of National Accounts has placed little importance on these intermediate transactions. The treatment of large parts of the economy as ‘margins’ (transport, trade, banking), rather than as creators of value and locations of technological progress is part of this problem. Measures of behaviour which can be integrated into National Accounts, as created here for the role of ICT, can play an important part in improving both the quality and the policy relevance of growth accounting and productivity analysis.

1.2.2 Summary conclusions on ICT impacts on economic performance

The results summarised in Section 1.1 above show that the evolution of ICT and its effect on business behaviour and performance still has some way to go within the EU, let alone the wider world. It seems clear that while the productivity effects associated with ICT in manufacturing are becoming standardised across Europe, possibly reflecting the greater success in creating a single EU market and international value chains for goods, there is much less commonality in services. The reasons for these differences still require further investigation, and since services make up such a large part of the EU economies this would be a natural next step.

We have also seen evidence – in comparing this work to studies of ICT investment – that measurement of ICT capital formation (especially software) poses significant difficulties for National Accountants. These come both from the differences in interpretation and implementation of OECD / UN guidelines for compiling software investment estimates, and from the difficulties in surveying ‘intangible assets’ such as software at firm level. These difficulties partly explain the results above – and why ICT use metrics calculated using distributed microdata do a better job than National Accounts ICT investment metrics in explaining economic impact.

These observed difficulties should not be seen as proof that measurement of ICT capital is unnecessary. But the evidence from this study that measures of ICT use can be compiled and used in comparable analysis shows that alternative measures are available, against which ICT investment survey estimates can be benchmarked. Linked firm level comparisons for the UK and Netherlands, two countries among the few for which firm level IT capital stock data have been developed, show that ICT use metrics (especially fast internet enabled employees) can provide a good proxy for IT investment.

A key conclusion from the statistical results is the relationship, in firm level data and more strongly in country / industry data between productivity and fast internet enabled employees. In this sense, the i2010 objective of creating a ‘single European information space’ can reasonably be tracked, over the decade and in economic terms, by the number or proportion of fast internet enabled workers. The productivity effect depends not only on this variable’s role as a proxy for IT capital, but also on the extent to which it also represents or embodies, for the time being at any rate, working practices of employees, investment in business processes, knowledge management and organisational capital.

10


It looks probable – given that ‘fast internet enabled workers’ does not yet show as a productivity driver in service industries in less ICT intensive countries - that it will remain a useful metric for some time to come. The split of four broad industries used in our analysis (ICT goods and services producing industries, non-ICT manufacturing, distribution and non-ICT services) is consistent with the KLEMS classification and also looks to be a useful analytical approach for ICT indicators, given the different results we see between them.

E-commerce (sales and purchases) and e-business links within and between firms also show productivity impacts, with effects of e-sales and e-purchases that are consistent with earlier work, and which reflect changes in business behaviour. We have also seen that productivity impacts can be associated with e-business process links, especially between firms, but an effective way to model this will probably have to await the more extensive data available from the 2007/8 Eurostat survey. It also looks probable that these will remain useful indicators of change in behaviour and performance.

But the most significant set of results from firm level analysis may be the relationships emerging from the three countries which have linked innovation and ICT use surveys. Taken together these suggest that a significant part of the productivity impact associated with ICT investment and use is channelled through innovation, broadly defined to include non-technical and business process innovation, by a range of mechanisms:

• by enabling knowledge exchange and management, through networks, which feeds the innovation process;

• by supporting ‘roll out’ of new goods, services and processes, through ICT enabled business systems which enable rapid scaling up and replication;

• by enabling more effective marketing of new products / services to new markets via e-commerce;

• by ‘being the innovation’ in business process improvement and redesign.

This evidence should be considered alongside the growing literature on ‘innovation accounting’, developing at both firm and national accounts level. This recognises a range of intangible inputs to innovation, including software, technology based R&D, non-technical expenditure on new products and services, skills, organisational and reputation capital. This framework is still under development with major measurement problems especially in the areas of non-technical service innovation (not least financial services) and organisational and business process change. But the central role of information and organisation in the ‘intangibles’ framework suggests that ICT hardware and software provide an important part of its infrastructure.

Looking ahead to a ‘next generation’ of ICT impact indicators, the changing patterns of innovation should perhaps be an organising framework for thinking about how measures should develop. The interaction between ICT as a knowledge management infrastructure, and the skills (ICT and general) of workers is worth further exploration. So to is the relationship between ICT and organisational / business process change or ‘re-engineering’.

Limited statistical work has been done so far on measuring the increasingly complex use of ICT to manage customer relationships and user input to innovation by firms, in user driven or ‘open innovation’. However, logical extension of the arguments above suggests that it should be on the agenda for future investigation of the ICT – innovation links. In the recommendations below we make a number of suggestions on how these links could feature in future surveys.

11


Section 1.3 Recommendations

1.3.1 Basis for recommendations Recommendations have been compiled by the UK Office for National Statistics as project manager, presented at the project conference in May 2008, and amended in the light of comments from all contributors. While they represent the views of the majority of NSIs in the network group, it cannot be assumed that every NSI supports all the recommendations.

The recommendations from this project spring from its success in meeting the milestones for survey linking and development of microdata analysis spelled out in the project plan, and the additional work to build a new method for development of indicators using linked datasets. This methodology is outlined in detail in Chapter 14.

In the intervening Chapters 2 - 13 we describe in detail the data used in the project, and outline the results of specific parts of the analysis, addressing particular aspects of ICT and complementary factors affecting performance of firms and industries.

The analysis at both firm and industry level identifies new conclusions for the relationships between economic performance and ICT investment and use, and employee enablement, which support the use of new indicators to track the impact of ICT across countries and industries.

1.3.2 Recommendations for Analysis

1.3.2.1 The project has shown that distributed micro-data analysis works, that it can be extended to new countries with investment in metadata and analytical capacity, and some central coordination, and that this approach can be used to develop effectively comparable ICT use indicators across countries / industries, using existing sources. We recommend this process, based on pooling detailed survey metadata from each country is extended to as many other countries as are willing to use the model. This could be coordinated by Eurostat, or, with Eurostat support, by the expert group created in this project.

1.3.2.2 The project has shown that new indicators, or new formulations of indicators to reflect interest in specific industries or firm types, can be generated by country / industry across the EU quite quickly, using existing surveys if the necessary metadata structure is put in place in advance. This provides a more flexible approach to indicator development and adaptation than the current Information Society Statistics regulation, and should be considered as a useful addition to the IIS regulation.

1.3.2.3 In order to conduct distributed micro-data analysis it is necessary to link ICT use surveys with production surveys. Therefore there is a good case for building up the metadata structure covering the ICT use survey (which is relatively easy as all countries undertake a similar survey), and as an absolute minimum, metadata for those parts of the production surveys and administrative data systems which help to weight linked surveys for industry level analysis. Developing the metadata structure to include all those areas of the production surveys and business registers used in this study (see Table 4.1, Chapter 4), would be the best option, and is recommended.

1.3.2.4 Linking through the metadata approach should be done between ICT and innovation surveys, to test the models developed in this project across the few countries where linking is already practical. These suggest that ICT is central to parts of the innovation cycle. More extensive analysis of the relationship

12


between ICT use in firms and steps in the innovation process, broadly defined to include non-technical innovation, would improve understanding of key elements in the working of ICT impacts. We recommend that this should form part of a further, limited, study. Much groundwork for this has already been done by OECD’s Working Party on Industry Analysis.

1.3.2.5 The analysis has shown that for those countries where data on firm level IT investment are available, as well as ICT ‘intensity of use’ indicators, there is a strong relationship between them. Some of these indicators are good proxies for IT investment as measured directly in surveys, and ICT use measures also proxy relatively well for the effects of IT/ CT and other forms of complementary investments. In part this is because firm surveys find it hard to capture some elements of IT spending (especially own account), and therefore use surveys may give more reliable indicators of ICT intensity. We recommend that these conclusions be taken into account in assessing the value, frequency and priority of firm level IT investment surveys.

1.3.2.6 The analysis of productivity effects of ICT, at both firm and industry level, has tended to confirm the view that impacts for ICT can best be measured through ICT in use (i.e. measures of what, and how much, firms do with ICT) rather than ownership measures of ICT assets. Such measures are preferred, in terms of explanatory power, to less effective statistics on ICT investment, in part because of the difficulty of measuring different elements of ICT capital. We recommend that in developing analytical frameworks for impact assessment this principle – that what matters is what and how much, firms do with IT, rather than whether they have it, is given priority.

1.3.2.7 Frameworks for assessing impacts for ICT should recognise the different roles that ICT plays in the different stages of the innovation process Especially in service firms, hardware, software and communications systems determine the adoption of new work practices and knowledge management systems, and may embody changes in business processes. We recommend that analytical frameworks for further indicator development should explore these areas in greater depth than we have been able to. The evidence from this project would justify further analysis, supported by new survey data if possible, on the role of ICT in, and between, firms as:

• infrastructure for knowledge gathering from outside;

• the basis for knowledge management and diffusion within the firm;

• the central element in commercialising new goods, services and processes and replicating them within and beyond the firm;

• the key driver of business process change, combining with complementary intangible investment to enable productivity improvements directly.

1.3.3 Recommendations on Surveys

1.3.3.1 Current surveys on which we have based this project are designed for existing ‘single survey’ compilation of ICT penetration and use aggregates. Few NSIs have yet designed ICT use surveys specifically to make the most of microdata analysis based on data linking, although all are conducted using business registers which permit linking. Also, within the European Statistical System ICT use surveys and Innovation surveys have grown up separately, which may make them more difficult to link, because of differences in some countries in business reporting unit definition. Given the conclusions reached in analysis for countries

13


where the two can be linked we recommend that further thought be given to examine the way ICT use and innovation surveys are done in order to improve the links between data sets. The options available might include

• an ‘innovation module’ in the next available ICT use survey to investigate how firms have used ICTs to change their products, services and processes, rather than a focus on ‘static’ questions about what business process links they use ICT for, as in the current survey;

• redesigning the ICT use and innovation surveys so they work better as complementary surveys, with linkable questions on how ICT is used to develop new products, services and processes, both in technology based and in non-technological innovation, including new organisational and business process developments;

• converge the two surveys, with less emphasis on ICT questions asking about specific technology and more emphasis on asking how ICT use changes firm behaviour, employee behaviour and the relationships with suppliers / customers (although merging the surveys could lose valuable information);

• include in both surveys questions that facilitate analysis toward the interrelationship between intensity of ICT and degree, and model, of innovation. This would preserve the advantages of each survey, but at the possible cost of additional compliance burden on respondent firms.

Careful thought will need to be given in survey development to the fact that ICT and innovation surveys are currently answered, in many firms, by different respondents, and require different knowledge. This is a strong argument for keeping the surveys separate, and starting future investigation with an ‘innovation module’ in the ICT survey. However, some of the innovation issues, on organisation and business processes, may well be accessible to ICT managers as easily as R&D managers who tend to answer current innovation surveys.

1.3.3.2 The recommendation above would improve understanding of the process of innovation in services, which is often organisational rather than technological, and enable us to get over the current ‘two survey limit’ which in most countries makes it very difficult to assemble a sample for analysis of ICT, innovation and business performance in any but the largest firms. This constraint needs to be overcome for effective analysis on larger numbers of firms, or for focusing on specific business types. The new questions on data exchange and on electronic supply chain management in the model 2008/9 survey could provide a base from which such questions could be developed.

1.3.3.3 An innovation module would also help build understanding of the relationships between ICT, innovation and business performance in smaller firms, where the current sampling and survey structure gives very little data to analyse. The analysis in this project has done enough to show that the innovation and productivity effects of ICT differ between broad business types – and are reasonably common in manufacturing firms across all 13 countries. We recommend that developing better data to support analysis for services (often comprising smaller firms) should be a priority as the ICT survey develops.

1.3.3.4 We recommend that consideration is given to asking about how, and how intensively ICTs are used to acquire and manage knowledge, to develop and commercialise new products and processes. Existing data in a few countries are adequate to suggest links between ICT and the innovation process. However,

14


there are questions which the Eurostat common survey does not currently ask about issues such as co-innovation through information exchange, ICT to support flexible working arrangements, and ICT to aid outsourcing, which could be interesting as impact indicators. In those countries which do ask such questions they turn out to be significant as determinants of business performance.

1.3.3.5 We recommend that Eurostat should consider paying equal attention in survey and sample design to generating aggregates in ICT use surveys (which currently dominates survey specifications), and to the design of surveys to support the exploitation of microdata for impact analysis (today not considered at all), which focuses on differences between respondents in a survey, and which often depends on intersection of linked surveys. This change would affect questionnaire design, sampling and the use of panels. In development of the Community Innovation Survey in a number of countries more attention is now being given to building panel datasets, to support better analysis of technology adoption and its impact on firms over time. Similar attention in ICT use surveys would be useful, especially for smaller firms where panel elements in existing surveys are very small.

1.3.4 Recommendations on Institutional Changes

1.3.4.1 The discussions on data sharing within the project demonstrate the need to develop a framework for sharing non-disclosive country / industry analytical data between NSIs in order to facilitate this type of work. We have used an ad-hoc arrangement for this study to permit access to study participants. To maximise research potential of this work it would help to allow appropriate access to non-disclosive distributed microdata to:

• growth accounting researchers for better understanding of macro productivity;

• specialist researchers interested in cross-country comparisons for policy purposes.

The use of data like this, in the project, has delivered research conclusions which do not in any way undermine the aggregate statistics published by individual NSIs, and Eurostat should look to a model to enable such research and analysis to continue.

1.3.4.2 As a starting point we suggest Eurostat should seek agreement with NSIs to allow the research dataset from this project to be used for specific research purposes. A model might be the conditions which currently apply to disaggregated data prepared for KLEMS – data based on official surveys, and fit for research, but not to the same quality standards as official statistics.

1.3.4.3 Eurostat should also consult with NSIs to develop a coordinated framework for the sharing and analysis of data such as that developed in the study, i.e. industry indicators at a lower level of aggregation than NSIs currently publish. One model could be the relatively restrictive framework which Eurostat applies already to structural business statistics, collected by Eurostat from member states in more detail than NSIs publish, but non-disclosive, and used for internal analysis by Eurostat under disclosure rules. A more liberal model could be closer to the approach which many countries apply to the Community Innovation Survey, which allows more detailed analysis to emerge from researchers, into the public domain (although some NSIs in the group would find this option difficult to accept).

15


1.3.5 Recommendations on Indicators

1.3.5.1 The project has produced a ‘complex’ indicator1, using two elements of the ICT use survey which is strongly linked to productivity at firm or industry level for this decade. The ‘fast internet enabled employees’ metric is a proxy, in economic terms, for the ‘common European information space’ over the last five years – and for some time ahead. It has greatest impact in more knowledge intensive industries, and is linked, in analysis, to the innovation process.

1.3.5.2 It has also produced very clear evidence of links between ICT intensity in EU industries, across the countries, and competitive dynamics. In more ICT intensive industries (measured by use) the tendency of firms to gain or lose market share is larger. There is also evidence that the combination of more dynamic markets (entry / exit / innovation) and ICT use is complementary for productivity improvement. Complex indicators which bring these two together will be worth considering.

1.3.5.3 Complex indicators representing the complementary use of ICT, and engagement in innovation (technological, organisational, or other), would also be worth considering on the basis of evidence from the study. This would require the CIS survey to be added to the metadata framework. The relationship between skills and ICT could be one candidate for a complex indicator, either through CIS linking (using CIS higher education questions) or using the ICT skills data from the 2007 model survey to test ICT use / skills complementary effects on productivity

1.3.5.4 In our analysis we have shown that greater intensity of ICT use ( on various measures) is related to other metrics of 'business change' - in the broadest sense defined by the innovation survey, and that at least some of the productivity gain attributed to ICT investment and use can be traced back to ICT through innovation. If this conclusion is right, and policymakers need indicators to help explain changes taking place in firms, and across industries, then it makes sense to look for measures which relate to the main sources of competitive growth. In this study we have shown evidence for several such indicators, and suggested a framework which relates ICT indicators to sources of growth. A summary is shown in Appendix B.

1.3.5.5 The NESIS project in 2004 recommended greater use of intensity indicators, some of which have been developed in Eurostat surveys. This study has shown that ‘use intensity’ indicators using existing surveys are on the whole more effective indicators of ICT impact than simple ‘possession’ or ‘expenditure’ metrics. Developing the conclusions of both these studies, we can suggest future benchmark indicators, some of which could be developed from Eurostat surveys in the pipeline, while others will require new survey questions, or recasting of the relationships between ICT use and other surveys to permit more extensive linking. Suggestions include attempting to gather information on firm spending on complementary investment to ICT, which has been successfully tried by industry experts in the UK.

1 Complex indicators are indicators constructed from more than one question in a survey, or from questions in linked surveys. The ‘% fast internet enabled employees’ is a complex indicator, because is constructed from separate firm responses on % employees using internet enabled computers, and on the speed of connection of firms’ networks to the internet.

16


17

Appendix 1A: Expected outputs contained in the project contract, and results achieved Outputs in Project Agreement Action / Results achieved; issues arising Phase 1 Identify by literature research the activities in enterprises for which ICT usage has a potential impact on business productivity and business performance.

Reviewed 18 studies in participating countries in 2006; updated with wider material May 2007; additional updates added for final report, and specialist reviews included for final report chapters on specific areas of interest.

An audit of surveys and sources for participating countries which can be use for firm level analysis of ICT, innovation and competitiveness.

Project completed audit for ICT use and firm performance data in Phase 1 countries, then updated for Phase 2; a parallel OECD study, to which some members of the project have contributed, has audited innovation data across most countries, and 3 project contributors have linked innovation and ICT surveys.

Review of data linking approaches in participating countries, including assessment of the data quality and properties of linked datasets.

Covered by literature review, and by metadata analysis; UK (Chesher and Nesheim) work on statistical properties of linked datasets has formed the basis; further review of data linking methods in Chapters 3 and 5 of this report.

Assessment of approaches in participating countries to firm level analysis of ICT and competitiveness, and comparison of their assessments of impacts.

Covered by literature review, and by work reviewed at project meetings hosted by Swedish, Italian, French and Czech NSIs, and by Amsterdam Free University, with country input of alternative approaches.

Proposed common framework for analysis, recognising the potential to use additional data where available in individual countries.

Agreed set of analytical variables used across all countries, at firm and industry level, with sub-groups to tackle specific impacts based on comparable local data. Firm level analysis framework extended to include linking with National Accounts aggregates via EU KLEMS database. Approaches for additional topics agreed in terms of variables to use, and results outlined in Chapters 7-13.

Initial analysis across the participating countries, using the common framework, to derive comparable assessments of ICT impacts.

Firm level analysis carried out for all Phase 1 countries, using common variables; groups of countries successful in developing (non-identical) analysis on additional variables Initial round of industry data done, subject to final quality checking. Raises issue of most efficient way to develop cross-country analysis and constraints of data sharing.

Comparison of results between countries, and if possible with OECD work, leading to proposals for indicators of ICT impact.

‘Impact’ review at OECD at WPIIS in May 2007, and access to Japan, Canada, Australia, China studies. Wide discussion of results in this project and elsewhere provides good basis for - comparisons of individual indicators; - understanding links from indicators to productivity / growth; - relating ICT indicators to market measures (e.g. contestability); - integrating National Accounts data for Phase 1 countries via EU KLEMS.

Tested framework to allow invitation for more countries to participate in Phase 2, and a network of experts to support it.

Framework for metadata applied across all countries, with user guide and network of lead country experts to help; used by NSIs with model surveys and micro data access. Modules for additional variables defined. Constraint on further extension is availability of analytical resource in some NSIs.


18

Outputs in Project Agreement Action / Results achieved; issues arising Phase 2 Framework for analysis, using common surveys and code.

Workshop of 11 experts from 6 Phase 1 countries in October 2007 to agree final analytical content, with observers from Phase 2 countries.

Application by Phase 2 countries, (and update / extension for Phase 1) with support from the experts group, providing comparable analysis.

Network of experts to support new countries joining analytical work. Metadata / definitions successfully applied in four Phase 2 countries. All are able to use metadata supported firm level regression analysis. Initial results used to review method at project meetings in Rome and London.

Review analytical conclusions, with comparisons from non – EU countries in OECD performing this type of work.

2007 reviews with OECD updated at WPIIS 2008, but US review was not possible because of lack of US firm level data in time.

Recommendations on indicators of ICT impact and on methodology.

Key indicators identified on ICT use, fast internet, business processes and mobility; analytical process for focussing on key indicators should - have explanatory power at firm and industry level; - discriminate on forward policy issues; - reflect innovation impacts and changes in business behaviour.

Recommendations on surveys to support indicators of ICT impact, innovation and competitiveness.

Survey recommendations based on: - evidence from this project on ICT and existing

contribution to innovation; - need to consider new developments in ICT effects

on innovation process.


19

Appendix 1B: Indicators in the project using existing data, and suggestions for the future Table 1: Framework for ‘sources of growth’ which can be described using existing surveys

Firm level Source of growth Possible role of ICT

Acquisition of knowledge from outside the firm Internet links to professional experts, 'customer driven' idea generation

Managing knowledge to formulate commercial ideas Internal networks between people Turning ideas to delivered products / services Replicating product / service through ICT based

'enterprise architecture' Marketing to target customers Use of CRM systems, or web marketing

Industry / economy level

Source of growth Possible role of ICT Faster spillovers of knowledge between firms Industry levels of internet exchanges More efficient markets, speeding up reallocation e-commerce penetration Supply chain management / optimisation e-procurement

Table 2: Indicators which can contribute as impact measures, related to sources of growth

Measure / indicator Element of growth contribution From existing data

Fast internet enabled employees, by firm and industry % e-sales, by firm and by industry % e-procurement, by firm and by industry ICT investment / employee use Dynamism (share change) in high ICT using industries

ICT investment, knowledge acquisition and management Marketing efficiencies / targeting Supply chain management, market efficiencies Enterprise architecture, to replicate processes Speed of reallocation of resources

From data in the 2007/8 and 2008/9 surveys

ICT skills and ICT use in firms Summation of e-business links Intensity / extent of information sharing (from 2008/9 survey)

Ability to absorb and manage knowledge Internal / external business systems, and ability to replicate new products / processes Internal / external coordination, and ability to replicate new products / processes

From survey extension / better linking

Skills and ICT overall capability - needing a linkable skills question Investments complementary to ICT, requiring some form of expenditure question (e.g. how much spent on process redesign compared to ICT)

Absorptive capacity of firm for new ideas Training, business process change etc; how much firms invest in 'business change'


Chapter 2 Introduction and Background

Tony Clayton UK Office for National Statistics

2.1 Outline of ICT indicator development 1. Indicators for the ‘Information Society’ have a relatively recent history within the European Statistical System. Nordic countries, and INSEE in France, began to take a significant interest in how information and communication technology (ICT) was being used in industry, and in wider society, in the mid 1990s, as email, e-commerce, and the use of networks began to impact on increasing numbers of firms and households. Eurostat too began to develop approaches based on best practice among its member states. In 1999, the US Bureau of Census undertook its first (and so far only) survey of computer network use in firms, to start measuring how firm behaviour was changing in response to new ICT opportunities.

2. A key step forward in measurement came at the OECD’s 1998 ministerial meeting on ICT and e-commerce, held to initiate both a policy framework and statistical approaches to track the challenges posed by new technology based business models which:

• made it possible to conduct business electronically without regard to international borders;

• created new commercial relationships which might bypass traditional channels of distribution and payment;

• provided opportunities for productivity improvements which did not appear (at least at first) to show up in the statistics.

3. The 1998 ministerial meeting committed OECD member states, including the majority of EU members, to develop common statistical approaches to measuring the information society, at work, in the home and in the wider community. The initial conceptual work on definition of the ICT industries and of ICT products and services, on e-commerce and on the measurement of ICT use in business and households was pioneered as a result by a relatively small group of countries in the late 1990s.

4. The approach used to develop metrics was focused on understanding the progressive transformation of economic and social relationships by ICTs. A simple linear model, aimed at understanding:

• the ‘readiness’ of economies and institutions, businesses, households and government, to accept or perform electronic transactions of various kinds,

• the ‘use’ of ICT, e-commerce and electronic business processes, and

• the ‘impact’ or change in behaviour and performance of economic and social actors

was developed as a way of setting indicators in a policy context.

5. This ‘S curve’ approach (borrowed from early work in Statistics Canada, see below) was prominent in thinking through the first few years of statistical development. It accompanied policy focus on building the foundations for internet use, through education, familiarisation, infrastructure in terms of equipment and the creation of networks. At the time there was little empirical evidence on the gains to be expected from the ‘impact’ of ICT. The assumption behind both policy and measurement was that the economic and social benefits of ICT would become evident, and that the Solow Paradox “You can see the computer age

20


everywhere but in the productivity statistics”i of a decade earlier would be resolved. As we shall see, this resolution took some years.

• Impact; changes in behaviour, and in economic structure and performance as result of use

• Intensity / use; measures of the proportion of firms, individuals who adopt, and the amount of use

• Readiness; ability of firms, individuals to adopt technology

Level ofElectronicCommerceActivity

Time

Readiness

Intensity

Impact

Figure 2.1 OECD ‘adoption’ framework

6. OECD developed in this initial phase, definitions of e-commerce in use today, model surveys and metrics which still form the basis of much international comparison work. Most of the metrics were aimed at the ‘use’ element of the framework, measuring the behaviour of firms and households, their ownership and use of specific items of ICT goods and services, and the proportion or value of sales and purchases executed electronically. There were also some measures of ‘readiness’, mainly gauging attitudes to and opinions of e-commerce or the internet, to try to understand barriers to firms’ or individuals’ use (for non-users), or the self assessed benefits (for users).

7. In retrospect, the self assessment questions, which covered both ‘readiness’ and ‘impact’ measures to help policymakers, were the least robust elements of the early framework. An example of the difficulty in interpreting questions relating to ‘barriers’ came from early results which suggested firms experiencing most barriers were also those which had done most in terms of adoption. Perceptions could not be used as a reliable guide to behaviour. Similarly, international comparisons of household surveys suggested that perceptions of risk (of fraud etc) were lower in the UK – where incidence was higher – than in other EU economies where incidence was lower. Thus, while the standard OECD surveys do a good job in measuring ‘use’ through ownership, transactions or behaviour, assessing the other parts of the framework require more complex analysis.

2.2 How indicators have been used for policy 8. The strongest boost to ICT measurement effort came as a result of adoption, in 2001, of the Lisbon strategy to promote Europe as a ‘dynamic knowledge based economy’. Under this policy initiative, the Council of Ministers committed to a programme in which innovation, based in part on ICT use, and also designed to promote inclusion and sustainability, would break the trend of poor relative productivity performance which the EU has seen during the 1990s compared to the US. The programme was designed to be managed through the ‘open method of coordination’, without strong central direction but with peer review of progress using relevant indicators, and this demanded a clear evidence framework at all levels.

21


9. The EU framework for indicators to support the Lisbon programme covered metrics for the overall economy, for employment, market reform, distributional and environmental effects, and, not least, for innovation. The evidence framework needed for open comparison of national policy environments was contained in the EU ‘Structural Indicators’ dataset, developed over 2001/2, and refined in subsequent reviews. This was designed to provide a comprehensive, but easy to assimilate, picture of the outcomes of policy (growth, productivity, employment) and of underlying drivers (skills, investment, entrepreneurial activity, technology, training).

10. Indicators used for innovation in the headline benchmarking dataset included ICT investment and e-commerce use. ICT investment was measured using private sector estimates, because national accounts estimates were not considered reliable, but consistent measures of use of electronic transactions across EU member states were delivered using a coordinated development of the OECD model survey, led by Eurostat.

11. With a coordinated programme of local experimentation and knowledge sharing among the EU statistics offices, experience in ‘what worked’ for ICT use surveys grew rapidly. The survey programme extended from initial concentration on business use of computers, networks, internet and e-commerce to more complex questions on e-business processes, barriers and benefits of use, employee engagement, security, skills, and other areas.

12. In addition to headline metrics, DG Information Society and Eurostat developed a range of benchmarking indicators specifically to monitor the ‘e-Europe’ programme which ran from 2001 to 2005. Most of the metrics created were designed to measure the ‘e-readiness’ and ‘use’ stages of ICT development in households, government and business. Attention in this set focused on individual / household measures of IT and internet use, on education and government services, but with the largest section on business metrics.

13. As an example of business metrics the ‘e-readiness index’ designed by the Joint Research Centre of the European Commission combined a number of technology ownership and use measures in a composite indicator, underpinning the capability of firms to bring together the elements of integrated e-business operation. But because of the lack of real data, such indicators were essentially conceptual rather than empirical.

14. Examples of broader integrated sets of indicators used for policy management include:

• the set of ‘e-government indicators’ designed by DG INFSO, and carried out for the Commission by Cap Gemini Ernst and Young,ii which tracked availability and sophistication of a specified range of government services in all EU countries, essentially by direct testing of specific government on-line activity in each member state, to classify the level of interactive service in the range ‘provides information’, through ‘provides forms for off line completion’ to ’complete transaction’. This is, in effect, a user centred test of e-business process use.

• the UK’s e-commerce benchmarking framework, assembled with Booz Alleniii in 2002, assembled official and unofficial / commercial data from a number of leading countries, on ‘readiness / use / impact’ of ICT for business, citizens and government, and was used to identify key areas for focus of government activity, as well as gaps in UK statistical capability. Its measures of ‘impact’ were limited to evidence of behaviour change measured through ICT use (e.g. consumer purchases over the internet) and macro-economic productivity estimates of ICT contribution to output growth from growth accounting.

22


2.3 Development of ICT impact analysis

2.3.1 Macro analysis 15. The first assessments of the economic impact of ICT investment and adoption on an international scale were largely based on macro economic measurement. A comprehensive review of evidence by OECD as late as 2003 concluded that ‘evidence on the role of ICT investment is primarily available at the macroeconomic level’. This was a result of the decision, in 1993, to treat investment in software as an asset under the System of National Accounts (SNA), which allowed analysis of the role of ICT investment (hardware and software) in growth accounting across the majority of developed economies.

16. Comparisons set out by the OECD in its 2003 reportiv show, for the 1990s, how ICT investment contributed to overall growth across 15 member states, and split out the productivity effects for ICT producing industries, and for ICT using manufacturing and for ICT using services. The study highlighted the strength of ICT investment in service industries – whereas much of the early impact analysis has been focused on manufacturing for measurement reasons. While it showed ICT investment as a contributor to output growth and productivity, the differences in impact between countries were striking.

17. Strong multi-factor productivity growth in the US associated in this study with ICT use was interpreted as a result of the US’ early lead in adoption of ICTs, overcoming adjustment costs and benefiting too from highly competitive markets in which entry, exit and adjustment were easier. There remained a number of EU economies for which the contribution of ICT use to output and productivity growth did not grow as ICT investment grew, and some in which it actually seemed to fall.

18. A major difficulty in early macro assessment of ICT impacts for policymakers was that estimates of ICT investment in macro-economic data were not consistent across countries. Principal reasons for the differences were:

• differences in estimation methods for software investment (first included as an asset in ESA 1995) with surveys in place in some countries but not others, and major variations in the treatment of ‘own account’ software written within firms for their own use;

• lack of consistent deflators for both hardware and software, and considerable international debate on the ‘hedonic’ approach used in the US to quality adjust computer and software prices.

19. Macro economic estimates through the late 1990s / early 2000s were also complicated by the distortions of the ‘dot com boom’ which changed market conditions to such an extent that productivity gains were attributed to this rather than to longer run structural or technological change. Productivity gains could be attributed to the (fast growing) ICT producing industries, but universal benefits from investment by ICT users were less clear.

20. Macro analysis, but at the industry level, has been the focus of more recent work in the US, and the EU. Brynjolfsson and colleagues, in ‘Scale without Mass’ (Brynjolfsson et al, 2006) looked at the relationships between industry ICT intensity, and the characteristics of competition across US industries and concluded that:

• greater ICT use in industries speeds up the diffusion of new, successful, business models by ‘winning’ firms, and so is associated with greater degrees of market share change within these industries;

• the effect of this process is to encourage increasing supply concentration, as successful firms supported by ICT grow, and others lose market share or exit the market.

21. This US analysis drew no specific empirical conclusions on productivity or on economic performance associated with technology. However the ‘KLEMS’ initiative starting in 2004 and funded by the EU was designed to take industry level National Accounts data –

23


starting with EU countries but also looking at other leading economies – and to develop growth accounting models by industry taking account of capital (K), labour (L), energy (E), materials (M) and services (S). Among the inputs separately identified as part of this programme is ICT capital (as part of K).

22. The final results are still under review, but interim outputs have been presented to the group undertaking this project, and show significant differences between countries, and between the EU and the US, in the growth accounting impact of ICT investment. However, the broad picture demonstrates that:

• differential gains in productivity in more intensive ICT using industries have been an important part of the US productivity advantage over the decade to 2004;

• distribution and business / financial services have shown the most substantial gains.

23. The data shows these differences largely in terms of TFP (i.e. unexplained) growth. This suggests that National Accounts data on ICT investment may not be sufficiently well developed to act as a good explanatory variable – essentially the same conclusion as that reached by the compilers of the EU structural indicators.

24. Among reasons why IT investment may be inconsistently recorded in official data in different countries are:

• the difficulty of extracting survey data from firms on software investment, both on own account work by their own employees which is usually not recorded in accounting systems as capital, and on software which is embedded in IT system purchases;

• differences in the detail with which deflators are applied to different elements of IT, hardware, purchased packaged (i.e. standard) software, purchased bespoke (one off project) software and own account software;

• differences in application of the internationally recommended method for calculating own account software investment.

25. These issues are explored in a detailed analysis by the UK Office for National Statisticsv. The implications are that ICT investment may be a useful directional indicator for change in technology investment. However, as the proportion of IT investment in national economies becomes more heavily weighted in favour of software rather than hardware, and the proportion of software writers outside the IT industry grows, the reliability of official IT investment may become harder to guarantee rather than easier.

26. An additional factor which affects the value of ICT investment, revealed by this project, is the growing importance of IT service outsourcing. Earlier studies have shown both the importance and the productivity effects associated with offshoring of IT enabled services. Finnish analysis for this project, using survey questions on outsourcing of IT services, shows that the productivity incentives for outsourcing IT are significantly positive. This may well influence the distribution of IT investment across industries, and make it unrepresentative of the pattern of ICT use, and of the impact of ICT on business operations. Direct measures of ICT use, from the Eurostat firm level survey, may be the best way of assessing this.

2.3.2 Firm level impact analysis 27. Early firm level use surveys, piloted by OECD, were implemented in Canada, Scandinavia, Australia and the US in the late 1990s. Starting in 2001 the EU began the most sustained programme of implementation and development of ICT use surveys anywhere in the world, in which first fifteen member states and then a larger group were supported in developing practical survey instruments around a common core of questions.

28. By 2002/3 enough experience had been gained to build confidence in the firm level responses to most (but not all) the questions included in EU ICT use surveys. Researchers then started work on linking these surveys to NSI business output and employment data to

24


test whether productivity differences between firms could be linked to their use of information technology or communications.

29. The use of firm level data to study the relationship between ICT and firm performance spread across a number of countries as soon as consistent surveys became available. Early studies drew on both official and private data sources and used different methodologies. Some examples of some of the different approaches adopted are:

• inclusion of ICT capital stock at firm level as a separately identified capital input in labour productivity or total factor productivity (TFP) analysis (Brynjolfsson & Hitt, 2001; Hempell, 2002);

• inclusion of ICT capital alongside other measures of ICT use, such as internet use or number of employees using ICT (Maliranta & Rouvinen, 2003);

• inclusion of ICT capital stock with measures on innovation and / or organisation change (van Leeuwen & van der Wiel 2003);

• inclusion of measures of computer network use (i.e. behaviour) as an additional determinant of labour productivity or TFP in a productivity regression equation (e.g. Atrostic and Nguyen, 2002).

30. In 2004 OECD was able to publish a portfolio of firm level studies, some comparing ICT impact in different countries, and using similar analytical methods, across 13 countries. In most cases the scope for cross-country comparison was limited to two or three NSIs, because of differences in ICT use surveys or in the scope to link to other sources of data. For several countries comparisons could only be drawn for manufacturing, and in some (e.g. Germany) links could, at that time, only be made outside the statistical system. But EU member states featured strongly in this first major review which also included Japan, the US, Korea, Australia and Canada.

31. In 2005 the UK Office for National Statistics published a set of studiesvi, based on firm level data, which took account of:

• firm level data on IT capital stock, both hardware and software;

• firm level measures of ICT use by employees, of computers and the internet;

• firm level use of e-commerce for both procurement and selling;

• firm level use of communications networks.

32. In combination these studies showed that while IT investment can be shown to increase firm productivity, the effects depend on a range of contingent factors, including whether or not the firm operates as a multinational, and in particular whether it has a US home base, its age, and whether it is a manufacturing or service operation. They also showed that:

• greater ICT use by employees has an additional association with higher productivity, over and above the effect of IT investment;

• e-commerce use for selling and for buying does not have equivalent productivity effects; for manufacturing at least, e-procurement has a much stronger productivity influence, partly based on price effects;

• organisational advantages associated with US ownership have a significant influence on productivity returns associated with investment in IT hardware for the UK affiliates of US firms;

• greater communications expenditure by firms can also enhance the productivity effects associated with IT investment;

25


• returns to IT investment are also influenced positively by firm level possession of skills (measured by employees with university qualifications) and by investment in fixed capital.

33. These studies were followed by a further analysis of investment in high speed internet by UK firms and the effects of broadband use on productivityvii. This appeared to suggest – on the basis of a relatively short time series of firm level data – that early broadband adoption by UK firms was biased towards those which already had high productivity, but that their subsequent performance showed benefits from adoption. The analysis went on to show that employee use of DSL or fast internet connections was a useful productivity indicator, in combination with others. It provides a possible definition, in economic terms, of the EU’s first objective for its i2010 programme of a ‘common European information space’.

34. Each of these studies has shown that ICT investment and use by firms has an impact of productivity levels or growth which is:

• dependent on the sector of the economy in which a firm operates, so is business model specific;

• depends on other inputs, related to skills, organisation, or innovation.

35. It is therefore worth considering, as part of the background to this project, the work which has been done on ‘complementary investment’ to ICT.

2.3.3 Increasing understanding of ‘complementary investment’ 36. More recent productivity analysis (so far only at whole economy level) has suggested an alternative view of traditional accounting approaches to definition of assets in economic aggregates, beyond the standard System of National Accounts, which treats software as investment but few other intellectual inputs to the production process. For the US and the UK, and now for an increasing number of other developed economies, this analysis has recast National Accounts to take account of ‘intangible investment’, including R&D, expenditure on ‘non technical’ innovation including financial services, training, organisational change expenditure and branding.

37. The data on some of these areas are still open to question, but the overall picture which emerges from these studies is a growth accounting analysis which is more intuitively sensible over the 1990s and the early part of the 2000s. It shows:

• first, that intangible investment is a rising part of overall investment by firms, even if it does not show up in their balance sheets, and that software and the associated business process / organisation investment have been the fastest growing elements;

• second, that intangible investment in total now rivals investment in fixed assets for the US and the UK;

• third, that much of this intangible investment represents ‘capitalised labour’ and so shows a very different picture of the relative returns to capital and labour from that in official economic statistics;

• fourth, that this treatment captures well the major investment by US and UK firms in the late 1990s associated with major business change.

38. To interpret this macro framework using firm level or industry level analysis requires us to treat ICT as one agent of innovation and growth. Applying this framework at firm / industry level, requires linking to surveys on:

• Innovation and R&D;

• Skills;

• Organisation / e-business links;

26


• Other ‘intangibles’ which show up as firm characteristics;

• Business performance, productivity and growth.

39. The project has set out to do this in specific countries where data are available and linkable to the ICT use survey, and to business output data. Inevitably this is easier to do in those NSIs which have more developed statistical infrastructure, and a local capability for data linking analysis. But a majority of the NSIs involved in the project have been able to provide evidence on at least one of these themes.

2.4 Overall approach to analysis 40. In addressing the objectives set for us by Eurostat in this project, in addition to linking surveys and data sources we have set out to link micro (firm level) and macro (industry or whole economy level) analysis. This has a range of advantages, but also poses challenges for the statistical system.

41. We have been able to build on earlier studies of surveys, undertaken by the EU and others, on the types of indicators which are most valuable in developing measures of ‘impact’ for ICTs. The NESIS project, in addressing the relationships between ICT use and business organisation and business processes, had as one of its key recommendations that ‘more intensity indicators should be developed on the way from readiness to impact indicators’viii. The recommendation recognises that the intensity of ICT use within firms is often indicative of how far they have changed processes and organisation, or their capability to change in the future. It also provides a useful link to the conceptual macro work on complementary investment in organisation mentioned above.

42. The NESIS conclusion has supported more precise Eurostat survey questions on the degree to which employees are engaged with ICT, and how far ICT is embedded in business processes and transactions, but the practical fruits for most countries will not be available until 2009. However, both our firm and industry level analyses have, in effect, used intensity measures now available as the best starting point, looking for relationships between usage and productivity or growth.

43. Firm level analysis of the drivers of productivity and growth is vitally important. It provides the foundation of our economic understanding of firm behaviour and performance, and of the influence of market conditions and technology change on competitive behaviour. It can be argued, in the debate over capitalisation of software and R&D in National Accounts, that we should only treat these activities as investment in the whole economy if we can show how they behave as assets for individual firms. This principle has driven much of the microdata work on innovation.

44. In addition, the insights gained from firm level analysis benefit from much more exhaustive use of data. The range of experience and performance captured in firm level data is much richer, and contains an additional order of magnitude in degrees of freedom, compared to industry level data – however detailed. Firm level analysis is where we should first expect to pick up signs of impact from use of a technology, by comparing successful and unsuccessful firms. This is usually well before the successful firms have a sufficient impact on industry performance to permit analysis to identify it at the higher level. However, we need to recognise that firm level analysis, focusing of productivity or growth performance of individual units, requires extending in order to pick up the ‘macro’ effects of resource reallocation as successful firm grow, and unsuccessful firms shrink or exit.

45. One drawback of firm level analysis is that – historically – it has not always been possible to achieve reliable outputs which can be compared cross countries. This is because of minor differences between surveys, in the precise wording of questions, or more usually in national sampling strategies and certainly in structural differences between countries. The project has adopted a ‘metadata’ strategy to address this common problem.

27


28

46. This has enabled us to conduct some relatively simple, but effectively comparable, analysis of the relationships between ICT use and firm level productivity across the countries, by major industry. This has proved a useful overall check on those industries / countries where the Solow paradox is resolved, and those where there is remaining variation or inability to demonstrate a link between computers and the productivity data of individual firms.

47. Industry level analysis is valuable because it is much more accessible to policy users of data and analysis. It permits impacts to be expressed in terms which relate to the economic aggregates through which most policy is managed. In addition, industry level analysis has access to data which are not available at firm level. Labour market data, pricing, market structure and change characteristics are among the measures which can only be measured at industry level – and which are often poorly reflected in firm level analysis.

48. Comparisons across countries are often more practical to do at industry level too, although there may be classification differences. International comparison is a key element of the Lisbon process, done in a way which allows policy makers to draw lessons from differences between effects of policy in different countries. It is apparent from the work done in this project that the ICT use surveys using the ‘Eurostat model’ are probably the most comparable sources of ICT intensity data available across countries and industries, so it is well worth the effort to integrate into other data frameworks.

49. Since indicators are almost always expressed and used at country / industry level, we have focused in the analysis which follows on using microdata from ICT use surveys in ways which permit it to be used directly with National Accounts data, for productivity and growth analysis. This presents a number of questions on the use of data, in particular the fact that ‘national accounts consistent’ estimates from ICT use surveys may not be consistent with the existing aggregates which Eurostat specifies in current regulations. There is also an overriding need to ensure that, in developing more detailed data to match the structure of national accounts, data confidentiality is not compromised. However, the evidence from the project to date shows that these challenges are worth addressing. This will need to be championed by Eurostat in a way which recognises the constraints on NSIs.

Endnotes

i Solow (1987). ii Web-based Survey on Electronic Public Services (2002) Cap Gemini, Ernst & Young, For DG Information Society e-Europe 2002. iii Booz Allen Hamilton (2002). iv OECD (2004). v Chamberlin et al (2006). vi Clayton et al (2005). vii Farooqui and Sadun (2006). viii Airaksinen (2004).


Chapter 3

Methods and Data Sources Mark Franklin UK Office for National Statistics

3.1 Summary 1. Firm level data from different surveys and sources has been assembled and integrated in a systematic fashion in each project country. This exercise is designed to investigate the impact of ICT through two separate but related lines of analysis. First, by generating consistent industry-level indicators of ICT usage and economic characteristics which can be integrated with pre-existing cross-country data, ICT impacts can be analysed at the conventional research level of industry groups. And second, the project framework serves as a platform to run identical regression analysis using the firm level data that is available in each project country.

2. The project also serves as a demonstration of a methodological approach which can be extended to other data sources, other research issues and other participants.

3.2 Introduction 3. As described in Chapter 2 above, at a fundamental level one can distinguish between two different types of analysis of ICT impacts:

• Analysis using firm-level (enterprise level) data;

• Analysis using “statistics”, that is, data which have been aggregated to some degree.

4. The methodology used in this project straddles these two forms, and is innovative in a number of respects:

• We have undertaken a metadata review and data preparation process allowing identical analytical code to be executed in all project countries;

• We have thus carried out identical comparative analysis of ICT impacts using linked firm level data in 11 European countries. We are not aware of any previous comparative analysis of firm-level data comparable in the degree of consistency or scope of application;

• We have built a suite of indicators from the firm-level data specifically designed to address policy issues related to ICT impacts. These indicators can be combined with existing cross-country data, notably the EU-KLEMS dataset. Chapter 14 reports some introductory analysis using this approach.

5. This chapter is organised as follows. We begin with a broad overview of the project methodology. We then discuss each stage of the process in more detail, highlighting issues that have arisen. This section draws on a User Guide that has been developed as part of the project1. The following section (3.4) discusses the rationale and feasibility of the two separate strands of analysis that we have conducted, which we term “co-ordinated firm-level analysis” and “distributed micro data” respectively.

1 The User Guide is attached as Appendix I to this Project Report.

29


3.3 Project methodology

3.3.1 Overview 6. The basic project methodology is illustrated in Figure 3.1. A metadata review was undertaken to establish what data are held in each National Statistical Institute (NSI) and to identify the set of variables which constitute the core dimensions of the project. The data are described in some detail in Chapter 4 below, but broadly speaking the variables include: Figure 3.1: Overview of Project Methodology

Secure Environment

Disclosure controls

Cross Country Datasets

Merge with all countries’ output datasets

Merged Datasets

Code execution

Metadata review and Data selection

Data assembly

Code development

ProductionSurvey E-Commerce

Analysis & Reporting

BusinessRegister

Outputdataset

Country-specific Source Surveys

Outputdataset

Outputdataset

• A set of variables pertaining to ICT-usage by firms, drawn from the Eurostat harmonised E-Commerce survey (EC);

• A set of variables which describe the economic characteristics and performance of firms, drawn largely from countries’ structural business surveys (denoted Production Surveys (PS) in this report);

• Some information on the overall population of firms in each project country, taken from Business Registers (BR).

7. Firm-level data drawn from Business Registers (BR), Production Surveys (PS) and E-Commerce surveys (EC) are assembled and processed to create a set of output datasets in each project country. This activity normally takes place in a secure environment, reflecting the confidential nature of the firm-level data. However, the processing is designed to generate a priori non-disclosive statistics2, being aggregates of firm-level data across one or more dimensions such as industry group, size class, year etc. The processing also generates certain direct project outputs such as regression statistics.

8. Recognising that some cells in the output datasets may be quite “thin” in terms of the number of underlying firms that they represent, the datasets are checked for disclosure before being released from their secure NSI environments and complied into multi-country datasets. Data from these multi-country datasets are described in Chapter 4 and analysed further in Chapters 5, 6, 7, 8 and 14. At the present time, access to the multi-country

2 That is, statistics which do not violate confidentiality constraints.

30


datasets is confined to a subset of project participants. One of the recommendations of the project is that consideration should be given to providing wider access for research purposes.

9. We now review the project methodology in more detail.

3.3.2 Metadata review and data selection 10. The choice of variables and detailed scope of the analysis has been an iterative process, informed by the knowledge and experience of project members, responses to a metadata survey conducted at an early stage of the project, issues that have arisen in the development of the analytical code and investigation of preliminary results, disclosure issues, the format and scope of the EU-KLEMS database and other factors. Some examples of issues arising out of the metadata phase are as follows:

• Unique firm identifiers. A central feature of the source data in each participating country is the existence of unique firm identifiers which provide the ability to match firms across different surveys and datasets. Typically this characteristic follows from the development of a longitudinal business register which (a) tracks firms over time and (b) is used as a sampling frame for surveys of the business sector, including the structural business surveys and E-commerce surveys.

• ICT investment and capital stocks. Firm level data on ICT capital is a priori highly relevant to the analysis of ICT impacts. However, the metadata survey revealed that only two of the project members (the United Kingdom and the Netherlands) have access to such data, and it was therefore decided to exclude ICT capital from the core dimensions of the project. Chapter 10 reviews analysis using the firm level ICT capital stocks data for the UK and the Netherlands3.

• Data on fast internet capacity (DSL) and the share of workers with access to such capacity (DSLPCT) were not originally included in the project specification. These variables were added in the light of preliminary results showing near saturation of PC-enabled workers (PCPCT) and the dissemination of research using fast-internet capacity in the UK4.

• Functionality to incorporate firm-level data on labour skills, such as the share of the firm’s employees with post-secondary level education, was added, to facilitate the analysis reported in Chapter 8. The specification requires that such data, if they exist, are included in the annual PS input datasets, although there is no particular reason to expect this to reflect data storage conventions in those NSIs with firm-level education data.

• On the other hand, certain questions from the E-Commerce model questionnaire pertaining to ICT business integration have not been included in the core specification, due to the relatively high costs of data assembly. These data have, however, been included in a variant of the core project analysis which has been run by a sub-group of 6 project countries, the results of which are reported in Chapter 9.

• The E-Commerce survey is co-ordinated by Eurostat and ostensibly harmonised across all countries. However, one outcome of the metadata review is that residual differences exist in terms of the detailed implementation of the survey across member countries. For example, translation of the model questionnaire can lead to differences in the precise wording of questions; there are differences in coverage (particularly of

3 Another problem with ICT capital stock data is the measurement of software capital; see Chamberlin & Chesson (2006), Chamberlin et al (2007). Note that, in contrast to the bottom-up, firm-level ICT capital stock approach used in the United Kingdom and the Netherlands, the EU-KLEMS database generates industry level ICT capital stock data for all countries, by partitioning national accounts business investment series. 4 See Clayton et al (2005).

31


optional elements in the harmonised questionnaire), differences in frequency of the survey and differences in the sampling methodology.

• As described in Chapter 5, business registers are used to provide a reference framework for re-weighting of sample variables, exploiting the property that the business register is register of whole population of firms, and all business registers carry basic information such as on firm employment. But as reported in Chapter 5, we find that employment on the business register is not routinely updated and typically is below employment reported by those firms sampled by the annual structural business survey.

• The project design assumes that business registers are also a source of additional information on firm characteristics, such as the age of the firm, whether the firm is owned by a multinational organisation, and whether the firm is an exporter.

• In some countries the business registers contain many millions of records. This can cause technical issues such as violation of memory constraints and slow program execution times.

11. A full set of the core variables used in the project is shown in Chapter 4, Table 4.1. Conceptually, there are two different methods of adding additional data sources to this project framework:

• Include additional variables that exist on (or can be added to) the input datasets accessed by the project specification. An example of this approach is the analysis of data on ICT business process integration (BPI) reported in Chapter 9. These data are drawn from the E-Commerce survey but are not included in the core project specification. It is comparatively straightforward to add functionality to address such additional variables.

• Include data from separate datasets, such as firm level data on ICT investment or capital stocks, data from Community Innovation Survey (CIS), and data on trade and foreign direct investment. As long as such data can be matched through a unique form identifier then we could in principle modify the project specification to incorporate such data. However, this is an order of magnitude more complex than addressing additional variables in the existing datasets and has not been attempted here. Some project countries have, however, matched project data with data from separate sources in an ad hoc fashion5.

3.3.3 Data assembly 12. It is unlikely that the raw firm-level data held by NSIs will be in a form and format that can be addressed directly by the analytical code developed for this project. Recall that a central feature of the project is that identical code is run in each NSI. It has not been a trivial exercise to achieve this when the coverage and scope of available data inevitably varies across project countries.

13. The general specification of the analysis assumes that EC and PS input data should be stored in annual datasets of the form “PRE_ECYYYY” and “PRE_PSYYYY” where “PRE_EC” and “PRE_PS” are common prefixes and “YYYY” is the year to which the data pertain (not the year when the survey was conducted, which was one of the ambiguities picked up at the metadata stage). The analytical code is specified to read a single BR input dataset, containing register information stacked over all years.

14. Although the analytical code is robust to missing variables, certain features of the input data are essential for the code to function properly. These include:

5 Matching to firm-level ICT capital stock data is reported in Chapter 10. Matching to trade data is reported in Chapter 11 and matching to CIS Innovation data is reported in Chapter 12.

32


• The existence of unique firm identifiers in each of the BR, EC and PS datasets. This is necessary to enable firms to be matched across different datasets.

• The year to which the data pertain is also a mandatory variable6.

• Every firm must be assigned to an industry code. Country specific industry codes are linked to the uniform project industry classification (see Chapter 4) through a user-defined concordance table, which assigns each country specific code to one of the project industry codes. Where country-specific industry codes exist in multiple input files and are not consistent, the project code uses a hierarchy to assign a single industry code to each firm.

15. Some issues that have arisen in preparing data are as follows:

• Inconsistent naming of variable in different vintages of the source data, for example, where survey question numbers have changed over time. The project code only allows a one-to-one mapping of variable names, and project members must therefore ensure that variables are named consistently in every annual input dataset.

• Missing variables in some survey years, for example, where new questions have been added to the E-Commerce survey. Users must ensure that all variables that are assigned to project variables are found in every annual input dataset read by the code. Where variables are missing in some of the source files, place-holders must be added and populated with missing values7.

• It may be necessary to pre-merge data from different source files into the input datasets to be read by the project code. For example, the code is specified to read multinational and foreign ownership flags, and the age of the firm, from the BR input dataset, and data on labour skills (if they exist) from the PS dataset. In practice, such data may exist outside the business registers and structural business surveys, in which case they must be merged in to create the input datasets.

• Firm level data in Germany are collected and held at regional (Länder) level. These data have been integrated into synthetic national firm-level datasets for the purpose of this project.

• There may be a need for data cleaning. Boolean variables such as PC should take the values 0, 1, or missing. Continuous variables such as PCPCT should be bounded 0-1 (or missing) in all survey years. Similarly, variables denominated in national currency units in the PS survey, such as value-added, should be in consistent units across all survey years. All EC variables, and some of the PS variables such as employment should be non-negative.

• Sample re-weighting (see Chapter 5 below) uses a PS variable WGT_PS which represents the weight of that firm within the PS sample frame. This variable should be ≥ 1, that is, standing for the number of firms represented by that record in the dataset. Where native data has probabilities (WGT_PS ≤ 1), users must substitute the inverse in the input files to be read by the code.

• Users must ensure that their concordance table links every industry code that appears in their BR, PS and EC datasets to the project industry classification, see Chapter 4.

16. One finding of the project is that the quality of firm-level data varies across NSIs.

6 All project data are annual, either flows over the year or snapshots at the year end. 7 All missing values must be represented by the appropriate symbol for the software used, typically a “.” character. Local symbols such as -9999 or zeroes must not be used.

33


3.3.4 Code development 17. Marrying identical code with varying national data availability has been achieved by building dynamic flexibility into the code. For example, apart from a small number of core variables without which the analysis cannot proceed (see above) the code is designed to allow complete flexibility with regard to data availability. In running the code, project members in each NSI assign country-specific variable names to each project variable, entering a null value if the project variable does not exist in their input datasets. The code then builds dynamic lists of variables that exist.

18. Similarly the code is dynamically flexible with respect to time periods and will read all annual EC and PS input datasets located in an input directory named by the user in setting up the program run control file. This feature also deals with non-consecutive surveys.

19. The outputs of the core project code are broadly twofold:

• A set of industry / country indicators which are described in more detail in Chapters 4, 5 and 14, and

• A set of results based on running identical regression specifications on the underlying matched firm-level data in each project country. Outputs from these regressions are described in Chapter 6.

20. The code runs in SAS with a duplicate version available in STATA8. The SAS code was developed by Eric Bartelsman and tested and trialled on UK firm-level data by ONS. The STATA version was developed by ONS. Outputs from the STATA version have been tested against the SAS outputs using ONS firm-level data.

3.3.5 Running the code 21. Once the input data have been assembled and checked as described in section 3.3.3 above, it is a comparatively straightforward process to execute the analytical code. For the most part, it is simply a matter of populating the program run file with local parameters to match project variables with local names. The relevant section of the program run file is shown in Figure 3.2, where, for example, “multifl” is the local name for the multinational flag on the business register, and sample weights on the production survey (wgt_ps) are missing.

22. Users are also required to specify in the parameter “typestr” which of up to four productivity metrics to use. These metrics are as follows:

• LPQ – log of real gross output per employee. Note that firm-level data are normally only available in nominal terms. Nominal values are deflated using EU-KLEMS industry level deflators, stored in “pfile”. These deflators are not necessarily the most disaggregated that are available for each country, but this methodology has the advantage of being common to all project countries and being consistent with the aggregation structure of the project. Note also that, since all regressions include time dummies, it can be shown that the regression results do not depend on the choice of deflator.

• LPV – log of real value-added per employee.

• TFP – a log index of real value added divided by weighted inputs of labour and capital, with weights derived from the average factor shares of labour and capital in each industry.

• MFP – a log index of real gross output divided by weighted inputs of labour, intermediate inputs and capital, with weights derived from the average factor shares of labour, intermediate inputs and capital.

8 Due to time and resource constraints the STATA version does not currently include the coding to run regressions on the matched firm-level data.

34


23. Generally users are expected to populate “typestr” as fully as possible given local data availability. Thus if the source data include gross output (“nq”), value added (“nv”), intermediate inputs (“nmat”) and capital stocks (“k”) then the code can generate all four productivity metrics. If “nmat” data are not available, then MFP cannot be computed and should be excluded from “typestr”. If there are no “k” data then neither TFP nor MFP can be computed, and so on. Figure 3.2: User-defined run parameters ccc = TST,

typestr = LPV LPQ TFP MFP,startyr = 2000,endyr = 2004,survlib = dirIN,unit = firm_id,year = yyyy, BusReg = br,ind_BR = indn,emp_BR = empment,frgn_own = foreign,mnc = multifl,age = age,pre_PS = survey,ind_PS = indn,nv = nvadd,nq = ngo,e = emp,pay = payroll,nm = nmat,k = kstock,hkpct = hk_emppct,

hkitpct = it_hkemppct, hknitpct = nonit_hkemppct, export = exportflag,

wgt_PS = ,pre_EC = ecommerce_survey,ind_EC = indn,PC = PC,PCpct = PCpct,Web = Web,Epurch = Epurch,Epurchpct = Epurchpct,Esales = Esales,Esalespct = Esalespct,Inter = inter,Interpct = interpct,Intra = intra,Intrapct = intrapct,DSL = DSL,pfile = TSTdefl,concfile = conc,

hierfile = althier, keepBR = false, DoReWgt = true, DoCov = true, DoIndd = true, DoStat = true, DoCorr = true, DoReg = true, DoTabs = false, debug =0

35


24. Users may also choose to limit the number of productivity metrics to speed up execution time, e.g. while testing that the input datasets are addressed and read correctly and that program outputs are as expected.

25. Towards the bottom of Figure 3.2, “concfile” is a user-defined concordance file used to link all local industry identifiers to the most disaggregated industry classification used in the project (“euk0”, see Chapter 4) and “hierfile” is a file provided by the project co-ordinators to map between “euk0” and higher level industry aggregations.

26. Lastly, the program run file includes a number of switchable modules which control the execution of the analytical code:

• “DoReWgt” – controls a module which generates statistics for each variable based on a range of weighting methodologies, including sample re-weighting which uses characteristics of firms in the sample and in the full business register. See Chapter 5.

• “DoCov” – controls a module which computes basic statistics from the local input datasets, such as numbers of firms and numbers of employees in each industry.

• “DoIndd” – controls a module which computes industry dynamics, including the measure of churn used in the analysis reported in Chapter 14.

• “DoStat” – controls a module which generates statistics from the PS, EC and linked PS-EC input datasets, such as means of input variables and derived variables (such as DSLPCT) by industry, size-class, year, and status (e.g. statistics of multinational firms, exporting firms etc).

• “DoCorr” – controls a module which computes correlations between lists of variables, including full sample correlations and quartile correlations (that is, correlations between x and y for each quartile of the distribution of y). Again, correlations are reported by industry, size class, year and status.

• “DoReg” – controls a module which runs regressions on the firm-level data and writes results to an output file. This module is described in more detail in Chapter 6.

27. Two further control flags, “KeepBR” and “DoTabs” are redundant features from earlier vintages of the program code.

3.3.6 Disclosure 28. As mentioned above and shown in Figure 3.1, the output of the project analysis in each NSI is a set of output files containing statistics derived from the input firm-level data and aggregated over industries, size classes and other categories. The statistics include means and totals, standard deviations, correlations and regression results.

29. The output datasets also include the number of firms represented by each cell. In many cases the number of firms is measured in hundreds or thousands but in some case the number may be quite small. An example might be for variables drawn from the linked PS-EC surveys for small industries and small size classes.

30. The process of disclosure control varies across project countries, depending on legal frameworks and custom and practice. Thus while some countries check the outputs and suppress certain results before releasing the outputs to the project co-ordinator (for example by suppressing all outputs where the number of underlying firms is less than, say, 10) other countries carry out no disclosure tests at this stage, but reserve the right to check for disclosure at the project reporting stage. In general, the number of results that are suppressed prior to release of individual country datasets to the project co-ordinator is fairly small.

31. Once country datasets have been approved for release, they are sent to the project co-ordinator where they are combined with the outputs from other countries and held

36


securely within the project framework. Given that some countries have reserved their rights to test for disclosure until the project reporting stage, access to the combined datasets is restricted to nominated and approved researchers from within the project, and is restricted to research conducted under the terms of reference of this project.

32. Lastly, it is important to re-emphasise that data held in the cross-country datasets are purely for research and do not have the status of national statistics.

3.4 Methodology: Rationale and Feasibility 33. The background to the data-linking approach used in this project is set out in Chapter 2 above. In summary, the approach is a combination of co-ordinated firm-level analyses carried out in each separate country, and a programme of analysis of country / industry datasets built from aggregation of comparable linked data in all (or most) countries in the project. This latter approach is referred to as a “distributed micro data” (DMD) approach.

34. It may be helpful to illustrate the two methodologies by comparing regression specifications:

(i) Regressions on industry/country (DMD) data (as in Chapter 14)

dummieshrsakakaICTaav NITijt +++++= 43210

Where: ijtv real value added per employee, in industry i, country j, year t

ICT indicator of ICT usage for industry i, country j, year t from E-Commerce survey, such as DSLPCT

kIT, kN, hrs IT capital stocks, non-IT capital stocks and hours worked (all three variables taken from EU-KLEMS dataset

dummies 2 of industry, country and time dummies.

(ii) Regressions on firm-level data (as in Chapter 6)

dummiesLNWbICTbKbbtfp oz ++++= 321

Where: ztfp total factor productivity for firm z

K capital stock for firm z ICT indicator of ICT usage for firm z LNW implied firm-level wage taken from firm employment and wage bill dummies industry, size-class, year and other dummies such as multinational status.

3.4.1 The Distributed Micro Data (DMD) approach 35. The rationale for the DMD approach is set out in Bartelsman and Barnes (2001)9. In the context of this project, DMD refers to the process of compiling conceptually identical indicators at a relatively disaggregated industry level across multiple countries and multiple time periods. Bartelsman and Barnes (2001) provide two arguments in favour of this approach:

• The DMD approach provides an improved trade-off between timeliness and comparability. It is more timely than, say, waiting for Eurostat to harmonise statistics at source, and more comparable than, say, EU-KLEMS data derived from disaggregation of higher level national statistics;

9 See also Bartelsman, Haltiwanger & Scarpetta (2004) which discusses why the DMD approach may be attractive for policy analysis.

37


• The DMD approach involves confronting basic policy questions with the data realities, and a process of making choices regarding the analyses that can be done. This is a subtle but important point – clearly there are limits on the data that are available to be collected, and equally there are policy and other research issues that cannot adequately be addressed by the data that are available. Effectively, the DMD approach involves an iterative process between policy questions and data realities.

36. This iterative process is clearly reflected in this project, firstly in the refinement of the scope of the core analysis and analytical sub-themes, and secondly in the development of the set of data to be collected. For example, fast internet usage was not originally included in the dataset but has been added as the project has evolved, while other variables that were initially viewed as conceptually important – such as firm profitability, international engagement and ICT investment – have either been discarded or confined to sub-themes.

37. In addition to the two arguments set out in Bartelsman and Barnes (2001), it can be argued that the DMD approach is attractive for policy analysis. The intuition here is that, in any single country, the impact of a policy event cannot be measured very precisely because there is, by definition, only one observation of that policy event. Cross country datasets can help by providing more observations of policy events. But in addition, the DMD approach allows summary statistics from the underlying firm-level data to be captured within the country / industry datasets. For example, the project has generated observations of (means of) fast internet usage and productivity metrics by industry, country and year which can be expressed as a scatter plot. But behind each observation in the scatter plot, the DMD dataset contains a suite of variables (“indicators”) describing the properties of the firm-level data, such as the variance of the firm-level data, and quartile correlations between each variable and other variables of interest such as wage levels, size of firm etc.

38. This integration of firm-level properties with the richness of comparable data by industry, county and time is a key feature of the DMD approach.

3.4.2 Co-ordinated firm-level analysis 39. As described in Chapter 6 below, the project has conducted co-ordinated firm-level analysis of productivity and core ICT metrics, building on previous work and exploiting the development of comparable linked firm-level datasets.

40. The rationale for this line of work is more pragmatic and opportunistic than the rationale for the DMD approach set out above. There is a natural interest in, say, the comparability of firm level relationships between ICT-usage and productivity across different countries, even if the “meaning” of these relationships is not thoroughly grounded in economic theory, and even if the relationships are not stable across countries.

41. An example is the employment effects of ICT adoption. Classical economic theory suggests there should be no such effects, but in practice this is an issue of interest to policy makers.

42. In this regard we see the project as highlighting questions for further research, rather than providing all of the answers.

3.4.3 Feasibility of the methodology 43. The twin track methodology has been challenging and has taken longer than anticipated in the original project timetable. Some of the issues that have arisen have been described earlier in this chapter. However, despite the challenges, the project has delivered distributed micro data outputs for all counties including Denmark, which withdrew from the project in early 200810. We have also conducted co-ordinated firm-level regression analysis

10 Outputs for Denmark are, however, only available from an early version of the project analytical framework and do not contain the full functionality of the final version.

38


39

on productivity and core ICT metrics for 11 countries. We do not have regression results for Denmark and Slovenia.

44. The experience of conducting the analysis across the fairly broad sample of countries represented in the project gives rise to a number of learning points:

• The basic requirements to engage usefully in the project are quite limited but include the ability to access firm level data from a range of sources, the ability to match firms across different surveys and to possess reasonable (but not necessarily comprehensive) coverage of the data used in the project).

• Having said that, there is a clear relationship between the maturity of NSIs in terms of their experience of working with firm-level data and the resource requirement (time and effort) for successful engagement in the project. Those NSIs with prior experience of the kind of analysis employed in the project have generally found it more straightforward to apply the analysis used here than have those NSIs with less experience.

• The computational and technical skills required to participate in the project are not especially pronounced. In practice we have found that a combination of basic technical proficiency in data manipulation and a reasonable understanding of economics has been most effective.

45. The common structure of the project code, and the metadata and data cleaning tasks which have preceded running the code, mean that analytical modules can be added relatively easily and that these modules can be replicated across countries. As an example, 6 countries have run identical code to incorporate additional ICT Business Integration variables and run regressions following the framework of the common firm-level regressions, as described in Chapter 6.


Chapter 4

Describing the Data Peter G. Stam, UK Office for National Statistics

4.1 Disclaimer 1. All data in this report complies with statistical disclosure measures and standards throughout all project member countries.

2. Furthermore data refer to experimental and provisional datasets. They should not be treated as or compared to any official national statistics.

4.2 Introduction 3. The cross country merged dataset has many characteristics which are important to understand before any useful analysis can take place. This chapter explains methods used in collecting data and provides a summary of the contents and characteristics of the cross-country and merged datasets.

4. The ICT impacts project makes use of firm level, cross country data. This is collected from each country’s microdata, aggregated to industry level and merged into a common dataset in a central location. Comparable data are achieved through a combination of rigorous metadata checks and the use of a common piece of code which is run in every country on the microdata. The code has been developed in collaboration with every member country and collects all of the data - where available - required for analysis. The development of the code is an important phase of the project, a discussion process with project members where representatives and academics finalise the content of core code and themes. The scope of this discussion includes the analysis that the project shall focus on and the variables that shall be collected and merged to facilitate this.

5. The agreed code is distributed to the project members who run it on their individual firm level microdata. The code runs through several datasets and collects the agreed information. It then maps the data to agreed aggregation levels and reports twelve output files. The output files follow the same format for each country and so can be easily merged later. These files are then reported to a central location via a dedicated, secure and encrypted electronic file transfer protocol for merging and distribution.

6. The chapter begins with a short review of the sources of data, briefly explaining the metadata process and the aggregation methods. Descriptive statistics then follow regarding the E-commerce data (including access to ICT, use of ICT, E-buying and E-selling) and production data (including productivity metrics, employment and wages). Finally the characteristics of the merged data are examined.

4.3 Sources of Data 7. The project makes use of variables from several different sources. Source surveys vary between countries but generally include a Business Register, Structural Business Survey and the Community Survey on ICT Usage (E-commerce). Countries may also use administrative data including tax returns or company accounts to gather some variables.

8. One of the fundamental challenges facing the project is the merging of the various source surveys and ensuring variable names, formats and units etc are compatible.

40


9. Permission has been granted to use National Accounts data from the EUKLEMS project from which the industry classification structure has been adapted. These data are now publicly available (from http://www.euklems.net/). This classification is used across the whole project alongside other systems of aggregation. Deflators have also been derived from EUKLEMS data.

10. Years that the project is interested in are 2000 – 2005. However it is not necessary for survey data to be available for all years and there is not total coverage over every year. It is normal practice to include at least one year of production and business data prior to the earliest available E-commerce data (this is to facilitate computation of lagged variables).

11. The ‘merged dataset’ is the resulting panel dataset containing both E-commerce and Production Survey variables for all countries and all years. Descriptive statistics are discussed later in this chapter. The key variables available in the merged dataset are listed in Table 4.1.

Table 4.1 – Key Variables Variable Description PC Firm uses computers Web Firm has website Epurch Firm purchases through Internet Esales Firm has sales through Internet Inter Firm has Internet Intra Firm has intranet DSL Firm has broadband Internet PCpct % of workers using computers Epurchpct % of purchases ordered through Internet (or EDI) Esalespct % of sales made through Internet (or EDI) Interpct % of workers with access to Internet Intrapct % of workers with access to intranet (derived variable) DSLpct % of workers with access to broadband (made from DSL and Interpct) COUNTRY Country code YEAR Year to which data pertain EUK Industry classification (EUK or ALT definitions) SZ_CLS Size class (based on employment) MNC Multinational Dummy 0 or 1 (1 = multinational firm) FRGN_OWN Foreign Ownership Dummy 0 or 1 (1 = owned by foreign firm) SRC Aggregation selection variable WGT_PS Sample weight variable NQ Nominal Sales NV Nominal Value Added NM Nominal material inputs PAY Payroll E Employment WAGE Derived mean wage (logged WAGE available) K Capital Stock LPQ Labour productivity (based on gross output) LPV Labour productivity (based on value added) TFP Total factor productivity (based on value added) MFP Multi factor productivity (based on gross output)

41

http://www.euklems.net/


4.4 Metadata 12. In order to usefully describe the data it is important to understand the process involved in gathering the data, especially the metadata checking that has been conducted.

13. Many of the Information & Communication Technology (ICT) variables are sourced from the Community Survey on ICT usage and E-commerce in enterprises (E-commerce survey). The E-commerce survey is an annual survey with a combination of Boolean and percentage variables. There are systems in place to ensure a certain amount of compatibility and consistency between countries’ data.

14. From its commencement in 2000 to 2005 the E-commerce survey was conducted by ‘gentleman’s agreement’ between individual NSIs and Eurostat. Eurostat provided model questionnaires with compulsory benchmark questions. If a country wished to include extra questions for local policy reasons, these were additional to the minimum recommended set. Most EU member and accession countries conducted the common survey, but with a certain amount of variation and experimentation. This means that metadata shows more variation up to 2005. It also means there are more gaps in data up to 2005 – including some countries which did not carry out the survey every year.

15. From 2006 the survey is now covered by regulation, which means that Eurostat requires a set of standardised questions to be covered in business surveys, with common minimum sampling and coverage across the EU. Individual NSIs are still free to add extra questions to meet local policy interest, and to add to sample size to meet local research needs. The standard Eurostat questionnaire, from 2007 onwards, contains a variable module each year to cover a topic of high policy interest across the EU, for example the 2008 survey contains a module on e-business.

16. Despite this, differences are still bound to occur; be they differing interpretation of standards, cultural differences and historical formatting of surveys or issues specific to a country’s statistical office. For example, France were unable to survey 2002 data. The model survey is written in English and linguistic complications can mean that responses do not always contain information quite as intended. Also some areas of potential interest are not classed as benchmark indicators, and therefore may not be included across multiple countries. For example only Finland has asked specific questions in some of their survey years about ICT mobility and employees using laptops.

17. Similarly other variables are gathered from a wide range of sources which may differ between participating countries and must be checked and adjusted to ensure total comparability across the cross country dataset. The project has audited all of the available survey data which has been used in firm level analyses, and then gone on to assemble a record of metadata across all thirteen countries which describe the source data.

18. Each country has provided information (described in Table 4.2) pertaining to variables used in analysis (both E-commerce variables and Production Survey).

19. These metadata tables are stored in a central location and are outcomes of an iterative process of establishing what data exist. The metadata tables are available upon request; please contact [email protected] for details about access. The metadata tables have been used both to ensure comparability but also to derive ICT use indicators modulated by industry / country and by a range of other business descriptors. This is also potentially a powerful tool for diagnosing causes of output errors and spurious results as it reveals where differences and gaps remain. It is the first place consulted when unexpected results appear in the output datasets as it is simpler to examine and change source data for one country than to change the core code and re-run it for every country.

42

mailto:[email protected]


Table 4.2: Example metadata questions asked of all variables Field Notes Quote of survey question Please identify which question in the survey the specified variable is

linked with. Translated into English. If multiple choice, please include the options given.

Format This will probably be either percentage or Boolean for the E-commerce variables.

Range State the numerical range. For example a Boolean variable might be 0 , 1. A percentage variable may be 0 – 100

Unit This is expected to be either % or Boolean, please confirm

Source Survey Name Where is the question linked to this variable found?

Source Survey Frequency How frequently is the question linked to this variable asked?

Timeliness If the survey assumes a point in time, when is this?

Level of firm to which the survey is aimed

At what level is this survey aimed? For example the UK's E-commerce survey is aimed at the 'Enterprise' level

Sampling Frame What sampling frame is used? (e.g. Business Register)

Additional metadata comments Please use this space to identify any additional metadata concerns or comments

20. The metadata framework employed by this project therefore permits the inclusion of variables which are completely comparable across countries and auditable back to source data. It also allows us to generate data which can be used for analysis across countries and industries of the differences in behaviour and performance associated with ICT use. This new resource has also been linked to National Accounts output and employment data, which provides another perspective on technology and productivity.

4.5 Aggregation 21. Firm-level data are aggregated across various dimensions. Data can be split across industry classification, year, size class etc.

22. Datasets for the ICT Impacts project are aggregated by two similar industry classification types. At the lowest level of aggregation, data are aggregated for seventy one industries, using standard Eurostat and National Account approved industrial breakdowns which are classified very closely to the International Standard Industrial Classification (ISIC).

23. The alternative (and related) structure - for purposes of this project named “Alt” - features bespoke grouping for the project and assembles industries into forty two classifications (see Table 4.3).

24. The advantage of using the alternative aggregation is that it splits ICT producing industries from ICT using industries. It allows analysis of ICT production, manufacturing using ICT, common services and differentiated services. For the purposes of this chapter data shall be described using the ALT aggregation (unless otherwise specified).

43


Table 4.3: ALT Industry definitions Alt Code Description ELECOM ELECTRICAL MACHINERY, POST AND COMMUNICATION SERVICES 30t Electrical and optical equipment 64 Post and telecommunications MexElec TOTAL MANUFACTURING, EXCLUDING ELECTRICAL Mcons Consumer manufacturing 15a6 Food products, beverages and tobacco 17t9 Textiles, textile products, leather and footwear 36a7 Manufacturing – Not elsewhere classified (NEC); recycling Minter Intermediate manufacturing 20 Wood and products of wood and cork 21a2 Pulp, paper, paper products, printing and publishing 23 Coke, refined petroleum products and nuclear fuel 24 Chemicals and chemical products 25 Rubber and plastics products 26 Other non-metallic mineral products 27a8 Basic metals and fabricated metal products Minves Investment goods, excluding high-tech 29 Machinery, NEC 34a35 Transport equipment OtherG OTHER PRODUCTION 10t4 Mining and quarrying 40a1 Electricity, gas and water supply 45 Construction 1t5 Agriculture, hunting, forestry and fishing MSERV MARKET SERVICES, EXCLUDING POST AND TELECOMMUNICATIONS DISTR DISTRIBUTION 50 Sale, maint. and repair of motor veh. and motorcycles; retail sale

of fuel 51 Wholesale trade and commission trade, ex motor veh and

motorcycles 52 Retail trade, ex motor veh and motorcycles; repair of household

goods 60t3 Transport and storage FINBU FINANCE AND BUSINESS, EXCEPT REAL ESTATE 65t7 Financial intermediation 71t4 Renting of m&eq and other business activities PERS PERSONAL SERVICES 55 Hotels and restaurants 90t3 Other community, social and personal services 95 Private households with employed persons NONMAR NON-MARKET SERVICES 75 Public admin and defence; compulsory social security 80 Education 85 Health and social work 70 Real estate activities

44


Table 4.4: Commonly used Industry codes Code Description DISTR Distribution Elecom Electrical machinery, post and communication

services FINBU Finance and business (except real estate) MexElec Total manufacturing (excluding electrical) OtherG Other production NONMAR Non-market services TOT Total economy

Table 4.5: Country codes Code Country AUT Austria CZE Czech Republic DNK Denmark FIN Finland FRA France GBR United Kingdom GER Germany IRE Ireland ITA Italy NLD Netherlands NOR Norway SLO Slovenia SWE Sweden

Table 4.6: Categories Category Tabulation by: 1 Euk, Year 2 Euk, Year, Sz_Cls 3 Euk, Year, Mnc 4 Euk, Year, Frgn_Own 4.2 Euk, Year, Hge 4.4 Euk, Year, Gzl 5 ALT, Year 6 ALT, Year, Sz_Cls 7 ALT, Year, Mnc 8 ALT, Year, Frgn_Own 9 ALT, Year, Hge 10 ALT, Year, Gzl

Table 4.7: Size classes Size class

Employment

0 emp=0 1 0≤emp<10 2 10≤emp<20 3 20≤emp<50 4 50≤emp<100 5 100≤emp<250 6 250≤emp<500 7 Emp≥500

Table 4.8: Productivity Metrics Variable Description LPQ Labour productivity (based on

sales) LPV Labour productivity (based on

value added) TFP Total factor productivity (value

added, capital and labour) MFP Multi factor productivity (gross

output with capital, labour and materials)

25. Data can be easily analysed at any industry or group of industries using the SRC and EUK variables. Table 4.9 gives an example of UK E-commerce data when src=5. This simply breaks the data into industry level aggregation.

45


Table 4.9: Screenshot – category = 5

country Euk Year Sz_Cls category pc Epurch Esales inter intra GBR 70 2004 5 98.04 62.75 11.76 98.04 92.16GBR 71t4 2004 5 99.43 71.07 14.56 99.24 71.29GBR 85 2004 5 100.00 63.64 6.82 97.73 56.82GBR 90t3 2004 5 97.92 65.73 32.17 97.22 54.17GBR DISTR 2004 5 100.00 68.22 34.42 99.94 67.98GBR Elecom 2004 5 100.00 73.33 33.33 100.00 81.90GBR FINBU 2004 5 99.54 71.07 14.56 99.39 75.11GBR Mcons 2004 5 99.33 63.55 28.76 99.00 63.67GBR MexElec 2004 5 99.78 68.28 27.86 99.67 74.15GBR Minter 2004 5 100.00 71.39 29.11 100.00 76.46GBR Minves 2004 5 100.00 69.16 24.30 100.00 84.58GBR MSERV 2004 5 99.76 68.58 29.97 99.64 69.08GBR NONMAR 2004 5 98.95 63.16 9.47 97.89 75.79GBR OtherG 2004 5 100.00 69.12 13.97 100.00 84.56GBR PERS 2004 5 98.73 65.53 34.04 98.31 59.75GBR TOT 2004 1 99.77 68.66 28.53 99.63 71.70

26. Table 4.10 demonstrates how the data can be carved further. For example, by examining data where category=6, a user introduces size class to the aggregation. In the table below data are displayed by broad industry sector where the average employment level is larger than 50 employees.

Table 4.10: Screenshot - category = 6 country euk year Sz_Cls src pc Epurch Esales inter intra GBR DISTR 2004 4 6 100.00 65.71 29.26 100.00 54.20GBR DISTR 2004 5 6 100.00 69.46 25.94 100.00 64.44GBR DISTR 2004 6 6 100.00 71.68 34.18 100.00 68.70GBR DISTR 2004 7 6 100.00 67.10 42.28 99.82 79.60GBR Elecom 2004 4 6 100.00 64.29 21.43 100.00 53.57GBR Elecom 2004 5 6 100.00 76.92 34.62 100.00 69.23GBR Elecom 2004 6 6 100.00 78.69 26.23 100.00 75.41GBR Elecom 2004 7 6 100.00 71.58 41.05 100.00 97.89GBR FINBU 2004 4 6 100.00 66.67 9.18 100.00 67.86GBR FINBU 2004 5 6 98.86 72.97 12.16 98.86 57.95GBR FINBU 2004 6 6 99.27 71.30 15.65 98.54 72.99GBR FINBU 2004 7 6 99.75 70.79 15.56 99.75 80.10GBR MexElec 2004 4 6 98.41 50.00 19.05 98.41 42.86GBR MexElec 2004 5 6 100.00 63.37 21.78 99.02 53.92GBR MexElec 2004 6 6 100.00 70.04 24.91 100.00 76.90GBR MexElec 2004 7 6 100.00 74.01 34.16 100.00 87.13

46


4.6 Descriptive statistics – Characteristics of the data

4.6.1 E-commerce data Coverage

27. All countries have contributed E-commerce data to this project. Data are available for the years as described in Table 4.11.

Table 4.11: E-commerce coverage

Year AUT CZE DNK FIN FRA GBR GER IRE ITA NLD NOR SLO SWE 2000 ● ● ● ● 2001 ● ● ● ● ● ● ● ● 2002 ● ● ● ● ● ● ● ● ● ● ● 2003 ● ● ● ● ● ● ● ● ● ● ● ● ● 2004 ● ● ● ● ● ● ● ● ● ● ● ● ● 2005 ● ● ● ● ● ● ● ● ● ● ●

Access to ICT

28. The time period covered by this project is 2000 – 2005. Year labels refer to the year that data pertain to. For example the 2006 E-commerce survey is labelled 2005. This is because the survey collects information about 2005 (or occasionally some variables could refer to a point in time in early January 2006). Over the project years the E-commerce survey covers a lot of information about firm and employer access to ICT. Because of saturation levels it quickly became obvious that some of these variables would not be relevant (for example data about access to PCs or workstations). The survey also paid particular attention to the speed of Internet connection that firms enjoy. Merging this with employment and PC access data allows information for analysis about access to Internet, speeds of connection and percentage of employees with access. This signals a shift of attention away from level measures of simple access and investment in ICT wares and towards the utilisation and employment of ICT.

29. Table 4.13 simply shows an index of all project countries ranked in order of the percentage of workers with access to high speed Internet.

Table 4.12: Indexing key Most intensive

Least IntensiveNo Data

47


Table 4.13: Index of fast-Internet enabled workers in 2004

Elecom MexElec DISTR FINBU TOT Sweden Norway Finland

U.K. Netherlands

Slovenia Germany

Czech Rep Austria

Italy Ireland

30. Country rankings tend to be fairly consistent across all the industries (e.g. if a country ranks low in manufacturing, it is likely to be ranked low in services). This review of the data reveals some notable points of interest, for example the United Kingdom appears to rank uncharacteristically low in the Financial and Business services category. A possible explanation for this is that in the years surveyed the E-commerce questionnaire did not cover the banking sector.

31. A simple ‘Manufacturing’ / ‘Services’ split is shown in Figure 4.1, where

• Manufacturing: Euk code: 15t37

• Market Services: Euk code: 50t74

Figure 4.1: Fast Internet - Manufacturing vs. Market services

0

10

20

30

40

50

60

70

AUT CZE DNK FIN FRA GBR GER IRE ITA NLD NOR SLO SWE TOT

ManufacturingServices

32. The data show that it is the manufacturing industry which consistently has lower percentages of employees connected to fast speed Internet compared to the service

48


industry. Figure 4.1 shows this clear pattern, taken from the E-commerce data set it shows the mean over the whole time period for the variable DSLpct.

33. Figure 4.1 also demonstrates well the variability between countries. Some countries have lower high speed connectivity throughout both industries than others. This could be an indicator that the country is still relatively immature in terms of its ICT infrastructure. This could be a factor to be taken into consideration when analysis is conducted on the cross-country data.

Use of ICT

34. The E-commerce survey is not restricted to the collection of data about the level of ICT held by businesses; rather it is increasingly becoming focused towards how firms use their ICT equipment and systems. For example how a firm utilizes its ICT to conduct business (connecting their sales and ordering systems etc). E-sales and percentage of E-ordering of materials through the Internet has been increasing over the recent time period.

Table 4.14: E-commerce variables Variable Description Epurch The percentage of firms within the specified industry who uses the Internet to

purchase goods Esales The percentage of firms within the specified industry who uses the Internet to sell

goods Epurchpct Purchases ordered through Internet as a percentage of total purchases Esalespct Sales made over the Internet as a percentage of total sales Ecpct Sum of Epurchpct and Esalespct

Table 4.15: Average E-commerce over time Variable Growth of selected E-

commerce variables over time (Mean of countries available 2000 – 2005)

Epurch 53.0% Esales 76.8% Epurchpct 29.7% Esalespct 64.5%

35. Early review of the data shows that E-commerce variables (described in Table 4.14) have not yet reached saturation point (i.e. there are no observations approaching 100%). The issue of saturation is discussed later in this chapter.

36. It is useful to consider means across the entire merged cross-country dataset for the whole period (2000 – 2005). Figure 4.2 does this using the E-commerce data. For nine of the thirteen countries the trend appears to be that the service industry is more ICT intensive - in terms of percentage of goods bought or sold online - than the manufacturing industry. For the remaining countries the inverse is true. For those countries where the service industry is more intensive in E-buying and E-selling though, the difference between the two industries is much more dramatic.

37. Patterns emerging from the data show that E-purchasing appears more popular amongst firms than E-selling. A further industrial breakdown allows additional investigation. Figure 4.3 breaks the data down into the four E-buying / selling variables and compares the data over four broad industries (and the total economy) looking at the mean of all countries.

49


Figure 4.2: E-commerce (%) - Manufacturing vs. Services

0

10

20

30

40

50

60

70

AUT CZE DNK FIN FRA GBR GER IRE ITA NLD NOR SLO SWE TOT

ManufacturingServices

Figure 4.3: E-commerce - Industry split (all countries)

E-Commerce

0

5

10

15

20

25

30

35

40

45

50

epurch esales epurchpct esalespct

Mean 2000 - 2005

%

DISTRElecomFINBUMexElecTOT

38. The “Electrical machinery, post and communication services” industry has the highest proportion of firms who either sell or purchase over the Internet. However this does not relate directly to the relative percentage of purchases made online. In fact it is the Finance and Business sector with the highest percentage of purchases made over the Internet. No industry average in the data is above 10% of sales of purchases made online. There is longevity potential for these variables.

50


39. E-commerce questions are standard across all of the E-commerce questionnaires so it can be expected that there will be good coverage across the member countries. Tables 4.16 and 4.17 show countries split into broad industries and ranked (see Table 4.12 for a key to the index).

Table 4.16: Index of purchases made through the Internet in 2004 Elecom MexElec DISTR FINBU TOT

Norway

Austria

France

Germany

Czech Rep

Finland

Ireland

Netherlands

U.K.

Sweden

Italy

Slovenia

Table 4.17: Index of sales made through the Internet in 2004 Elecom MexElec DISTR FINBU TOT

Norway

France

Austria

Germany

Sweden

Finland

Czech Rep

Ireland

U.K.

Netherlands

Italy

Slovenia

40. It is not necessarily the case that those countries who rank higher in purchases made over the Internet are the same as those who rank highly in sales made over the Internet. There also appears to be quite a large amount of industrial variation within countries in terms of use of the Internet to buy and sell.

41. However, those who rank higher in E-selling and E-purchasing appear to do so across all major industries, those ranked lower tend to have fairly consistent ranking across all industries, with only the occasional exceptional industrial performance in relation to the total economy.

51


4.6.2 Production Survey data 42. Breaking down the data can give a different picture of the distribution of the observations available to the project. For optimal coverage, the figures that follow in this section are based on the Production Survey data for 2004 covering all available countries’ data (unless otherwise specified).

43. Firstly an industrial breakdown (Figure 4.4) shows that the majority of firms are classified within the “Market services” classification. This pattern holds true for every year in the merged dataset (2000 – 2005).

Figure 4.4: Industrial split of underlying firm observations in the merged dataset

2004

Eleco Other m5% G

9% NONMAR

1% Elecom= Electrical machinery, post and communication services MexElec= Manufacturing (excluding electrical) MSERV= MexElec

34% Market services NONMAR= Non-market services OtherG= Other production

MSER V51%

Productivity metrics

44. Breaking the productivity metrics into quartiles allows more detailed results to be obtained. The distribution of productivity is fairly consistent. Figures 4.5 to 4.8 examine productivity metrics relative to the mean (where the mean = 1). They show the country with the widest dispersion about the mean, the country with the narrowest dispersion about the mean and a typical country with middling dispersion. Total factor productivity and labour productivity have been chosen for quartile breakdown because they are most commonly used in analysis.

52


Figure 4.5: Dispersion about the mean. Total factor productivity. Quartile mean relative to industry mean (=1). Manufacturing

MexElec : TFP

0

0.5

1

1.5

2

2.5

3

1 2 3 4

Qrt

Mea

n (r

elat

ive

to q

rt m

ean)

CZEGBRNLD

Figure 4.6: Dispersion about the mean. Labour productivity based on value added. Quartile mean relative to industry mean (=1). Manufacturing

MexElec : LPV

0

0.5

1

1.5

2

2.5

3

1 2 3 4

Qrt

Mea

n (R

elat

ive

to q

rt m

ean)

GBRNORFIN

53


Figure 4.7: Dispersion about the mean. Total factor productivity. Quartile mean relative to industry mean (=1). Finance and Business services

FINBU : TFP

0

0.5

1

1.5

2

2.5

3

1 2 3 4

Qrt

Mea

n (R

elat

ive

to q

rt m

ean)

CZENORITA

Figure 4.8: Dispersion about the mean. Labour productivity based on value added. Quartile mean relative to industry mean (=1). Finance and Business services

FINBU : LPV

0

0.5

1

1.5

2

2.5

3

1 2 3 4

Qrt

Mea

n (R

elat

ive

to q

rt m

ean)

GBRNORCZE

45. There appears to be considerable variability around the mean when considering quartiled productivity. However, even in the contrasting industries of ‘Manufacturing’ and ‘Financial & business services’ the dispersion of productivity metrics follow broadly the same pattern. All observations show a keen increase in the fourth quartile (that is to say that those

54


55

firms in the fourth productivity quartile hold a disproportionate share of TFP and LPV; when compared to the mean of the industry).

Employment and Pay

46. Industrial breakdown of number of firms in the Production survey matches that of number of employees (as shown in Figure 4.4). Again, it is the ‘Market Services’ that holds the largest proportion of employees in this sample.

Figure 4.9: Industrial split of employee observations in the merged dataset

2004

Other G6%

Value Added

47. The Production survey holds data on the ‘nominal tends to be dominated by the ‘Market Services’ industrythe smallest proportion of the value added reported in th

NONMAR Elecom11% 1%

MexElec28%

MSER V54%

Elecom= Electrical machinery, post and communication services MexElec= Manufacturing (excluding electrical) MSERV= Market services NONMAR= Non-market services OtherG= Other production

value added’ indicator. Value added , with ‘Non-market services’ holding e survey.


56

Figure 4.10: Industrial breakdown of Nominal Value-Added in the merged dataset

2004

Other G10%

NONMAR 2% Elecom

23%

Elecom= Electrical machinery, post and communication services MexElec= Manufacturing (excluding electrical) MSERV= Market services NONMAR= Non-market services OtherG= Other production

MSER MexElecV45% 20%

48. In order to understand the wage dynamics within industries for the sample it is again useful to consider the data relative to the industry mean. In the charts that follow the data have been carved into size classes by broad industry (see table 4.7). The average country wage for that industry is displayed relative to the industry average and normalised to 1. The reason for doing this is to abstract from differences in currency units etc. Select data are shown for the most and least dispersed countries (and a typical country). The bold line displays the mean for all countries.

Figure 4.11: Wage dispersion about the mean by size class. (Mean=1) Manufacturing

Wage : MexElec

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

0 1 2 3 4 5 6 7

Size Class

Wag

e (r

elat

ive

to m

ean

over

all

size

cla

sses

)

CZENLDFRATOT


Figure 4.12: Wage dispersion about the mean by size class. (Mean=1) Electrical machinery, post and communication services

Wage : Elecom

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

0 1 2 3 4 5 6 7

Size Class

Wag

e (r

elat

ive

to m

ean

over

all

size

cla

sses

)

GBRSWETOT

49. The ‘Manufacturing’ and ‘Electrical machinery, post and communication services’ industries display the same general pattern. Smaller firms tend to pay below average wages to their employees. As the firms employ more workers they tend to pay more and it is the firms in the largest size classes who pay most (relative to the industry average). This does not hold true, however, for the ‘Financial and business services’ or ‘Distribution’, as displayed in Figures 4.12 and 4.13.

Figure 4.13: Wage dispersion about the mean by size class. (Mean=1) Financial and business services

Wage : FINBU

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

0 1 2 3 4 5 6 7

Size Class

Wag

e (r

elat

ive

to m

ean

over

all

size

cla

sses

)

ITASWETOT

57


Figure 4.14: Wage dispersion about the mean by size class. (Mean=1) Distribution

Wage : DISTR

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

0 1 2 3 4 5 6 7

Size Class

Wag

e (r

elat

ive

to m

ean

over

all

size

cla

sses

)

TOTCZEGBR

50. Similar to the ‘Manufacturing’ and ‘Electrical machinery, post and communication services’ industries, the data show that smaller firms tend to pay below industry average wages. It does not hold true, however, that it is the largest firms who pay - relatively - the most. The middling size classes (firms with approximately 50 – 250 employees) pay most above the industry average, with the larger firms tending to pay below industry average wages.

4.6.3 Merged data 51. The project has achieved good data coverage, using responses from every participating country. Because of rigorous metadata procedures, the data are comparable and accessible in a single, merged dataset.

52. The merged E-commerce survey and Production survey dataset consists of aggregated, industry level data. The industry data though, is made up from a large number of many underlying firm level observations across six consecutive years. The following results are taken from the merged data (unless otherwise specified).

58


Table 4.18: Coverage statistics – Sum of available underlying firm observations aggregated into merged dataset (All years available: Total economy)

Country Number of observations in production survey

Number of observations in the merged dataset

Overlap with production survey

AUT 1,145,116 14,970 1.3% CZE 93,253 15,422 16.5% FIN 751,500 16,033 2.1% FRA 5,975,993 41,319 0.7% GBR 324,202 19,461 6.0% GER 675,075 20,843 3.1% IRE 63,293 11,244 17.8% ITA 186,919 24,833 13.3% NLD 206,756 13,689 6.6% NOR 326,082 10,428 3.2% SWE 3,773,595 14,309 0.4% TOTAL 13,521,784 202,551 1.5%

53. Table 4.18 displays the overlap that the E-commerce survey has with the production survey. It can be seen that there is a large amount of variability between countries (ranging from 0.4% to 17.8%. Predictably those with the least percentage overlap are those countries with the largest production survey samples. Namely France and Sweden. This is likely to be because in France the Production survey is conducted on a census basis and in Sweden the data are collected from compulsory tax returns. Despite the variability, the project has succeeded in collecting a large sample in the cross country merged dataset.

54. As well as this large number of merged observations, the project has achieved good coverage of the features required for comprehensive analysis of the data. Reported variables held across the project countries in the merged dataset are shown in Table 4.19.

Table 4.19: Availability of cross country data in merged dataset

Description AUT CZE DNK FIN FRA GBR GER IRE ITA NLD NOR SLO SWE Employment ● ● ● ● ● ● ● ● ● ● ● ● ● Sample Weight ● ● ● ● ● ● ● Sample Reweighting ● ● ● ● ● ● ● ● ● Multinational ● ● ● ● ● Ownership flag ● ● ● ● ● ● ● ● High growth firms ● ● ● ● ● ● ● ● ● ● ● Gazelles ● ● ● ● ● Human capital ● ● ● Gross output ● ● ● ● ● ● ● ● ● ● ● ● Value added ● ● ● ● ● ● ● ● ● ● ● ● ● Nominal materials ● ● ● ● ● ● ● ● ● ● Payroll ( wage) ● ● ● ● ● ● ● ● ● ● ● ● Capital Stock ● ● ● ● ● ● ● ● ● ● ● Productivity – LPQ ● ● ● ● ● ● ● ● ● ● ● Productivity – LPV ● ● ● ● ● ● ● ● ● ● ● ● Productivity – MFP ● ● ● ● ● ● ● ● ● Productivity – TFP ● ● ● ● ● ● ● ● ● ● ●

59


Saturation of ICT variables across the economy

55. An important question facing the project is how long will the ICT variables be relevant? Reason suggests that workers’ access to computers and workstations could be a relevant variable for the timeframe that this project examines; but is it today and will it be in the future? ICT variables can lose explanatory power over a short amount of time as usage reaches saturation levels quickly. For example in the near future ‘fast Internet’ variables are likely to become less relevant as high speed connections become the standard.

Figure 4.15: Cross-country mean of ICT variables over time

Saturation of ICT variables

0

10

20

30

40

50

60

70

80

90

100

2000 2001 2002 2003 2004 2005

Cro

ss-C

ount

ry M

ean

(Tot

al E

cono

my)

PCepurchesalesIntraInterWebDSL

56. Figure 4.15 demonstrates how PC and Internet level variables are now close to saturation point in the merged dataset. If fast Internet (DSL) follows its historical pattern of rapid growth then it shall soon reach saturation as high speed Internet becomes the standard. The discriminating factor can change quickly in the area of E-commerce and ICT. There are always new ICT variables that are relevant, which is why the survey is updated often.

Finding a proxy for skills

57. Only three of the twelve project countries have explicit human capital data (Finland, Norway and Sweden). It is, therefore, pertinent to ask whether a robust proxy exists that the majority of countries hold which can reliably by used as a measure of skills. The obvious data to investigate is payroll (the hypothesis being that higher skilled workers will earn higher wages, IT skilled workers higher wages still).

58. Analysis will centre around three main concepts:

60


Table 4.20: Human capital variables Variable Description Hkpct Percentage of employees with post upper secondary education : All HkItpct Percentage of employees with post upper secondary education : IT related HkNitpct Percentage of employees with post upper secondary education : Non IT related

Figure 4.16: Correlation of human capital variables Correlation coeffecient (Skills vs Wage)

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

HkPct HkItPct HkNitPct

FINNORSWE

59. The results shown in Figure 4.16 appear to be encouraging. Where both wage and skills data are available: wages are most highly correlated with “highly skilled, IT literate employees” and least correlated with “non-IT literate employees” suggesting that this could be taken further for analysis (see chapter 8: Employment, skills and information technology).

Identifying and developing potential indicators – Relationships between ICT variables and productivity

60. The relationship between high speed Internet and productivity can be investigated further by examining quartiles of productivity (in this case TFP) against DSLpct (percentage of workers with access to high speed Internet). Refer to Table 4.1 for details about variables used. Figures 4.17 to 4.20 do this by broad industry (using the ALT classification described in Table 4.3).

61


Figure 4.17: Quartile Total Factor Productivity vs. High speed Internet. Industry: Manufacturing (excluding electrical goods)

0

0.1

0.2

0.3

0.4

0.5

0.6

ITA AUT CZE GER SLO NLD GBR NOR SWE Mean

DSL

pct (

1=10

0%)

Productivity QRT1Productivity QRT2Productivity QRT3Productivity QRT4

This graph shows mean dslpct, quartiled into TFP distributionYear==2004euk==MexElec

Figure 4.18: Quartile Total Factor Productivity vs. High speed Internet. Industry: Financial and Business services

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

CZE GBR ITA NLD NOR SLO SWE Mean

DSL

pct (

1=10

0% Productivity QRT1Productivity QRT2Productivity QRT3Productivity QRT4

This graph shows mean dslpct, quartiled into TFP distributionYear==2004euk==FINBU

62


Figure 4.19: Quartile Total Factor Productivity vs. High speed Internet. Industry: Electrical Machinery, Post and Communication Services

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


DSL

pct (

1=10

0%)


This graph shows mean dslpct, quartiled into TFP distributionYear==2004euk==Elecom

Figure 4.20: Quartile Total Factor Productivity vs. High speed Internet. Industry: Distribution

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

ITA AUT CZE SLO NLD GBR NOR SWE Mean

DSL

pct (

1=10

0%)


This graph shows mean dslpct, quartiled into TFP distributionYear==2004euk==DISTR

61. An overview of the data shows encouraging results. There is good variability between countries: a healthy range of values represent economies in different stages of ICT maturity. An overview of the means display patterns consistent with reasonable expectations, that is, the firms in the lowest quartile of productivity also tend to have the lowest percentage of high speed Internet connected employees, while those who are most productive tend to have the

63


highest proportion of employees connected to high speed Internet. This holds true within the manufacturing (excluding electrical), financial and business services and distribution sectors.

62. This concept weakens when TFP is considered against Esalespct (the percentage of sales made over the Internet) and Epurchpct (the percentage of purchases made over the Internet).

Figure 4.21: Quartile Total Factor Productivity vs. E Purchasing. Industry: Financial and business services

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

AUT CZE FIN FRA GBR ITA NLD NOR SLO SWE

Epur

chpc

t (1=

100%

)


This graph shows mean epurchpct, quartiled into TFP distributionYear==2004euk==FINBU

63. When considering the ‘Financial and Business Services’ industry in Figure 4.21: E-purchasing appears to hold the same overall relationship, that those firms in the higher quartile of productivity are also those who purchase relatively more online (and vice versa for those in the lower quartiles). The relationship, however, is not as obvious when compared to that of (for example) high speed Internet as discussed earlier.

64. A review of the same variable in the ‘Electrical machinery, post and communication services’ industry reveals a much weaker relationship. In fact in some countries the relationship is inversed and it is the less productive firms who are engaged in E-purchasing. Figure 4.22 demonstrates that, on the whole, there is no discernible relationship between E-purchasing and productivity in the ‘Electrical machinery, post and communication services’ industry.

64


Figure 4.22: Quartile Total Factor Productivity vs. E Purchasing. Industry: Electrical machinery, post and communication services

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

AUT CZE FRA GBR GER ITA NLD NOR SLO SWE TOTAL

Epur

chpc

t (1=

100%

)


This graph shows mean epurchpct, quartiled into TFP distributionYear==2004euk==Elecom

65. Figure 4.23 investigates the relationship between firms (quartiled into productivity) and their respective E-selling data. It shows a similar tale to that of E-purchasing in the ‘Financial and business services’ industry. Some countries display an inverse relationship; some show a broadly positive relationship between the most productive firms and their respective level of E-purchasing. However the overall result is a haphazard and indiscriminate connection between productivity and E-commerce.

66. The distribution of E-selling is also somewhat random. The overall mean of Figure 4.24 displays a declining relationship in the ‘Electrical machinery, post and communication services’ (it is the less productive firms who are engaged in E-selling, those in the second productivity quartile actually engage in the most E-selling). Examining the underlying country relationships though shows no evident relationship with a seemingly random display of E-selling.

65


Figure 4.23: Quartile Total Factor Productivity vs. E Sales. Industry: Financial and business services

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

AUT CZE FIN FRA GBR ITA NLD NOR SLO SWE TOTAL

Esal

espc

t (1=

100%

)


This graph shows mean esalespct, quartiled into TFP distributionYear==2004euk==FINBU

Figure 4.24: Quartile Total Factor Productivity vs. E Sales. Industry: Electrical machinery, post and communication services

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

AUT CZE FIN FRA GBR GER ITA NLD NOR SLO SWE TOTAL

Esal

espc

t (1=

100%

)


This graph shows mean esalespct, quartiled into TFP distributionYear==2004euk==Elecom

67. E-commerce and productivity is explored further in chapter nine (ICT business integration).

66


67

4.7 Disclosure 68. Statistical disclosure control methodology is applied to all the data contained within this report. This ensures that information attributable to an individual is not disclosed in this or any related publication. The report complies with statistical disclosure measures and standards throughout all member countries. This includes The National Statistics Code of Practice of the United Kingdom, and specifically the Protocol on Data Access and Confidentiality. The Protocol includes a guarantee to survey respondents that "no statistics will be produced that are likely to identify an individual unless specifically agreed with them". For more information on ONS statistical disclosure control methodology, see:

http://www.statistics.gov.uk/about/data/methodology/general_methodology/sdc.asp

4.8 Further Information For further information please contact:

Peter Stam: Office for National Statistics, UK E-mail: [email protected] Telephone: +44 (0) 1633 455982

http://www.statistics.gov.uk/about/data/methodology/general_methodology/sdc.asp



CHAPTER 5

Properties of Linked Data: Evidence from the ICT Impacts Project

Eric J. Bartelsman

Vrije Universiteit Amsterdam, Tinbergen Institute, and IZA

5.1 Abstract

This Chapter provides background information on the statistical properties of the linked datasets used in the ICT-Impacts project. The paper starts by describing the nature of the linked sub-samples used in the analysis. Next, a quick overview is provided on the various methods described in the literature for producing statistical estimates from combined sample surveys and population registers, including a simple yet practical method, which we call sample re-weighting. Finally, we analyze the properties of the linked samples across the countries involved in the ICT-Impacts project, by comparing versions of indicators built up from different sub-samples. The paper concludes with practical tips for practitioners and with suggestions for future research.

5.2 Introduction

The ICT Impacts Project makes use of information pertaining to ICT that is available at National Statistical Institutes (NSIs). In particular, the project explores information available at the firm level from the ‘E-Commerce Survey’ (EC), linked to firm-level data from the Structural Business Survey, or Production Survey (PS), and to information from the Business Register (BR). The various statistical and analytical exercises conducted for the chapters of this report use different samples and sub-samples of the underlying data sources. This Chapter is concerned with highlighting the statistical properties of the exercises, with special attention to problems associated with linked datasets.

This Chapter will remain practical, both in terms of the presentation of the underlying theory and in the presentation of evidence collected during the ICT Impacts project. A good place to start exploring the theoretical issues is a ‘policy’ report by Chesher and Nesheim (2006) or a paper by Ridder and Moffitt (2003). However, not all the issues presented in these reports are relevant for the topics of our project. Based on the Eurostat regulation on Business Registers (Council Regulation (EEC) No 2186/93), longitudinal linking of particular surveys and linking of different surveys in the cross-section is made relatively easy because the BR is the sample frame for business surveys and linking thus may be done on the basis of a unique firm identifier. In theory, one need not worry about errors in record linkage, either omitted links or erroneous links. In practice, not all business surveys in all EU countries are designed as required, nor are all implemented as desired, but in the countries participating in our project the EC and PS generally are linkable to the BR over the period 2000-2005.

68


The statistical issues relevant to this project are those that fall under the heading ‘direct record linkage’ in the paper by Chesher and Nesheim (2006). When a particular firm-level survey is linked to another survey (or to the same survey in a different year), the resulting dataset may be regarded as a sample conducted under a complex survey design that could in principle be derived from the designs of the underlying surveys. In practice, the underlying survey designs, along with actual record- and item-level response, is not available in machine-readable form. Instead, for each statistical exercise on a linked dataset, one has to consider whether selectivity may bias the produced result, and one has to consider whether or not unbiased results may be extrapolated to regions outside the sample.

The Chapter will start with summary statistics of employment and firm counts for the BR, the PS, and the linked PS_EC sample for selected countries and industries. Next, a method for calibration of a linked dataset used in our project will be described (see Renssen and Nieuwenbroek (1997), or Renssen et al. (2001), Särndal (2007)). Using this method, aggregates of various economic indicators will be made using either no weights, the sample weights from the PS or the weights computed via the re-weighting method. The aggregates from the PS and from the PS_EC linked sample then are compared to evaluate the various weighting schemes. As an added exercise, regression coefficients from a relationship between aggregated variables (using a panel by countries, industries, and time) may be checked across sub-samples under various weighting strategies. Finally, comparisons are made of the cross-sectional distribution of growth rates of selected variables for ‘continuer’ firms in longitudinally linked PS and PS_EC samples. The Chapter ends with practical advice on the robustness of indicators based on linked datasets, with cautions to users of such data and with recommendations on improving the quality of the data.

5.3 Descriptive Statistics of Linked Datasets

Using the common code that was distributed to all NSIs, one of the harmonized cross-country datasets that was collected contains information about the BR, PS, and EC datasets, as well as about various combinations of linked datasets. This dataset, labelled ‘coverage’, contains for each detailed industry (from the EU KLEMS industry list), each country, and each year, the total employment and total number of firms of the underlying samples. For more aggregated industries (the EU KLEMS ‘Alternate’ industry definitions), the tabulations also are split by firm-size. When available, weighted aggregations were made for the PS sample, using the sample weight supplied for this survey by the NSI.

The Business Register forms the backbone for firm-level statistics, and is the source for firm-level survey design. The BR contains the ‘universe’ of firms, with information on firm-size (employment), location, and industry. As such, it is the ultimate source of, and serves as the benchmark for population totals. However, in many countries, the business register is not retro-actively maintained or updated to reflect improved information on variables such as employment or industry collected by surveys or through government (tax) registers. Table 5.1 shows some summary information from the ‘coverage’ dataset from Finland in 2004 for a selection of industries. (Appendix Table 5.A.1 shows the information for all countries, and the ‘coverage’ dataset contains all underlying information for all countries). As may be seen,

69


Finland had 346 thousand employees in the Distribution sector in 2004, according to the Business Register (the BR column). The Production Survey (the PS column), shows 103 percent of this total, i.e. about 355 thousand employees. Because the production surveys disproportionately cover large firms, the PS only has 77 percent of the number of firms in the BR in the Distribution sector. Coverage in the sample combining the PS and the Ecommerce Survey (the PS_EC column) is 37 percent of employees and only 2 percent of firms.

Table 5.2 displays some coverage statistics from Italy, split by size class (determined by employment in the BR). In Italy, in 2004, the BR shows 3.4 million employees in the sector ‘manufacturing excluding electrical machinery’. The subsequent rows show the distribution of these employees by size-class of firms, in percent. Firms with more than 500 employees account for nearly 20 percent of employment in the sector. The second column (W_PS), shows the tabulations from the PS sample, using sample weights on the firm-level PS data available at the national statistical institute (NSI). Weighted aggregate employment in the sector is 3.68 million, and the weighting of the PS sample tends to provide employment by size-class with a similar distribution as that of the BR. In the third column (PS), we see that employment at firms available in the PS dataset only covers 47 percent of the total, and that the coverage ranges from 4 percent of total employment in firms with fewer than 10 employees to 100 percent coverage for firms with one hundred of more employees. The combined PS_EC dataset is even more skewed towards large firms.

The highly skewed nature of the PS_EC dataset needs to be considered when analyzing features of this dataset. However, some indicators may be more affected than others. The main purpose of this Chapter is to understand which indicators may be used directly from the PS_EC dataset, and which indicators required sample weighting or other correction methods.

5.4 A method for sample re-weighting

Following the discussion of Chesher and Nesheim, we consider a sample of firms obtained by linking two separate samples, each drawn from the BR using a certain sample design. If the design were known and described, it would in principle be possible to describe the effective sampling rules for the combined sample. In practice, this information is difficult to retrieve, especially in an electronic form, for the historical surveys at NSIs. Further, issues of non-response would cloud the relationship between the design and the available micro-data.

Instead, information from the sample frame, BR, may be used to assess how the combined sample compares with the universe. In our project, we use a method by Renssen and Nieuwenbroek (1997), to generate new weights for the combined sample. We call this method ‘sample re-weighting’.

The methodology is simple to follow and easy to implement in the harmonized code that was sent to the participating NSIs in this project. In short, the method provides a new weight for each firm in the linked sample, based on characteristics of firms in the sample and the same characteristics of all the firms in the BR. A technical description of the algorithm is provided in Appendix B.

70


The intuition for the method is rather straightforward. Assume that the characteristic ‘firm-size’ is available on both the sample and the population. By dividing the number of firms per size group in the population by the number of firms per size group in the sample, we get a ‘weight’ to apply to each firm in each particular size group which allows weighted aggregation of the number of firms. In this case, the weighted sum of firms in the sample will equal the number of firms in the population.

In practice the situation is slightly more complex. First, multiple characteristics may be used, say size-class, industry, and/or region, or any other characteristics that are available for the population (because the firms in the sample are linkable at the firm-level to the BR, all information in the BR is available for firms in the sample). Next, the weights to be computed will apply to all firms in a ‘cell’, or a unique occurrence of values of the Cartesian product of the characteristics (say, the matrix of size-class and industry). Unfortunately, the sample may not have many observations, or may have none, for a particular set of characteristics. One could ‘thicken’ the cells (i.e. take a more aggregated industry classification, or join certain size-classes), but this might cause loss of important information. Another option is to use a regression technique whereby the population totals are regressed on the sample total interacted with a ‘design matrix’. For example, one could use an indicator variable for each size class (with the value 1 if the cell belongs to that size class) and an indicator variable for each industry, and generate a regression coefficient for each product of an indicator variable times the sample total. The weight for each cell would then be the sum of the appropriate indicator coefficients that each had a value of 1 for that cell. With multi-dimensional characteristics, one could use a dummy variable indicator for each dimension, or use sums of Cartesian products of dummy indicators for each dimension. If one takes a Cartesian product of all dimensional indicators, the regression generates weights that are equal to the first-mentioned case where cells from the population are divided by cells of the sample.

A further complication pertains to the variable for which one wants to generate weights. In our initial example, the variable of interest was the number of firms. Using the sample re-weighting method, one could also generate weights for aggregating employment, or other variables in the sample. Ideally, one would ‘benchmark’ or calibrate to the same variable from the population. In our case, the BR has employment and number of firms, so that one needs to make a choice which weight to apply when aggregating, for example, value added, or e-commerce variables. In our analysis we use weights generated for employment to aggregate the ecommerce variables. More complex methods, for example maximum likelihood estimators, can generate weights with multidimensional calibration targets.

Using sample re-weighting techniques, we generate weights for each firm in the PS sample and for each firm in the PS_EC sample. These weights are then used to aggregate variables from both samples. In some countries, the original sample weights for the PS also were available. We explore using these weights to aggregate variables from both the PS and the PS_EC samples and can compare and contrast the various results.

Table 5.3 shows results comparing different aggregate measures of the ratio of value added to employment (a proxy for productivity). Among the participating countries, the sample-re-

71


weighting procedure and the PS sample weights were available in Austria, Czech Republic, Italy, the Netherlands, and the U.K. For each industry, year, and size class in the PS and PS_EC samples we generate aggregate value added and aggregate employment using both the PS weights and the ‘sample re-weights’. The first row (total) in column 1 (wPS_rPS) shows the correlation coefficient between the weighted and the re-weighted productivity measure for the total industry, across industries, countries, and years, which is 0.93. Each following row shows the correlation of the two measures across industries, countries and years, by size class. In theory the correlation should be high for each size class because both the PS weight and the re-weight attempt to do the same thing. In practice, however, the correlation in the smallest size classes is 0.67. More interesting is to see how productivity from the PS_EC sample compares to productivity from the PS sample, after weighting. The columns wPS_wEC and wPS_rEC show the correlation between the (weighted) productivity measure from the PS sample, with respectively, the weighted and re-weighted productivity measures from the PS_EC sample. While the correlation for the aggregate industry/country/years is reasonable, 0.89 and 0.82 respectively, the correlations are rather small for the smallest size classes but very high for the larger classes. The final column is a check on differences between using the PS sample weight in the PS_EC sample, instead of the appropriate re-weight. Overall, the correlation between the two productivity measures is high, and especially so for the large size classes.

While no ‘metrics’ are provided to evaluate whether these correlations are ‘high’ or ‘low’, it appears that for the productivity measure, the skewed size distribution in the population works in our favour. Given that large firms contribute strongly to aggregate productivity and that large firms have high coverage in the PS_EC sample, the sensitivity to the weighting schemes is less for aggregate industry productivity than it would otherwise be.

Some caveats on sample re-weighting While the sample re-weighting method is straightforward and has been programmed to run with the harmonized code, it generally requires more ‘hand work’ than the results presented here. In typical applications at NSIs, skilled statisticians spend much time in setting up weights to be used for tabulations of a single survey. Even when the analyst has access to pre-programmed code to generate weights, much time is spent determining values for parameters that are required in the programs, and in reviewing results. In our algorithm, a choice must be made regarding the dimensions, the level of aggregation within each dimension, and the ‘design matrix’ of interactions between dimensions. The sensitivity of the results to these choices has not been tested, nor has the robustness of the results (e.g. because resulting cells may be too thin).

Finally, the method we use requires that some information is available in the sample for each of the unique values in a dimension, otherwise the resulting aggregate will not reflect the total population. For example, if no firms in the sample fall within the size class of firms with fewer than 10 employees, the aggregate produced will not provide any information about this area of the population, either. Whether this is a shortcoming of our method and other methods of calibration and imputation that in principle can be overcome, is a contentious

72


issue of debate. For example, Särndal (2007) suggests that no statistical inference may be made in areas of the data that are not covered by the sample. Others (e.g. Elbers et al. 2003, Lewbel 2005), think that modelling relationships in the data may allow extrapolation into areas for which no data are available in the sample. In any case, our simple method of re-weighting does not generate weights for cells for which no data exist in the sample.

5.5 Moments of longitudinally linked samples

The ultimate goal of this study, namely measuring the impact of ICT, often requires measuring the change in some variable at the firm level, for example the growth of firm-level value added. At the industry level, value added growth is a weighted sum of value added growth of firms that operate both in some start and end period, and appropriately weighted value added of entrants minus the value added of exiting firms. Properly determining entry and exit effects, using PS samples, is beyond the scope of this chapter.1. We will focus on estimates of the distribution of firm-level growth among continuing firms from the PS and the PS_EC samples.

Given the sample design of the PS and the resulting sample characteristics of the PS_EC sample, the coverage of small firms in continuing firms will be rather small. For example, assume that the probability of a small firm to be in the sample is 0.1. Then the chance of being sampled in both periods is 0.01. In some countries, sampling designed to minimize administrative burdens may even reduce the chance to be taken up in two subsequent periods. Nonetheless, results from Table 5.4 show that the mean and variance of the distribution of firm growth is fairly similar between the PS and the PS_EC sample with reasonably sized EC samples. In France, where the PS_EC sample only has one-twenty-fifth of the number of firms of the PS sample, the means and variances move similarly across industries. In Norway, where the PS_EC sample has very few observations, the relationship becomes more problematic.

Table 5.5 digs a bit deeper into the firm growth distribution, by showing means and variance by quartile of the firm growth distribution. Here as well, the results from the PS sample and the PS_EC samples are quite similar, for the example of Ireland where the PS_EC continuers have about one third of the observation of the sample of the PS continuing firms.

5.6 Regression coefficients from different samples

In the ICT impacts project, many results are derived from the sign and significance of regression coefficients, for example the effect of broadband use on productivity at the firm level. Also, Chapter 14 presents results from regressions at the industry level. Here, we will compare regressions from industry/time panel datasets that have been built up from different samples of the firm-level data. In our example, we will compare regression results from a dataset aggregated up from the PS sample with the results from data aggregated up from the PS_EC sample.

1 See Martin (2005)

73


The main objective of this project is to assess the impact of ICT use. The main indicators used are from the E-Commerce survey, and thus not available from the PS dataset alone. However, we may be worried about the bias in regressions done at the industry level when the variables are built from the smaller PS_EC sub-sample. To test this, we take as a basis an industry level regression that is similar to a regression run at the firm-level, namely productivity regressed on a human capital measure. For Finland, Norway, and Sweden, human capital variables are available on the PS survey. This allows running an industry/time regression of productivity on human capital, for data aggregated up from firms in the PS sample as well as for an aggregate based on firms in the PS_EC sample. Table 5.6 shows the results from a regression of the (log) level of labour productivity and of the growth of labour productivity on the share of employees with IT skills and fixed effects. In Finland and Norway, the regression coefficients are very similar, both in magnitude and significance, both in levels and in first-differences.

5.7 What have we learned?

• In all countries, it was possible to link production surveys to the business register.

• Because a resulting sample can be linked to the business register at the individual level, statistical procedures are available to do sample ‘re-weighting’.

• Productivity measures from the production survey are similar to the measures built up from the sub-sample linked to the E-commerce survey. This holds for the mean of the productivity distribution, and also the variance.

• Regression coefficients to explain productivity seem fairly robust to the nature of sampling and weighting schemes.

74


Table 5.1. Coverage: Employment Number of firms

Country Sector BR

(000s) PS (%)

PS_EC (%)

BR (000s)

PS (%)

PS_EC (%)

FIN 1 Electrical Mach. and Comm. 105 104 80 2 79 8

2 Manuf. excl Mach. 331 101 57 23 71 4 3 Other Goods 160 103 24 42 70 1 4 Distribution 346 103 37 69 77 2 5 Financial and Business Services 215 91 25 39 68 1 6 Personal Services 88 103 19 29 80 1

Table 5.2.

Coverage: Employment 1 BR 2 W-PS 3 PS 4 PS_EC Country Sector Size class Emp% Emp% % % Italy Total 3400 3681 47 18 e<10 15 19 4 0 10<=e<20 15 15 8 1 20<=e<50 18 17 20 2 50<=e<100 12 11 37 8 100<=e<250 14 13 100 24 250<=e<500 8 7 100 51

2 Manuf. excl. Elec Mach.

500<=e 19 17 100 60

Table 5.3. Correlations wPS rPS wPS wPSEC wPS rPSEC wPSEC rPSEC Total 0.93 0.89 0.82 0.88 e<10 0.67 0.37 0.40 0.74 10<=e<20 0.87 0.65 0.56 0.81 20<=e<50 0.93 0.56 0.51 0.85 50<=e<100 0.94 0.76 0.66 0.87 100<=e<250 0.97 0.79 0.69 0.88 250<=e<500 1.00 0.80 0.75 0.93 500<=e 1.00 0.94 0.93 1.00

Table 5.4.

Mean Growth Rate of Value Added MEAN Std. Dev. Sample Size Country Sector PS PS_EC PS PS_EC PS PS_EC FRA 1 Electrical Mach. and Comm. 0.086 0.070 0.432 0.438 12487 458 2 Manuf. excl. Elec Mach. 0.027 0.024 0.408 0.315 121208 3597 3 Distribution 0.021 0.025 0.435 0.306 294252 3212 4 Financial and Business Ser 0.044 0.036 0.541 0.374 98799 1428 NOR 1 Elecrtical Mach. and Comm. 0.200 0.185 0.810 0.712 141 34 2 Manuf. excl. Elec Mach. 0.004 0.020 0.466 0.351 1068 375 3 Distribution 0.074 0.070 0.744 0.324 3560 390 4 Financial and Business Ser 0.125 0.075 0.881 0.380 1782 174

75


Table 5.5.

Mean Growth Rate of Value Added MEAN Std. Dev. Sample Size

country Sector Qrt PS PS_EC PS PS_EC PS PS_EC

1 -1.283 -1.494 2.146 2.402 90 422 -0.007 0.003 0.065 0.068 89 413 0.190 0.197 0.068 0.070 90 41

1 Electrical Mach. and Comm.

4 1.565 1.700 2.053 2.708 90 411 -1.064 -1.065 1.591 1.645 955 3152 -0.033 -0.041 0.055 0.063 957 3153 0.146 0.164 0.060 0.067 957 314

2 Manuf. excl. Elec Mach.

4 1.247 1.338 1.660 1.802 956 3141 -0.942 -0.592 1.761 0.574 393 1072 -0.059 -0.049 0.059 0.055 397 1073 0.126 0.125 0.059 0.056 401 108

3 Distribution

4 0.916 1.372 1.553 2.878 396 1071 -0.522 -0.423 0.649 0.448 170 592 -0.037 -0.034 0.048 0.045 173 593 0.120 0.101 0.048 0.039 173 59

IRE

4 Financial and Business Ser

4 0.591 0.702 0.823 1.457 170 59

Dep var: (log) Labour Prod Growth Labour Prod Country PS PS_EC PS PS_EC

Table 5.6.

FIN HK_IT 1.51 1.82 0.64 0.60

(t) (3.0) (4.9) (3.5) (6.1) R-sq 0.09 0.12 0.26 0.23 Nobs 276 276 28 128

NOR HK_IT 0.76 0.57 0.36 0.27 (t) (3.5) (2.7) (1.7) (2.3) R-sq 0.04 0.03 0.02 0.02 Nobs 258 258 210 210

SWE HK_IT 1.32 0.79 -0.04 -0.02 (t) (2.5) (1.9) (1.7) (0.8) R-sq 0.03 0.01 0.01 0.00 Nobs 260 260 201 201

76

49102.2005.017-2006.128 Chapter 5

77

Appendix A: Additional TablesTable A1:

1 BR 2 W-PS 3 PS 4 PS_ECCountry Sector Size class Emp /pct Emp /pct pct pct

Total 111 113 100 79e<10 . . . .10<=e<20 3 3 100 520<=e<50 5 4 100 1150<=e<100 5 5 100 20100<=e<250 12 14 100 46250<=e<500 8 8 100 69500<=e 67 66 100 99

Total 386 377 100 57e<10 . . . .10<=e<20 . . . .20<=e<50 . . . .50<=e<100 13 13 100 23100<=e<250 25 25 100 24250<=e<500 22 22 100 74500<=e 40 39 100 80

Total 349 340 100 18e<10 . . . .10<=e<20 23 24 100 520<=e<50 27 27 100 650<=e<100 16 16 100 19100<=e<250 20 20 100 22250<=e<500 14 14 100 61500<=e . . . .

0 100 50. .. .

18 100 713 100 2016 100 2413 100 7040 100 84

Coverage: Employment

AUT 1 Electrical machinery, post and communication services

2 Manufacturing (excluding electrical)

3 Distribution

Total 294 27e<10 . .10<=e<20 . .20<=e<50 1750<=e<100 13100<=e<250 16250<=e<500 13500<=e 42

4 Financial and business services

Eurostat Agreement No.


1 BR 2 W-PSCountry Sector Size class Emp /pct Emp /pc

Total 238e<10 6 310<=e<20 3 320<=e<50 5 550<=e<100 6 6100<=e<250 11250<=e<500 13500<=e 56

Total 1038e<10 6 410<=e<20 6 520<=e<50 1150<=e<100 11100<=e<250 19250<=e<500 15500<=e 33

Total 748e<10 2610<=e<20 1220<=e<50 1250<=e<100 8 9100<=e<250 8 9250<=e<500 7 8500<=e 26

Total 280e<10 2610<=e<20 1220<=e<50 1450<=e<100 12100<=e<250 14250<=e<500 8 9500<=e 15

CZE 1 Electrical machinery, post and communication services


3 Distribution


Coverage: Employment 3 PS 4 PS_ECt pct p

215 9420438890

11 9713 9959 100

967 851632

11 6911 7619 9415 9634 99

606 6818 811 2814 70

749499

30 100

218 6321 812 2414 7913 8115 99

9016 93

ct

7235

1525369490

50238

15198789

43138

29389292

28027

26348070

78



Total 876 822 77 37e<10 7 6 1 010<=e<20 3 3 10 020<=e<50 6 7 18 050<=e<100 5 6 34 4100<=e<250 9 9 55 6250<=e<500 7 6 100 43500<=e 62 63 100 54

Total 2952 2695 61 30e<10 11 10 1 010<=e<20 6 7 11 020<=e<50 11 12 19 050<=e<100 10 11 37 3100<=e<250 15 16 59 4250<=e<500 12 10 99 35500<=e 34 34 100 75

Total 5483 5196 62 44e<10 19 20 2 010<=e<20 7 7 10 020<=e<50 8 8 20 050<=e<100 5 5 37 11100<=e<250 6 6 48 12250<=e<500 4 4 100 73500<=e 51 51 100 79

Total 4803 3956 50 28e<10 20 23 1 010<=e<20 6 7 6 020<=e<50 8 9 11 050<=e<100 6 7 21 1100<=e<250 8 10 36 4250<=e<500 7 6 100 21500<=e 45 38 100 72

GBR 1 Electrical machinery, post and communication services


3 Distribution



79



Total 555 581 73 44e<10 7 10 7 010<=e<20 8 8 14 120<=e<50 10 10 30 250<=e<100 7 7 47 9100<=e<250 8 8 100 24250<=e<500 6 6 100 47500<=e 53 51 100 74

Total 3400 3681 47 18e<10 15 19 4 010<=e<20 15 15 8 120<=e<50 18 17 20 250<=e<100 12 11 37 8100<=e<250 14 13 100 24250<=e<500 8 7 100 51500<=e 19 17 100 60

Total 2120 2368 43 18e<10 24 29 2 010<=e<20 13 15 7 120<=e<50 13 12 20 250<=e<100 8 7 35 9100<=e<250 9 8 100 23250<=e<500 6 6 100 39500<=e 26 23 100 57

Total 1764 1419 58 16e<10 13 20 2 010<=e<20 7 10 7 120<=e<50 8 10 19 350<=e<100 7 7 26 5100<=e<250 11 11 100 21250<=e<500 8 8 100 31500<=e 46 34 100 32


ITA 1 Electrical machinery, post and communication services


3 Distribution


80



Total 224 92 88 36e<10 8 9 15 110<=e<20 3 3 32 620<=e<50 5 5 62 2350<=e<100 4 2 96 51100<=e<250 5 3 100 52250<=e<500 6 2 100 66500<=e 69 75 100 40

Total 858 129 66 23e<10 11 24 17 010<=e<20 7 16 30 220<=e<50 13 16 86 1350<=e<100 11 13 98 28100<=e<250 16 11 100 47250<=e<500 12 7 100 82500<=e 30 12 100 45

Total 1612 706 65 25e<10 30 19 12 010<=e<20 10 11 30 120<=e<50 14 16 46 250<=e<100 9 8 79 10100<=e<250 9 8 100 31250<=e<500 5 6 100 58500<=e 23 31 100 58

Total 1189 474 62 28e<10 26 28 15 010<=e<20 7 9 38 420<=e<50 10 13 49 850<=e<100 7 7 82 25100<=e<250 10 8 96 46250<=e<500 7 6 100 68500<=e 33 27 100 61

NLD 1 Electrical machinery, post and communication services


3 Distribution



81



Total 148 204 100 24e<10 4 3 100 010<=e<20 3 2 100 920<=e<50 5 4 100 850<=e<100 4 3 100 10100<=e<250 7 5 100 37250<=e<500 8 5 100 60500<=e 69 44 100 42

Total 610 581 100 42e<10 8 8 100 010<=e<20 6 6 100 620<=e<50 10 10 100 750<=e<100 9 9 100 8100<=e<250 13 13 100 41250<=e<500 10 10 100 59500<=e 43 42 100 69

Total 623 586 100 23e<10 25 26 100 010<=e<20 12 12 100 620<=e<50 14 14 100 650<=e<100 8 8 100 7100<=e<250 9 9 100 40250<=e<500 6 5 100 60500<=e 27 23 100 60

Total 484 412 100 27e<10 19 22 100 010<=e<20 8 8 100 620<=e<50 10 11 100 750<=e<100 8 8 100 9100<=e<250 10 9 100 36250<=e<500 8 8 100 51500<=e 37 29 100 61

SWE 1 Electrical machinery, post and communication services


3 Distribution



82


Table A2:

Country Sector BR (000s) PS (pct) PS_EC (pct) BR (000s) PS (pct)

1 Electrical machinery, post and communicaiton services 111 102 80 1 1002 Manufacturing (excluding electrical) 386 98 55 2 1003 Other Goods . . . . . .4 Distribution 349 97 18 11 1005 Financial and Business Services 294 92 46 3 96 Personal Services 147 63 3 6 75

1 Electrical machinery, post and communicaiton services 238 85 65 332 Manufacturing (excluding electrical) 1038 79 46 1333 Other Goods 308 72 34 1384 Distribution 748 55 35 2945 Financial and Business Services 280 49 21 1976 Personal Services 164 34 16 114

1 Electrical machinery, post and communicaiton services 105 104 80 2 72 Manufacturing (excluding electrical) 331 101 57 23 713 Other Goods 160 103 24 42 704 Distribution 346 103 37 69 775 Financial and Business Services 215 91 25 39 686 Personal Services 88 103 19 29 80


Employment Number of Firms

AUT

FIN

FRA

CZE

COVERAGEPS_EC (pct)

1835

79 16

2

4 16 13 02 01 01 0

9 841211

320110

83


Country Sector BR (000s) PS (pct) PS_EC (pct) BR (000s) PS (pct) PS_E





COVERAGE Employment

GER

GBR

IRE

ITA

Number of FirmsC (pct)

110000

120000

72821

111

220110

84


85

Country Sector BR (000s) PS (pct) PS_EC (pct) BR (000s) PS (pct) PS_EC (pct)


1 Electrical machinery, post and communicaiton services . . . 3 912 Manufacturing (excluding electrical) . . . . . .3 Other Goods . . . 16 904 Distribution . . . 37 885 Financial and Business Services . . . 22 896 Personal Services 0 . . 2 37



NOR

SWE

SLO

NLD

COVERAGE Employment Number of Firms

100000

4

1112

361111

220000


86

Table A3:

country Sector PS PS_EC PS PS_EC PS PS_EC

AUT 1 Electrical Mach. and Comm. 0.057 0.049 0.58 0.512 946 922 Manuf. excl. Elec Mach. 0.014 0.022 0.48 0.465 10315 7493 Distribution 0.005 0.032 0.63 0.564 31217 6174 Financial and Business Ser 0.017 0.054 0.59 0.526 20215 338

1 Electrical Mach. and Comm. 0.149 0.162 0.421 0.431 762 2482 Manuf. excl. Elec Mach. 0.035 0.041 0.440 0.411 5191 12873 Distribution 0.052 0.057 0.484 0.452 3199 7554 Financial and Business Ser 0.019 0.043 0.430 0.421 1442 338


1 Electrical Mach. and Comm. -0.003 0.010 0.636 0.634 178 712 Manuf. excl. Elec Mach. -0.012 -0.020 0.522 0.516 948 2453 Distribution 0.041 0.033 0.626 0.576 1169 3634 Financial and Business Ser 0.026 0.021 0.512 0.491 937 133

1 Electrical Mach. and Comm. 0.025 0.016 0.340 0.377 1328 2262 Manuf. excl. Elec Mach. -0.003 0.009 0.315 0.310 9406 16033 Distribution4 Financial and Business Ser


1 Electrical Mach. and Comm. -0.013 -0.033 0.445 0.427 833 2152 Manuf. excl. Elec Mach. -0.008 -0.020 0.410 0.380 7149 19313 Distribution 0.018 0.012 0.473 0.406 2683 6704 Financial and Business Ser 0.021 -0.004 0.447 0.385 1626 412

1 Electrical Mach. and Comm. 0.062 0.031 0.587 0.514 162 412 Manuf. excl. Elec Mach. -0.008 -0.017 0.458 0.447 1142 1733 Distribution 0.011 0.030 0.427 0.372 4381 3784 Financial and Business Ser 0.000 -0.001 0.569 0.507 2621 480



SWE

IRE

ITA

NLD

NOR

CZE

FRA

GBR

GER


49102.2005.017-2006.128 Chapter 5

87

Table A4

Country Sector Quartile PS PS_EC PS PS_EC

1 -0.533 -0.442 0.453 0.4712 -0.058 -0.023 0.058 0.0493 0.131 0.127 0.064 0.0494 0.688 0.555 0.582 0.468

1 -0.485 -0.400 0.456 0.4242 -0.065 -0.048 0.047 0.0403 0.087 0.081 0.050 0.0414 0.520 0.458 0.451 0.528

1 -0.658 -0.505 0.596 0.4802 -0.083 -0.052 0.059 0.0463 0.100 0.089 0.060 0.0444 0.660 0.597 0.587 0.634

1 -0.594 -0.362 0.531 0.3842 -0.086 -0.034 0.057 0.0303 0.100 0.072 0.061 0.0384 0.638 0.546 0.580 0.721

1 -0.321 -0.281 0.353 0.3472 0.061 0.062 0.049 0.0483 0.231 0.230 0.055 0.0584 0.628 0.646 0.336 0.408

1 -0.454 -0.398 0.365 0.3142 -0.053 -0.046 0.053 0.0483 0.121 0.111 0.053 0.0464 0.526 0.496 0.372 0.400

1 -0.493 -0.437 0.392 0.3812 -0.034 -0.020 0.056 0.0493 0.140 0.137 0.055 0.0524 0.594 0.550 0.414 0.407

1 -0.442 -0.382 0.380 0.3372 -0.058 -0.039 0.043 0.0373 0.090 0.100 0.049 0.0524 0.488 0.497 0.381 0.434

CZE 1 Electrical machinery, post and communication services


3 Distribution


AUT 1 Electrical machinery, post and communication services


3 Distribution


Mean Growth Rate of Value Added MEAN Std. Dev.

PS PS_EC

234 24238 23239 23234 23

2579 1882579 1872580 1872577 187

7687 1548021 1547797 1547711 154

4934 855131 855173 854978 84

191 62190 62191 62190 61

1298 3221298 3221298 3221297 322

800 189800 189800 189799 188

361 85361 84361 84360 84

Sample Size



Country Sector Quartile PS PS_EC PS PS_EC PS PS

1 -0.543 -0.348 0.496 0.324 3172 -0.064 0.005 0.057 0.041 3173 0.106 0.129 0.051 0.036 3174 0.588 0.583 0.515 0.623 317

1 -0.496 -0.317 0.493 0.311 30262 -0.060 -0.031 0.048 0.035 30243 0.095 0.082 0.048 0.036 30254 0.501 0.367 0.437 0.354 3025

1 -0.459 -0.299 0.492 0.336 89512 -0.043 -0.009 0.041 0.031 89513 0.093 0.096 0.044 0.037 89534 0.495 0.446 0.441 0.441 8951

1 -0.592 -0.371 0.516 0.354 35722 -0.088 -0.045 0.053 0.038 35723 0.079 0.082 0.053 0.041 35724 0.562 0.601 0.495 0.645 3572

1 -0.353 -0.332 0.382 0.390 29242 -0.005 -0.007 0.034 0.036 31593 0.127 0.118 0.047 0.044 32744 0.545 0.502 0.452 0.473 3130

1 -0.384 -0.276 0.389 0.315 281552 -0.041 -0.024 0.034 0.030 313493 0.071 0.072 0.039 0.030 322514 0.446 0.324 0.440 0.329 29453

1 -0.418 -0.258 0.424 0.288 669932 -0.054 -0.023 0.033 0.024 747403 0.058 0.062 0.041 0.027 803164 0.465 0.319 0.466 0.340 72202

1 -0.527 -0.303 0.492 0.320 222492 -0.056 -0.018 0.040 0.028 236783 0.077 0.077 0.053 0.031 273354 0.598 0.389 0.561 0.461 25538

FRA 1 Electrical machinery, post and communication services


3 Distribution


FIN 1 Electrical machinery, post and communication services


3 Distribution


Mean Growth Rate of Value Added MEAN Std. Dev. Sample Siz_EC

39383838

234233233233

225224225224

102102102101

115115115114

898900900899

802804804803

356358358356

e

88


89

Country Sector Quartile PS PS_EC PS PS_EC PS PS_EC

1 -0.676 -0.642 0.478 0.464 45 182 -0.114 -0.101 0.075 0.069 44 183 0.101 0.101 0.066 0.064 45 184 0.690 0.714 0.635 0.641 44 17

1 -0.582 -0.577 0.527 0.525 237 622 -0.082 -0.079 0.057 0.051 237 613 0.092 0.078 0.053 0.048 237 624 0.524 0.503 0.419 0.421 237 61

1 -0.604 -0.582 0.562 0.556 292 912 -0.054 -0.038 0.056 0.050 292 913 0.129 0.121 0.061 0.054 293 914 0.692 0.634 0.618 0.504 292 91

1 -0.468 -0.419 0.475 0.554 235 342 -0.050 -0.034 0.043 0.034 234 333 0.093 0.088 0.044 0.042 235 344 0.528 0.458 0.554 0.423 234 33

1 -0.330 -0.344 0.336 0.319 332 572 -0.021 -0.021 0.035 0.030 332 563 0.086 0.070 0.033 0.028 333 574 0.368 0.364 0.309 0.431 330 56

1 -0.326 -0.282 0.316 0.340 2352 4012 -0.048 -0.035 0.033 0.029 2352 4003 0.056 0.056 0.032 0.028 2352 4014 0.307 0.298 0.296 0.303 2351 400

1234

1234

GER 1 Electrical machinery, post and communication services


3 Distribution


GBR 1 Electrical machinery, post and communication services


3 Distribution




90


1 -1.283 -1.494 2.146 2.402 90 422 -0.007 0.003 0.065 0.068 89 413 0.190 0.197 0.068 0.070 90 414 1.565 1.700 2.053 2.708 90 41

1 -1.064 -1.065 1.591 1.645 955 3152 -0.033 -0.041 0.055 0.063 957 3153 0.146 0.164 0.060 0.067 957 3144 1.247 1.338 1.660 1.802 956 314

1 -0.942 -0.592 1.761 0.574 393 1072 -0.059 -0.049 0.059 0.055 397 1073 0.126 0.125 0.059 0.056 401 1084 0.916 1.372 1.553 2.878 396 107

1 -0.522 -0.423 0.649 0.448 170 592 -0.037 -0.034 0.048 0.045 173 593 0.120 0.101 0.048 0.039 173 594 0.591 0.702 0.823 1.457 170 59

1 -0.486 -0.469 0.449 0.492 209 542 -0.065 -0.063 0.043 0.038 208 543 0.063 0.047 0.041 0.030 208 544 0.438 0.361 0.384 0.330 208 53

1 -0.417 -0.393 0.435 0.420 1788 4832 -0.057 -0.056 0.037 0.034 1787 4833 0.058 0.047 0.035 0.030 1787 4834 0.382 0.325 0.391 0.359 1787 483

1 -0.440 -0.366 0.489 0.441 671 1682 -0.038 -0.030 0.036 0.030 671 1683 0.078 0.065 0.037 0.030 671 1684 0.473 0.382 0.479 0.415 670 167

1 -0.417 -0.386 0.438 0.420 407 1032 -0.040 -0.040 0.036 0.031 406 1033 0.078 0.055 0.039 0.030 406 1034 0.462 0.356 0.454 0.356 406 103

ITA 1 Electrical machinery, post and communication services


3 Distribution


IRE 1 Electrical machinery, post and communication services


3 Distribution




91


1 -0.478 -0.483 0.487 0.606 41 112 -0.011 -0.007 0.044 0.064 40 103 0.131 0.155 0.047 0.051 41 104 0.619 0.496 0.662 0.270 40 10

1 -0.463 -0.435 0.476 0.520 285 42 -0.065 -0.058 0.041 0.036 286 43 0.062 0.052 0.040 0.036 286 44 0.434 0.382 0.445 0.417 285 4

1 -0.425 -0.330 0.429 0.332 1094 952 -0.046 -0.022 0.039 0.032 1095 943 0.080 0.081 0.039 0.030 1096 954 0.434 0.392 0.405 0.415 1095 94

1 -0.597 -0.504 0.525 0.435 650 1202 -0.079 -0.075 0.052 0.045 658 1203 0.084 0.063 0.052 0.043 658 1204 0.590 0.517 0.542 0.550 654 119

1 -0.601 -0.476 0.619 0.648 342 -0.001 0.016 0.052 0.047 363 0.240 0.251 0.148 0.208 374 1.160 1.021 0.679 0.318 34

1 -0.457 -0.335 0.510 0.313 266 92 -0.052 -0.041 0.041 0.034 267 93 0.073 0.066 0.040 0.036 268 94 0.447 0.392 0.425 0.338 267 9

1 -0.707 -0.219 0.658 0.268 820 92 -0.056 0.010 0.059 0.025 923 93 0.130 0.099 0.064 0.030 930 94 0.872 0.393 0.764 0.365 887 9

1 -0.933 -0.265 0.683 0.303 332 42 -0.115 -0.008 0.096 0.029 494 43 0.153 0.101 0.091 0.040 505 44 1.133 0.480 0.890 0.426 451 4

NOR 1 Electrical machinery, post and communication services


3 Distribution


NLD 1 Electrical machinery, post and communication services


3 Distribution



4333

9898

4444

8777

4343



1 -0.669 -0.471 0.676 0.588 354 222 -0.073 0.002 0.067 0.033 512 223 0.135 0.124 0.066 0.046 519 224 0.751 0.866 0.769 1.357 408 21

1 -0.626 -0.299 0.592 0.362 3610 1972 -0.080 -0.023 0.062 0.033 5238 1963 0.114 0.090 0.064 0.037 5675 1974 0.699 0.415 0.687 0.448 4025 196

1 -0.603 -0.245 0.590 0.287 10107 1722 -0.064 0.003 0.057 0.028 13243 1723 0.117 0.102 0.061 0.032 14831 1724 0.716 0.442 0.675 0.474 11628 172

1 -0.921 -0.364 0.746 0.366 6823 972 -0.131 -0.028 0.085 0.040 8294 973 0.124 0.102 0.084 0.042 9157 974 0.953 0.642 0.834 0.808 7517 97




3 Distribution



92


1 -0.669 -0.471 0.676 0.588 354 222 -0.073 0.002 0.067 0.033 512 223 0.135 0.124 0.066 0.046 519 224 0.751 0.866 0.769 1.357 408 21

1 -0.626 -0.299 0.592 0.362 3610 1972 -0.080 -0.023 0.062 0.033 5238 1963 0.114 0.090 0.064 0.037 5675 1974 0.699 0.415 0.687 0.448 4025 196

1 -0.603 -0.245 0.590 0.287 10107 1722 -0.064 0.003 0.057 0.028 13243 1723 0.117 0.102 0.061 0.032 14831 1724 0.716 0.442 0.675 0.474 11628 172

1 -0.921 -0.364 0.746 0.366 6823 972 -0.131 -0.028 0.085 0.040 8294 973 0.124 0.102 0.084 0.042 9157 974 0.953 0.642 0.834 0.808 7517 97




3 Distribution


92


49102.2005.017-2006.128 Chapter

93

^

1

for firm is

, where D

S

d

th

nn

dd

f

D

D n n

γ=

=∑

the population dataset

Appendix B - A technical description of sample ‘re-weighting’

^

^

1

is 'design matrix' with rows and cols

is regression estimate from:

var var , where

var = var is t

S

d

DN

d

i

smpl D

dn

npop smpldi i i

d

pop popi f

f Ncube

D

γ

γ ε=

∈

= +∑

∑ he aggregate value of from

for all firms in element of categorical thf i Ncube

weight

f

n

rwgt

5


Chapter 6

Productivity and Core ICT Metrics at Firm Level Mark Franklin UK Office for National Statistics

6.1 Summary 1. We have run comparable firm-level regressions in 11 of the 13 project countries and in so doing have demonstrated proof of the feasibility of the method employed. But this is work in progress. We have had limited time available to analyse the results and iterate towards more robust and meaningful specifications, and to make sense of sometimes conflicting results.

2. Nevertheless there are some suggestive findings even at this early stage:

• We find strong positive relationships between average wages and productivity in all countries and across all specifications. This is hardly surprising, but important confirmation nonetheless.

• Our results support prior work focussing on the manufacturing sector. We find significant positive coefficients on fast internet connectivity for several, though not all, countries in the study.

• However we find only a limited role for firm maturity, at least in these initial and preliminary specifications. We also find only weak evidence of a relationship between productivity and multinational status.

• Identification of relationships between ICT metrics and productivity is much less apparent outside the manufacturing sector, and particularly elusive in the ICT-producing parts of the economy.

• In all sectors there are substantial differences across countries, both in terms of the magnitude of empirical relationships, and even in some cases in terms of direction.

3. No endogeneity testing has been carried out in this Chapter and we recognize that none of the results and findings demonstrates causation. Much further work remains to be done.

6.2 Introduction

6.2.1 Productivity theory 4. The focus of this chapter is on productivity at the level of the individual firm. This stands in contrast to mainstream work on productivity, which centres on the growth accounting framework first developed by Jorgenson and Griliches (1967) and set out, for example, in ONS (2007). The growth accounting framework also lies at the heart of the EU-KLEMS project, which focuses on deducing multi-factor productivity from production function specification with “wide” definitions of capital and labour inputs (capital services and quality-adjusted labour inputs). In this chapter, by contrast, firm-level capital and labour inputs (and intermediate inputs in multi-factor productivity specifications) are simple measures based on capital stocks and headcount employment.

5. As noted in Chapter 14, a central conclusion of the EU-KLEMS project is that the growth accounting framework cannot “explain” productivity. That is, changes in output cannot be accounted for wholly (or even predominantly) by changes in measured inputs. A

94


substantial part of productivity is “Total Factor Productivity” or the unexplained residual in the growth accounting framework1. Although unexplained in accounting terms, there is no shortage of competing hypotheses for the residual source of productivity, including organisational capital and measurement error. Several authors have noted a role for ICT2. It should be noted, however, that EU-KLEMS explicitly includes ICT capital services as an input. There may be question marks over its measurement3, but in principle the EU-KLEMS framework accounts for the direct productive contribution of ICT inputs.

6. In principle, one could apply the growth accounting framework at firm level, but there are a number of practical problems:

• Firm-level data are in nominal terms. No firm-level deflators are available, either for outputs or inputs. This matters because we know that firms are heterogeneous both in terms of their physical productivity (again, this is conditional on our ability to measure productive inputs at the firm level) and in terms of pricing4.

• In addition, there is a paucity of firm-level data at the level of detail used in the EU-KLEMS project. For example, some project members have a detailed breakdown of capital inputs at the firm level, including a split by asset types, and some countries have detailed information on labour inputs, including educational attainment and hours worked. But no country in this project has both of these.

7. In effect there are trade-offs between productivity analysis at the firm level, which is the subject of this chapter, and at the more aggregated level reported in Chapter 13. Data deficiencies at firm-level mean that it is not possible to replicate the growth accounting framework which is now the standard approach at the aggregate level. It is generally not possible to match detailed information on labour and capital inputs with firm level outputs. And the volume/price decomposition of firm-level output is under-researched5.

8. On the other hand, firm-level data provide a means whereby the researcher can analyse firm characteristics that may have a bearing on productivity. One such characteristic is the firm’s use of ICT, which could be considered to be proxy for innovation and responsiveness to change, or perhaps a proxy for technology inputs that are not captured in measured capital inputs.

9. In essence, the motivation of this chapter is a simple one: whether the observed heterogeneity among firms can be systematically related to their use of ICT, conditional upon other measured productive inputs. The intuition goes both ways – everyone can cite examples of ICT usage which have spurred productivity (e.g. by speeding up access to information) and equally examples of botched ICT projects which have consumed resources to no useful end.

10. A key point is that we are at the stage only of identifying relationships, not causality. Finding that, say, firms with high proportions of their workforce with access to fast internet connections have higher productivity, other things equal, than other firms does not establish

1 Total Factor Productivity (TFP) and Multi Factor Productivity (MFP) have specific definitions in this report (see Chapter 3) but are sometimes used interchangeably in the general productivity literature. 2 OECD (2004) and Clayton et al. (2005) provide overviews of ICT impacts on productivity. Both are described in Chapter 2. Links between ICT and organisational capital and organisational change are reviewed in Bloom, Sadun and Van Reenen (2007) and Crespi, Criscuolo and Haskel (2007). Maliranta and Rouvinen (2004) examine relationships between ICT and firm demographics. For linkages between ICT and broader intangibles currently not measured as productive inputs, see Crespi et al (2008). 3 For example, the EU-KLEMS include own-account software only to the extent that it is included in the underlying national accounts data, since EU-KLEMS ICT capital services data are derived by partitioning national accounts data on business investment. The reconciliation of these series with firm level data on ICT capital stocks (see Chapter 10) is beyond the scope of this project. 4 Syverson (2004) examines physical productivity in US firms. For price heterogeneity see Foster, Haltiwanger and Krizan (2001), Fitzgerald and Haller (2008). 5 This is not meant to imply that there are no data issues at the aggregate level, only that data generation systems and processes are more developed at that level.

95


that fast internet access is the driver and productivity the result. It is perfectly possible (a) that causality is the other way round and (b) fast internet access is only a proxy for some other feature of the firm that is not measured.

11. Figure 6.1 uses some data from the country/industry datasets to illustrate the endogeneity issue. The data are for the manufacturing sector (excluding electronics) and show the level of fast internet penetration (DSLPCT) by quartile of the wage distribution, for a single year (2004). In almost all countries for which data are available, there is a strong pattern of DSLPCT rising with wage quartiles, such that on average, DSLPCT is around 3 times greater in firms in the highest wage quartile than in firms in the lowest wage quartile.

12. Not surprisingly, our regressions confirm that measures of revenue productivity also correlate strongly with average wage levels. Indeed, one of the strongest findings of the regression analysis presented later in this chapter is that coefficients on wages are positive and significant in all specifications with productivity as the dependent variable. What Figure 6.1 tells us is that fast internet penetration is also correlated with wages. It is not straightforward, therefore, to identify the separate contributions to productivity of wages and DSLPCT.

Figure 6.1 – Fast internet penetration and wages

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

AUT CZE GBR GER IRE ITA NLD NOR SLO SWE TOTAL

DSL

pct (

1=10

0%)

Wage Qrt1Wage Qrt2

Wage Qrt3Wage Qrt4

This graph shows mean dslpct, quartiled into wage distributionYear==2004euk==MexElec

6.2.2 Choice of core ICT metrics 13. As described in Chapter 4, some ICT usage metrics such as access to PCs and access to the internet are already approaching saturation. For this reason, and reflecting previous work in some project countries (e.g. the UK) the analysis in this chapter focuses on the proportion of workers with high speed internet access (DSLPCT – in project mnemonics).

14. DSLPCT is a derived or “complex” variable. The standardised E-Commerce survey does not ask for this information directly. We have calculated DSLPCT as the product of the

96


Boolean variable DSL6 and the continuous variable INTERPCT7. The focus on DSLPCT is not fixed or immutable; it would be relatively straightforward to redo the analysis using other E-Commerce metrics. And the focus on DSLPCT is not exclusive. The regression analysis also includes ECPCT, another derived variables computed as the sum of the percentage shares of E-sales (in total sales) and E-purchases (in total purchases).

15. The methodology uses continuous ICT usage indicators rather than Boolean variables8. Again there is nothing sacrosanct about this; it is certainly possible to construct models (such as principal component models) using Boolean variables. Recall that at this stage we are not testing a specific hypothesis about the role of ICT in the production process.

16. One specific methodological issue that needs to be mentioned is that matching firm-level data from sample surveys leads to heavy attrition of first difference variables in the matched datasets. This arises because in many jurisdictions NSIs specifically avoid sampling small firms in consecutive years. Datasets in first differences are, therefore, even more biased towards larger firms (who are sampled every year) than are merged data for a single year. We report a regression in first differences below, but most of the results in this and other chapters are in levels with no dynamic properties.

17. On the other hand it is certainly plausible that ICT impacts productivity with a time lag. And analysis of UK data by Farooqui and Sadun (2006) supports the first mover advantage argument, namely that early adopters of ICT tend to display higher productivity than later adopters9. Thus it would be preferable to be able to analyse a balanced firm-level dataset over a number of time periods. One possible route would be to include survey questions that ask about dates of adoption of particular ICTs. Another possibility is to build a representative panel of firms which can be followed over a number of years.

6.2.3 Outline of rest of chapter 18. The following section describes the regression framework and some preliminary results. These results should be interpreted as a demonstration of the ability to run identical regressions across a group of countries, rather than a definitive analysis of the underlying data. As mentioned in Chapter 1, we have had comparatively little time to analyse all of the data that has been assembled for this project.

19. Section 6.4 presents some informal descriptive statistics drawn from the industry/country datasets. The main chapter concludes with some suggestions for subsequent research.

20. Appendix A to the chapter reports some more further econometric analysis using data for the Netherlands and the UK. This work covers:

• Informal analysis of relationships between productivity and date of adoption of fast internet connectivity;

• Comparisons of productivity regressions including firm-level data on ICT capital stocks alongside ICT-usage metrics;

• Industry level regressions to test for differential pathways in different sectors.

6 The 2006 model questionnaire (pertaining to 2005 data) asks (Question B3) “Did your enterprise have the following types of external connection to the internet, during January 2006 …… (c) DSL connection?” In addition, Question B5 of the model questionnaire asks about maximum download speeds. 7 The suggested wording from the 2006 model questionnaire is “Please indicate an estimate of the percentage of the number of persons employed using computers connected to the World Wide Web at least once at week, during January 2006.” 8 DSL is an example of a Boolean variable, taking values 0 or 1. 9 See also Appendix A to this chapter by Shikeb Farooqui which compares adoption effects in the UK and the Netherlands.

97


6.3 Firm-level results

6.3.1 Framework 21. The firm-level regressions are hard-wired into the identical analytical code which has been run by all project members except Denmark and Slovenia. The regressions run on the same firm-level data that underlies the industry/country statistics described in Chapter 4 and used in the EU-KLEMS analytical framework described in Chapter 14.

22. The regressions run on up to 4 productivity metrics (see Chapter 3) and their first differences. Two of the productivity metrics are in terms of value added, and two in terms of gross output. Our preferred specifications use the value-added metrics (LPV and TFP) rather than gross output which is used in much of the productivity literature. This is for the pragmatic reason that data on intermediate inputs are not available in some project countries.

23. TFP (which is our preferred dependent variable in the following exposition of results) is a proxy index of real value-added, after taking account of labour and capital inputs using industry average weights. This can be linked to the traditional Cobb-Douglas production function:

)1( αα −= itititit KLAY (1)

Where Y is value added in constant prices, L is labour inputs measured as employment, K is a measure of total capital inputs, A represents TFP, i and t refer to firms and time respectively, and α is the labour factor share, computed as the industry share of wages in total value added.

24. Re-arranging equation (1) and taking logarithms give equation (2):

)ln()1()ln()ln()ln( itititit KLYA αα −−−= (2)

Where ln(A) is ln(TFP). Some properties of TFP are described in Chapter 4. In general, we observe significant variation in TFP among firms in the same industries.

6.3.2 Regression specifications 25. The primary specification hard-wired into the code uses firm-level human-capital indicators, as follows:

dummiesHKNITPCTbHKITPCTbRHSVbbLHSV oz ++++= 321

Where LHSV is one of the productivity metrics (levels or first differences), RHSV is a vector of independent variables dependent on LHSV and varying across four separate specifications, HKITPCT is the share of the firm’s employees with IT-related post-secondary skills and HKNITPCT is the share of the firm’s employees with non-IT related post-secondary skills. The choice of dummies depends on the specification and on data availability. But all regression specifications include industry, size-class (based on employment) and year dummies.

26. However, only 3 project countries (Sweden, Finland and Norway) have firm-level data on skills. Some results for the hard-wired regression specifications for these countries are reported in Chapter 8.

27. For the purposes of this chapter we will focus on a second specification which uses the log of average wages (LNW) as a proxy for labour skills. Note that LNW is computed as the firm wage bill divided by the number of employees. Note also that, for obvious reasons, the wage bill tends to be highly correlated with gross output or value-added.

28. Each regression is run over 6 different levels of coverage, as shown in Table 6.1.

98


Table 6.1 – Regression coverage All firms Electronics and telecomsManufacturing sector Manufacturing (excl electronics)Market services Market services (excl telecoms) 29. There is one further feature of the regression framework to mention before describing the results. The regression code carries out a pre-regression of the fullest available productivity specification based on data availability in each country, and removes outliers based on the 1 percent tails of the distribution of residuals. In other words, firms with the highest and lowest productivity are removed.

6.3.3 Results 30. Table 6.210 shows regression results for the whole sample for the most basic specification with LPV as the dependent variable. Note that MNC is a dummy variable for whether the firm is part of a multi-national enterprise. Other dummies for industry, size-class and year are unreported. Significant t-statistics (at the 5% level) are shown in bold. The ordering of the countries is completely arbitrary, based only on the timing of delivery of results.

31. Points to note:

• The number of observations varies from around 6,000 (for Italy) to 3.9 million (for France). This reflects the feature that this specification uses the full PS sample (which in France and Sweden covers all firms). No data from the EC survey is included in this specification.

• The co-efficient on LNW is positive and significant for all countries. The overall range of coefficients is 0.17 (Finland) to 1.11 (Italy). Apart from the unusually low coefficient for Finland, the next smallest co-efficient is 0.51 (Austria).

• For those countries with such data, coefficients on MNC are positive and significant, ranging from 0.03 (Netherlands) to 0.10 (France).

• The equation r-squared varies from 0.19 (Finland) to 0.96 (Ireland).

32. Table 6.3 shows the same regression specification but with DLPV, the first difference of LPV, as the dependent variable. Points to note:

• The number of observations falls sharply compared with Table 6.2, for two reasons. Firstly because of survey sampling - for this specification we need firms that appear in consecutive PS samples. And secondly because we lose the first year.

• Coefficients on LNW are again universally positive and significant, suggesting that higher wages are associated with faster growth of productivity, albeit that the explanatory power of this specification is very low (equation r-squared ranging from zero to 0.08).

• The MNC dummy has a small but significant negative effect on DLPV. Together with the results in Table 6.2, this suggests that multinational firms have higher productivity levels, but lower productivity growth, than non-multinational firms.

33. Table 6.4 shows results for a fuller specification with TFP as the dependent variable and the regression confined to the manufacturing sector (excluding the IT producing electrical and optical sub-sector). LNKL is the log of (total) capital inputs per employee; AGE is the age of the firm in years and AGE2 is the age squared (to test for non-linear effects). DSLPCT measures the percentage of the firm’s workers with access to high-speed internet; ECPCT is the sum of the shares of E-sales (in total sales) and E-purchases (in total

10 Tables of regression results are at the end of the main chapter text.

99


purchases). Some descriptive statistics of the data used in this regression are shown in Chapter 4.

34. The number of observations in this regression is much lower than in Tables 6.2 and 6.3 because this specification uses variables from the EC survey and is therefore restricted to firms that appear in both the EC and PS surveys in a particular year. For reasons set out in Chapter 5, this biases the sample towards larger firms.

35. There are no results for this specification for France or Ireland. For France, the reason is that the composition of their EC survey is such that there are no responses for fast internet access and E-sales/E-purchases in the same survey year; hence there are zero observations in the regression set. Ireland has not been able to provide firm-level capital input data and hence we cannot compute TFP.

36. The results of this regression are mixed. Coefficients on DSLPCT are not universally significant, and the significant coefficients range from 0.10 in Germany to 0.52 in Norway. The coefficient on DSLPCT is marginally insignificant in the Netherlands and not significant in the UK and Italy. The coefficient on ECPCT registers as significant only in Germany11. These results broadly confirm the findings from single-country micro-data studies, namely that ICT usage is positively related to productivity in the manufacturing sector.

37. The addition of ICT usage variables does not affect the sign of LNW which remains significant in all countries and strongly significant (at the 1% level) in all countries bar Norway (where the number of observations is comparatively low). The multinational status dummy remains significant in all countries where such data are available. However, the coefficient on the multinational dummy changes sign (compared with Table 6.2) in Sweden and the Czech Republic. Age variables do not exist for most countries, and where data do exist, the coefficients are insignificantly different from zero.

38. Note finally the negative coefficients on LNKL (capital per employee). This is not an unusual finding and reflects the fact that the output elasticity with respect to capital inputs is less than the factor share of capital that is implied in the calculation of TFP.

39. The R-squared of this specification ranges from 0.32 in the Netherlands to 0.77 in Italy and is 0.6 or higher for 6 of the 9 countries for which we have results.

40. Table 6.5 repeats the specification of Table 6.4 but in a regression confined to the ICT producing sectors12. The striking feature of these results is that there is no longer a positive relationship between the ICT usage variables and TFP. The coefficient on DSLPCT is significantly negative for Sweden and Italy, and the coefficient on ECPCT is significantly negative for Finland. In this industry there is some evidence that firm-age has an impact on performance – coefficients on AGE and AGE2 are significant in Finland and Sweden, though with opposite signs. In Finland, older ICT-producing firms are marginally less productive than younger firms, other things equal, while in Sweden older ICT-producing firms are more productive.

41. The wage proxy for skills remains generally significant but becomes insignificant in Norway and Austria, and the coefficient is implausibly large in Italy. Another feature of this table is that the number of observations is becoming very thin for some countries (Norway, Italy)13. Despite the smaller sample size, there is no clear pattern in the explanatory power of the specification compared with Table 6.4. In 5 of the 9 countries which have generated results, the r-squared is lower than in the broader manufacturing sector, while in the other 4 countries the r-squared in higher than in Table 6.4.

11 It is possible that adding together esalespct and epurchpct may mask separate relationships for the two indicators, see also Chapter 10. 12 The electrical and optical equipment industry within the manufacturing sector can be thought of as the producer of IT, and the telecoms sub-sector of services as the producer of CT. 13 Strictly speaking the number reported as observations is the number of degrees of freedom, that is, the number of observations minus the number of regressors.

100


42. Table 6.6 repeats the specification for the market services sector (excluding post and telecommunication services). Apart from Finland and Sweden, where the coefficients on DSLPCT are positive and significant, there is little evidence of a systematic relationship between the ICT usage variables and TFP. Coefficients on DSLPCT are negative in several countries and significantly negative in the Czech Republic. The coefficient on ECPCT is significantly positive in Austria and in Sweden (where it is close to zero) but significantly negative in Italy.

43. Another interesting feature of Table 6.6 is that the coefficients on the multinational dummy turn significantly negative in the Netherlands and Sweden and are not significant elsewhere. The age variables display similar characteristics to those in the ICT-producing sector (Table 6.5). As before, we find a significant positive role for wages, apart from in Germany, where the number of free observations is very low.

44. For completeness, Table 6.7 repeats the specification of Tables 6.4 through 6.6 for all firms. The coefficients on DSLPCT are generally positive but significant only for Finland, Sweden and Germany. Coefficients on ECPCT are positive and significant for the Netherlands, Austria and Germany. Comparisons between Table 6.7 and the sectoral regressions in Tables 6.4-6.6 clearly point to differences in the relationship between ICT usage (measured here by DSLPCT and ECPCT) and productivity, with the strongest relationship evident in manufacturing (excluding ICT) and the weakest relationship in the ICT-producing sector.

45. Table 6.8 reverts to the manufacturing sector (excluding ICT), applying a dynamic version of the specification shown in Table 6.4. As expected, the number of observations is lower, but generally of the same order of magnitude as Table 6.4. This reflects the bias towards larger enterprises, which are more likely to be surveyed in consecutive years. This dynamic specification tests for a one-year relationship between ICT usage and productivity, and it is perhaps not surprising that such a relationship is largely absent14. But it is interesting to note that the coefficient on DSLPCT is significantly negative in Finland and Sweden. Combined with Table 6.4, this suggests that firms with high levels of fast internet penetration enjoy higher levels of productivity than their counterparts in these countries, but lower rates of productivity growth.

46. Lastly in this section, Tables 6.9 and 6.10 present results for a slightly different specification. LNWDSL is the interaction of LNW and DSLPCT, and ECPCT is dropped. Table 6.9 shows results for the manufacturing sector, and Table 6.10 shows results for market services, in both cases excluding ICT-producing sub-sectors. The interaction term tests for complementarity between wages (as a proxy for skills) and ICT intensity, measured by the penetration of fast-internet access15. Note that interpretation of coefficients on DSLPCT in this specification depend on the level of LNW, which can vary by orders of magnitude across countries. Broadly speaking, the reduced form coefficients on DSLPCT are very close to those in Tables 6.4 and 6.6. Intuitively, this is because the coefficients on ECPCT in the previous specification are generally not significant. Note also that by dropping ECPCT we now obtain results for France.

47. For the manufacturing sector (Table 6.9) the coefficient on the interaction term is positive and significant for the Netherlands and Austria, but negative and significant for France. For market services (Table 6.10) the coefficient on the interaction term is positive and significant for the United Kingdom, Sweden and France, but significantly negative for Italy. Together these results suggest that the interaction between skills (as measured here by wages) and ICT usage is not straightforward, but varies both across countries and across sectors.

14 By contrast, the specification in Table 6.4 implicitly allows for multi-year effects. Note however that the rapid take-up of new technologies described in Chapter 4 possibly implies that time lags are comparatively short. 15 Some results using an interaction between “true” skills data and DSLPCT are shown in Chapter 8.

101


48. At this stage we should repeat that there is nothing sacrosanct or definitive about the regression specifications reported in this chapter. In principle it is possible to run many alternative specifications within the data parameters of the project, subject only to the practical constraints of distributing and running revised versions of the project code. Some suggestions for alternative specifications are discussed in the final section of this chapter.

6.4 Some Descriptive Statistics 49. Some informal insights into ICT usage and productivity can be obtained from examining information from the country/industry datasets on the properties of joint distributions. Figure 6.2 is similar to Figure 6.1 but shows the mean level of fast-internet penetration across the quartiles of the productivity distribution for firms in the manufacturing sector (excluding ICT) in 2004. Countries are ranked by the average of DSLPCT. In general, firms in the higher quartiles of the productivity distribution tend to have higher average levels of DSLPCT, although the variation across countries is somewhat greater than the variation across the quartiles for most countries.

Figure 6.2 – Fast internet penetration and productivity

0

0.1

0.2

0.3

0.4

0.5

0.6


DSL

pct (

1=10

0%)


This graph shows mean dslpct, quartiled into TFP distributionYear==2004euk==MexElec

50. Figure 6.3 shows corresponding information on the correlations between DSLPCT and TFP, again across the quartiles of the TFP distribution using the same underlying data as Figure 6.2. For clarity, data in Figure 6.3 represent correlations between firm-level observations on DSLPCT and TFP, with a separate correlation for each quartile of the TFP distribution. It is interesting to note that, while on average, correlations become more positive as we move up through the productivity distribution, among firms in the lowest quartile there is, on average, a negative correlation between DSLPCT and TFP. For firms in the second quartile the average correlation is close to zero, and in fact only one country (the UK) has a monotonically increasing correlation coefficient over the four quartiles.

51. At first sight it may seem difficult to reconcile Figure 6.3 with Figure 6.2. The explanation is that Figure 6.2 shows that the average level of DSLPCT tends to increase as we move up the quartiles of the TFP distribution. By contrast, Figure 6.3 shows that, within the quartiles, it is far from the case that DSLPCT is positively correlated with TFP. On the

102


contrary, among firms in the lowest productivity quartile, the data show that on average, higher DSLPCT correlates with lower TFP, while Figure 6.2 shows that, across all firms in the lowest TFP quartile, the average level of DSLPCT is generally lower than among firms in higher productivity quartiles.

52. This is clear evidence that there is no simple relationship between DSLPCT and productivity, a finding which is also supported by the regression results reported in this chapter and elsewhere in the project report.

Figure 6.3 – Quartile correlations

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5


Cor

rela

tion

Coe

ffici

ent


This graph shows correlation coefficient between dslPCT and TFP, quartiled into TFP distributionYear==2004euk==MexElec

6.5 Next Steps 53. This project has assembled a great deal of data but has had comparatively little time for analysis, (for example, of the coefficients on the fixed effects) and no time at all for the process of iteration between results and specifications that is common in an econometric investigation. Avenues for further investigation might include the following:

• Analysis of descriptive statistics and cross-correlations of ICT metrics;

• Test alternatives to DSLPCT as an indicator of ICT usage across different countries; test EPURCHPCT and ESALESPCT separately rather than as a combined metric;

• Test LPV specifications including capital inputs, and TFP specifications excluding capital inputs;

• Test across finer industry breakdown than that in Table 6.1, e.g. including distribution and differentiated services;

• Test some form of ICT adoption specification;

• Test the impact of outlier adjustment.

103


Table 6.2 Dependent Variable REAL VALUE-ADDED PER EMPLOYEECoverage ALL FIRMS

GBR FIN NLD SWE FRA ITA CZE NOR IRE AUT GERRegressor coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-statlnw 0.75 297.0 0.17 168.1 0.59 280.9 0.63 679.7 0.79 1803.1 1.11 91.3 0.91 220.1 0.69 162.8 0.66 107.7 0.51 345.2 0.81 204.8mnc dummy 0.09 17.1 0.03 7.6 0.08 22.3 0.10 10.6 0.04 5.5[other dummies]Observations 128229 388950 188573 861907 3877391 6086 90150 46930 36106 512681 84380R-squared 0.50 0.19 0.36 0.43 0.49 0.68 0.47 0.51 0.96 0.34 0.57

Table 6.3 Dependent Variable CHANGE IN REAL VALUE-ADDED PER EMPLOYEECoverage ALL FIRMS

GBR FIN NLD SWE FRA ITA CZE NOR IRE AUT GERRegressor coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-statlnw 0.11 24.0 0.05 42.0 0.23 69.0 0.22 183.0 0.25 444.6 0.20 10.3 0.06 11.5 0.10 18.5 0.27 21.3 0.04 26.0 0.06 13.4mnc dummy -0.02 -2.8 -0.02 -5.5 -0.03 -8.7 -0.10 -9.3 0.00 -0.47[other dummies]Observations 46276 261343 86796 614299 2543056 3248 42265 17617 14560 351148 44775R-squared 0.02 0.01 0.06 0.06 0.08 0.05 0.03 0.03 0.04 0.00 0.01 Table 6.4 Dependent Variable REAL TOTAL FACTOR PRODUCTIVITYCoverage MANUFACTURING EXCL ELECTRICAL AND OPTICAL EQUIPMENT

GBR FIN NLD SWE FRA ITA CZE NOR IRE AUT GERRegressor coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-statlnkl -0.30 -16.6 -0.45 -48.3 -0.22 -26.8 -0.21 -28.0 #N/A #N/A -0.29 -9.0 -0.51 -98.9 -0.30 -7.4 #N/A #N/A -0.26 -10.2 -0.16 -28.7lnw 0.87 18.3 0.57 10.2 0.62 25.6 0.94 22.4 #N/A #N/A 1.17 6.9 1.09 45.8 0.52 2.2 #N/A #N/A 0.72 8.6 0.78 39.8AGE 0.01 1.1 0.00 0.8 #N/A #N/A 0.01 1.3 #N/A #N/A 0.00 -0.1 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Aage2 0.00 0.7 0.00 0.8 #N/A #N/A 0.00 -0.4 #N/A #N/A 0.00 0.6 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Adslpct 0.11 1.6 0.32 8.4 0.06 2.0 0.29 9.1 #N/A #N/A 0.15 0.7 0.17 4.7 0.52 2.4 #N/A #N/A 0.22 2.5 0.10 4.3ECpct 0.03 0.4 0.06 1.0 0.03 1.1 0.00 -0.2 #N/A #N/A 0.44 1.9 0.04 1.7 0.11 1.0 #N/A #N/A 0.01 0.2 0.06 3.9mnc dummy 0.09 2.6 0.03 2.2 -0.07 -2.9 #N/A #N/A -0.04 -2.2[other dummies]Observations 1451 3165 3970 3807 #N/A 196 5508 92 #N/A 492 7283R-squared 0.48 0.68 0.32 0.60 #N/A 0.77 0.75 0.65 #N/A 0.34 0.70

104


Table 6.5 Dependent Variable REAL TOTAL FACTOR PRODUCTIVITYCoverage ELECTRICAL & OPTICAL EQUIPMENT, POST & TELECOMMUNICATION SERVICES

GBR FIN NLD SWE FRA ITA CZE NOR IRE AUT GERRegressor coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-statlnkl -0.28 -8.0 -0.39 -14.1 -0.27 -11.9 -0.16 -7.0 #N/A #N/A -0.38 -1.7 -0.51 -51.6 -0.37 -3.1 #N/A #N/A -0.14 -2.3 -0.24 -7.0lnw 0.84 9.4 0.78 4.9 0.49 6.5 0.74 6.6 #N/A #N/A 2.35 3.3 0.86 20.6 1.67 1.5 #N/A #N/A 0.20 1.0 0.67 6.3AGE 0.01 0.4 -0.02 -2.8 #N/A #N/A 0.06 2.4 #N/A #N/A 0.05 0.7 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Aage2 0.00 0.5 0.00 2.3 #N/A #N/A 0.00 -2.4 #N/A #N/A 0.00 -0.6 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Adslpct -0.09 -0.9 0.02 0.1 0.07 1.0 -0.37 -3.3 #N/A #N/A -1.66 -2.4 0.02 0.4 0.47 0.8 #N/A #N/A 0.17 0.8 -0.17 -1.7ECpct 0.27 1.4 -0.64 -4.1 0.17 2.0 0.00 0.1 #N/A #N/A 0.22 0.2 0.14 3.2 -0.07 -0.1 #N/A #N/A -0.03 -0.2 0.10 1.4mnc dummy 0.02 0.3 0.03 0.6 0.17 1.7 #N/A #N/A -0.03 -0.7[other dummies]Observations 412 483 599 402 #N/A 14 1029 5 #N/A 99 1050R-squared 0.38 0.48 0.48 0.39 #N/A 0.69 0.83 0.95 #N/A 0.42 0.13 Table 6.6 Dependent Variable REAL TOTAL FACTOR PRODUCTIVITYCoverage MARKET SERVICES EXCL POST & TELECOMS

GBR FIN NLD SWE FRA ITA CZE NOR IRE AUT GERRegressor coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-statlnkl -0.34 -33.1 -0.37 -33.2 -0.14 -12.4 -0.20 -47.3 #N/A #N/A -0.38 -12.9 -0.51 -126.2 -0.35 -15.2 #N/A #N/A -0.22 -11.2 -0.40 -2.2lnw 0.95 41.4 1.07 17.5 0.92 37.2 1.11 49.9 #N/A #N/A 1.26 10.7 0.94 55.5 1.19 10.7 #N/A #N/A 0.98 15.5 -0.58 -1.0AGE 0.00 0.2 -0.01 -2.1 #N/A #N/A 0.01 3.3 #N/A #N/A 0.01 0.8 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Aage2 0.00 0.5 0.00 2.0 #N/A #N/A 0.00 -3.0 #N/A #N/A 0.00 -0.4 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Adslpct 0.06 1.6 0.31 6.3 0.01 0.2 0.08 4.6 #N/A #N/A -0.03 -0.2 -0.10 -4.1 -0.05 -0.5 #N/A #N/A -0.10 -1.8 0.77 1.4ECpct 0.02 0.2 -0.05 -0.6 0.07 1.6 0.00 2.5 #N/A #N/A -1.40 -2.8 -0.01 -0.5 0.02 0.4 #N/A #N/A 0.12 2.8 0.56 0.3mnc dummy 0.04 1.5 -0.08 -2.9 -0.05 -2.7 #N/A #N/A -0.05 -1.8[other dummies]Observations 4316 5005 6773 5816 #N/A 282 6145 203 #N/A 606 12R-squared 0.40 0.39 0.24 0.63 #N/A 0.81 0.80 0.67 #N/A 0.49 0.93

105


Table 6.7 Dependent Variable REAL TOTAL FACTOR PRODUCTIVITYCoverage ALL FIRMS

GBR FIN NLD SWE FRA ITA CZE NOR IRE AUT GERRegressor coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-statlnkl -0.32 -40.1 -0.40 -56.9 -0.17 -23.8 -0.25 -73.2 #N/A #N/A -0.34 -16.1 -0.52 -195.1 -0.33 -17.4 #N/A #N/A -0.24 -15.7 -0.15 -29.1lnw 0.93 49.1 0.92 23.8 0.86 50.3 1.04 53.7 #N/A #N/A 1.20 13.4 0.97 81.1 1.03 10.6 #N/A #N/A 0.82 16.6 0.72 39.3AGE 0.00 0.4 0.00 -1.5 #N/A #N/A 0.01 3.2 #N/A #N/A 0.01 1.3 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Aage2 0.00 1.8 0.00 2.1 #N/A #N/A 0.00 -2.7 #N/A #N/A 0.00 -0.5 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Adslpct 0.03 1.1 0.29 9.2 0.02 0.7 0.11 6.5 #N/A #N/A -0.08 -0.6 -0.02 -1.3 0.08 1.0 #N/A #N/A 0.01 0.2 0.06 2.7ECpct 0.05 0.9 -0.06 -1.3 0.06 2.2 0.00 1.7 #N/A #N/A -0.17 -0.7 0.03 1.8 0.07 1.5 #N/A #N/A 0.09 2.8 0.07 4.3mnc dummy 0.04 2.1 -0.03 -1.9 -0.03 -2.0 #N/A #N/A -0.02 -1.31[other dummies]Observations 6641 9881 12595 11392 #N/A 555 14849 343 #N/A 1254 10342R-squared 0.45 0.62 0.27 0.81 #N/A 0.79 0.83 0.69 #N/A 0.40 0.84 Table 6.8 Dependent Variable CHANGE IN REAL TOTAL FACTOR PRODUCTIVITYCoverage MANUFACTURING EXCL ELECTRICAL AND OPTICAL EQUIPMENT

GBR FIN NLD SWE FRA ITA CZE NOR IRE AUT GERRegressor coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-statdkl 0.24 3.9 -0.39 -25.3 -0.19 -13.2 -0.10 -7.4 #N/A #N/A -0.38 -10.4 -0.53 -46.6 -0.02 -0.2 #N/A #N/A -0.30 -6.7 -0.16 -15.8lnw 0.12 3.0 0.22 6.6 0.17 7.3 0.37 11.8 #N/A #N/A 0.40 3.4 0.07 3.3 0.30 2.7 #N/A #N/A 0.17 2.4 0.04 2.8AGE 0.01 1.7 0.00 -0.7 #N/A #N/A 0.00 -0.6 #N/A #N/A 0.01 0.9 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Aage2 0.00 -1.5 0.00 0.8 #N/A #N/A 0.00 0.3 #N/A #N/A 0.00 -0.8 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Adslpct -0.05 -0.8 -0.09 -3.9 -0.03 -1.1 -0.08 -3.1 #N/A #N/A 0.27 1.9 -0.02 -0.8 -0.06 -0.7 #N/A #N/A 0.03 0.4 0.00 -0.2ECpct 0.09 1.2 0.02 0.7 0.06 2.0 0.00 1.6 #N/A #N/A -0.06 -0.4 0.02 1.1 0.05 1.0 #N/A #N/A -0.05 -1.0 -0.01 -0.9mnc dummy 0.04 1.1 -0.01 -1.0 -0.02 -1.1 #N/A #N/A 0.01 0.3[other dummies]Observations 1157 2370 3276 2877 #N/A 155 3852 87 #N/A 464 4737R-squared 0.06 0.23 0.07 0.08 #N/A 0.50 0.39 0.22 #N/A 0.12 0.06

106


Table 6.9 Dependent Variable REAL TOTAL FACTOR PRODUCTIVITYCoverage MANUFACTURING EXCL ELECTRICAL AND OPTICAL EQUIPMENT

GBR FIN NLD SWE FRA ITA CZE NOR IRE AUT GERRegressor coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-statlnkl -0.30 -16.7 -0.45 -50.7 -0.22 -28.1 -0.21 -27.4 -0.22 -28.0 -0.30 -9.2 -0.51 -98.1 -0.24 -29.2 #N/A #N/A -0.20 -26.5 -0.16 -28.3lnw 0.87 14.0 0.53 7.9 0.56 20.4 1.02 16.8 0.94 30.5 1.22 6.3 1.08 38.9 1.04 15.5 #N/A #N/A 0.77 26.1 0.77 33.6AGE 0.01 1.1 0.00 1.3 #N/A #N/A 0.01 1.6 0.00 1.5 0.00 -0.2 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Aage2 0.00 0.7 0.00 0.5 #N/A #N/A 0.00 -0.7 0.00 -1.3 0.00 0.7 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Adslpct 0.10 0.2 0.99 0.8 -1.28 -3.9 1.47 2.1 1.15 3.3 2.58 0.4 -0.12 -0.3 1.16 2.0 #N/A #N/A -0.62 -1.9 -0.06 -0.1lnwdsl 0.01 0.0 -0.06 -0.5 0.35 4.0 -0.19 -1.7 -0.32 -3.5 -0.23 -0.3 0.05 0.7 -0.17 -1.8 #N/A #N/A 0.19 2.3 0.02 0.3mnc dummy 0.09 2.7 0.03 2.1 -0.07 -2.8 -0.12 -2.2 -0.04 -2.3[other dummies]Observations 1452 3543 4285 3867 3903 196 5508 1930 #N/A 3829 7283R-squared 0.48 0.68 0.34 0.58 0.51 0.77 0.75 0.61 #N/A 0.35 0.70

Table 6.10 Dependent Variable REAL TOTAL FACTOR PRODUCTIVITYCoverage MARKET SERVICES EXCL POST & TELECOMS

GBR FIN NLD SWE FRA ITA CZE NOR IRE AUT GERRegressor coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-statlnkl -0.47 -17.6 -0.38 -36.8 -0.14 -12.9 -0.19 -47.5 -0.28 -37.9 -0.38 -13.0 -0.51 -125.0 -0.31 -66.3 #N/A #N/A -0.15 -18.2 -0.41 -2.3lnw 0.93 12.6 1.16 12.8 0.87 28.5 1.03 31.6 0.96 32.4 1.44 10.3 0.92 40.5 0.96 32.0 #N/A #N/A 0.82 31.0 -0.64 -1.0AGE 0.04 1.4 -0.01 -2.2 #N/A #N/A 0.01 3.7 0.01 3.3 0.01 0.9 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Aage2 0.00 -0.8 0.00 2.0 #N/A #N/A 0.00 -3.4 0.00 -3.5 0.00 -0.6 #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/A #N/Adslpct -1.87 -5.2 1.64 1.4 -0.18 -1.0 -0.72 -2.8 -0.87 -3.9 7.36 2.5 -0.41 -2.1 0.11 0.5 #N/A #N/A -0.15 -0.9 -2.82 -0.1lnwdsl 0.67 6.0 -0.13 -1.1 0.05 1.1 0.14 3.1 0.25 4.3 -0.71 -2.5 0.06 1.6 -0.02 -0.4 #N/A #N/A 0.07 1.6 0.34 0.1mnc dummy 0.12 1.8 -0.08 -2.8 -0.05 -2.9 0.04 0.3 -0.05 -1.8[other dummies]Observations 4467 5675 7746 6011 3250 282 6145 2717 #N/A 4858 12R-squared 0.50 0.41 0.24 0.63 0.57 0.81 0.80 0.75 #N/A 0.38 0.93

107


Appendix 6.A: Productivity, ICT Capital and Fast Internet Adoption in the UK and the Netherlands

Shikeb Farooqui UK Office for National Statistics

6A.1 Introduction A1. This Chapter lays down the basis for identifying and evaluating the relationships between ICT use and productivity. The analysis centres on two core variables that define ICT use intensity within a firm: the proportion of the firm’s workforce with access to high-speed internet connection and the proportion of a firm’s total trade undertaken through electronic mediums, such as the internet.

A2. The results suggest that, on average, ICT usage has a positive relationship with productivity across the different countries considered. However, differences in individual country coefficients, and between ICT use types suggest that the impact of ICT use has had a varying impact according to use and across sample countries.

A3. This raises a fundamental question as to whether ICT use really serves as a general-purpose technology, or whether it is use and environment specific. In order to answer this question we dissect the data deeper for the Netherlands and UK and conduct industry level regressions.

A4. The primary conclusions from this analysis are qualitatively similar for both countries and show that the benefits from ICT use are industry specific, with certain types of businesses benefiting from a particular use of ICT and other types benefiting to a greater degree from other sources of ICT use. In fact the analysis will show, for both Netherlands and UK, that:

• Manufacturing firms primarily benefit from e-procurement; • Distribution firms source higher productivity from implementing e-sales; • Firms in Differentiated Services evidence gains from increasing the proportion of their

workforce with access to high-speed enabled internet.

6A.2 Productivity and Broadband Adoption A5. Certain issues, related to measuring the relationship between ICT use and productivity, were raised in the main body of this Chapter. Before presenting the results from the industry level regressions, we will touch upon these issues further in the form of a discussion, in this section.

A6. A single equation model, although good for assessing correlations, fails to capture the nature of causality. Although it can tell us firms with higher intensity of electronic trade are more productive, it might be that these firms are operating at a higher intensity of ICT use precisely because they were more productive to begin with, or because they have complementary investments in organisational and managerial practices in place.

A7. For example, Figure 6.3 (from the main body of the Chapter) suggests that the lowest productivity firms lose from investing in a high speed internet enabled workforce. This may be because these firms take longer to learn the most effective way of using the new investment or do not have the necessary complementary investments in place, which themselves are correlated to better productivity.

108


A8. Other work has looked at this assertion in more detail, but this is a phenomena also exhibited in our data.

A9. The following figures plot a short time series of average firm-level productivity by year of adoption of high speed Internet. The analysis makes use of a binary (0,1) indicator of broadband adoption taken from the E-commerce survey. Each year firms are asked whether they were using high speed Internet in that particular year. We convert the yes/no response into a Boolean variable. Due to the fact that firms fall in and out of the sample every year, we restrict our analysis to those firms that were consecutively sampled in all years.

A10. For these firms we track the first year of high-speed Internet use and label this the year of adoption. We then follow the evolution of real value added per employee (y-axis) for all firms starting in 1999 for the UK and 2002 for the Netherlands. For example, the blue line labelled 2001 in Figure 6.A1 tracks average productivity levels for all firms that first started using high-speed Internet in 2001.

A11. The pattern, a lot more palpable for the UK than for Netherlands, clearly shows that firms with higher productivity adopted earlier and subsequently maintained their productivity advantage. There is weak evidence in the UK of small drops in average productivity in the year of adoption but these are recouped through productivity improvements in subsequent years after adoption.

6A.3 Causality And Correlations A12. We will not deal with the issue of causality here. We use a simple OLS regression framework to analyse and uncover how sensitive the productivity impacts of ICT use are to the inclusion of IT capital stocks and skills measures and whether the productivity impact of ICT use is industry specific. Chapter 10 provides a more detailed analysis. Using the same set of variables from this Appendix it explicitly deals with causality and endogeneity and reaches fairly similar conclusions to those presented here regarding the relationships between ICT use and productivity.

Figure 6.A1

Broadband Adoption and Labour Productivity Dynamics - UK

25

30

35

40

45

50

55

60

1999 2000 2001 2002 2003 2004 2005

Year

(£'0

00s)

2001

2002

2003

2004

2005

109


Figure 6.A2

Broadband Adoption and Labour Productivity Dynamics - Netherlands

40

45

50

55

60

65

70

75

2002 2003 2004 2005

Year

(E 0

00's

) 200320042005

A13. Accepting that we are not making causal statements, what can be said about these relationships? The first point of concern is identifying the channels through which ICT use may impact productivity. The main body of this Chapter assumes that both types of ICT use under consideration impact total-factor productivity (TFP).

A14. It is now widely assumed that ICT capital constitutes a general purpose technology in the sense that investment in ICT capital is characterised by pervasiveness, inherent potential for technical improvements, and innovational complementarities. Improvements in the quality of ICT filter into the R&D and operational capabilities of using sectors increasing productivity as a consequence. As ICT improves, its use spreads throughout the economy, bringing about generalized productivity gains16.

A15. There are two ways in which we can capture ICT investment. We can look at the total monetary amount spent by a firm, to date, on ICT infrastructure including hardware and software and neglect how this investment is used. On the other hand by focusing on specific uses we disregard the pervasiveness and capital deepening effect of the technology.

A16. This goes back to a point made originally in the Chapter: ICT hardware and software constitute types of capital and their contribution to productivity should be measured alongside other forms of capital. Failure to do this will overstate the relationship between ICT use and TFP. Chapter 10 elaborates on why this is so in the methodology section.

A17. In the main body of this Chapter we have seen the impacts of various ICT use metrics on productivity and how these relationships vary across sectors. Are these relationships robust to the inclusion of ICT capital stocks? We expect the relationships to be more modest in terms of magnitude, but is it possible their significance disappears outright?

A18. Netherlands and UK are the only two countries in our sample for which we have firm-level data on ICT capital stocks and we can use this to analyse the relationship between ICT use and productivity controlling for the capital impact.

16 The notion of General Purpose Technologies was first introduced by Bresnahan and Trajtenberg (1995).

110


A19. If the relationship between ICT use and TFP disappears once we include the effect of ICT capital, it may suggest that ICT is in fact general purpose in the sense of providing generalised productivity gains which are not use specific. This would suggest that as long as a firm is more ICT intensive i.e. it has a larger ICT capital stock per employee, it does not matter how this firm directs its extra Euro of investment: the returns from improving electronic sales systems are the same as the returns from giving an extra unit of labour access to high-speed internet.

A20. If the impact of ICT use is controlled for by including ICT capital stock in the productivity model and additional measures of ICT use are significant in the productivity regression then this suggests that ICT is a use specific – rather than a general purpose - technology.

A21. There is no reason why this should not be the case. Even though ICT is clearly pervasive nowadays firms with the same level of ICT capital stock may direct the use of their ICT capital in different ways giving rise to within firm innovational complementarities. Some firms may use it to automate internal production processes while others may concentrate supply chains, still others may direct their investment towards knowledge management. Each firm through its decisions on the hardware, software and organisational mix will be at a different point on its technology frontier and this should appear as differences in TFP.

Table 6.A1 NETHERLANDS Manufacturing Distribution

Services Differentiated

Services 2002 2005 2002 2005 2002 2005 DSL 48 96 54 92 59 94 E-sales 35 34 44 51 36 35 E-purch 43 56 49 62 62 68 PCpct 43 54 57 63 69 79 DSLpct 13 34 21 39 34 61 E-salespct 5.3 7.4 8.0 9.9 3.0 3.8E-purchpct 2.5 3.4 9.7 11.0 6.7 5.3 UK Manufacturing Distribution


Services 2001 2005 2001 2005 2001 2005 DSL 76 77 67 83 75 89 E-sales 47 60 55 63 29 45 E-purch 62 86 64 88 53 85 PCpct 50 56 52 59 51 55 DSLpct 27 34 26 36 31 39 E-salespct 21.0 29.0 15.0 16.0 5.5 11.0E-purchpct 1.2 19.0 1.5 27.0 1.7 18.0Figures reported are sample averages. A22. Table 6.A1 shows the diffusion pattern of the core ICT use variables, used in this study, for different sectors in the Netherlands and UK. Although we cannot say anything about the monetary proportion of a firm’s ICT investment directed to the various ICT uses, we see a clear tendency for different sectors to prefer different types of ICT use.

A23. The first three variables give a sense of within-industry dynamics, whereas the last four (with suffix-pct) describe within-firm dynamics. Substantially more firms in Dutch differentiated services (59%) had adopted high speed internet in 2002 compared to other sectors. An even larger proportion (62%) of firms in Dutch differentiated services were running electronic procurement systems, but were using these systems to facilitate only a

111


very small proportion of overall procurement (6.7%). In contrast less than half the firms in the Dutch distribution services sample were running e-procurement and e-sales systems, but these firms used their systems much more proactively securing almost 10% of their trade through electronic routes.

A24. An employee working in the Dutch differentiated services sector in 2002 was much more likely to have access to a computer (69%) and high-speed internet (34%) than in any other sector. And although firms within all sectors have been increasing employee access to computers and high speed internet this pattern was still visible in 2005.

A25. In the UK as early as 2002 there was already a high proportion of firms in each sector connected to high-speed internet and operating electronic sales and procurement systems. Within industry growth between 2002-05 has been steady at a rate slightly lower than in the Netherlands. Even though 62% of the 2002 UK manufacturing sample, and 64% of the UK distribution sample were running electronic procurement systems only 1.2% (1.5%) of total procurement was secured through electronic means. From this starting point there has been tremendous growth in the UK in within-firm electronic procurement and sales with electronic sales, in 2005, amounting up to 29% of total sales in manufacturing firms and 27% of all procurement in distribution services carried out through electronic routes.

A26. In contrast to the Netherlands, however, employee access to computers and high speed internet does not seem to have been too sector dependent. In 2002 approximately half of the employees in all three sectors were using computers and almost a third of employees in each sector had access to high speed internet, with the service sector leading the way. Growth in computer and high speed internet use within the workforce has been moderate in all three sectors and in 2005 it was still the case that firms within UK differentiated services were giving a larger proportion of their workforce access to high speed internet (39%).

A27. Having established that certain industries may prefer certain types of ICT use that better complement their business model we ask whether all forms of ICT use directly impact TFP? Whereas the decision to automate a production line or e-trade will impact all factors involved in the production line, thus raising TFP, enabling an employee with a more efficient spreadsheet program or faster internet will only really impact the marginal productivity of the employee in question. This is a point made by Maliranta & Rouvinen (2003).

A28. Following this insight we may consider modelling the correlation between DSLPCT and productivity as being driven by increases in the marginal productivity of those employees that are connected to high speed Internet.

A29. Econometrically it does not matter which of the two channels, TFP vs. marginal product of labour, we believe to be correct. Conditional on including a measure of IT capital the baseline specification remains the same, only the interpretation of the coefficients changes.

A30. So if we were to consider DSLPCT as a driver of marginal labour productivity, lacking a formal measure of ICT capital we should include the proportion of the workforce with computers (PCPCT) as a proxy measure of ICT capital. Otherwise, as in the core analysis, DSLPCT will take on this role and its coefficient will be overstated capturing both the impact of ICT capital and ICT use17.

A31. However, as mentioned, DSLPCT is derived from PCPCT and the high correlation between the two variables will lead to inconclusive results. Instead we opt for IT capital measures derived from investment surveys.

17 Although similar concerns apply to the coefficient of the E-commerce variable, these are not so pronounced and in the core analysis it still primarily acts as a proxy for TFP.

112


6A.4 ICT Capital A32. Firm level ICT capital stocks, for both countries, are constructed by applying the perpetual inventory method on a time series of investment survey returns. ICT capital stocks for Netherlands represent hardware stocks, whereas UK ICT capital stocks are a mixture of hardware and purchased software stocks18. Here we highlight two of the major problems faced in constructing ICT capital stocks.

6A.4.1 Under reporting of IT investment A33. Hardware investment series are based on capitalisation of items such as mainframes and computers and are less likely to be subject to measurement error. Many of these hardware items come bundled with built-in software packages and programs. Work in the UK has shown that the lion’s share of software used by firms is produced in-house19. Because this strand of software development is a slow and evolving process, it is easily expensed by firms and is very hard to quantify and capitalise. For this reason firms are only really able to return purchases of pre-packaged software in surveys. As most of this category of software is already captured in the hardware series, returns for purchased software generally tend to be small.

6A.4.2 Hedonic IT price deflators A34. Another prominent concern that applies equally to both types of IT investment is the use of appropriate price deflators. Unlike other forms of capital investment the evolution of hardware and software prices does not reflect the dramatic improvements witnessed in the quality of IT goods. Holdway (1999) reports that while observed prices for desktop computers were fairly stable between 1993-98, CPU processing speeds increased by 1263%, hard drive capacity increased by 3700% and system memory increased by 1500%. These improvements are not just restricted to hardware; software programs are also much more evolved and more efficient in terms of running time.

A35. Price deflators that attempt to capture the real value of IT investment in constant prices should capture not only current price trends but also the improvements in quality. As Wallis (2007) points out for the purposes of deflation it is changes in ‘pure price’ that we are interested in. Changes in ‘observed price’ are a composite of changes in quality and ‘pure price’. When quality changes are slow or non-existent changes is observed price mimic changes in pure price. But when quality change is consistently dramatic, observed prices need to be adjusted for quality changes.

A36. There are many ways to construct deflators that quality adjust. The two most prominent methodologies include matched models and hedonic regressions, but these two estimation procedures are also the most resource intensive requiring the finest quality data on both price and technology trends20.

A37. Until recently the UK statistical office quality adjusted price deflators using an ‘option cost’ approach. Efforts initiated by the Review of Short Term Output Indicators (STOIR) have led to a shift in policy towards the implementation of hedonic deflators.

A38. Estimating quality-adjusted deflators is a demanding task, especially for software. Software investment can take three different forms. Pre-packaged software, for which a quality-adjusted deflator can be constructed using hedonics. Own-account software, which generally needs to be priced through the wage costs of labour hired for the production of in-house software. And custom software: purchased software that is further customised in-

18 Further details on construction methodology and descriptive statistics for IT capital stocks can be found in Chapter 10. 19 See Chamberlin et al (2007) for more details on work undertaken at the Office for National Statistics, UK on improving estimates of Software investment in the National Accounts. 20 See Allen & Ball (2003).

113


house and therefore needs to be priced through a weighted combination of own-account and pre-packaged software deflators.

A39. Constructing quality-adjusted deflators for hardware encounters other problems. Choosing the definitive characteristics of hardware quality that affect prices can lead to a trade-off between parsimony and missing characteristics. The relationship between computer components and their prices may differ across components, and may not necessarily be linear, a simplifying assumption used in hedonic regressions. Finally hedonic pricing is essentially a forecasting exercise based on past information. What happens if there is a structural break - for example a market shock - that leads to hardware and software being priced in a different way? Hedonic regressions will not necessarily be able to pick up these breaks in real time and will have to re-adjusted and re-estimated once it has become clear that there has been a change in market conditions.

A40. A full discussion of quality adjustment methods, their pros and cons and price index specific (e.g. PPI vs. CPI) issues that may arise in the implementation of hedonic regressions is beyond the scope of this report. The UK experience is covered in greater detail by Brand (2001), Vaze (2001), Allen & Ball (2003), Ball et al (2004).

A41. Despite the big strides made by the UK statistical office in developing UK hedonic deflators for both hardware and software capital, for purposes of consistency with the rest of the report, we use hedonic price deflators sourced from NIESR-EUKLEMS, which are more closely linked to the US hedonic deflator series.

6A.5 E-Commerce A42. It is reasonable to assume that e-trade leads to improvements in TFP i.e. it impacts all aspects of the business equally. E-procurement has been a relatively stable phenomena for more than a decade, especially so in manufacturing, with medium to large businesses running automated procurement systems well before the advent of fast internet connections. On the other hand e-sales, especially over the internet, is a relatively new phenomenon and as identified in Chapter 12 constitute a specific form of innovative practices.

A43. We argue in the main body of the Chapter that TFP may include organisational learning, flexibility and adoption of innovative practices. Here we aim to capture this by splitting our e-trade variable into two separate IT use variables that may also proxy for other factors: An e-procurement (efficiency and effectiveness) variable and e-sales (flexibility and innovation) variable.

A44. Data from Table 6.A1 also suggest within-industry dynamics summarised in Boolean variables (without suffix-pct) that capture ICT use through simple yes/no firm-level responses are reaching saturation. Whereas within-firm intensity variables are far from saturation and capture a firm’s positioning on the technology frontier with finer gradations.

6A.6 Estimation Strategy A45. Taking into account the above considerations, but not making any inferences on causality we run the following single equation specification, in per employee terms:

),()ln()1()ln()ln()ln( 2121 itititit

I

it

R

it HCECFLLK

LK

LY

+−−++= αααα

As mentioned above ( IK ) contains hardware capital for Netherlands, and hardware and purchased software for the UK. A46. F(EC ) is a function of our core ICT use variables: % of employees with high speed enabled internet, % of e-purchases and % of e-sales. We also include a measure of a firm’s

114


ICT maturity based on whether a firm has a website, provides customer support for its e-sales and runs a different number of automated processes. With this we hope to capture any additional returns to TFP not accounted for by the core EC variables.

A47. F(HC ) is our control for skills which is defined as in the main body of this Chapter: log of wage costs per employee.

6.6 Sector level regressions A48. The cross-country comparison controls for industry level fixed effects. However, by pooling all industries for each country into a single regression sample, we are estimating the average effect of IT use across all industries, by country. This estimation strategy gives a good idea of an overall effect but obscures any industry specific effects. Therefore the regression sample is divided into individual regressions for the manufacturing and electrical equipment and communication sectors. These highlight that differences in coefficient estimates can arise across sectors.

A49. We extend this analysis and split our overall country sample into three broad industry categories. By running our specification separately for each broad industry we aim to distinguish industry specific impacts. Within each regression we still control for narrow industry level fixed effects in a similar way to the main analysis.

A50. Our industrial classification departs somewhat from the EU-KLEMS categorisation used throughout this report. In order to differentiate business types defined over production, supply chains and knowledge creation and management we define three broad sectors:

• Manufacturing: All Manufacturing Sectors (SIC 15-37) • Distribution Services: Wholesale, Retail and Transport (SIC 51,52,60-64) • Differentiated Services: Services (SIC 70-85)

6A.7 Results A51. The results for Netherlands and UK are presented in Tables 6.A2 & 6.A3. The first column of each table presents results for a pooled regression using a measure of total capital stock i.e. we do not distinguish between types of capital, including all sectors and controlling for wages. The second column introduces a measure of IT capital stock to the pooled regression while still controlling for wages. The third column runs the same specification as column 2 but removes wages as a control.

A52. Three major insights obtain from these three columns. For both the Netherlands and UK, comparing columns 1 and 2, we see that the coefficients on the e-trade variables remain largely unaffected by the inclusion of IT capital. However, the coefficient on DSLPCT witnesses a drop21 suggesting that alongside a productivity impact DSLPCT is also capturing an IT capital effect.

A53. Secondly, comparing columns 2 and 3, the coefficient on DSLPCT increases with the exclusion of wages. In fact for the UK it becomes significant again. This suggests that DSLPCT may also be a proxy for skills.

A54. If this is so, then its coefficient actually reflects the impact on the marginal productivity of labour, as suggested above, and not TFP. As we will see, in Appendix C to Chapter 12, there is some evidence at least for the UK that DSLPCT may be a proxy for knowledge management.

A55. Thirdly, we notice that by not including skills (wages) in our specification we crowd out the impact of the e-trade variables. The coefficients on these variables are roughly

21 For the Netherlands we see a slight increase in the coefficient of DSLPCT, but notice the weak explanatory impact of the IT capital stock. It is likely that the NLD IT capital stock suffers from higher mis-measurement relative to DSLPCT, than the UK IT capital stock. We will come back to this issue in Chapter 10.

115


comparable between columns 2 and 3, but now the coefficient on DSLPCT ends up explaining all the significance. It is therefore necessary to retain a measure of skills as an extra control22. Because of these reasons the regression specification of column 2 remains our preferred option.

A56. The final three columns of Tables A2 and A3 run this regression specification by broad industry breakdowns. The results show evidence of sector specific benefits from ICT use. Whereas manufacturing firms, in both Netherlands and UK, primarily derive productivity gains from e-procurement, the primary source of benefit for distribution services is e-sales. Differentiated services, both in the Netherlands and UK benefit principally from connecting more of their employees to high speed internet suggesting the major gains in differentiated services is from improved knowledge management.

- A 10% increase in e-procurement by a UK manufacturing firm increases its productivity, on average, by 2.6%. A similar increase by a Dutch firm increases its productivity, on average, by 1.2%23.

- A 10% increase in e-sales by a UK distribution firm increases its productivity, on average, by 3.1%. A similar increase by a Dutch firm increases its productivity, on average, by 2.7%24.

- A 10% increase in its high-speed internet enabled workforce by a UK differentiated services firm increases its productivity, on average, by 0.9%. A similar increase by a Dutch firm increases its productivity, on average, by 1.3%.

6A.8 Conclusion A57. For the Netherlands we find an insignificant coefficient on IT capital in all three sectors25. We also note that the coefficient on DSLPCT is only significant for differentiated services. E-sales intensity is significant in distribution and differentiated services and E-purchases intensity remains insignificant in all three sectors, although it is only marginally insignificant in manufacturing.

A58. For the UK we note the insignificant coefficient on IT capital in manufacturing and distribution services. Although the coefficient on DSLPCT in differentiated services is not very large, this is the only sector that enjoys positive returns from investment in DSLPCT. The coefficient on DSLPCT is, surprisingly, negative and significant in manufacturing and insignificant in distribution services. E-sales intensity is only significant for distribution services and E-purchases intensity is significant in all three sectors, but has a negative impact in distribution services.

A59. The individual coefficient on DSLPCT tells us what happens to productivity if we increase DSLPCT but hold IT capital and all other uses of IT at a constant level. What would happen to productivity if we increase IT capital and all IT uses simultaneously by 10%? Table A5 summarises all individual effects and the total effect obtained by adding together all statistically significant individual effects.

A60. Consider differentiated services in the UK. Increasing IT capital by 10% improves productivity by 0.8%. Increasing DSLPCT by 10% improves productivity by 0.9% and increasing e-purchases by 10% improves productivity by 1.2% leading to a total improvement of 2.9%. The results show how the total effect differs across sectors and countries but is positive and in the range of 0.8-2.9%.

22 Wages are not the best measure of skills as they are highly correlated with current productivity. This in itself may lead to biased results. We will correct for this thoroughly in Chapter 10. 23 The NLD e-procurement effect is not significant at conventional levels. 24 For further analysis of the e-commerce variables and automated business processes in Distribution Services please refer to Chapter 9. 25 This may be an artefact of the limited sample size and/or an indication of the weak explanatory power of the Dutch IT capital stocks. We investigate this issue further in Chapter 10.

116


117

Table 6.A4 NETHERLANDS IT Capital DSLPCT % e-sales % e-purch TFP Manufacturing ns ns ns 1.2(ns) 1.2(ns)Dist Services ns ns 2.7 ns 2.7 Diff Services ns 1.3 0.5 ns 1.8 UK IT Capital DSLPCT % e-sales % e-purch TFP Manufacturing ns -1.6 ns 2.6 1.0 Dist Services ns ns 3.1 -2.3 0.8 Diff Services 0.8 0.9 ns 1.2 2.9 A61. This analysis shows that not only is ICT use industry specific but also particular uses of ICT may have opposing impacts in different industries, which may lead to the average effect cancelling out in pooled regressions. We could therefore erroneously conclude that a particular ICT use has no impact on productivity, even though it has a strong impact in a particular industry, but weakly impacts all other industries.


Table 6.A2 – Regression output for Netherlands (1) (2) (3) (4) (5) (6)

Sample Whole

EconomyWhole

EconomyWhole

Economy Production Distribution


Services

Total Capital 0.147 (0.007)*** Non IT Capital 0.082 0.122 0.066 0.06 0.125 (0.006)*** (0.007)*** (0.008)*** (0.012)*** (0.012)***Labour 0.012 0.023 0.031 0.046 0.017 0.0001 (0.005)** (0.006)*** (0.007)*** (0.008)*** (0.011) (0.011)IT Capital 0.0001 0.046 -0.003 -0.012 0.005 (0.005) (0.006)*** (0.008) (0.01) (0.014)% of Employees using Broadband Computers 0.063 0.082 0.253 0.042 0.03 0.131 (0.026)** (0.027)*** (0.031)*** (0.043) (0.047) (0.047)***% e-sales (intensity) 0.148 0.165 0.153 0.045 0.27 0.054 (0.042)*** (0.043)*** (0.050)*** (0.057) (0.065)*** (0.163)***% e-procurement (intensity) -0.001 0.007 0.323 0.126 -0.016 -0.078 (0.043) (0.044) (0.052) (0.09) (0.051) (0.098)ICT Maturity (+)* (+)* (+)** Yes (+)** Yeslog (wages) Yes Yes No Yes Yes YesObservations 3852 3852 3852 2426 730 696R-squared 0.52 0.5 0.32 0.41 0.55 0.69SIC2 and Time - Fixed Effects Yes Yes Yes Yes Yes Yes

118


119

Table 6.A3 – Regression Output for the UK (1) (2) (3) (4) (5) (6)

Sample Whole

Economy Whole

Economy Whole

Economy ProductionDistribution


Services

Total Capital 0.132 (0.008)*** Non IT Capital 0.112 0.235 0.091 0.226 0.068 (0.008)*** (0.008)*** (0.021)*** (0.023)*** (0.009)***Labour 0.010 0.002 -0.048 -0.017 0.002 -0.014 (0.007) (0.007) (0.008)*** (0.018) (0.021) (0.010)IT Capital 0.051 0.150 0.022 0.032 0.078 (0.007)*** (0.008)*** (0.016) (0.021) (0.009)***% of Employees using Broadband Computers 0.055 0.038 0.189 -0.164 0.056 0.093 (0.027)** (0.027) (0.031)*** (0.068)** (0.072) (0.034)***% e-sales (intensity) 0.138 0.128 0.096 -0.054 0.311 -0.035 (0.035)*** (0.034)*** (0.040)** (0.062) (0.079)*** (0.061)% e-procurement (intensity) 0.037 0.030 0.006 0.269 -0.230 0.123 (0.045) (0.087) (0.052) (0.112)** (0.128)* (0.067)*ICT Maturity (+)*** (+)*** (+)*** Yes (+)** (+)*log (wages) Yes Yes No Yes Yes YesObservations 6872 6872 6872 1596 1250 2669R-squared 0.65 0.66 0.54 0.62 0.43 0.77SIC3 and Time - Fixed Effects Yes Yes Yes Yes Yes Yes


Chapter 7

ICT characteristics of fast growing firms

Yoann Barbesol, INSEE

Simon Quantin, INSEE

7.1 SCOPE OF THE CHAPTER The simultaneity of increasing use and diffusion of ICT among firms in developed countries and the upsurge of their productivity has fuelled beliefs that, contrary to the so called Solow paradox, ICT use had a strong impact on productivity. Many studies have been carried out successfully to identify that link (e.g. Lehr & Lichtenberg (1999), Oliner & Sichel (2000)).

In the last decade, a growing literature has appeared on the role of Fast Growing Firms (“Gazelles”), in promoting economic growth. The ICT characteristics of these firms is the main focus of this chapter. Moreover, as suggested by Brynjolfsson et al (2006), ICT enables firms to replicate business process innovation throughout an organization, thereby also increasing market share and market value. ICT investment could thus foster firm-level growth.

The link between fast growing firms and ICT demands careful attention. More specifically, we will examine whether the fast growing firms that have appeared since 2000 are ICT intensive users. Any empirical study is however strongly limited by the lack of appropriate data. In the next section, we will present one definition for fast growing firms. Section 7.3 gives some descriptive results in the French case, while the last section presents a cross country analysis.

7.1.1 Definition used for fast growing firms The definition of Fast Growing Firms is not unambiguous. In fact, two dimensions are to be considered.

- Fast Growing Firms as High Growth Employment firms (hereafter HGE). The criterion according to which firm size should be measured is not unique. Most studies focus on employment or value added. We prefer to use the former criterion as employment is a prior objective for public authorities, whereas increasing the profitability of a firm much more concerns the entrepreneur. One drawback of this criterion based on employment growth is that the whole labour force used by a firm is not observable in our data. In particular, it could ignore high-growth firms resorting to outsourcing, or firms using intensively temporary work.

- Second, how should growth be measured? Employment growth is highly sensitive to initial size of firms: hiring ten more employees is not the same for a ten employee firms (thus doubling his staff) and a 100 000 employees firms. Because of the small size of our sample, it is however unrealistic to define fast growing firms in each class size. We thus rely on an absolute criterion: a HGE firm is a firm whose growth in four years exceeds 20%, but we choose to exclude smaller firms (less than 30 employees in 1994)1.

1 See Schreyer (2000) for a discussion of the impact of relying on an absolute or relative criterion on the final selection.

120


Some points to note:

- The analysis is made on a firm level basis. This may not be the most appropriate level of analysis. For instance if a firm splits into two new smaller companies, the whole group should be taken into account rather than the two resulting structures. More and more SMEs resort to this internal reorganization process2, especially the HGE, as they belong more often to a group. A more accurate measure of the employment growth of a firm is given if one could distinguish between external growth, which mostly stems from restructuring processes (mergers and acquisitions in particular), and internal growth, which is more the growth we are interested in. Mergers and acquisitions also induce fast growth. According to Picart (2006), “between 1998 and 2001, external growth accounts for one half of the growth of the gazelles” in France. Schreyer (2000) finds that only one third of the Swedish gazelles’ growth can be regarded as internal growth. Dealing with this group dimension requires data on the whole group. We can pick out firms affiliated to a group at one date, but longitudinal data on the whole groups are not available: it requires surveys dedicated to restructuring3.

- We thus may select firms whose growth just stems from a merger or an acquisition, and set aside real HGE for the same reason (as they were merged with another entity to form a new structure, or were purchased by a big company).

- In order to analyse ICT impacts we select firms present from 2001 to 2004 ICT surveys. This condition drastically reduces the panel on which HGE are selected. This could cause serious attrition bias, as it leads to the overrepresentation of the biggest firms, more likely to be surveyed. In the last section, we will illustrate this point with the French case.

7.2 ANALYSIS ON AGGREGATE DATASETS If ICT does indeed foster high growth firms, we would expect to observe a positive correlation between the share of ICT equipment and HGE firms. The first basic test is to check that in cross country comparisons, HGE firms are more frequent in countries where firms are more ICT equipped. Besides, we expect that HGE firms are more ICT-equipped. We address these two issues using the aggregated dataset created from each NSI (hereafter Metadata table). The e-commerce coverage varies markedly across countries (e.g. 2000/2005 for Austria; 2003/2004 for Denmark). This prevents any use of the temporal dimension of the data.

As ICT probably impacts services and manufacturing sectors differently, we distinguish these two sectors in the descriptive analysis. We use four different measures of ICT diffusion. Two refer to ICT usage: the share of firms with broadband internet; and the share of firms selling over the internet. The latter two correspond to the intensity of ICT use: the share of firms purchasing through the internet and finally how ICT-intensive is the firm.

7.2.1 HGE firms and IT use Manufacturing firms

Graphic 7.1 presents the percentage of HGE firms (right scale) and the percentage of firms with broadband internet access for HGE firms and all firms (left scale), in manufacturing sectors. Countries are classified according to the decreasing share of HGE firms in the manufacturing sectors. This proportion is the highest in the United Kingdom (6.2%). By contrast, only 0.8% of firms in the manufacturing sector are HGE in the Netherlands.

Firms connected to internet in the manufacturing sector usually have broadband access (the coverage is higher than 50% in all countries). HGE firms more often have broadband access than others firms (except in Norway), but in many countries the gap is tiny. It 2 See Picart (2006) for an analysis for France. 3 In France, a yearly administrative dataset retraces the links between firms and groups and provides key data on the whole group.

121


exceeds 10 points in Ireland, the Netherlands and Italy (and almost in Finland also). Thus no clear pattern appears between the share of HGE firms and broadband access.

Graphic 7.1: Fast Growing Firms and Broadband Internet users in Manufacturing Sectors (excluding Electrical)

6.2

6.1

3.9

3.0

1.9

1.7

1.6

1.6

0.8

United Kingdom

Czech Republic

Ireland

Norway

Italy

France

Finland

Gemany

Netherlands

% of firms with broadband internet % of HGE

HGE

All firms

100 80 60 40 20

Source: Metadata table Reading note: HGE firms represent 1.7% of the firms in France. 84% of French firms use broadband Internet and 86% of French HGE firms use broadband internet. Graphics 7.2 and 7.3 present estimates for e-sales and e-purchases. In the Netherlands, and to a smaller extent in France and Finland, HGE firms clearly appear as more intensive users of internet for business purposes. The positive link is observed in all countries except Norway.

122


Graphic 7.2: Fast Growing Firms and e-sales in Manufacturing Sectors (excluding Electrical)

6.2

6.1

3.9

3.0

1.9

1.7

1.6

1.6

0.8

United Kingdom

Czech Republic

Ireland

Norway

Italy

France

Finland

Gemany

Netherlands

% of sellers through internet % of HGE

HGE

All firms

100 80 60 40 20

Source: Metadata table Reading note: HGE firms represent 1.7% of the firms in France. 32% of French firms use broadband Internet and 42% of French HGE firms use broadband internet. Graphic 7.3: Fast Growing Firms and E-Purchasers in Manufacturing Sectors (excluding Electrical)

6.2

6.1

3.9

3.0

1.9

1.7

1.6

1.6

0.8

United Kingdom

Czech Republic

Ireland

Norway

Italy

France

Finland

Gemany

Netherlands

% of purchasers through internet % of HGE

HGE

All firms

100 80 60 40 20

Source: Metadata table Reading note: HGE firms represent 1.7% of the firms in France. 32% of French firms are e-purchasers and 42% of French HGE firms are e-purchasers.

123


Market Services

Graphics 7.4, 7.5 and 7.6 show the percentage of HGE firms (right side) and, on the left side, respectively the percentage of broadband internet users, e-purchasers and e-sellers for both HGE firms and all firms, in market services.

Similarly to the manufacturing sector, the UK and the Czech Republic have among the highest percentage of HGE firms in the service industry. The UK is way ahead; the rate of HGE firms in the UK is more than twice as large as in Czech Republic (respectively 13.6 and 5.6%). Italy has the second largest HGE rate in market services.

As in the manufacturing sector, broadband access is widespread in all countries (note that the coverage is even higher, more than 80% in all countries, except Ireland). In all countries, HGE firms are more likely to have broadband access than others firms.

Graphic 7.4: Fast Growing Firms and Broadband Internet users in Market Services (excluding Post and Telecommunications)

13.6

6.6

5.6

5.4

4.8

4.0

3.6

3.3

2.8

United Kingdom

Italy

Czech Republic

Netherlands

Ireland

Norway

Gemany

France

Finland

% of firms with broadband internet % of HGE

HGE

All firms

100 80 60 40 20

Source: Metadata table In all countries firms are more often e-purchasers than e-sellers. E-sellers never exceed 40% of the firms except in Ireland (roughly 40%) whereas the percentage of e-purchasers is greater than 40% (except for Italy, 30%), especially in Ireland (cf. chapter 4). In many countries, HGE firms are slightly more e-sellers than firms on average, except for UK. This positive link doesn’t hold true for e-purchasing activities. In fact no clear pattern stands out: in Finland, Germany and Czech Republic, HGE firms purchase more on the internet than do non HGE firms, whereas in the other countries, the gap is tiny.

124


Graphic 7.5: Fast Growing Firms and E-Purchasers in Market Services (excluding Post and Telecommunications)

13.6

6.6

5.6

5.4

4.8

4.0

3.6

3.3

2.8

United Kingdom

Italy

Czech Republic

Netherlands

Ireland

Norway

Gemany

France

Finland

% of purchasers through internet % of HGE

HGE

All firms

100 80 60 40 20

Source: Metadata table Graphic 7.6: Fast Growing Firms and e-sellers in Market Services (excluding Post and Telecommunications)

13.6

6.6

5.6

5.4

4.8

4.0

3.6

3.3

2.8

United Kingdom

Italy

Czech Republic

Netherlands

Ireland

Norway

Gemany

France

Finland

% of salers through internet % of HGE

HGE

All firms

100 80 60 40 20

Source: Metadata table

125


7.2.2 HGE firms are more intensive IT-users Graphics below present for each country, for both the market services sector and manufacturing:

- The percentage of HGE (on the right) - The intensity of IT use (sum of % of e-purchases and % of e-sales) (on the left).

Graphic 7.7: Fast Growing Firms and IT intensive use in Manufacturing Sectors (excluding Electrical)

6,2

6,1

3,9

3,0

1,9

1,7

1,6

1,6

0,8

United Kingdom

Czech Republic

Ireland

Norw ay

Italy

France

Finland

Germany

Netherlands

% of IT intensive use % of HGE

HGE

All firm s

60 40 20

Source: Metadata table Graphic 7.8: Fast Growing Firms and IT intensive use in Market Services (excluding Post and Telecommunications)

13,6

6,6

5,6

5,4

4,8

4,0

3,6

3,3

2,8

United Kingdom

Italy

Czech Republic

Netherlands

Ireland

Norw ay

Germany

France

Finland

% of IT intensive use % of HGE

HGE

All firm s

100 40 206080

Source: Metadata table Note: we do not have information for Irish HGE firms

126


In the manufacturing sector, HGE firms are clearly more IT intensive users, except for Norway (the difference is tiny though). However, there is no positive link between the IT use intensity and the percentage of HGE firms. For instance, on the one hand Norwegian, French and German firms are the most IT intensive users, whereas on the other hand, the UK and the Czech Republic present the highest percentages of HGE firms in the manufacturing sector.

In the market services sector, HGE firms are more IT intensive users in Norway, the Netherlands and the Czech Republic (the gap exceeds 20 points for Norway). But in the other countries, the gap is tiny, either slightly positive (Germany) or negative (Finland, France, UK, Italy). Note that there is no apparent link between the rate of HGE firms in market services and IT intensive usage emerging from this graphic. Firms in the UK (which is way ahead in terms of HGE firm rate), are definitely not IT intensive users; their IT intensity use rate is even lower than that of firms from Finland, which country ranks last in terms of HGE firms. Norway appears as a very singular observation; Norwegian firms are very IT intensive users (65%; the second highest rate is lower than 20%) though the rate of fast growing firms is rather average compared to that of the other countries.

Conclusions on IT-use:

To conclude, the cross-country analysis show that IT use does not unambiguously influence the percentage of HGEs. The percentage of HGE firms appears not correlated to broadband penetration in the economy nor to the proportion of e-purchasing or e-selling. This does not mean that there is no link whatsoever, as the data are missing for identifying such a link. There are indeed many parameters we obviously do not and cannot control for (no country fixed effect in the absence of any temporal dimension, different business cycles between countries etc); they may explain such variations across countries at least partially.

However the comparison between HGE firms and the other firms clearly show that, in each country, HGE firms are both more likely to use IT and to be more intensive users of IT (which are two complementary aspects of IT usage).

- HGE firms are more IT users than the other companies in the manufacturing sector. This is less clear concerning in Market services.

- Concerning IT use intensity, HGE are clearly more IT intensive users than the other firms in the manufacturing sector. Results however strongly differ between countries.

In market services, no such a conclusion can be held as the IT intensity rate is generally very low (less than 20%) for both HGE firms and the rest of the firms: the Norwegian case stands out though.

The next section focuses on the French microdata, as they contain some more detailed information and helps to investigate this issue further.

7.3 RESULTS FOR FRENCH FIRMS

7.3.1 Attrition bias The sample is first restricted to firms present in the 2001 ICT survey, and still active in 2004. It represents 7,521 firms. Only 1,623 firms of this sub-sample were again surveyed in 2004. The condition of presence both in 2001 and in 2004 ICT surveys thus leads to a drastic attrition of the panel (78%). Not surprisingly, the number of Fast Growing Firms is quite low, as we observe only 53 in the final sub sample. As stated before, the attrition causes a selection bias in favour of the largest firms (Graphic 7.9): the mean size by the FGF is 412, which is twice the mean size of the first sub-sample.

127


Graphic 7.9: Attrition bias In E-commerce (EC) survey in

2001

In EC-survey 2001 still active in 2004

Fast Growing Firms (I) In EC-surveys 2001 & 2004 Fast Growing

Firms (II)

Number of firms 8 322 7 521 139 1 623 53

Mean of employment in 2001 258 264 212 845 412

Mean of value-added per employee in 2001 62 62 75 62 94

Source: INSEE EC-surveys 2001 & 2002 Reading note: E-commerce survey contains 8,322 firms in 2001. Only 1,623 firms are also in E-commerce survey in 2004.

7.3.2 IT use and French HGE firms On that selection basis, we do not find, as far as the French case is concerned, significant correlations between IT use of a firm in 2001 or 2002 and the fact it is a HGE in 2004.

Correlations between HGE (in 2004) and 2001 IT variables are clearly not significant:

Graphic 7.10: Correlations between HGE dummy and 2001-IT variables

S a le o n in t e r n e t P u r c h a s e o n in t e r n e tP e a r s o n 's c o r r e la t io n c o e f f ic ie n t s 0 ,0 1 1 0 ,0 1 8

P r o b > |r | u n d e r H 0 : R h o = 0 0 ,3 4 0 ,1 3

N u m b e r o f o b s e r v a t io n s 7 2 8 8 7 2 8 8

Source: INSEE EC-surveys 2001 & 2002 Correlations between HGE (in 2004) and 2002 IT variables are again clearly not significant:

Graphic 7.11: Correlations between HGE dummy and 2002-IT variables

web dsl reseau% of employees with

internet% of employees that use micro

Pearson's correlation coefficients 0,019 0,022 0,029 0,028 0,026

Prob > |r| under H0: Rho=00,11 0,05 0,02 0,01 0,03

Number of observations6 674 7 365 6 833 7 365 7 365

Source: INSEE EC-surveys 2001 & 2002

7.3.3 IT use intensity impacts the probability of being a HGE firm However, the correlation becomes significantly positive if:

- – IT use intensity for business purposes is introduced instead (% of sales through internet plus % of purchases through internet)

- – IT usage by employees (% of employees connected to the net)

128


In order to control for the firms’ characteristics, we implement a probit model, the specification of which is:

ii

i

i

iii ort

LVA

LK

ITuseEcomHGE exp)1( ++++==Ρ

where iEcom is the E-commerce intensity of the firm

iITuse is the % of employees using internet

iortexp is a dummy for exporter. Graphic 7.12: Probit estimations

P a r a m e t e r E s t i m a t i o n

I n t e r c e p t - 1 , 6 4 8 * * *

E - c o m m e r c e i n t e n s i t y i n 2 0 0 1 0 , 2 6 4 * *

D S l ( d u m m y ) 0 , 0 1 6

% o f e m p l o y e e s u s i n g i n t e r n e t 0 , 1 2 4

% o f e m p l o y e e s u s i n g P C 0 , 0 2 4

E v o l u t i o n o f c a p i t a l i n t e n s i t y - 0 , 2 8 2 * * *

E v o l u t i o n o f v a l u e -a d d e d p e r e m p l o y e e - 0 , 2 2 0 * * *

E x p o r t e r ( d u m m y ) - 0 , 1 2 3

I n d u s t r i e s - 0 , 3 3 7 * * *

S e r v i c e s - 0 , 4 3 2 * * *

T r a d e - 5 , 5 5 1

O t h e r s R e f .

Source: INSEE EC-surveys 2001 & 2002 Note: size class dummies have been added but not shown here. In France, IT use intensity for business purposes (which is captured by the e-commerce intensity variable) increases the probability for a firm to be a Fast Growing Firm, i.e. to grow fast. But IT use by employees does not (whatever variable is considered; DSLPCT, PCPCT or INTERPCT).

In the probit estimation, we added some useful controls;

- industry dummies. - exporter dummy; firms engaged in international trade may grow more rapidly

than firms that are not. - evolution of value-added per employee variable, which is supposed to capture

labour productivity effects. - evolution of capital intensity.

This result is crucial as it extends and refines the conclusions we drew from the cross country comparison. Indeed, we found a positive link between IT usage (DSL Boolean

129


variable) and IT intensive usage and HGE firms. The first link does not hold true should we believe the results obtained from the previous causality analysis (probit estimation).

The second relationship seems to be confirmed from the probit analysis. However, concluding remarks (see below) should also be taken into account.

7.3.4 Concluding remarks Using French data, we do not find evidence for an impact of IT use per se on firm-level growth, whereas IT intensity does seem to correlate significantly with the incidence of rapidly growing firms. This is in line with results of our cross country analysis.

As underlined in the first part of the chapter, it is crucial to distinguish external growth from internal growth4 (see Schreyer (2000): only one third of the Swedish HGE firm’s growth can be regarded as internal growth and Picart (2006) on French data: external growth accounts for 50% of the growth of the HE). In order to refine our results, a next step could be to add the “group level of analysis” in the ICT study.

4 External growth stems from restructuring processes (e.g. mergers and acquisitions) contrary to internal growth of employment which is not due to outside help (funding, intra-group restructuring etc.)

130


APPENDIX A

Results for manufacturing industry (Graphics 7.A.1, 7.A.2)

Patterns emerging for the data show that, for all countries, HGE firms in manufacturing industry appear to be more e-sales intensive. In most countries, they are also more e-purchasing intensive.

That is, IT use intensity for business purpose may certainly increase the probability to be HGE.

Results for services industry (Graphics 7.A.3, 7.A.4)

The trend appears to be that the service industry doesn’t present the same results as the manufacturing industry. It demonstrates well the variability between countries and also within countries. In some countries, fast growing firms appear to be more IT intensive e-sellers but less intensive in e-purchasers.

131


Graphic 7.A.1: Fast Growing Firms and Intensity of E-Purchases in Manufacturing (excluding Electrical)

6.2

6.1

3.9

3.0

1.9

1.7

1.6

1.6

0.8

United Kingdom

Czech Republic

Ireland

Norway

Italy

France

Finland

Gemany

Netherlands

% of purchases order through internet % of HGE

HGE

All firms

20 15 10 5

Graphic 7.A.2: Fast Growing Firms & Intensity in E-Sales in Manufacturing (excluding Electrical)

6.2

6.1

3.9

3.0

1.9

1.7

1.6

1.6

0.8

United Kingdom

Czech Republic

Ireland

Norway

Italy

France

Finland

Gemany

Netherlands

% of sales order through internet % of HGE

HGE

All firms

40 30 20 10

132


Graphic 7.A.3: Fast Growing Firms & Intensity in E-Purchases in Market Services (excluding post and telecommunication services)

13.6

6.6

5.6

5.4

4.8

4.0

3.6

3.3

2.8

United Kingdom

Italy

Czech Republic

Netherlands

Ireland

Norway

Gemany

France

Finland

% of purchases order through internet % of HGE

HGE

All firms

60 40 20

Graphic 7.A.4: Fast Growing Firms & Intensity in E-Sales in Market Services (excluding post and telecommunication services)

13.6

6.6

5.6

5.4

4.8

4.0

3.6

3.3

2.8

United Kingdom

Italy

Czech Republic

Netherlands

Ireland

Norway

Gemany

France

Finland

% of sales order through internet % of HGE

HGE

All firms

60 40 20

133


Chapter 8

Employment, skills and information technology Eva Hagsten, Statistics Sweden Patricia Kotnik, University of Ljubljana1

8.1 General Introduction The labour force and its skills are not only closely intertwined but also important factors of firm production. While the number of employees could indicate both the type of firm as well as its relative success so can the skills of the employees give a hint of the firm productivity level. Many studies suggest a relationship between skill levels and productivity while some evidence also has been found for the importance of type of skills as well as the possible complementarity between skills and other factors such as investments in and the use of information technology (IT). Less empirical evidence has been found for a relationship between information technology and employment at disaggregated levels. Availability of data of course plays a part in this.

This chapter examines both the possible relationship between employment and IT as well as its counterpart between productivity and IT intensive skills. The analysis has been performed on both firm level and cross country industry level data. The main IT variables are derived from the firm ICT usage survey while the skills data originate from the education registers. However, the latter linked information is scarce over Europe, allowing this analysis to be performed only on data from Finland, Norway and Sweden.

8.2 Employment

8.2.1 Introduction The debate on technology and employment touches upon the fears of the impact of technology, captured by the problem of “jobless growth” or even the neo-Luddite expectations of technology massively displacing jobs. Supporting empirical evidence for the latter - what is sometimes called the “end-of-work” literature – is mostly based on cases focusing on direct labour-saving effects of technology only and using the static perspective which does not take into account the indirect effects of innovation, resulting in a strong bias of this literature (Spiezia, Vivarelli, 2002). Opposed to this is the approach that takes a macroeconomic perspective that can integrate all the indirect effects through which technological change can affect employment, in the tradition of “theory of compensation”. This view takes into account different market compensation mechanisms that can counterbalance the initial labour-saving impact of process innovation: via decrease in prices, new machines, new investments, decline in wages, increase in incomes and via new products (Pianta, 2005).

Information technology and its use represent an important part of technological change. In manufacturing, we can expect that the use of IT will show mainly in labour-saving process innovations. The expectations are different for the service sector since IT has reshaped the ways in which most traditional services are produced and delivered and also generated new 1 Section 8.2 on employment has principally been prepared by Patricia Kotnik. Section 8.3 on skills has principally been prepared by Eva Hagsten,

134


services. Product innovations could thus be expected to prevail and have a positive effect on employment. The available IT use data do not capture the uses of IT for product and process innovations separately. Because of this and the market compensation mechanisms there are no clear-cut answers about the employment effects of IT, as will also be shown in the literature review. Therefore, studies at aggregate, sectoral and microeconomic level, analysing manufacturing and services, are needed in order to build up more empirical evidence. Only macro analyses would give a proper estimate since they would take into account the direct effect of technological change on labour and the various compensation mechanisms that follow. However, their feasibility is limited due to the complex models they demand, problems in specifying all relevant relationships and the lack of adequate data (Pianta, 2005).

Few studies focusing on IT as a determinant of employment changes have been done so far. One of the many problems of empirical work in this area was due to the fact that IT diffusion is not easy to measure and that the statistics used vary in different economic contexts, in different countries and in different sectors within the same country (Spiezia, Vivarelli, 2002). The data gathered and harmonized within this project allows us to contribute to the empirical evidence on IT effects on employment. We use cross-country dataset with industry level data to explore the relationship between employment growth and IT use. After a short literature review, some descriptive statistics are shown. This is followed by the results of a regression analysis, estimating an equation for employment growth with IT intensity as one of the explanatory variables and controlling for the effects of changes in output and labour costs. The section ends with concluding remarks and suggestions for further research.

8.2.2 Literature review A review of the most relevant empirical evidence emerging from firm level studies mostly points to a positive correlation between innovation and employment. Innovating firms (introducing product innovations and also those introducing process innovations) grow faster and are more likely to expand their employment than non-innovating ones, regardless of industry, size and other firm characteristics (Pianta, 2005). Most of these studies focused on manufacturing firms only. Harrison et al. (2005), conducting a study funded by the European Commission, using comparable micro-data from four EU countries, included the service sector and pointed out the differences between the sectors. In the report they concluded that “in manufacturing, although process innovation tends to displace employment, compensation effects are prevalent, and product innovation is associated with employment growth. (…) In the service sector, there is less evidence of displacement effects from process innovation, and though less important than in manufacturing, growth in sales of new products accounts for a non-negligible proportion of employment growth” (Jaumandreu et al., 2005).

Microeconomic empirical evidence will overestimate the positive effects of technology change (Spiezia, Vivarelli, 2002): the firms introducing product innovations as a result of new technology will increase employment due to the market share they gain as a result of innovation, and even where innovation is labour-saving the estimated impact on employment might be positive since it does not take into account the effect on competitors that are crowded out of the market by innovative firms, thereby losing jobs. These considerations can be accounted for by conducting the analysis at the industry level, where the data captures the new hires by innovative firms, the indirect effect on competitors and the number of employees at the end of diffusion process. However, even sectoral-level studies do not capture all of the direct and indirect effects of technology on employment since they do not take into account all compensation mechanisms that operate outside the sector originally affected by IT use. As already mentioned, only aggregate studies could do that.

135


Studies on industries showed that employment impact of innovation is positive in industries characterized by high demand growth and an orientation toward product (or service) innovation, while process innovation leads to job losses (Pianta, 2005). Not many of the studies focus specifically on ICT. Matteucci and Sterlacchini (2003) used disaggregated data for Italian industries for 1997-2000 to test whether industries with greater ICT intensity experience higher rates of employment growth. They concluded that employment changes are not only associated with changes in output and unit labour costs but also with ICT intensity – with a negative relationship in manufacturing and a positive relationship in services.

The lack of empirical studies can probably be attributed to the methodological issues that add to the issues of measuring ICT. Problems include the availability of data on omitted variables that affect both the presence of ICT and employment characteristics, difficulties in distinguishing between cause from effect and various empirical anomalies (Handel, 2003).

8.2.3 Descriptive statistics The data used for the purposes of this analysis are industry-level data from linked E-Commerce and Production Survey files. The following variables were used: growth of employment, gross output, wages, and labour productivity, together with two measures of ICT use – percentage of workers with access to internet (PCPCT) and percentage of workers with access to broadband (DSLPCT). Whereas the growth variables refer to their average growth in 2001-2005 period (or the years available in this period)2, the ICT use variables refer to the average of their lagged value in the same period (with a one-year lag as compared to production survey variables). Nine countries are included in the dataset: Austria, Czech Republic, France, Germany, Great Britain, Italy, Netherlands, Norway and Sweden. Data for 26 sectors of manufacturing and services are considered (the classification of industries corresponding to the one used in EUKLEMS databases) and Agriculture, Hunting, Forestry and Fishing as well as Non-Market Services were excluded from the analysis.

Figure 8.2.1 and Figure 8.2.2 provide descriptive evidence of the relationship between ICT use and employment growth in European industries, showing the intensity of ICT use measured by DSLPCT and the average annual growth rate of number of employees. There seems to be no discernible relationship between these variables, either in manufacturing or in services. A very similar relationship would be observed if PCPCT were used as an indicator of ICT intensity.

Examining correlations between employment growth and ICT use variables gives the same conclusion. Tables 8.2.1 and 8.2.2 report cross-industry correlations for these variables and also include other variables of interest, important for employment growth: output growth, labour productivity growth and growth in wages. Correlation coefficients are not statistically significant where ICT use variables are considered. However, we find a positive and significant correlation between employment growth and growth in output and a negative and significant correlation between employment growth and growth in labour productivity and labour costs. These relationships hold for both sectors.

2 For some of the countries, not all of the years were covered in the dataset. For France, Czech Republic and Germany, a three year period was used (taking into account the available data).

136


Figure 8.2.1: Annual rate of change of employment and ICT use by industry and country – Manufacturing:

-.1-.0

50

.05

.1

0 .2 .4 .6 .8Average % of workers using broadband

Average annual rate of change of employment Fitted values

Figure 8.2.2: Annual rate of change of employment and ICT use by industry and country – Services:

-.05

0.0

5.1

0 .2 .4 .6 .8Average % of workers using broadband

Average annual rate of change of employment Fitted values

137


Table 8.2.1: Correlation matrix – Manufacturing: Variables 1 2 3 4 5 6 1. Growth of employment . 2. Growth of output 0.2080** .

3. Growth of labour productivity -0.4849***

0.6770*** .

4. Growth of wages -0.3721*** 0.1614*

0.2979*** .

5. Average percentage of workers using broadband 0.0189 0.1247 0.0799 0.1232 . 6. Average percentage of workers using internet 0.0069 0.1122 0.0811 0.1045

0.8479*** .

Notes: *** significant at the 1% level; ** significant at the 5% level; * significant at the 10% level. Table 8.2.2: Correlation matrix – Services: Variables 1 2 3 4 5 6 1. Growth of employment . 2. Growth of output 0.5474*** . 3. Growth of labour productivity -0.2287* 0.6710*** .

4. Growth of wages -0.5296*** 0.0141 0.4261*** .

5. Average percentage of workers using broadband -0.0098 0.0527 0.0566 0.2193 . 6. Average percentage of workers using internet -0.0290 0.0490 0.0744 0.2255 0.9912*** .

Notes: *** significant at the 1% level; ** significant at the 5% level; * significant at the 10% level.

8.2.4 IT and employment growth - Industry-level evidence To test for the strength of the relationship between employment growth and ICT use, employment changes have to be considered within a theoretically-based specification where the influence of other variables apart from ICT use is controlled for. Following the models used for example in Antonucci and Pianta (2002) or Mateucci and Sterlachinni (2003), we model the determinants of employment change as changes in demand (proxied by output), changes in labour costs (wages), and changes in technology (proxied by ICT use intensity):

Lij = α0 + α1 Yij + α2 Wij + α3 ICTij + e, where, for sector i and country j, L is annual rate of change of employment, Y is annual rate of change of output, W is annual rate of change of wages, ICT is average ICT intensity, and e is the error term. Country dummies and a sector dummy for manufacturing were also included. Our aim is to test whether industries with higher ICT use experience higher (or lower) employment growth.

Table 8.2.3 shows the OLS estimate of the model. The first column shows the results where the average percentage of workers using broadband was used as a proxy for ICT use whereas the second uses average percentage of workers using internet. The same data as for descriptive statistics were used – a pool of 26 manufacturing and service industries for 9 European countries in 2001-2005 period. Country dummies were included (for all countries except one) as well as a dummy for manufacturing.

138


Table 8.2.3: The determinants of employment change, 2001-2005: Change of employment Change of output 0.302 0.300 (7.50)*** (7.48)*** Change in wages -0.809 -0.814 (-7.42)*** (-7.52)*** Average percentage of workers using broadband 0.017 (1.29) Average percentage of workers using internet 0.021 (1.65) Manufacturing dummy -0.016 -0.015 (-4.21)*** (-4.05)*** Constant 0.034 0.030 (3.39)*** (3.43)*** R-squared 0.5427 0.5449 Number of cases 227 227

Notes: T-statistics between brackets. *** significant at the 1% level; ** significant at the 5% level; * significant at the 10% level. The results of the regressions suggest that industry-level employment changes in this period are explained by changes in output (positive and significant relationship) and changes in labour costs (negative and significant). The dummy for manufacturing industries is significant which suggest lower employment growth as compared to service industries. The impact of ICT use on employment growth is not significant, regardless of the variable we use to measure ICT use.

8.2.5 Conclusions The results of regression analysis confirm the conclusions suggested by descriptive statistics – that ICT use does not have a statistically significant effect on employment growth at the industry level. The results of this study should be considered with some limitations in mind. To study the employment effects of the new technology embodied in ICT, it would be optimal to have the data that would connect the ICT use with product and process innovations. Without the data on innovations we are missing an important part of the link between ICT and its effects. Also, the framework used for the regression does not take into account the endogeneity of output which might give biased estimates. This issue should be tackled with a study estimating a simultaneous system of labour demand and supply.

Other areas of research connected to employment would also be very interesting for research. One of them is the impact of ICT on skills. The estimates of the impact of ICT use on quantity of employment do not provide any information on the quality of created versus suppressed jobs – skill content of jobs will probably not be similar. Technology may not significantly reduce overall labour demand but it may affect the composition of the type of labour demanded therefore having an impact on the quality of jobs. A research question would thus be whether the relative importance of skilled/unskilled jobs changes with increased ICT use. Another research area refers to the impact of ICT use on wages: are wages higher and/or do wages grow faster in industries and firms with higher ICT use? Also, an interesting research question is the importance of skills (in general and ICT skills) for the effect of ICT on productivity. The latter will be explored in the second part of the chapter.

139


8.3 Skills

8.3.1 Background and Introduction When studying the effects on firm productivity from information technology (IT) the subject can be approached from at least two angles. The one considers the intensity of IT in the factor inputs of the firm production, the other the dissemination and the ability to use the technology, and secondly, the dissemination of and the ability to use the technology.

In a production function approach, including the inputs capital, labour and intermediates each could be considered to have certain IT intensity through the share of IT capital, proportion of IT intensive higher educated employees and the kind of intermediate goods or services used. Information on both IT investments and details on intermediates are scarce but data on investments in IT human capital are available for at least some countries in the cross country dataset accumulated within the frame of the Eurostat ICT impacts project. Indicators of the firm’s use of IT are also available. Even if human capital cannot be separated into type, it still plays an important role as a control variable when studying the effects on productivity from the dissemination of IT or from its interaction with IT.

In this section the purpose is to shed some light on the effects on productivity from the quality and type of the labour input to the firm production for a selection of countries and industries. The effects will be investigated both on their own and together with the firm IT usage on the matched production survey datasets in Finland, Norway and Sweden. Firstly some literature references will be briefly described together with the expected firm behaviour. Thereafter follow some results from the regressions run on the national firm level datasets. These results will be presented for different productivity measures and datasets. Finally there is a comparison between the use of real skills and the wage bill as the proxy for skills in the regressions, followed by some concluding remarks.

8.3.2 Literature Draca et al (2006) have summed up a wide range of literature on effects from information technology on productivity and find out that the Solow paradox no longer holds, if it ever did. They found evidence of an important role for IT in accounting productivity both from growth accounting and econometrics. They also discuss the investments in organisations as something that needs to be further investigated, just like Abramovsky and Griffith (2005) touch upon the issue of the slower or lower uptake of IT in Europe compared with the United States. They consider that the differences may be referred to fewer investments in the organisational changes needed to allow the IT investments full blown effects. Crespi et al (2007) find evidence of an interaction between organisational changes and IT in their effects on productivity.

Doms et al (1997) conclude that plants using a considerable amount of new technology employ more educated workers and pay higher wages. Those firms high in capital seem to pay high wages both pre and post adoption of new technologies, as well as being high productivity plants. Galindo-Rueda and Haskel (2005) find evidence that firms high in college educated staff tend to be more productive and according to Ilmakunnas and Maliranta (2005) non-technical higher education affected productivity positively, and stronger than technical ones in Finland during the period 1988 to 1998.

Gunnarsson et al (2001, 2004) show that an upgrade in the skills level with given IT affected productivity more than an upgrade in IT with given level of skills in Sweden during the period 1986-1995. They also found that human capital is complementary to IT capital and when IT was included in the regressions the direct effects on productivity were substituted by

140


indirect ones over human capital. The largest indirect effect came from workers having upper secondary education (12 years) relative to workers with only 9 years at school. Disaggregated by field of study the next largest effect came from university educated engineers in relation to upper secondary school engineers. Romer (1996) concludes that not only higher education in itself but the right type of education is of importance for productivity.

8.3.3 Expectations The literature presented tells that there is a relationship between human capital (skills) and productivity. There is also a link between IT and productivity. IT can affect productivity either directly or through the human capital. If assumed that these relationships hold, human capital is expected to affect productivity at least to some extent. However this does not necessarily encompass all types of higher skills. Another series of literature found links between the IT effects on productivity over the organisational changes. This gives further fuel to the importance not only of human capital, but of the right type of human capital for productivity. There also seems to be a link between the firm investments in human capital on the one hand and wages and productivity on the other. This means that firms high in IT investments or IT use are also likely to be high in human capital.

8.3.4 Descriptive data Information on skills of firm employees is mainly available in the Nordic countries, France and Slovenia. In other countries these types of data are either not matched at all to the production surveys or are not available at the firm level. When data on skills are not available the wages bill is often used as a proxy. Wages as a variable is a much more encompassing measure than real skills and does not only include education but also experience and whether a person is liked or not by his or her employer (the latter does not necessarily correlate to the level of education and ability). Table 8.3.1: Employees with access to PCs or fast broadband connections at work

2005 Finland Norway Sweden % % %

All firms 70 60 66 Services 76 68 73

Proportion of employees with PC access

Manufacturing 56 51 60

All firms 62 54 55 Services 66 62 64

Proportion of employees with fast broadband access

Manufacturing 43 46 43 Source: Statistics Finland, Statistics Norway and Statistics Sweden Finland is the country with the highest proportion of employees with access to a PC or to fast Internet connections at work. Norway is lagging behind somewhat and Sweden has the highest share of manufacturers with PC access. In general, the services firms have a stronger use of these IT tools in all countries.

In this study IT intensive higher education (or skills or human capital) is equalised with post upper secondary education in mathematics, physics or information technology. The proportion of employees with higher education differs among the countries studied, and seems to be lowest for Sweden. Moreover the IT intensive employees are far smaller groups

141


in all countries and amounts to about a third of the total number of employees with higher education.

While there is a difference among industries, those firms high in skills, IT or general, on average and independently of country do have more employees with access at work to computers or fast broadband. These same firms are also most often high in capital, wages and productivity, just like the findings by Doms et al (1997) and Galindo-Rueda and Haskel (2005). However, the causality of the possible links is still unclear.

Table 8.3.2: Average share of employees with post upper secondary education in 2005

2005 Finland* Norway Sweden % % %

All firms 10 17 4Services 11 19 5

Proportion of employees with IT intensive post upper secondary education Manufacturing 11 15 4

All firms 17 6 11Services 22 7 11

Proportion of employees with general post upper secondary education Manufacturing 10 4 10 *Education data for Finland are at present only available up to the year of 2004. Sources: Statistics Finland, Statistics Norway and Statistics Sweden

8.3.5 Real skills and productivity For the countries with real skills available some ordinary least square production function regressions in the natural logarithm (ln) have been set. One aim is to study the effect on productivity from a change in the proportion of highly skilled employees, while another is to illustrate the differences between real skills and the proxy wages.

Several measures of productivity have been calculated for the project cross country dataset, both single and multifactor ones. Pointing out what is most relevant for the impacts regression including human capital is not uncomplicated. Earlier studies by for instance Hagsten et al (2007) on Swedish industry subsets reported that there are no major differences between using value added and gross production, but Bartelsman and Doms (2000) favour the gross values on the basis that the shift in the use of intermediate inputs relative to capital and labour over time may create bias in the productivity measure based on value added. This is also emphasised by Bailey (1986). Despite this, the attention here is focussed on the value added based productivity metrics. The decision to do so follows from the fact that intermediate inputs are available only in the form of goods, leaving out the services inputs which could easily be IT intensive.

The Ordinary Least Squares regressions are performed step by step on the unbalanced pooled panels of firms including the years 2001 to 2005 (for Finland skills data are only available up to the year of 2004). Beginning with the two skills variables (Hkitpct) and (Hknitpct), meaning the share of employees with IT intensive and general (non IT intensive) post upper secondary education, the effects on productivity can be observed as well as whether the addition of new variables and changes of the datasets affect the skills estimates. Included also are dummies controlling for fluctuations in productivity over time, differences among firms, size and affiliation. Labour productivity is calculated as value added per person and total factor productivity (TFP) also includes capital in the denominator.

142


All countries receive labour productivity (Lpv) bonuses from an increase in the share of employees with higher IT intensive education. For Finland the estimates are close to an elasticity of one and are stable over the different datasets used. When adding further variables like firm age (Age) and age squared (Age2), the latter controlling for non-linearity, the Swedish and Finnish estimates are kept almost untouched, that is, significant and positive with stronger effects from high IT skills than from high general skills. The Finnish coefficient estimates are clearly higher than those for Norway and Sweden. The final matching of the E-commerce survey to the production data set leaves the major part of firms behind and also affects the estimates for Sweden and Norway in particular. In Sweden the effects on labour productivity from all kinds of higher education become stronger while high IT skills are still more important. However, Norwegian labour productivity is not affected at all by improvements in higher general skills in the smaller sample of firms, while the effects from IT skills improve strongly.

Table 8.3.3: Effects on firm labour productivity from higher skills in Finland, Norway and Sweden. Ordinary Least Squares estimations on pooled and unbalanced panels of firms 2001-2005 LnLpv FIN* SWE NOR Hkitpct 0.767 0.734 0.766 0.127 0.129 0.262 0.454 0.749 (0.014)*** (0.014)*** (0.039)*** (0.006)*** (0.006)*** (0.045)*** (0.001)*** (0.137)***

Hknitpct 0.596 0.581 0.610 0.118 0.117 0.217 0.111 0.190 (0.011)*** (0.011)*** (0.038)*** (0.004)*** (0.004)*** (0.036)*** (0.021)*** (0.238)Age 0.004 0.002 0.052 0.020 (0.000)*** (0.001)* (0.001)*** (0.003)*** Age2 0.000 0.000 -0.002 -0.001 (0.000) *** (0.000) (0.000) *** (0.000)*** DSLpct 0.085 0.194 0.255 (0.018)*** (0.014)*** (0.064)***

ECpct 0.012 -0.001 -0.080 (0.023) (0.000) (0.067)R-squared 0.21 0.22 0.27 0.14 0.16 0.28 0.27 0.43Observations 59635 59617 7746 776737 775778 11398 38491 663

*Skills data for Finland are only available up to the year 2004. Robust standard errors are shown within brackets with ***, ** and * meaning significant at the one, five and ten per cent level. Unreported time, industry, size and affiliation dummies are used to control for firm specifics and changes over time. Sources: Statistics Finland, Statistics Sweden and Statistics Norway There is a difference among industries in the labour productivity effects from improvements in the skills level. Generally, services firms have more to gain than manufacturers. Some of the firms in the latter group completely lack the link between improvements in the firm skills share and labour productivity.

The proportion of employees with access to fast broadband (DSLpct) and the firm share of E-commerce (ECpct), which is the sum of the firm shares of buying and selling over the Internet, are the two IT variables meant to capture different phases of IT maturity. The analysis of these is not the main focus of this chapter, but since the estimates may be affected by the choice of skills variables or the interaction with IT, even the direct effects are briefly discussed.

Firms in all three countries were positively affected in labour productivity by an improvement in the share of employees with access to fast broadband. The most visible effect appears in Norway and the weakest in Finland. This follows quite logically from

143


Finland being the country with the highest access rates to PCs and fast broadband at work and therefore may already have depleted possible productivity gains, while Norway, lagging behind somewhat, still has bonuses to pick up. Selling and buying over the Internet was found to be of no importance at all for labour productivity in any country.

Table 8.3.4: Effects on firm total factor productivity from higher skills in Finland, Sweden and Norway. Ordinary Least Squares estimations on pooled and unbalanced panels of firms 2001-2005 LnTfp FIN* SWE NOR Lnkl -0.386 -0.387 -0.378 -0.169 -0.169 -0.221 -0.280 -0.260 (0.003)*** (0.003)*** (0.008)*** (0.001)*** (0.001)*** (0.004)*** (0.002)*** (0.013)***

Hkitpct 1.095 1.081 1.034 0.279 0.282 0.253 0.535 0.881 (0.028)*** (0.028)*** (0.089)*** (0.007)*** (0.007)*** (0.056)*** (0.014)*** (0.140)***

Hknitpct 1.110 1.107 1.027 0.234 0.231 0.289 0.194 0.334 (0.023)*** (0.023)*** (0.087)*** (0.005)*** (0.005)*** (0.045)*** (0.021)*** (0.244)Age -0.002 -0.005 0.059 0.022 (0.001)** (0.002)* (0.001)*** (0.004)*** Age2 0.000 0.000 -0.003 -0.001 (0.000)*** (0.000)** (0.000)*** (0.000)*** DSLpct 0.238 0.316 0.249 (0.041)*** (0.018)*** (0.066)ECpct -0.060 0.001 0.034 (0.052) (0.001)* (0.069)R-squared 0.63 0.63 0.58 0.61 0.61 0.77 0.68 0.61Observations 59634 59616 7745 776675 775716 11397 38490 662

*Skills data for Finland are only available up to the year 2004. Robust standard errors are shown within brackets with ***, ** and * meaning significant at the one, five and ten per cent level. Unreported time, industry, size and affiliation dummies are used to control for firm specifics and changes over time. Sources: Statistics Finland, Statistics Sweden and Statistics Norway The regressions with total factor productivity as the dependent variable give far higher R-squared values indicating that the variables included explain this productivity better than the labour productivity. All skills estimates now become stronger than in the labour productivity regressions. The patterns also differ, showing that the type of education is of lesser importance, except for Norway, where a relationship between improvements in high general skills and TFP does not exist. Changes in the pattern are not surprising since TFP is a much more complex measure capturing also spillover effects from human capital as well as many other things. If the total factor productivity could be considered to reflect even the organisational uptake of IT, discussed by Draca et al (2006) and Abramovsky and Griffith (2005), the estimates here indicate that the Finnish adoption of IT may be on another level than in Sweden and Norway. Increases in capital per employee (Lnkl) reduced productivity independently of dataset and country.

Separate regressions by type of industry reveal that the effects are still stronger on total factor productivity than on labour productivity. While Norwegian groups of firms often seem to lack the link between productivity and high general skills, both Finland and Sweden show similar patterns where the effects from both high IT and high non IT skills give stronger effects on total factor productivity, stronger than for firms in general and particularly so for the services firms. Nevertheless, the exception to the rule appears in the form of Finnish manufacturers, who gained more from improvements of general skills in the exhaustive data

144


set while no effects could be found for them in the regressions on the smaller sets. This partly coincides with earlier findings by Ilmakunnas and Maliranta (2005).

The share of broadband access exerts a significantly positive effect on Finnish firm TFP. Similar but stronger effects are received by Swedish firms while Norway stands out with no significant estimates. Moreover, the effects on productivity from the E-commerce are still to be found. Even if there is a significant estimate for Swedish firms it is so small that it could hardly be separated from zero. The very stable Finnish estimates over the different datasets may reveal that Finland is suffering sample bias to a lesser degree than other counties.

In order to study whether IT works through the skills in its effect on productivity, an interaction variable has been created. This variable (Hkdsl) is based on the firm share of employees with post upper secondary education multiplied with the firm share of employees with access to fast broadband. Only the Swedish firms received significant estimates for this variable, indicating that this complementarity or indirect effect does not exist in either Finland or Norway. The results also show that the effect is stronger on the total factor productivity and adding this variable weakens, or takes away, the direct effect on productivity from skills. This mirrors the findings by Gunnarsson et al (2001, 2004) on the complementarity between skills and IT investments. The broadband variable only suffers marginally from the operation. This indirect effect does not appear as often (or as strongly) in the Swedish manufacturing firms as in the services firms.

Table 8.3.5: Interaction between skills and IT use in Finland, Sweden and Norway. Ordinary Least Squares estimations on pooled and unbalanced panels of firms 2001-2005 LnLpv LnTfp Variable FIN* SWE NOR FIN* SWE NOR Lnkl -0.383 -0.219 -0.249 (0.008)*** (0.004)*** (0.005)*** Hkpct 0.712 0.116 0.932 1.129 -0.010 0.905 (0.050)*** (0.052)* (0.086)*** (0.109)*** (0.065) (0.090)*** Age 0.002 0.020 -0.004 0.023 (0.001) (0.003)*** (0.002)* (0.004)*** Age2 0.000 -0.001 0.000 -0.001 (0.000) (0.000)*** (0.000)** (0.000)*** Dslpct 0.099 0.175 0.171 0.308 0.277 0.126 (0.026)*** (0.016)*** (0.028)*** (0.056)*** (0.019)*** (0.030)*** Hkdsl -0.047 0.196 -0.138 -0.188 0.449 0.047 (0.064) (0.068)** (0.100) (0.140) (0.085)*** (0.105) R-squared 0.26 0.27 0.37 0.60 0.76 0.63 Observations 8791 11685 5139 8790 11684 5138

*Skills data for Finland are only available up to the year 2004. Robust standard errors are shown within brackets with ***, ** and * meaning significant at the one, five and ten per cent level. Unreported time, industry, size and affiliation dummies are used to control for firm specifics and changes over time. Sources: Statistics Finland, Statistics Sweden and Statistics Norway

8.3.6 Wages and productivity When comparing wages with real skills and using data on firms in Sweden as an example, it becomes clear that wages differ in estimates, in this case with far stronger effects on productivity, even stronger than the two skills variables together. The estimates also seem to be clearly significant and positive, with lesser risks of turning negative or not significant due to the high correlation between wages and productivity. As opposed to the real skills

145


variable, the estimates of the wages hardly differed at all between labour productivity and total factor productivity. However, they are stronger for the smaller datasets, which is only partly true when real skills are used because the effect on labour productivity solely increases while the effects on TFP are if anything redistributed rather than increased in favour of the general non IT high skills.

While IT seemed to operate through the real skills in its effects on labour productivity, this is not the case for the proxy wages when interacted with the broadband variable (Lnwdsl). Nevertheless, although not reported here, there is a significant estimate for the manufacturing firms, but with a negative sign. These negative effects also appear for the total factor productivity in manufacturing firms while the services firms receive a boost in productivity somewhat larger than for firms in general. The effect on TFP over wages is much smaller than with the real skills.

Table 8.3.6: Effects on firm labour productivity in Sweden from real skills and wages Ordinary Least Squares estimations on pooled and unbalanced panels of firms 2001-2005 Lnlpv SWE SWE Hkitpct 0.127 0.129 0.262 (0.006)*** (0.006)*** (0.045)*** Hknitpct 0.118 0.117 0.217 (0.004)*** (0.004)*** (0.036)*** Hkpct 0.116 (0.052)*

Lnw 0.626 0.666 1.000 0.998 (0.001)*** (0.001)*** (0.014)*** (0.021)*** Age 0.007 0.009 0.010 0.052 0.020 0.020 (0.000)*** (0.002)*** (0.002)*** (0.001)*** (0.003)*** (0.003)***

Age2 -0.000 -0.000 -0.000 -0.002 -0.001 -0.001 (0.000)*** (0.000)* (0.000)* (0.000)*** (0.000)*** (0.000)***

DSLpct -0.012 -0.064 0.194 0.175 (0.012) (0.182) (0.014)*** (0.016)***

ECpct -0.001 -0.001 (0.000)* (0.000) Lnwdsl/Hkdsl 0.008 0.196 (0.030) (0.068)**

R-squared 0.43 0.46 0.50 0.49 0.14 0.16 0.28 0.27Observations 861907 773853 11393 11679 776737 775778 11398 11685

Robust standard errors are shown within brackets with ***, ** and * meaning significant at the one, five and ten per cent level. Unreported time, industry, size and affiliation dummies are used to control for firm specifics and changes over time. Source: Statistics Sweden

It can also be seen that the e-commerce variable is significant in the wages labour productivity regression, but the value is so small that it in practice means no effect. When using the skills proxy the improvement in the firm employee share of broadband access did not affect labour productivity while it is clearly significant in the real skills regressions. Nevertheless, it still plays a role for TFP, but the magnitude in its effect on productivity is smaller than with real skills. Generally, the effects from the IT maturity variables are either significant to a lesser degree or smaller than in the real skills regressions, despite the fact that R-squared ends up at more or less the same level.

Unreported Norwegian estimates of the skills proxy resemble the Swedish ones in pattern as well as in sign and magnitude. The interaction variable is also significant both for labour and

146


total factor productivity, but with small values and negative signs. Even the Finnish firms receive far stronger effects from wages on productivity in the matched dataset than in the exhaustive one, on similar levels as Sweden and Norway. However, here the spread in estimates between the datasets is wider and the initial estimates are much smaller compared with both Sweden and Norway as well as with its own real skills. The Finnish interaction estimates are weakly significant, showing small and negative values, just like for Norway.

Table 8.3.7: Effects on firm total factor productivity in Sweden from real skills and wages Ordinary Least Squares estimations on pooled and unbalanced panels of firms 2001-2005 LnTfp SWE SWE Hkitpct 0.279 0.282 0.253 (0.007)*** (0.007)*** (0.056)*** Hknitpct 0.234 0.231 0.289 (0.005)*** (0.005)*** (0.045)V Hkpct -0.010 (0.065)Lnkl -0.194 -0.194 -0.247 -0.246 -0.169 -0.169 -0.221 -0.219 (0.000)*** (0.000)*** (0.003)*** (0.003)*** (0.001)*** (0.001)*** (0.004)*** (0.004)***

Lnw 0.649 0.691 1.043 0.983 (0.001)*** (0.001)*** (0.019)*** (0.028)*** Age 0.011 0.011 0.012 0.059 0.021 0.023 (0.000)*** (0.003)*** (0.003)*** (0.001)*** (0.004)*** (0.004)***

Age2 -0.000 -0.000 -0.001 -0.003 -0.001 -0.001 (0.000)*** (0.000)** (0.000)*** (0.000)*** (0.000)*** (0.000)***

DSLpct 0.105 -0.643 0.316 0.277 (0.016)*** (0.245)** (0.018)*** (0.019)***

ECpct 0.001 0.001 (0.000) (0.001)* Lnwdsl/Hkdsl 0.126 0.449 (0.041)*** (0.085)***

R-squared 0.73 0.72 0.81 0.81 0.61 0.61 0.77 0.76Observations 861852 773798 11392 11678 776675 775716 11397 11684

Robust standard errors are shown within brackets with ***, ** and * meaning significant at the one, five and ten per cent level. Unreported time, industry, size and affiliation dummies are used to control for firm specifics and changes over time. Source: Statistics Sweden Results for all countries from the regressions using wages suggest that firms on average receive stronger effects on firm productivity in the matched dataset than in the production survey. Just like for Sweden the results are almost identical over labour and total factor productivity. However, while the direction prevails, the spread in the size of the effects varies.

8.3.7 Conclusions Although the causality is not known, firms with many highly educated employees also tend to have high capital, high wages and a high proportion of employees with access to PCs and fast Internet connections. As expected, high skills affect productivity both directly and to a certain degree indirectly. For the labour productivity the type of skills seems to be of greater importance, in this case with stronger effects from high IT skills than for the total factor productivity where general high skills are almost more or equally important. Services firms do generally have more to gain from improvements in skills than manufacturers. The use of wages as a proxy for real skills has its obvious limitations since it gives more or less the same

147


148

estimates for both labour and total factor productivity, tends to increase when estimated on a sample biased towards larger firms (which in turn have higher production, productivity, capital, skills and wages) and risks missing out on skills that affect productivity negatively or not at all. Therefore, real skills data need to be made available in or for simple matching to the production surveys. The bias in the matched E-commerce sample could be dealt with by either widening the sample or reducing the rate of non-response, or hopefully both.


Chapter 9

ICT and Business Process Integration Tony Clayton / Mark Franklin UK Office for National Statistics

9.1 Summary 1. This Chapter focuses on the “C” in ICT. New technology allows firms to integrate their processes more efficiently, both internally across the links of the value chain (for example, by electronic tagging of materials, work-in-progress and finished goods) and externally, with their suppliers and customers. Examples of external links include the point-of-sale technology now widely used by retailers which automatically re-orders supplies from wholesales or indeed direct from manufacturers, and the systems in professional service firms which log time spent on individual projects, bill clients and integrate with credit and performance management systems.

2. A feature of these applications of technology is that they are less focussed on tangible inputs, and more concerned with how firms use the technology that is now widely available to them.

3. This Chapter describes an approach which uses the infrastructure of the project to develop a separate, but related, line of enquiry. By following this route, and notwithstanding that the analysis is very preliminary, we are able to report results at a level of consistency across multiple countries that would otherwise be very difficult to achieve. These (preliminary) results suggest that, at least in manufacturing, firms with external ICT links with either their suppliers or their customers tend to be more productive than similar firms with no such links. Moreover, this positive relationship is additive to the relationship between fast-internet penetration and productivity.

9.2 Introduction

9.2.1 Origin of data on Business Process Integration 4. The first attempts to collect data on electronic business process integration date from 1999, when the US Bureau of Census carried out a 50,000 firm ‘Computer Network Use Survey’ (CNUS) together with the manufacturing census. This was based on a ‘Porter model’ approach to business process definition in firms, which best fits the value chains of manufacturing operations (Porter (1985)). The survey provided evidence on the types of processes which were electronically linked in US manufacturing firms, and on the extent of ‘Fully Integrated Enterprise Resource Planning’ (FIERP) – or the use of electronic linkages to coordinate operations and planning across all business processes in a firm.

5. US efforts to identify productivity gains associated with elements of CNUS have gone through a number of iterations. Initial work by Atrostic and Nguyen (2002) showed some impacts associated with use of networks, but it took some time for details to emerge. The later work has tended to show that productivity impacts can be industry and process specific. The US survey has not yet been repeated.

6. EU efforts have been based on simplified version of the US survey, with the restriction that it was, in its UK pilot and initial EU extensions, applied only to firms which linked electronic business processes to e-commerce (buying and / or selling). This restricts the effective sample. But the EU questionnaire used during the period 2003-2005 essentially asks:

149


• which of a short list of processes (logistics, production, marketing, finance) are electronically linked to buying / selling);

• whether firms’ business processes are electronically linked to those of suppliers or customers.

7. Figure 9.1 shows the UK version of this part of the E-Commerce survey which applies to the year 2006, identifying:

• the processes in the firm which are electronically linked to e-buying or e-selling computer systems;

• the presence of Enterprise Resource Planning systems;

• the use, and integration, of customer relationship management (CRM) systems.

8. In addition, the survey carries the standard Eurostat questions on the use of intranet / extranet, LAN and WLAN.

9.2.2 Analysis of Electronic Business Process and their Impact 9. A number of studies have been carried out on penetration of e-business process use. Several of these were reported to an OECD workshop on e-business definition in 2004, and some EU work on productivity effects has been done.

10. While the focus of Clayton and Criscuolo (2002) is on e-commerce rather than explicitly on e-business integration, the authors provide a helpful taxonomy of business changes that may be brought about by ICT:

• Technology can reduce costs of search for buyers and cost of access to market for suppliers;

• Technology facilitates innovation in new methods of purchase and forms of delivery of services, such as digital downloads;

• Technology facilitates “vertical disintegration” of value chains, reducing co-ordination costs and allowing firms to specialise in specific parts of the value chain;

• Increased availability of information changes the balance of power between suppliers and customers. The authors conjecture that e-commerce generally moves the odds in favour of buyers.

11. In the UK, Goodridge and Clayton (2004) have looked at simple labour productivity statistics, defining e-business in terms of its intensity in individual firms. The analysis (for 2002 and 2003) treats the number of linked processes in the UK questionnaire as indicative of sophistication, and sets out to compare firms with multiple links to those with few or none. The findings from this work can be summarised as follows:

• little difference in labour productivity between those firms with no links and those with few, but some advantage for those with multiple links (4 or more);

• a difference in productivity effects between manufacturing and services. For manufacturing firms the highest productivity is achieved by those with multiple links including links to suppliers (back up the supply chain) while for the service sector the highest productivity is achieved by those with multiple links including links to customers (forward down the supply chain).

12. These differences have some theoretical explanations, in terms of price effects of electronic networks for commodities bought by manufacturers, or of customer information benefits for the complex relationships managed by services firms. However, we are not aware of any further development of this work, despite the development of the questionnaire in other countries.

150


Figure 9.1: UK extract of E-Commerce survey dealing with e-business processes

13. Statistics Sweden has undertaken some work on business process use, as part of the impact work presented at an OECD workshop in May 2007 (Hagen, 2007). More recent work using just the business process integration part of the index suggests that:

• a composite indicator of multiple linked processes can be effective as a discriminator of both labour productivity and multifactor productivity (using data from 2002 and 2004);

• many of the individual components of the composite also show up as significantly related to labour productivity, particularly those covering logistics, intermediate ordering and suppliers’ business systems.

14. The Swedish analysis also suggests that, in a hierarchy of ‘interconnectedness’, correlations with productivity are more pronounced for those firms at the top of the hierarchy, which may represent firms at the technology frontier. When the vast majority of firms display

151


some measures of ‘interconnectedness’, such as having a web-site, it is not surprising that it is difficult to distinguish relationships.

15. Another interesting features of the Swedish analysis is that differences between firms in terms of ‘interconnectedness’ in 2002 correlate much more strongly with productivity differentials of firms in 2005 than with productivity differentials in 2002. This suggests that measures of integration proxy dynamic rather than static features of firm behaviour.

16. De Graaf et al (2007) report results of a special survey of Dutch companies designed to provide insights into organisational practices surrounding ICT standardisation and centralisation. Internal and external integration are elements of this model. The authors find that ICT maturity has a significant level effect on productivity, and their model also shows that firms that combine a high level of ICT intensity with a high level of ICT maturity are closest to the technological frontier1.

17. Motohashi’s work – also reported to the OECD workshop in May 2007 (Motohashi, 2007) – also suggests evidence for electronic business process linkage effects for Japan. His formulation of the practical effects is summarised in Figure 9.2.

Figure 9.2: Motohashi model of increasing e-business effectiveness

9.2.3 Options for this project 18. There appears to be enough national evidence to suggest that e-business integration has productivity effects which can be detected at firm level, but that the impacts are industry / business model dependent. Also, the role of external links appears crucial in securing productivity gains for a firm – as the end point of a process in which e-business links have progressively more effect.

1 Indicators of business integration are also used in the UK and Netherlands research reported in the Appendix to Chapter 6 and, in a systems modelling context, in Chapter 10.

152


9.2.4 Outline of rest of this chapter 19. The following section (9.3) defines an index of ICT business process integration based on multiple questions from the harmonised E-Commerce survey. The method employed is to modify the core code used elsewhere in the project, retaining only the elements necessary to link the data and conduct the analysis reported in this chapter. We have suppressed the generation of the full suite of distributed micro data (DMD) statistics in order to speed up computational time, since the full version of the code takes several hours to run.

20. The principal advantage of this route is that it maintains much of the core data infrastructure and therefore maintains comparability of results. The main disadvantage is that we lose DMD statistics for the ICT variables2. There is also some programming overhead involved in setting up the analysis in this way, and some constraints on the form of analysis that can be undertaken, for example in terms of the regression specifications.

21. Execution of the modified code was optional among project members. At the present time we have results from 6 project countries: the United Kingdom (GBR), the Netherlands (NLD), Sweden (SWE), France (FRA), the Czech Republic (CZE) and Austria (AUT).

22. Section 9.4 reports some basic descriptive statistics and correlations between different indicators of e-business integration and productivity. In this section we also look at complementary data from the core analysis on e-commerce, to illustrate the differential impact of upward and downward integration links across different industry groups.

23. The following section (9.5) presents some results of simple OLS regressions. The regression specification is identical to that reported in Chapter 6 with the addition of 2 alternative measures of e-business integration as dummy variables.

24. The chapter concludes with some suggestions for extending and developing the analysis.

9.3 Methodology 25. Our review of the literature and of earlier work in some project countries led us to focus on question 9 in Figure 9.1 above. This question follows a filter question, the form of which, in the Eurostat model questionnaire (Question A5 of the version dates 11 May 2004) is as follows:

“Did your enterprise have IT systems to manage the placement or reception of orders, during [time period]?”

26. Firms answering “yes” to this question then answer a series of questions about the linkages between their order management systems and other systems within and beyond their organisation. The model questionnaire wording is shown in Table 9.1 alongside mnemonics used in this project3.

Table 9.1: E-Commerce survey questions Variable mnemonic Description Intlink_1 Internal link to systems for re-ordering replacement supplies (0/1) Intlink_2 Internal link to invoicing and payment systems (0/1) Intlink_3 Internal link to systems for managing production, logistics or service

operations (0/1) Extlink_s External links to suppliers’ business systems (0/1) Extlink_c External links to customers’ business systems (0/1) 2 For example, we cannot generate statistics describing the relationship between integration variables and productivity such as those shown below for E-commerce variables and productivity. 3 See also Figure 9.1, Question 9.

153


27. For this chapter we have analysed the data from these questions. This is a microcosm of the metadata and data assembly phases for the core analysis described in Chapter 3 above. An example of the sort of issue that can arise is where individual countries have chosen to combine multiple elements of the model questionnaire into a single question on the survey, or conversely to separate a single question in the model questionnaire into multiple survey questions4. It is not unusual for the mappings of survey questions to variable mnemonics to change from one survey to another. Care also has to be taken to ensure a consistent treatment of the filter question, for example to ensure that zeroes are imputed in each of the follow-up questions when the answer to the filter question is “no”.

Table 9.2: Integration Index

Index Link ecom , logistics

extlink_s, extlink_c

0 0 0 & 0 0 & 0 1 1 0 & 0 0 & 0 2 0 or 1 0,1 or 1,0 0 & 0 3 0 or 1 1 & 1 0 & 0 4 0 or 1 0 & 0 0,1 or 1,0 5 0 or 1 0,1 or 1,0 0,1 or 1,0 6 0 or 1 1 & 1 1 & 1 “link” =1 if intlink_1 or intlink_2 or intlink_3 =1 “ecom” = 1 if esales or epurch = 1 “logistics” = 1 if intlink_3 =1

28. Individual responses to the follow-up questions are converted into an “index” of business integration using the logic in Table 9.2. This logic is based on previous work carried out in Sweden, the Netherlands and the UK. It can be seen that the index combines responses from the follow-up questions with data on ESALES and EPURCH collected as part of the core analysis.

29. The first iteration of the analysis used the full 7-stage index of integration to create dummy variables that were then added to the regression framework described in Chapter 6. Analysis of the results suggested that this approach was not appropriate. The integration dummy variables were highly correlated with one another, and there was poor identification of relationships with firm performance at this level of detail.

30. Consequently we focus in the rest of this chapter on two further transformations of the index shown in Table 9.2. BUSINT2 is a dummy variable set to zero for firms with no links (that is, INDEX = 0) and 1 otherwise. BUSINT3 is a dummy variable set to zero for firms with no external links (INDEX ≤ 3), 1 otherwise.

9.4 Descriptive statistics 31. We have used the data management and merging routines built into the core project code to generate some simple descriptive statistics of the integration variables for those countries that have opted in to this analysis (Table 9.3)5. These statistics are based upon merged (unbalanced) samples of firms that appear in both the EC and PS input datasets in a particular year. There are some odd results for the Swedish data for 2001, but generally the data show the expected result that measures of integration tend to increase over time. Unsurprisingly, mean values of BUSINT2 are higher than BUSINT3 and indeed approaching saturation in some cases.

4 This reinforces the point made in Chapter 3 that differences arise in the wording of the E-Commerce survey notwithstanding that each country’s survey is based on the Eurostat model questionnaire. 5 Table 9,3 and subsequent tables are at the end of this chapter.

154


32. Previous work on UK data by Goodridge and Clayton (2004) suggests that the impact of integration varies across different industries. To throw light on this hypothesis, Table 9.4 shows simple correlations between BUSINT2, BUSINT3 and the separate measures of external links with suppliers and customers, in each case against value added per employee. The correlations have been run separately for firms in manufacturing and in services,

33. Most of the measures of integration have significant positive correlations with labour productivity. There is some evidence of stronger correlations among firms in the manufacturing sector, for example in Sweden and Austria. Four of the six countries that have conducted the analysis (GBR, NLD, FRA and CZE) demonstrate the “expected “ result that external links with suppliers are more correlated with productivity than links with customers among the manufacturing sector, while in the service sector the relative size of the correlation coefficients is reversed. Of the remaining countries, external links with customers is strongly correlated with labour productivity for Austrian service sector firms, while BUSINT3 (which takes the value 1 for external links with either suppliers or customers) is significantly correlated with productivity for manufacturers but not for service sector firms. This suggests that the significance of the direction of links differs between sectors.

34. Sweden is an exception. This may reflect the unusual data for 2001 and the relatively low penetration of external links in other survey years (Table 9.3).

35. Another perspective can be obtained from the distributed micro data generated by the core analysis. In principle it would be possible to include the additional E-Commerce variables in the full DMD framework but we have not done so due to time constraints6. Accordingly we do not have DMD indicators for the integration variables. We can, however, get a flavour from the core EC variables ESALESPCT and EPURCHPCT, respectively, the shares of E-sales and E-purchases in the total sales and purchases of the firm.

36. Figure 9.3 shows mean levels of EPURCHPCT by quartile of the productivity distribution for the manufacturing sector (excluding ICT-producing electronics). In general there is a tendency for EPURCHPCT to be slightly higher for firms in the higher quartiles of the productivity distribution, although there is considerable variation across countries.

37. Figure 9.4 shows the same data but for the distribution sector. Again there is a good deal of variation between countries, but in several countries there is a tendency for EPURCHPCT to be negatively related to the quartile distribution of productivity.

38. Figures 9.5 and 9.6 show the equivalent data for ESALESPCT. In the manufacturing sector (Figure 9.5), there is no clear relationship between ESALESPCT and the productivity distribution across all countries. Italy shows a positive relationship whereas France shows the opposite.

39. In all cases, shares of E-purchases and sales in Norway are considerably higher than other countries, and apart from Norway, there is a fairly general positive relationship between ESALESPCT and the productivity distribution among the distribution sector (Figure 9.6).

6 The full DMD analysis takes several hours of computer time due to its extensive interaction between samples and the full business register. By contrast, the code used for this chapter addresses the business register only once and runs in a few minutes.

155


Figure 9.3 – E-Purchasing and Productivity in Manufacturing

0

0.05

0.1

0.15

0.2

0.25

SLO ITA NLD GBR CZE SWE GER AUT FRA NOR Mean

Epur

chpc

t (1=

100%

)


This graph shows mean epurchpct, quartiled into TFP distributionYear==2004euk==MexElec

Figure 9.4 - E-Purchasing and Productivity in Distribution

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

SLO ITA NLD GBR CZE SWE AUT FRA NOR Mean

Epur

chpc

t (1=

100%

)


This graph shows mean epurchpct, quartiled into TFP distributionYear==2004euk==DISTR

156


Figure 9.5 - E-Sales and Productivity in Manufacturing

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

SLO ITA NLD GBR CZE FIN SWE GER AUT FRA NOR Mean

Esal

espc

t (1=

100%

)


This graph shows mean esalespct, quartiled into TFP distributionYear==2004euk==MexElec

Figure 9.6 - E-Sales and Productivity in Distribution

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

SLO ITA NLD GBR CZE FIN SWE AUT FRA NOR Mean

Esal

espc

t (1=

100%

)


This graph shows mean esalespct, quartiled into TFP distributionYear==2004euk==DISTR

157


9.5 Regression results 40. The regression methodology follows the same approach as that set out in Chapter 6, adding BUSINT2 and BUSINT3 as (separate) dummy variables. By utilising common code for data assembly and linking, the regression dataset is virtually identical, comprising firms appearing in both the Structural Business Survey and the E-Commerce surveys in a particular year, and trimmed for outliers in the productivity distribution identified from a prior regression7.

41. Regression results in Table 9.5 for firms in the manufacturing sector can be compared with those in the core analysis (Table 6.4 in Chapter 6). This table uses the BUSINT3 dummy variable which distinguishes those firms with external links with either their suppliers or customers. This proved to be more significant than BUSINT2 (which distinguishes all firms with any form of integrated link). Generally the coefficient estimates in Table 9.5 and the “core” results in Table 6.4 are very similar. Of the countries that have conducted this variant analysis, all those where DSLPCT was significant in Table 6.4 also have significant estimates on DSLPCT in Table 9.5. DSLPCT is not significant in either specification for the UK, and there is no comparable result in Table 6.4 for France.

42. Another noteworthy feature of Table 9.5 is that ECPCT is not significant for any of the 6 countries in this comparative analysis. Again this is consistent with the results reported in Chapter 6. However, in 3 of the 6 countries (the UK, Sweden and the Czech Republic), BUSINT3 is significantly related to TFP, and in the latter 2 countries this result is additive to a significant coefficient on DSLPCT. This suggests 2 possible interpretations:

• BUSINT3 captures different information about firms’ use of ICT than that captured or reflected in DSLPCT8;

• BUSINT3 seems to convey more information about firm-level productivity than ECPCT9.

43. Table 9.6 shows results for the equivalent regression for firms in the ICT-producing sector. It can be compared with Table 6.5 in Chapter 6. As in Chapter 6, regression results are less coherent for this sector, and there is little evidence of a systematic relationship between DSLPCT and TFP. The only significant coefficient is -0.35 for Sweden (compared with -0.37 in Table 6.5). BUSINT3 is not significant in any of the 6 countries in the analysis and its addition to the specification leads to insignificance for ECPCT across the board, compared with significant estimates for the Netherlands and the Czech Republic in Table 6.5.

44. For completeness, Table 9.7 shows results for market services (excluding telecoms and postal services). The comparable table in Chapter 6 is Table 6.6. Most parameters are similar between the two specifications, apart from in Sweden, where DSLPCT and ECPCT become insignificant. Sweden is also the only country with a significant estimate for BUSINT3, although with the ‘wrong’ sign. This is further evidence that the relationship between ICT usage and productivity differs across different sectors.

7 Minor differences in the datasets may arise because some countries have suppressed the multinational dummy for computational reasons. The Czech Republic only has data on the ICT integration variables for a sub-set of years. We have also suppressed the ECPCT variable in France to generate regression results. In the core specification reported in Chapter 6 there are no regression results for France because the French E-Commerce survey does not include questions on fast internet and e-sales/e-purchases in the same survey year. 8 For example, it could be hypothesised that DSLPCT is a proxy for ICT capital services while BUSINT3 is a proxy for TFP or the unexplained portion of productivity. See also Chapters 6, 10 and 14, and Maliranta and Rouvinen (2003). 9 Recall that ECPCT is the sum of shares of E-sales and E-purchases. We report above on the likely differential impact of E-sales and E-purchases in different sectors. Although the regression results in Table 9.5 are confined to manufacturers, it is possible that the results would be different if we included E-sales and E-purchases separately.

158


9.6 Next steps 45. The results presented in this chapter are preliminary, and are based on simple OLS regressions with no testing for endogeneity. Accordingly, the motivation is not about identification of cause and effect, but only about identification of relationships in a multi-country, multi-variate framework.

46. Despite the preliminary nature of the work, there are some promising indications that this type of information about applications of ICT is relevant for how firms perform. Possible avenues for further research include:

• Broaden the scope of the analysis to include all countries in the project.

• Incorporate ICT business process integration variables from the E-Commerce survey into the core disaggregated microdata framework. This would allow parallel testing of industry/country aggregate specifications.

• Examine alternative specifications including, for example, dynamic specifications to explore relationships between business process integration and future productivity.

• Endogeneity tests to establish, whether ICT integration is a driver of performance, or an effect, of performance, or of some other unobserved factor.

• Further development of inter-relationships between ICT use / intensity, ‘interconnectedness’ and innovation, for example by broadening the application of the analysis reported in Chapter 12.

159


Table 9.3: Descriptive Statistics Mean of extlink_s Mean of extlink_c

YEAR GBR NLD SWE FRA CZE AUT YEAR GBR NLD SWE FRA CZE AUT2001 0.23 0.49 2001 0.23 0.322002 0.19 0.14 0.08 0.58 0.28 2002 0.17 0.15 0.06 0.64 0.202003 0.22 0.18 0.11 0.70 0.33 2003 0.18 0.20 0.09 0.29 0.212004 0.23 0.18 0.11 0.73 0.05 0.35 2004 0.18 0.17 0.08 0.38 0.06 0.232005 0.31 0.18 0.10 0.13 0.37 2005 0.27 0.16 0.09 0.15 0.24

Mean of BusInt2 Mean of BusInt3YEAR GBR NLD SWE FRA CZE AUT YEAR GBR NLD SWE FRA CZE AUT

2001 0.70 0.84 2001 0.32 0.562002 0.61 0.76 0.73 0.82 0.95 2002 0.26 0.18 0.10 0.70 0.192003 0.76 0.82 0.81 1.00 0.96 2003 0.29 0.24 0.14 0.61 0.182004 0.83 0.84 0.80 0.97 0.58 0.98 2004 0.29 0.23 0.14 0.79 0.07 0.192005 0.89 0.87 0.80 0.59 0.98 2005 0.38 0.24 0.14 0.18 0.20

Table 9 4: Correlation Coefficients Correlation with REAL VALUE-ADDED PER EMPLOYEECoverage MANUFACTURING

GBR NLD SWE FRA CZE AUTLPV corr prob corr prob corr prob corr prob corr prob corr prob

extlink_s 0.19 <.0001 0.09 <.0001 0.05 0.0010 0.17 <.0001 0.17 <.0001 0.08 <.0001extlink_c 0.06 <.0001 0.08 <.0001 0.06 <.0001 0.10 <.0001 0.15 <.0001 0.10 <.0001BusInt2 0.12 <.0001 0.12 <.0001 0.12 <.0001 0.09 <.0001 0.17 <.0001 0.06 0.0028BusInt3 0.10 <.0001 0.08 <.0001 0.06 <.0001 0.14 <.0001 0.17 <.0001 0.14 <.0001

Correlation with REAL VALUE-ADDED PER EMPLOYEECoverage SERVICES

GBR NLD SWE FRA CZE AUTcorr prob corr prob corr prob corr prob corr prob corr prob

extlink_s 0.08 <.0001 0.03 0.0026 0.01 0.5802 0.08 <.0001 0.11 <.0001 -0.03 0.1333extlink_c 0.13 <.0001 0.05 <.0001 0.02 0.0546 0.12 <.0001 0.13 <.0001 0.15 <.0001BusInt2 0.14 <.0001 0.09 <.0001 0.07 <.0001 0.07 <.0001 0.15 <.0001 0.02 0.2545BusInt3 0.09 <.0001 0.04 <.0001 0.02 0.1670 0.09 <.0001 0.13 <.0001 -0.01 0.4441

160


Table 9.5: Dependent Variable REAL TOTAL FACTOR PRODUCTIVITYCoverage MANUFACTURING EXCL ELECTRICAL AND OPTICAL EQUIPMENT

GBR NLD SWE FRA CZE AUTRegressor coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-statlnkl -0.30 -15.3 -0.21 -27.0 -0.20 -21.2 -0.23 -29.4 -0.50 -68.1 -0.22 -10.8lnw 0.88 16.9 0.63 26.9 0.89 16.6 0.89 30.3 1.08 31.1 0.70 10.7dslpct 0.10 1.4 0.05 1.5 0.33 8.7 0.06 1.6 0.20 3.8 0.19 2.8ECpct -0.06 -0.6 0.02 0.8 0.00 -0.3 #N/A #N/A 0.04 1.1 0.02 0.4BusInt3 0.16 4.0 0.01 0.7 0.06 2.1 0.01 0.3 0.07 2.3 0.01 0.4[other dummies]Observations 1456 4028 3881 4049 2881 509R-squared 0.51 0.32 0.80 0.54 0.75 0.38 Table 9.6 Dependent Variable REAL TOTAL FACTOR PRODUCTIVITYCoverage ELECTRICAL & OPTICAL EQUIPMENT, POST & TELECOMMUNICATION SERVICES

GBR NLD SWE FRA CZE AUTRegressor coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-statlnkl -0.27 -7.1 -0.33 -14.8 -0.20 -9.4 -0.18 -5.9 -0.48 -39.2 -0.22 -4.7lnw 0.78 8.0 0.43 5.9 0.68 5.1 0.93 9.0 0.81 15.8 0.57 4.1dslpct -0.22 -1.9 0.06 0.8 -0.35 -3.3 -0.08 -0.8 0.13 1.8 0.07 0.5ECpct 0.39 1.8 0.12 1.5 0.00 -0.8 #N/A #N/A 0.07 1.4 -0.04 -0.3BusInt3 0.08 1.0 -0.07 -1.4 0.11 1.2 0.09 1.1 0.07 1.4 0.05 0.5[other dummies]_EDF_ 430 621 436 452 616 98_RSQ_ 0.44 0.57 0.34 0.35 0.83 0.66

161


162

Table 9.7 Dependent Variable REAL TOTAL FACTOR PRODUCTIVITYCoverage MARKET SERVICES EXCL POST & TELECOMS

GBR NLD SWE FRA CZE AUTRegressor coef T-stat coef T-stat coef T-stat coef T-stat coef T-stat coef T-statlnkl -0.36 -44.0 -0.20 -40.3 -0.13 -36.6 -0.27 -39.8 -0.53 -94.1 -0.14 -9.0lnw 0.91 48.1 0.88 81.4 0.99 52.0 0.99 41.5 0.95 41.9 0.83 16.1dslpct 0.05 1.7 0.04 2.8 0.02 1.4 0.01 0.3 -0.14 -4.2 0.02 0.5ECpct -0.06 -1.1 0.07 3.5 0.00 1.8 #N/A #N/A -0.06 -1.6 0.09 2.3BusInt3 -0.03 -1.3 -0.02 -1.2 -0.04 -3.1 0.02 1.3 0.02 0.8 0.02 0.5[other dummies]Observations 4706 6951 5905 3303 3455 719R-squared 0.48 0.60 0.62 0.63 0.80 0.48


CHAPTER 10: ICT INVESTMENT AND PRODUCTIVITY Shikeb Farooqui, UK Office for National Statistics George van Leeuwen, Statistics Netherlands Summary: Using UK and Dutch data on ICT use, ICT investment and firm performance, this Chapter considers whether the ICT usage variables collected in Ecommerce surveys are feasible predictors for missing ICT capital inputs. Furthermore, by embedding an augmented production function in a structural model, we investigate the contributions to productivity of ICT capital deepening and ICT use related TFP. Similar to Chapter 6 special attention is given to fast internet connectivity and Ecommerce. The results of the modelling exercise show that combining ICT usage variables collected in the Ecommerce survey and proxy measures for (total) capital inputs in a structural modelling framework can be seen as a useful approach for predicting missing ICT capital stocks, as well as for disentangling the productivity contribution of the two forms of ICT use in a direct (TFP) and indirect (through ICT capital deepening) component. For both countries we find evidence that, after controlling for IT investment, the productivity contribution of fast internet connectivity is channelled through ICT capital deepening. By contrast electronic selling seems to assert a direct impact on TFP. 10.1. Introduction

It is widely acknowledged that ICT has proven to be one of the most dynamic areas of investment

as well as a very pervasive technology. The use of ICT can make firms more productive along

different channels. It enables firms to customize services offered, to reduce inefficiency in the use of

other inputs (e.g. by reducing inventories or by streamlining other business processes), or to seize

spillover effects generated by ICT networks. These innovation-enabling characteristics make ICT a

potentially important driver of productivity growth (see e.g. OECD, 2003 and OECD, 2004).

Notwithstanding the increasing and overwhelming empirical evidence presented on the impact of

ICT on productivity (growth) in the past decade, the relative importance of different channels through

which ICT affects productivity still leaves several questions open to debate. A recurrent question is

how to disentangle the contributions of investing in ICT from its use given that ICT is often seen as a

general-purpose technology embodied in a special type of physical capital and that making it

productive also calls for complementary investments and special skills of its users.

A prerequisite for assessing the importance of the many factors that may play a role in explaining

the contribution of ICT to productivity (growth) is the availability of sound data. Traditionally,

growth accounting techniques have been employed to assess their contribution. Despite their virtue of

simplicity, growth accounting methods – in general – cannot go further than quantifying the

contribution of ICT capital deepening (i.e. the changes in ICT capital stocks per unit of labour) to

TFP-growth at the industry level (see e.g. Van Ark et al. 2008 for a recent application). Moreover,

applying growth accounting methods requires the availability of time series of ICT investment in

hardware and software to construct ICT capital stock data.

Today, much of the empirical literature focuses on the contribution of ICT to TFP-growth, i.e. the

unexplained residual of productivity (growth) that remains after capital deepening has been taken into

account. Here, micro-econometric approaches come into play. By going down to the firm level and

163


by using additional data sources this strand of research attempts to shed light on the ’ultimate causes’

of TFP-growth. In principle, such a ‘residual purifying’ can only be achieved if data on capital inputs

are available at the firm level. Otherwise one can have doubts whether these additional data variables

really represent TFP contributions to productivity growth.

A problem often encountered in micro-econometric research concerns the lack of suitable data on

capital stocks at the firm level. Constructing capital stock data at the firm level is seriously hampered

by sampling in investment surveys or by the lumpiness of the investment process.1 In addition,

accounting for software investment in productivity research raises serious difficulties for several

reasons (see Clayton et al (2005)). This lack of data on ICT and other capital stocks on the preferred

level of analysis hampers a fair comparison between econometric and growth accounting approaches.

Amongst other concerns, it may be questioned whether alternative data are valid substitutes for

missing data on (ICT) capital stocks.2

This Chapter elaborates on the above question by linking firm-level data on ICT use collected in

the harmonized EUROSTAT survey (henceforth labelled EC survey) with estimates of ICT and other

capital stocks derived from firm-level investment time series and accounting data on firm

performance.

This Chapter seeks to address two main questions:

• Do particular forms of ICT use have an impact on productivity after ICT capital stock data

are accounted for in the productivity models?

• Are the ICT use variables collected in the EC surveys feasible predictors of ICT capital

stocks?

The linked data are used to investigate the relation between ICT use variables and ICT capital

stock as well as the productivity impact of ICT use and ICT capital stocks. This is achieved by

embedding an augmented standard production function framework in a structural model. As there are

similar data in both countries this approach is feasible both for UK and for the Netherlands.

The plan of this Chapter is as follows. In sections 10.2 and 10.3 we discuss the research strategy

followed and the empirical specifications applied in econometric part of this research. Section 10.4

discusses the data used. Section 10.5 presents the estimation results and elaborates on the feasibility

of the EC variables for predicting (missing) ICT capital stocks. Finally, section 10.6 summarizes the

most important findings. Appendices A and B present some further descriptive statistics of the

Netherlands and UK datasets, and Appendix C briefly describes the results of a simulation exercise to

test out ability to predict firm level ICT capital stocks.

1 Industry investment time series hide the fact that many firms do not invest every year. If firms were surveyed every year and data on initial stocks were available, this would not raise a problem as zero investment implies zero addition to the (gross) stock of capital. However, estimating initial stocks at the micro level is problematic because of incomplete and short investment histories. 2 It goes without saying that this problem also applies to non-ICT capital inputs.

164


10.2. Motivation of the modelling exercise

A natural starting point for the discussion is a productivity model that accounts for ICT capital.

Following Marilanta and Rouvinen (2003) and Farooqui (2005), we augment the traditional Cobb-

Douglas production function model with a labour augmenting factor E, which is made ICT dependent,

and Ecommerce Survey variables (denoted by EC) as possible determinants of TFP. In the absence of

data on the composition of capital stocks this leads to βα= itititit ELKAY ).( (1)

),()ln( ititit HCECFA = , (2)

where Y is value added in constant prices, K is a measure of total capital inputs, L is labour inputs. A

represents TFP and i and t refer to firms and time respectively. Equation (2) expresses that TFP (in

logarithmic form) depends on a vector of firm and time specific EC and skill variables (labelled HC).

Marilanta and Rouvinen (2003) argue that the marginal productivity of labour should depend on

whether employees use computers or not. In the Ecommerce Survey we have data available on the

share of employees that use computers (PCpct). Thus, following their reasoning, we can specify E in

(1) as

PCpctE θ+= 1 (3a)

Then, inserting (3a) in (1) and expressing (1) in per employee terms, yields after taking logarithms:

),()ln()1()ln()ln( itititititit HCECFPCpctLLK

LY

+βθ+−β+α+α≈ , (4)

where the approximation follows from .)1ln( PCpctPCpct θ≈θ+

Another more direct route, conditional on the availability of IT capitals stocks, is to split up total

capital inputs into ICT (KI) and non-ICT capital inputs (K R). Allowing their productivity contribution

to vary and assuming constant returns to scale, yields the following linear specification for the

productivity model:

),(~)ln()1()ln()ln( 11 ititit

I

it

R

it HCECFL

KL

KLY

+−+= αα . (5)

Equation (5) represents the model usually applied to mimic growth accounting at the firm level

with the aim to disentangle the ICT productivity contributions into capital deepening and TFP

components. Equation (4) is used to clarify the main concern of this Chapter. In (4), represents the

productivity premium from using ICT. If ICT is not a special type of capital and does not have an

associated productivity premium then PCpct is redundant and F( ) can be interpreted as the

contributions of ICT use to TFP conditional on including some measure of total capital inputs (K

θ

P) in

the model. This would also imply that it is no longer necessary to split total capital inputs into (KI)

and (K R) in (5).

165


Now, let us assume that ICT is a special type of capital. Then, omitting data on the relative

importance of (past) ICT investment boils down to estimating (4) whilst excluding PCpct. In general,

this is the approach followed in the regression part of the core analysis and in most other sub themes

of this Project. Notice, that if PCpct is (erroneously) omitted from (4), then F( ) can no longer be

interpreted as a ‘pure’ TFP contribution to productivity of ICT use. In this case, estimates of

indicators that refer to ICT use (lumped into EC) may be biased upwards as it is likely that they

capture the productivity impact of an important omitted variable.

This problem will be more severe for those components of F(EC, HC) that are highly correlated

with (missing) IK . An example is the fast internet connectivity indicator DSLpct. This indicator is

constructed as PCpctDSL ⋅DSLpct = . It can be added to (3a) to account for the additional labour

augmenting productivity impact of broadband enabling employees that use computers:

PCpctDSLPCpctE ⋅++= 211 θθ .3 (3b)

If (3b) instead of (3a) is used in (1) and PCpct is not included in the empirical specification

(i.e.θ is assumed), then it is easily seen that the estimate01 = 2θ is likely to be biased upwards as it

will capture the underlying impact of IT investment (either measured by PCpct or IK ). Thus, it is

questionable whether the estimates of DSLpct reported in Chapter 6 of this Report really represent a

‘true’ TFP contribution to productivity4. Notice further, that PCpct seems to be redundant if data on

ICT capital stocks per employee are available. In this case we can directly estimate (5) using our

estimates of (KI) and (K R). The coefficient on DSLpct will then only capture the productivity

premium 2θ and not the capital deepening effect.

However, introducing IT investment as a separate input into production raises other problems.

Several forms of ICT usage are likely to be highly correlated with IT investment. Furthermore, to get

the best out of their IT investment, firms also need to have (unobserved) complementary investments

in place, which, in turn, are also likely to be correlated with productivity (see e.g. Bresnahan et al.,

2002, and Van der Wiel and Van Leeuwen, 2004).

Going beyond ICT measurement, a similar reasoning applies to other potential determinants of

productivity differences across firms. Here, one of the ‘usual suspects’ is labour input. If the

successful application of ICT requires better qualified labour inputs, then we face another

simultaneity problem as variations in high qualified labour inputs will simultaneously explain

productivity differences as well as differences in ICT intensity (use). In a similar fashion to many of

the other regression models used in this Project, we are forced to use the wage rate as a proxy measure

for skill differences between firms. Wages and productivity, measured using valued added, are highly

3 DSL is a binary indicator for having fast internet or broadband connections or not. 4 As mentioned in the appendix of Chapter 6 because DSLpct is a derivative of PCpct, correctly including both in a regression specification can lead to problems of multicollinearity that can incorrectly influence the significance of the variables. This is another reason for including IT capital stocks, if our real interest lies in assessing the impact of DSLpct.

166


correlated. Thus, the use of this HC indicator exacerbates the already existing simultaneity and

causality problems.

In order to cope with these complications, we embed the productivity equation in a structural

model that attempts to account for the simultaneity of productivity, IT investment, ICT use and skills

and reverse causality for skills and ICT use. In addition, the system also enables us to look at the

predictive power of the model for explaining per employee capital stocks.

10.3. Empirical implementation

The availability of firm level data on ICT and non-ICT capital stocks and proxy measures for total

capital inputs offers an interesting opportunity to investigate several of the issues discussed above. As

in Chapter 6, the structural model will focus in particular on high speed internet, connectivity

(DSLpct), electronic buying (Epurchpct) and selling (Esalespct). In addition, we will also use the

same ICT maturity indicators (BI) as given in the single equation framework presented in the

Appendix of Chapter 6.5 With these data, a system of estimation equations is constructed as follows:

(I) Productivity equation (5) is augmented with the three EC variables, mentioned above and

wages to proxy unobserved skills. After relaxing the familiar constant-returns-to-scale assumption

used in growth-accounting, using small letters to denote logarithmic transformations and omitting

firm and time subscripts, this yields:

wllklkly IR42121 )1()()( ααααα +−−+−+−=−

(6)

+ , jj

j BIEpurchpctEsalespctDslpct ∑=

+++13

8765 αααα

with w the logarithm of wages per employed person;

(II) next, we added equations for other capital inputs per employee inputs (7) and ICT capital

inputs per employee (8):

)(1 lklk PR −γ=− (7)

DslpctDSLWEBPCpctlk I5432 η+η+η+η=−

(8)

, jj

J BIEpurchpctEsalespct ∑+++=

13

876 ηηη

with a proxy measure for total capital inputs and WEB a binary indicator for website use; Pk

(III) finally, the system is completed after including wage equation (9) to control for simultaneity

and reverse causality between productivity, wages and ICT use:

DSLpctlywLw 321 )()( φ+−φ+φ= (9)

where L(w) represents lagged wage rates. 5 BI is a set of dummy variables representing increasing stages of ICT maturity. The underlying Business Integration Index is constructed from a set of binary response variables referring to automated business processes. Further details can be found in Chapter 9.

167


Equation (6) is a straightforward implementation of a standard Cobb-Douglas production

specification that accounts for IT capital and IT use. As mentioned above equations (7) and (8) are

capital stock prediction equations with a major focus on thoroughly describing possible determinants

of IT capital stock. The wage equation (9) captures the idea that firms face sticky wage schedules, so

that current wages depend on past wage bills and improvements in current productivity. We include

DSLpct to reflect the fact that computer enabled employees embody higher skills and therefore

command a wage premium.

By estimating (6)–(9) simultaneously, we attempt to free the parameters of the productivity

equations from simultaneity biases as well as to predict ICT capital stocks per employee from the

most important EC variables through equation (8). Notice further, that (7) is a prediction equation for

non-ICT capital inputs per employee. We keep this model simple by using the proxy measure of total

capital inputs ( ) as the only explanatory variable.Pk 6 For identification purposes we do not include

EC variables in this equation.

In closing, we mention that the simultaneous-equation approach opens several possibilities for

hypothesis testing. An interesting question concerns the importance of the continuous EC variables

for TFP in (6) as well as for explaining differences in ICT capital inputs in (8). For instance, if

differences in DSLpct mirror differences in ICT capital inputs per employee, then one can expect its

estimate in (6) to be smaller in a simultaneous model than when using a single-equation approach.

Similar assertions can be made regarding the importance of Esalespct and Epurchpct. Thus, for

instance, testing 55 αη ≠ , 66 αη ≠ and 77 αη ≠ is an important objective of using the systems

approach. Furthermore, this approach enables us to decompose the ‘reduced-form’ impact of single-

equation estimation into a direct effect and an effect that is channelled through ICT investment. Take

for instance DSLpct. The direct effect of an increase of DSLpct is given by 5α and the indirect effect

by . 25αη

10.4. Data 10.4.1 The construction of unbalanced panels

In the empirical part of this Chapter we will use unbalanced panels consisting of firm-level data

that constitutes the overlap of EC and PS surveys of the core analysis and the capital stock panel

constructed with the help of investment surveys. The panel consists of about 11800 observations for

UK and about 3900 for the Netherlands (firm x year, covering the period 2001–2005 for UK and

2002–2005 for the Netherlands). The accounting data cover, among others, the following key

6 For the Netherlands this proxy measure is deflated depreciation costs. For UK it is a mixture of depreciation rates and total capital stocks. Estimates presented in Table 10.3 use industry level depreciation rates. The systems approach presented in Table 10.4 uses firm level total capital stocks.

168


variables: gross output, total turnover, employed persons in full time equivalents, intermediate inputs,

wage costs (including social security charges), depreciation costs and before-tax profits. The data

enable the construction of value added as the measure of output.7 In order to consider real outputs and

inputs in our analyses, we use detailed price indices from the EUKLEMS database to construct value

added in 2001 prices at lower levels of aggregation.

The panels contain interesting features, but are not completely perfect. Regular issues of

sampling, coverage, and missing variables are at stake. The problem of missing variables arises for

firm-level prices which are not available.8 Problems of coverage arise because the average size of

firms in the panel is considerably higher than actually measured for the total population of firms. In

particular, in both countries the trade and service sector consists of many small firms and, due to

sampling designs, many of them are only occasionally covered in the PS and EC surveys. The

sampling probability increases with firm size and larger firms are sampled every year, in principle.

Nevertheless these larger firms may also disappear in the course of time because of bankruptcy,

merging with other firms etc. Despite these complications, due to unique firm identifiers one can

easily construct panel data linking the yearly surveys over time.

10.4.2 The construction of capital inputs

One of the objectives of this Chapter is to investigate the predictive power of EC variables for

missing ICT capital stock data. For this reason we had to construct ICT capital stock data with the

help of other sources. Figure 10.1 presents the data sources used for UK and the Netherlands.9 This

Chapter distinguishes two types of capital inputs: ICT and other capital.

Netherlands

Contrary to earlier research (see e.g. Van Leeuwen and Van der Wiel, 2003, for the Netherlands),

we used the full detail of the investment surveys, combined with National Account price indices that

are available for the different asset types, to construct real expenditures on ICT- and non-ICT

investment. The ICT investment deflator used resembles the hedonic ICT price index used e.g. in the

USA.

7 We could also opt for gross output (or total sales) as the measure of output, but we have chosen not to do so. The reason for this is that many firms belong to wholesale and retail trade. For these branches the data on intermediate inputs consist for a very large part of purchases of trading goods and this makes these data incomparable with the intermediate inputs of other branches. 8 This issue may lead to problems when measuring the impact of e-commerce on TFP. It has been shown (Clayton & Criscuolo (2002)) that the use of e-procurement leads to competitive pressures on prices that result in lower prices of intermediates. Using industry level price deflators, as opposed to firm level, means that more cost efficient procurement is reflected in higher TFP. Similar concerns with the final price of outputs may lead to a negative coefficient on e-sales. A positive coefficient on e-sales therefore indicates that the true impact on TFP outweighs the pricing impact. 9 IT investment in UK also covers purchased software.

169


Figure 10.1: Construction of IT capital stocks

Survey Investmentin Fixed Assets

NLD

ARD Business SurveyInvestment Survey (Q+A)

UK

IT CapitalHardware + Software

UK

IT CapitalHardware Stocks

NLD

Initial ConditionsDeflatorsPIM method

Subsequently, capital stocks are constructed only if we have available at least five consecutive

observations on investment in constant prices in 1995–2000. We use this period to estimate initial

stocks as a starting point for applying the perpetual inventory (PIM) method in later years.

Accordingly, the capital stock of type k in period t reads: ktK

11)1( −− +δ−= ktitkkt IKK (10a)

Estimates for the unknown initial levels of the stocks in 1995 were obtained by using the approach of

Hall and Mairesse (1995):

kk

kk g

IK

δ+= 1

1 , (10b)

in which represents the pre-sample growth rate of real investment for type k, and is real

investment in the base year.

kg 1kI

The implementation of (10a) and (10b) requires a number of assumptions concerning the pre-

sample growth of investment and its depreciation. Estimates for were taken from industry time

series and for the depreciation schedule we used the inverse of the weighted average service life of the

asset types considered. For the Netherlands weights were taken for the National Accounts estimates

for Productive Capital Stocks by industry.

kg

10

UK

UK IT capital stocks are constructed following closely the methodology laid out in Bloom, Sadun,

Van Reenen (2007). In addition we improve the estimates of software capital by looking deeper into

the issue of underreporting and missing values.

10 For ICT hardware a service life of four years was used.

170


The ONS runs three separate surveys from which information on IT investment can be sourced.

The Production Survey introduced a question on Purchased Software and another on Own Account

Software in 2000. The Annual Business Survey on Capital Items (BSCI), which has a sampling frame

of the largest 2500 firms, inquires about investment in hardware and software (both purchased and

own account). Like the BSCI, the quarterly Capital Expenditure Survey (Capex) also collects

information on hardware and software (purchased and own account) investment. It is sent out to

30,000 firms every quarter, and firms are rotated out of the sample every 5 quarters.

The sampling frame of all three surveys is designed so that the largest firms i.e. firms with more

than 250 employees, are sampled by all three surveys every year. A small proportion of medium

sized firms return more than one survey every year. This allows us to check the consistency of

responses, across surveys, and use the best responses to construct the capital stock.

Comparing software investment responses in the Production Survey shows a general tendency of

firms dramatically to understate own account software production. Of total software expenditure

reported in the Production Survey, on average only 5% pertains to own account. Using this average,

which we allow to vary across SIC4 industry and year, we adjust software investment responses from

the BSCI and Capex to account only for purchased software.

Responses from the Capex are annualized and all three surveys are combed for missing values and

data manipulations conducted by the survey teams, so that only clean firm responses are used in the

construction of the capital stock. The three different surveys are then used to construct a separate data

set for hardware capital and another dataset for software capital.

The hardware and software investment series are adjusted using economy wide asset specific

hedonic price deflators provided by the National Institute of Economic and Social Research (NIESR),

which are based on Jorgensen’s US price deflators and have been harmonized to feed into the EU-

KLEMS project.. Like the Netherlands we employ (10a) to calculate capital stocks from the real

investment series.11

However our initial conditions (10b) differ. Instead of adopting the approach suggested in Hall

and Mairesse (1995), we calculate industry level capital to investment ratios for each type of IT

capital.12

The initial capital stock of a firm that has been sampled for the first time is determined through its

industry level capital to investment ratio. Allowing this ratio to vary yearly, allows us to circumvent

the influence of firm-specific investment cycles. The initial capital stock of a firm is calculated as

follows:

11

11 . k

k

kk I

IKK = , (10c)

11 We chose depreciation rates, of 36% to reflect an average shelf life of 4-5 years. 12 Industry level data on investment and capital stocks are also sourced from NIESR, see Timmer, O’Mahony, Van Ark (2007).

171


where 11 , kk KI represent industry level real investment and capital stocks for type k, and is real

firm level investment in the base year.

1kI13

Table 10.1: Summary statistics for UK and the Netherlands UK NLD A All firms a Labour productivity (x 1000) 65b 68b

Employed persons (fte) 2236 319 Wage rate (x 1000) 21b 44b % % DSLpct 32 28 Esalespct 14 6 Epurchpct 6 5 B By sector (2005) Manufacturing and Construction DSLpct 32 34 Esalespct 26 8 Epurchpct 18 3 Distributive services

DSLpct 37 37 Esalespct 16 11 Epurchpct 24 8 Other business services DSLpct 46 62 Esalespct 8 4 Epurchpct 17 6 a UK average for 2001–2005, the Netherlands average for 2002–2005. b UK figures reported in GBP, Netherlands reported in Euro.

10.4.3 Some descriptive measures

Table 10.1 presents descriptive measures for some variables used in the model analyses. These

statistics are calculated from the overlap between EC/PS panels and the capital stock panels. Thus,

these descriptive statistics cannot be compared directly with the results of the core analysis. What can

be concluded is that the additional linking of capital stock data leads to data that are biased to larger

firms. Furthermore, looking at how ICT is used, there are striking differences between the two

countries. In particular, both electronic buying and selling seems to be much more important in the

13 Total capital stock is constructed using investment series from the Production Survey and Capex. Initial conditions are imposed using the VIX-industry level dataset for all asset types. See Martin (2002) for further details.

172


UK than in the Netherlands in all sectors considered.14 By contrast, the figures for fast internet

penetration seem to be more comparable, except that broadband connectivity in other business

services seems to be considerably higher in the Netherlands than in UK.

Because we link data on IT investment, our data consist of a subset of the firms used in the core

analysis of Chapter 6. Nevertheless, we find very similar patterns for the correlation between

broadband connectivity and labour productivity. Figure 10.2a shows that productivity is increasing

with DSLpct on average, although the increase is more monotonous for NLD than for UK. The same

conclusion also applies to the correlation between broadband connectivity use and wages (see figure

10.2b), which is not a huge surprise, taking into account that wages and labour productivity are

correlated too.

Figure 10.2a DSLPCT & PRODUCTIVITY 2005

0

20

40

60

80

100

DSLPCT < 25% 25% < DSLPCT <50%

50% < DSLPCT <75%

75% < DSLPCT <100%

UKNLD

14 We used the same industry breakdown as in the Appendix of Chapter 6.

Figure 10.2b DSLPCT& WAGES 2005

01020304050607080

DSLPCT <25%

25% <DSLPCT <

50%

50% <DSLPCT <

75%

75% <DSLPCT <

100%

UKNLD

173


Table 10.2: Heckman model for ICT capital stocks NLD UK I) Outcome equation Est. SE Est. SE Proxy total capital inputs per employee Wage rate 0.725 0.060*** 1.174 0.021***Pcpct 1.131 0.082*** 1.045 0.048***Web 0.238 0.055*** 0.104 0.048***Dsl 0.001 0.053 0.046 0.033 Dslpct 0.423 0.090*** 0.297 0.052***Epurchpct -0.079 0.121 -0.002 0.069 Esalespct 0.153 0.110 0.104 0.044***Stage of ICT maturity (sellers perspective) No No Business Integration dummies Yes *** Yes *** Constant -2.879 0.253*** SIC 2 digit dummies included Yes Yes Time dummies included Yes Yes N 3852 11717 II) Selection equation Proxy total capital inputs per employee 0.019 0.013 0.032 0.010***Employment 0.349 0.010*** 0.386 0.009***Dslpct -0.029 0.046 0.174 0.033***Esales 0.055 0.033 0.070 0.026***Esalespct -0.094 0.093 -0.018 0.046 Epurch 0.031 0.029 0.142 0.025***Epurchpct 0.069 0.094 0.095 0.069 Constant -1.999 0.087*** SIC 2 digit dummies Yes Yes Time dummies Yes Yes Rho -0.114 -0.523 Sigma 1.092 1.265 Mills ratio -0.124 0.072 -0.662 0.060*** N probit 13573 17216

10.5. Estimation results 10.5.1 Econometric issues

The linking of three data sources yields smaller data sets that seem to be biased to larger firms.

This raises the possibility (but does not necessarily guarantee) that the firms that can be used to

estimate key relationships between ICT investment, use and productivity, were performing

significantly better in terms of productivity and/or were on average significantly more advanced in

their ICT use. To safeguard against possible selectivity biases, the following route has been chosen:

• First, we estimate a Heckman selection model on each equation separately;

174


• Second, we evaluate the equation specific selection bias and capture in a new “mills bias”

variable;

• Finally, we re-estimate all four equations jointly, in a system, adding the “mills bias” variable

to the model specifications.

The procedure can be explained by looking at the ICT capital stock equation (8).15 This prediction

equation gives us the relationship between ICT use and ICT capital stocks per employee:

).()(),,,,,,( controlszskillshBIEsalespctEpurchpctDSLpctDSLWEBPCpctflk I ++=− (8’)

This equation can be estimated for a sub-sample of all Ecommerce firms (the firms with linked IT

data). To correct for selectivity we have to use a selection equation:

),(),,,,,,( controlszKLEsalespctEsalesEpurchpctEpurchDSLpctfD DP

sel += (12)

with

1=selD if ICT capital stock data are available (otherwise 0=selD ). The Heckman procedure boils

down to estimating (8’) and (12) in 2 stages. The results of this procedure, the first step of the overall

estimation process, are presented in Table 10.2.

The main equation of interest is (8’). The prediction equation gives us the relationships between

the various forms of ICT use and how they impact IT investment decisions. For example, the positive

and significant coefficient on PCpct in (8’) would suggest a capital deepening effect – giving a larger

proportion of the workforce access to computers requires buying more computers or improved

software that allows for a more time-efficient allocation of existing computers amongst the workforce.

An insignificant coefficient on WEB would suggest that firms are able to create and maintain websites

without necessarily increasing IT capital stocks.

But the relationships we uncover in (8’) are specific to the sample we are using. Are they

representative of the larger EC/PS sample? Firms are selected for the estimation because we have

data on capital stocks readily available. For all the other firms we have all the necessary data apart

from capital stocks. What factors affect the probability of being a selected firm?

In the case of the Netherlands we know that we will have capital stocks for a firm only if that firm

was sampled for five consecutive years between 1995–2000. But only the largest firms, in terms of

employment, get sampled every year. Therefore, employment seems to be a factor that affects the

probability of selection.

In order to run the productivity regression (6), we require data on both IT and Non-IT capital. For

the Netherlands this overlap is perfect. The UK capital data come from two separate survey strands.

Therefore the probability of being selected depends on whether a total capital stock figure exists for

the firm in question. In this way (12) captures all the variables that may affect a firm’s selection

status. If larger firms really do use ICT differently from smaller firms then the core EC variables

15 The same procedure is followed for the other equations of system (6)–(9). To save space we do not discuss their results here (results are available upon request).

175


should also be a part of the selection equation. Other controls such as firm industry, region and

sampling year in the EC survey is also taken into account.

Now that we have a way to estimate the probability of being selected16 is there a way we can use

this probability to re-weight the estimates in (8’) so that the results are more representative? We can

think of the problem as akin to estimating an equation with a missing variable. Using the probabilities

we have calculated in (12) we can construct a new variable that captures the omitted selection bias.

Adding this variable as an extra control in (8’) therefore attempts to correct for the selection bias.17

10.5.2 Estimates for the simultaneous equation model

The appendix of Chapter 6 discusses the results of estimating (6) in a single equation framework.

It shows how coefficients on the ICT use variables change as IT capital and wages (skills) are

included in the regression framework. The changing pattern of estimates makes it difficult to draw

conclusions on the role of the EC variables in explaining productivity differences. It is tempting to

label the specification without wages as the preferred one, but this may not be wise, given that this

specification omits the only available control for human capital.18

One can expect that the level of ICT use and (cumulated) ICT investment are broadly correlated by

definition. This may also be the reason for the unstable results for the two capital elasticities. The

best way to investigate this issue is by using a structural model for the relation between ICT capital

inputs and the ICT use variables. In what follows we discuss the results of using a simultaneous

modelling approach that allows ICT use variables to have an impact on ICT capital inputs as well as

on productivity.

The results of the simultaneous model are presented in Table 10.3. As mentioned above before

estimating this model, we first investigate the selectivity issue by using Heckman models for each

equation of the simultaneous model.19 After adding Heckman selectivity correction terms (the inverse

of the Mills ratio) to each equation, the system is estimated with the help of 3-SLS. Doing so, we

allow for a possible correlation between the error terms of the equations.

16 Estimation of (12) runs on the whole EC/PS sample accounting for both selected and non-selected firms. 17 Significance of this new variable – the mills bias – in (8’) indicates whether the smaller estimation sample suffers from selectivity problems. 18 Moreover, for the Netherlands, the sum of the two capital elasticities in this last result (Col. 3 – Table A3, Chapter 6) remains rather small whereas the impact of other explanatory variables that refer to how ICT is used has increased at the same time. 19 For the sake of brevity we omit the Heckman stage results for the productivity and wage equations. Tables are available on request.

176


Table 10.3: Estimation result structural models1 Equation NLD (N = 2015) UK(N = 6384) Est. SE Est. SE(6) Productivity Non-ICT capital per employee 0.242 0.021 0.116 0.011ICT capital per employee 0.080 0.028 0.062 0.026Employment 0.086 0.057 -0.029 0.013Wage rate 0.673 0.044 0.857 0.037Dslpct 0.050 0.045 -0.058 0.033Epurchpct -0.029 0.056 -0.015 0.029Esalespct 0.126 0.059 0.019 0.021Constant 0.878 0.758 Stage of ICT maturity Yes (+) Mills ratio Heckman model 0.332 0.266 -0.177 0.098R2 adjusted 0.399 0.573 (7) non-ICT capital per employee Proxy for total capital inputs per employee 0.629 0.021 1.025 0.001Constant 3.185 0.089 Mills ratio Heckman model 0.147 0.079 0.043 0.006R2 adjusted 0.569 0.996 (8) ICT capital per employee Pcpct 1.462 0.116 1.345 0.064Web 0.350 0.085 0.150 0.071Dsl -0.020 0.082 -0.006 0.046Dslpct 0.367 0.126 0.483 0.069Epurchpct 0.254 0.167 -0.023 0.088Esalespct 0.096 0.150 0.138 0.060Business Integration Yes Yes Constant -0.046 0.229 Mills ratio Heckman model -0.056 0.111 -1.520 0.102R2 adjusted 0.311 0.310 (9) Wage rate Productivity 0.656 0.018 0.389 0.031Wage rate t -1 0.202 0.021 0.393 0.014Dslpct 0.047 0.017 0.052 0.017Constant 0.348 0.071 Mills ratio Heckman model 0.063 0.016 0.134 0.028R2 adjusted 0.764 0.750 1All equations use time and industry dummies; Estimates printed in bold are significant at the 1 % or 5 % level.

A most notable result of the structural modelling approach compared to the result of the single-

equation model (see Tables A2 and A3 from the Appendix to Chapter 6) is that the capital elasticities

increase for the Netherlands, but remain relatively stable for the UK. Taking into account the

177


standard errors of the estimates, this increase appears to be more substantial for the estimate of non-

ICT capital inputs than for ICT capital. Furthermore, the sum of the two elasticities now allows a

better comparison with the capital deepening effect found in growth-accounting studies. However,

the estimate for Broadband intensity use (DSLpct) becomes insignificant, whereas the estimates for

the two Ecommerce variables are fairly comparable to the results of the single equation approach,

with again only Esalespct significantly different from zero.

Looking next at the ICT capital prediction equation of the model, it can be seen that the most

important EC variables are also the most important determinants for explaining differences in ICT

capital stocks per employee. The estimate for the share of PC enabled personal (PCpct) is sizable and

highly significant, for both countries. The same conclusion applies to the estimates for the business

integration indicators and for broadband connectivity (DSLpct). This last result suggests that the

productivity contribution of broadband use is channelled through ICT capital deepening, rather than

through TFP.20

The structural system of equations reveals that the impact of DSLpct on productivity is routed

through its indirect effect on ICT capital. This leads us to conclude that ‘investment’ in DSLpct is a

form of ICT capital deepening. However, we draw attention to the fact that the significant coefficient

on DSLpct in the ICT capital prediction equation represents a form of ICT capital deepening over and

above that already captured by PCpct. In Chapter 12, we will see how DSLpct plays an integral role

as an innovation input and how firms achieve a higher degree of knowledge sharing and integration

with external parties through improvements in DSLpct. This suggests that DSLpct is in fact capturing

a deepening of knowledge capital.

Once the ICT capital stock has been quality adjusted for this deepening in knowledge capital, the

direct impact of DSLpct is no longer significant. These results support the idea that improvements in

DSLpct should be viewed primarily as improvements in a labour augmenting technology, with a

secondary impact on overall TFP. Secondly, the results prescribe the use of DSLpct as a proxy for

knowledge capital when the IT capital stock under consideration does not account for this particular

aspect.

By contrast, the estimate for electronic sales (Esalespct) in the ICT capital equation is insignificant

for the Netherlands, and although it is positive and significant for the UK, other variables seem to

have a bigger impact on ICT capital deepening. Comparing this result with the corresponding

estimate of the productivity equation, this suggests that the positive productivity impact of electronic

sales found in the productivity equation represents a TFP effect primarily.

The simultaneous model uses the wage rate as a control for skill differences in the productivity

equation and uses a wage equation to account for joint endogeneity of skills and productivity and the

possible reverse causality for broadband adoption and productivity. The results of Table 10.4 show, 20 The estimate of DSL is in the ICT capital equation is insignificant. Which may not be very surprising as its impact is captured by broadband intensity use.

178


as one could expect, that wages (skills) are highly correlated both with productivity and broadband

intensity use. This last result confirms the existence of computer wage premiums found in earlier

research (see e.g. Krueger, 1993, and Muysken et al., 2006).

Table 10.4: Estimation result structural models after using predicted capital stocks1 Equation NLD (N = 6016) UK (N = 9645) Est. SE Est. SE(6) Productivity Non-ICT capital per employee 0.181 0.01 0.128 0.006ICT capital per employee 0.068 0.01 0.061 0.005Employment 0.006 0.00 0.019 0.015Wage rate 0.831 0.01 0.812 0.013Dslpct 0.008 0.01 -0.000 0.027Epurchpct -0.021 0.03 -0.015 0.041Esalespct 0.112 0.03 0.079 0.027Constant 0.572 0.04 Stage of ICT maturity Yes Yes R2 adjusted 0.601 0.574 1All equations use time and industry dummies; Estimates printed in bold are significant at the 1 % or 5 % level..

10.5.3 Imputing values for missing capital stock data

An interesting question of this research project concerns the predictive power of EC variables for

estimating missing data on (cumulated) ICT investment or ICT capital stocks. This has been

investigated further with the help of the estimates of the ICT capital prediction equation (8’). Recall

that capital stock data were available for 3852 Dutch firms. However, as a result of using lagged

wages, the structural model could only be estimated on the data of 2015 Dutch firms.21 A complete

set of predictors (covering all equations of the structural model) is available for 6016 Dutch firms.

Similarly the structural model could only be estimated using 6384 UK observations, but a complete

set of predictors is available for 9645 UK observations.

Thus, the capital equations of the model can be used for the imputation of ICT and non-ICT capital

stocks per employee for 2164 Dutch and 3261 UK firms. Notice that these observations are imputed

once corrections have been made for selection bias and the endogeneity between skills, ICT use and

productivity. Because these stocks have been cleaned, productivity assessments that use the new

imputed stocks can now be made through a single equation specification.

21 This loss of data illustrates that many (small) firms are not sampled consecutively in the Ecommerce and PS surveys.

179


To assess the predictive power of the model we use the results of Table 10.3 for the imputation of

missing capital stock data (ICT and non-ICT) and re-estimate the productivity equation (6) using the

larger sample of 6016 Dutch and 9465 UK observations.

The results of this exercise are presented in Table 10.4. A comparison of Table 10.4 and 10.3

shows the predictive power of the model. The elasticity of the new ICT capital stock remains robust

to the inclusion of the originally missing observations and at around 6% is strikingly similar for both

countries. For reasons mentioned above, the coefficient on DSLpct is no longer significant in either

set of data. However, the impact of electronic sales, although slightly higher in the Netherlands, is

also remarkably similar for both countries and in the range of 8%-11%.

This last result underlines that combining ICT use variables collected in the Ecommerce survey

and proxy measures for (total) capital inputs in a structural modelling framework can be seen as a

useful approach for predicting missing ICT capital stocks, as well as for identifying the productivity

contributions of the different forms of ICT use and capital deepening.

10.5.4 Comparing reduced-form impacts with results from simultaneous models

Taking stock, and using the estimates of the various models, we can look at the direct and indirect

effect of increasing several ICT use variables. From the continuously measured EC variables of

interest Epurchpct remains insignificant in all models. For this reason, we restrict the discussion to

DSLpct and Esalespct. Table 10.5 summarizes the results for an increase of DSLpct of 0.25 and

Esalespct of 0.05.22

In the single-equation (OLS) approach of the Appendix to Chapter 6 there is no distinction

between a direct and indirect effect. Here the estimates for DSLpct and Esalespct can only be

interpreted as the contribution of broadband connectivity to TFP. The table shows that these

contributions can be very substantial for DSLpct, ranging from 1.6% to 2.1% for the Netherlands, and

0.95% to 1.4% for the UK. However, using the system estimates, it is shown that this conclusion is

invalid. This is so because the estimate of DSLpct in the single-equation approach captures the effect

of ICT capital deepening if no information concerning the composition of capital stocks is taken into

account in the estimation procedure.23 The indirect effect of DSLpct on ICT capital deepening is

surprisingly similar for both countries and in the range of 0.7%.

By contrast, and irrespective of using a simultaneous or single-equation approach, it is shown that

using ICT for electronic sales lead to higher TFP-levels. Again the impact of electronic sales on TFP

is surprisingly similar for both countries and in the range of 0.5%. For the Netherlands, electronic

22This amounts to a 50% increase of broadband intensity use (say from 0.5 to 0.75) and a doubling of the relative importance of electronic sales (from 0.05 to 0.10). 23 The OLS results for fast internet connectivity in the Appendix of Chapter 6 are almost the same as obtained in the core analysis and after using wages to control for a skill related technological bias.

180


sales directly impacts TFP, whereas in the UK a very small but significant proportion is channelled

through ICT capital deepening.

Table 10.5a: Direct and indirect contributions to productivity (NLD)1

Broadband connectivity(DSLpct) α(5) α(2) η(5) α(2).η(5) ∆X CD TFP TotalOLS-12 0.063 0.25 1.6 1.6OLS-23 0.082 0.25 2.1 2.13-SLS N = 2015 0.050(ns) 0.080 0.367 0.029 0.25 0.7 ns 0.73-SLS N = 6016 0.008(ns) 0.068 0.367 0.025 0.25 0.6 ns 0.6 Electronic sales(Esalespct) α(6) α(2) η(6) α(2).η(6) ∆X CD TFP TotalOLS-12 0.148 0.05 0.8 0.8OLS-23 0.165 0.05 0.8 0.83-SLS N = 2015 0.126 0.080 0.096(ns) 0.000 0.05 ns 0.6 0.63-SLS N = 6016 0.112 0.068 0.096(ns) 0.000 0.05 ns 0.6 0.61 ns denotes not significant at the 1%, 5% or 10% level; 2 OLS model that does not account for differences in the composition of capital stocks; 3 OLS model that uses data on the composition of capital stocks but ignores the simultaneity of wages and productivity. Results taken from appendix of Chapter 6. 4 Based on the single equation OLS estimation of productivity using predicted capital stocks from the 3-SLS estimation technique. Results taken from Table 10.4.

Table 10.5b: Direct and indirect contributions to productivity (UK)1

Broadband connectivity(DSLpct) α(5) α(2) η(5) α(2).η(5) ∆X CD TFP TotalOLS-12 0.055 0.25 1.4 1.4OLS-23 0.038(ns) 0.25 0.95(ns) 0.95(ns)3-SLS N = 6384 -0.05(ns) 0.062 0.483 0.029 0.25 0.75 ns 0.753-SLS N = 96454 -0.00(ns) 0.061 0.483 0.029 0.25 0.75 ns 0.75 Electronic sales(Esalespct) α(6) α(2) η(6) α(2).η(6) ∆X CD TFP Total OLS-12 0.138 0.05 0.7 0.7OLS-23 0.128 0.05 0.65 0.653-SLS N = 6384 0.019(ns) 0.062 0.138 0.008 0.05 0.04 ns 0.043-SLS N = 96454 0.079 0.061 0.138 0.008 0.05 0.04 0.4 0.451 ns denotes not significant at the 1%, 5% or 10% level; 2 OLS model that does not account for differences in the composition of capital stocks; 3 OLS model that uses data on the composition of capital stocks but ignores the simultaneity of wages and productivity. Results taken from appendix of Chapter 6. 4 Based on the single equation OLS estimation of productivity using predicted capital stocks from the 3-SLS estimation technique. Results taken from Table 10.4.

181


Taking into account simultaneity and joint endogeneity, in both countries, we obtain similar results

(close to 0.5%) for the contribution of both forms of ICT use to productivity. Whereas the impact of

DSLpct is primarily channelled through ICT capital deepening, electronic sales directly impacts TFP

10.6. Conclusions This Chapter elaborates on the use of ICT capital stock data for improving our understanding of

the contribution of EC variables to productivity. A structural modelling approach has been compared

with the results of reduced-form single equation estimation in order to disentangle the contribution of

EC variables to ICT capital deepening and TFP. ICT can raise productivity either by (ICT) capital

deepening or more directly via improving TFP. The variables collected in the Harmonized

Ecommerce Surveys may point in both directions. Some of the variables seem to be closely related or

complementary to ICT investment, others focus on how ICT is used and assessing their role calls for a

simultaneous approach which models their relation with ICT investment as well as productivity. As a

by-product we can also investigate the predictive power of EC variables for explaining differences in

the stock of ICT capital per employee.

The results show that using a simultaneous model makes sense. This model yields plausible values

for the elasticities of the estimated capital stock. Moreover, it is shown that some variables are more

closely linked to ICT investment than others. For instance, enabling more employees with broadband

connections seems to assert a productivity impact via capital deepening, because – after controlling

for endogeneity and reverse causality – it is insignificant in the productivity equation, but very

significant in the ICT capital equation. In conjunction with the results from Chapter 12, we posit that

the capital deepening captured by high-speed broadband connections is in fact deepening in

knowledge capital.

By contrast, the E-commerce variables are not important for explaining differences in ICT capital

per employee. Nevertheless, an increase in electronic sales appears to have a strong and significant

contribution to labour productivity via TFP, a result that appears to be rather robust to the various

specifications applied. Another robust result is that no productivity impact of electronic buying could

be found, irrespective of the model used. The level of ICT integration also seems to be more

important for explaining ICT capital stocks than for explaining productivity. Finally, an attempt is

made to estimate ICT capital stocks from the EC variables collected in the Harmonized Surveys.

Simulations are used to show the close correspondence between the distributions of predictions and

actual realizations, suggesting an alternative means to estimating ICT capital stocks (see Appendix C).

182


Appendix A1: Summary statistics for capital stock data for the Netherlands

Table A1 reports selected summary statistics for the Netherlands concerning the construction of

capital inputs used in the econometric part of this Chapter. The unbalanced ‘Capital Stock’ panel

consists of 3852 firms for which capital stock data could be constructed in 2000–2005 and that could

be linked to the EC/SBS panel. In terms of output (value added in 2002) the balanced panel represents

nearly 46% of all firms used in the core analysis. This relatively low coverage ratio is mainly due to

the fact that the smallest firms have low inclusion probabilities and, thus, are not surveyed

consecutively. However, their contribution to aggregate capital stocks appears to be rather stable as

the growth rates for capital inputs are comparable to those found at the aggregate level.

According to Table A1, ICT stocks grew at an average rate of 11.3% per year in the short period

considered, a result that is much higher than for other capital inputs. Furthermore, it can be verified

that the shares of ICT capital in total capital are markedly higher in the trade and service sector than in

manufacturing and construction.

Table A1: Summary statistics for ICT and total capital inputs 2002 2005A Share of ICT in total capital stocks a Manufacturing and Construction 4.0 4.8Wholesale and retail trade 7.6 11.0Business services 5.4 9.7All firms 4.8 6.6 B Annualized growth of capital stocks, 2002–2005a

ICT capital stocks 11.3Total capital stocks 1.1a Calculated on the bases of totals Appendix A2: Summary statistics for capital stock data for the UK

Table A2 reports selected summary statistics for the UK capital stocks used in the econometric

estimations of this Chapter. The UK IT capital stock consists of two individual series that capture

hardware and purchased software capital separately. The overlap between the two series is large and

for years 2000-2005 we have approximately 40,000 firm-year observations for which a total IT capital

stock, i.e. hardware and purchased software, can be constructed. However, only 11717 of these firm-

year observations can be linked to the EC survey.

The overlap shrinks even more once we include total capital stock, which is constructed from

another series of survey returns. Eventually we are left with 6872 observations, for years 2000-2005,

for which we have data on IT and Non-IT capital stock and EC returns.

183


In terms of value added in 2002, these 6872 observations represent nearly 60% of all firms used in

the UK core analysis. According to Table A2, ICT stocks in the UK sample grew at an average rate

of 18.5% per year over the short period considered, compared to growth rate of 10.3% for other

capital inputs. These aggregate growth rates seem to be fairly representative: not only are shares of

ICT capital in total capital higher in the trade and service sectors compared to manufacturing and

construction, but also these shares have been growing apace.24

Table A2: Summary statistics for ICT and total capital inputs 2002 2005A Share of ICT in total capital stocks a Manufacturing and Construction 1.0 1.1Wholesale and retail trade 4.3 5.5Business services 1.7 2.2All firms 2.1 2.7 B Annualized growth of capital stocks, 2002–2005a

ICT capital stocks 18.5Total capital stocks 10.3a Calculated on the bases of totals Appendix B1: Distributions for core variables for the Netherlands

Table B1 presents some descriptive measures for the variables used in this model analyses. Most

of the variables are calculated from the overlap between EC and PS surveys. Thus, these distributions

can be compared with the results of the core analysis. The distribution of the variables that are also

based on the capital stock panel adds to the insight of what is behind the weighted averages of Table

A1. In spite of the fact that computers are everywhere nowadays, there still appears to exist a large

dispersion in their relative importance, measured either by the share of ICT capital in the total capital

stock or by ICT capital stocks per employee.

Of the firms that apply Ecommerce only relatively few appear to have realized substantial degrees

of electronic sales and/or electronic buying. This general picture also emerges for the distributions of

the same variables after matching the EC/PS panel with the Capital Stock panel. However, even

though there are relatively more large firms in this smaller data set, it cannot be concluded that these

firms, that are used in the modelling exercise, were performing significantly better in terms of

productivity or were, on average, significantly more advanced in their ICT use.25

24 In absolute terms the shares of ICT capital in total capital are much lower in the UK than in the Netherlands. This is likely to be a reflection of a greater degree of under-reporting, in the UK, of IT investment relative to investment in other types of capital. 25 For PCPct, Interpct and DSLpct the means of the distributions of the matched EC/PS/Capital stock panel for 2005 were 61%, 41% and 40% respectively. This is lower than the comparable scores of Table 10.1. The mean of Esalespct was 7.2% in 2005, which is slightly higher than the comparable figure of Table 10.1.

184


Table B1: Distribution statistics core variables 2002 and 2005 (NLD) p25 Mean Median p75 I) 2002 (N EC/SBS = 3969) Employment 19 174 44 138 Productivity 34.7 68.1 46.8 64.6 Share ICT capital 1.7 12.1 4.7 14.6 ICT capital per employee x 1000EUR 1.1 5.4 2.9 6.4 WEB 74.7 DSL 49.4 Esales 35.1 Epurch 44.6 Pcpct 20.0 53.1 50.0 98.0 Interpct 5.0 35.2 20.0 60.0 DSLpct 0.0 23.6 0.0 35.0 Epurchpct 0.0 5.1 0.0 1.0 Esalespct 0.0 5.0 0.0 0.0 II) 2005 (N EC/SBS = 3216) Employment 27 230 77 212 Productivity 38.3 72.1 52.1 71.9 Share ICT capital 2.8 15.6 7.5 20.7 ICT capital per employee x 1000EUR 1.1 8.5 5.4 11.2 WEB 90.2 DSL 91.0 Esales 34.4 Epurch 57.3 Pcpct 30.0 61.3 65.0 100.0 Interpct 15.0 46.6 36.0 80.0 DSLpct 10.0 44.0 33.0 80.0 Epurchpct 0.0 5.5 0.0 2.0 Esalespct 0.0 6.0 0.0 0.1

Appendix B2: Distributions for core variables for the UK

Table B2 presents descriptive statistics for the EC variables used in the UK analyses. Firms in the

UK sample are on average much larger than in the Netherlands. However trends in ICT use seem to

be roughly comparable. According to the un-weighted averages, both countries have witnessed a

sharp increase in ICT intensity. Almost all firms in both countries have a website. More striking

differences arise when we consider individual uses of ICT. Growth in high-speed internet use has

been more pronounced in the Netherlands, whereas a much larger proportion of UK firms have

introduced e-sales and e-procurement into their business processes.

185


By 2005, 55% of firms in the sample had e-sales systems in place. Nearly 85% of firms were

operating e-procurement systems.26 Firms using e-commerce systems in 2005 were selling, on

average, 17% through e-sales and procuring almost 20% through e-purchases.27

Similar patterns emerge in the distributions of the core variables after matching the EC/PS panel

and the Capital Stock panel.28 However, this smaller data set is slightly more biased towards larger

firms, and unlike the Netherlands, this does lead to issues of selection bias in the UK data.

Table B2: Distribution statistics core variables 2002 and 2005 (UK) p25 mean median p75 I) 2002 (N EC/SBS = 2847) Employment 150 1520 404 1278 Productivity 17.5 52.9 28.9 47.0 Share ICT capital 0.4 3.2 1.3 3.6 ICT capital per employee x 1000GBP 0.2 2.0 0.6 1.7 WEB 89.7 DSL 63.9 Esales 43.6 Epurch 55.2 Pcpct 20 49.0 45 80 Interpct 5 32.9 20 55 DSLpct 0 25.4 8 40 Epurchpct 0 1.5 0 0.1 Esalespct 0 13.6 0 10 II) 2005 (N EC/SBS = 3587) Employment 179 1594 438 1297 Productivity 17.3 55.7 30.4 51.9 Share ICT capital 0.6 3.9 1.7 4.2 ICT capital per employee x 1000GBP 0.3 2.1 0.8 1.8 WEB 94.8 DSL 83.0 Esales 55.0 Epurch 84.5 Pcpct 25 57.8 60 90 Interpct 12 45.2 38 80 DSLpct 5 37.7 25 70 Epurchpct 0.1 20.3 5 25 Esalespct 0 17.0 0.5 20

26 Compared to 34% and 57%, respectively, for the Netherlands. 27 Compared to 6% and 5.5%, respectively, for the Netherlands. 28 Sample period, 2001-2005, averages for DSLpct, Esalespct and Epurchpct in the EC/PS overlap were 30%, 13.9% and 2.9% respectively. In the EC/PS/IT overlap these were 32%, 14.3% and 5.9% respectively.

186


Table B2 also presents a more detailed distribution of the UK IT capital stock adding more insight

to what is behind the weighted averages of Table A2. The similarities in ICT use intensity between

the two countries contrast sharply with the differences between measures of IT capital stock,

suggesting again that UK investment surveys suffer more from under-reporting of IT investment.

Appendix C: Predicting IT capital stock

One of the key elements of this analysis is the attempt to construct a quality adjusted ICT capital

stock using core ICT use variables from the EC survey. It is shown that using an ICT capital stock

that is quality controlled for high-speed internet and e-sales, amongst other ICT uses, in a productivity

regression diminishes the direct impact of high-speed internet enabling on productivity. However the

indirect impact of DSLPCT on ICT capital stock remains strong and significant.

This raises an interesting concern about the predictive power of the core EC variables in estimating

missing data on ICT capital stocks. In order to establish the stability and usefulness of the predicted

ICT capital stocks we randomly sample 50% of firms repeatedly29 and re-estimate the ICT prediction

equation in a 3-SLS setup. More precisely, this is investigated using the estimates of the ICT capital

equation (8) for predicting ICT capital per employee as follows:

1) The sample of 3852 Dutch observations and 6384 UK observations has been split up in two

equal parts by drawing 100 random samples of 50% of all firms;

2) For each random sample the 3SLS method has been applied, yielding estimates for the ICT

capital per employee model;

3) The estimates of stage 2 were used for out-of-sample predictions of ICT capital employee for

the non-sampled firms;

4) Finally, and for each replication, the mean of predicted and realized values and the relative

difference between predictions and realizations for the non-sampled firms were calculated.

The results of this simulation exercise are summarized in Figures 1 and 2. It can be seen that the

shape of the distributions are reasonably comparable. There are relatively few outliers in the

predicted means and the average prediction of all replications is very close to the same score for the

realized data.30 In general the difference between predicted ICT stocks and realized values lies in a

10% range centred around 0%, for the Netherlands. For the UK, predicted stocks tend to be slightly

lower than the actual returns, but the discrepancy is on average less than 5%. We therefore conclude

that the EC variables used have some validity in explaining missing data on ICT capital stocks.

29 This exercise was undertaken 50 times on the UK data and 100 times on the Netherlands data. 30 Figures 1 & 2 present the distribution of the geometric means of ICT capital per employee, which is lower than the arithmetical means of Tables B1 and B2.

187


Figure 1a: Out-of-sample geometric means of ICT capital per employee - Netherlands

6 6

4 4

2 2

0 0 3.3 3.4 3.5 3.6 3.7 3.3 3.4 3.5 3.6 3.7 data predictions

Figure 1b: Relative difference between predicted and realized values - Netherlands

.15

.1

Density

.05

0 0 5 -10 -5 10

rdiff

188


189

Figure 2a: Out-of-sample geometric means of ICT capital per employee - UK

05

1015

20D

ensit

y

4.14 4.16 4.18 4.2 4.22Data

0

5

10

15

20

25

Den

sity

4 4.02 4.04 4.06 4.08Predictions

Figure 2b: Relative difference between predicted and realized values - UK

0

10

20

30

40

50Den

sity

.94 .95 .96 .97 .98 .99rdiff


Chapter 11 - Offshoring and outsourcing

11.1 Global Dimension Eva Hagsten, Statistics Sweden

11.1.1 Background This chapter adds a global dimension to the Eurostat ICT Impacts Project. Comparing several EU countries based on firm microdata studies is of course already a highly international undertaking, but the interaction between countries, or globalisation, is only dealt with to a limited extent taking into account firm affiliation, ownership and international trade behaviour (partly). The global dimension added here comes in the shape of the firm offshoring (sourcing abroad) behaviour and its effects on productivity as well as its possible links to information technology (IT).

Recently, the firm sourcing activities have been highlighted both by media and researchers. IT has often been mentioned as a facilitator when the firms look out for new markets, rare skills or lower costs of production. Earlier studies on for instance French, British, Irish and Swedish firm data show that both labour demand and productivity can be affected by offshoring.1 The effects on labour demand have mainly been negative, but small, while there are productivity bonuses to receive for certain firms from offshoring of both intermediate goods and services. However, the relationships between IT and offshoring on the one hand and productivity on the other have been poorly investigated.

Data on exports and imports are to a certain degree collected in similar fashions in most countries, allowing offshoring to be equalised with imports. However, many use sample surveys for the international trade in services while the international trade in goods comes closer to total surveys. In some countries there are also different authorities responsible for the datasets needed, impeding linking exercises to the production surveys. Not only the international trade variables may differ or be less easy to deduce at firm level for different countries but also some other variables of importance like IT, skills, ownership and multinational affiliation may suffer from various compositions or lack of overlapping. Therefore it quite early became clear that this dimension had to be pursued via a countrywide firm level approach.

Based on a Cobb Douglas style production function specification initially developed for the Swedish dataset, France and Italy have followed with similar but not identical analyses of the effects on productivity from the offshoring behaviour of firms when IT maturity is taken into account. In the comparison of the results this restriction is important to keep in mind, but earlier studies, as referred to above, have showed that the effects of offshoring on productivity are quite robust over different methods as well as variations in the variables used.

In order to avoid confusion over the extent of comparability the studies performed on dataset from Sweden, Italy and France are presented as separate sections in this chapter. The studies follow directly after this introductory section including also a brief discussion of the results and sought after improvements of statistics. 1 See for instance Aubert and Sillard (2005), Criscuolo and Leaver (2006), Görg, Hanley and Strobl, (2008), and Hagsten, Karpaty and Svanberg (2007).

190


191

11.1.2 Results The analyses show that there are productivity premiums to win from offshoring of intermediate goods for both manufacturing and services firms. However, the effects are somewhat stronger in France and in Italy than in Sweden. France also shows some significant negative effects from offshoring. Meanwhile firm productivity in both Sweden and France only seems to be marginally affected by offshoring to low wage countries, this is of larger importance for Italy. These results can be explained by different structures of industries, various reasons behind offshoring as well as the prevailing phase of economic activity in the country. When taking human capital into account for the Swedish dataset it showed that there are most often high general human capital that affects productivity, but a sub-group of large firms high in IT intensive skills seems to gain more than most in productivity from offshoring of intermediate goods to low wage countries.

The addition of different IT maturity variables to the offshoring regressions gave that both the share of employees with PCs and the share of employees with fast broadband generally receive significant and positive estimates in the offshoring regressions, that is, they affect productivity while buying and selling over the Internet only did so to a certain degree, and in the case of Swedish firm with clear negative effects on productivity. The question can also be posed whether IT maturity has an indirect or direct effect on productivity over the offshoring of intermediate goods. In trying this no proofs of indirect IT maturity effects could be found in any of the countries.

11.1.3 Conclusions The ideal analysis would have covered several IT aspects like capital, human capital and IT usage over all types of firms in connection with their offshoring effects on productivity. Unfortunately, it quite early became clear that severe problems may arise when linking different sample surveys like the international trade in services and the ICT usage and E-commerce survey to the production statistics. This partly follows from the fact that the surveys often are small and designed not to overly burden the firms. In the case of microdata analyses this is devastating. Having many sample surveys to deal with is a huge drawback in itself, bur when the samples also become more or less mutually exclusive the usage of the data sets is drastically reduced. Anyway, this means that only the offshoring of intermediate goods could be dealt with while the offshoring of services, which partly has been made possible by improved technology, could not be investigated further. This is despite the fact that services offshoring could well be expected to show a completely different pattern in its relationship with productivity and IT.



11.2 Offshoring, IT Maturity and Productivity: Evidence from Swedish firm level data

Eva Hagsten, Statistics Sweden Stefan Svanberg, Statistics Sweden

11.2.1 Abstract In this study the firm IT human capital and IT maturity are considered in the examination of the effects on firm productivity from offshoring of services and intermediate goods. The study is based on unbalanced panels of firms active in Sweden during the years 2000-2005. Both services firms and manufacturers gain in productivity from offshoring of services. Similar results come from the offshoring of intermediate goods, but with smaller economic values. The offshoring of intermediate goods to low wage countries give very marginal effects on productivity. When taking type of human capital into account firms tend to gain more in productivity from an increase in the share of employees with high general skills than high IT skills. Moreover, firms high in IT human capital are to a far lesser degree affected in productivity by an increase of the offshoring intensity than those firms high in all human capital. However there seems to be a link between large firms high in IT human capital and offshoring of intermediate goods to low wage countries. The IT maturity variables indicate small or no effects on productivity from the firm employee share of PCs and fast broadband. Nevertheless, the services firms seem to lose in productivity from an increase of the buying and selling activities over the Internet. Unfortunately, due to the sampling methods the IT maturity variables could not be linked to the international trade in services survey.

11.2.2 Introduction Globalisation has been facilitated by improvements in technology among other things, like the development of information technology (IT). Certain international activities of firms are now far easier performed with the aid of IT. This has sparked the debate on what effects the sourcing of production abroad may lead to. Up to now most studies have emphasised the effects on labour demand while fewer have focussed on the effects on productivity. The role played by IT for the effects from offshoring is an additional area not thoroughly examined.

The purpose of this paper is to investigate whether IT is important for the offshoring effects on productivity. In doing this the effects on productivity from offshoring in relation to different types of IT maturity will be studied on a pooled sample of firms for the years 2000 to 2005. This angle arose within the work on the Eurostat IT impacts project, focussing on links between IT and productivity at the firm level.1 Offshoring will be defined as intermediate goods or services inputs purchased from affiliated or independent firms abroad.

In the next section some references are presented together with the results expected from this study. Then a description follows of the methods used as well as the data sources. The results presented include descriptive pieces of information about offshorers in general, the firm IT human capital, the firm IT use and the effects on productivity. Finally there are some concluding remarks.

1 Eurostat project number 49102.2005.017-2006.128.

192


11.2.3 Literature and expectations Criscuolo (2006), Amiti and Wei (2006),and Görg et al (2008) find effects on productivity from offshoring in studies using UK, Irish and US data when equalising offshoring with imports and using the Feenstra et al (1996) measure (imports of the input in question over total purchases). For the UK a positive correlation between offshoring of services and productivity at the firm level was found, while international engagement does not seem to be important. The strongest effects were derived from non-exporting domestic firms. In the Irish case on the other hand robust positive effects on productivity from offshoring of services inputs appeared mainly for internationally engaged firms.

The US study showed effects on productivity from both types of offshoring. Daveri and Jona-Lasinio (2007) find positive relationships between offshoring of intermediate goods and productivity when using a somewhat different approach based on the input-output tables. However, this effect disappears with the Feenstra et al (1996) measure. Ekholm and Hakkala (2006) show significant but small negative effects on labour demand from offshoring of inputs. In Hagsten et al (2007) positive relationships between offshoring and productivity could be seen for both intermediate goods and services. Even the literature on vertical integration touches upon offshoring like Grossman and Helpman (2005), but puts more emphasis on the determinants of subcontracting (search, customizing and incomplete contracts). But of course the reasons behind offshoring can well affect the effects from it.

The last few years the effects on firm or aggregated productivity from IT have been vastly investigated and a concise description and assessment of available methods used are found in Draca et al (2006). They conclude that there seems to be evidence of strong firm level association between information technology and productivity. According to Bartel et al (2005) the final goods producers are more likely to purchase outside services when they increase their reliance on IT. They also found that it is not IT in itself that affects services offshoring but the frequency of its improvements. In a study on UK manufacturing firms Acemoglu et al (2004) point out that downstream vertical integration is more likely when the supplier of inputs is less technology intensive than the producing industry. The effects get even stronger when the supplying industry accounts for a larger fraction of the producer’s costs.

Abramovsky and Griffith (2005) find in their study on UK firm level data that ICT reduces transaction as well as adjustment costs of moving activities outside the firm. More ICT-intensive firms also purchased a larger amount of services on the market and were more likely to offshore than less ICT intensive firms. Ilmakunnas et al (2005) found that non-technical higher education affected productivity positively and stronger than technical ones in Finland during the period 1988 to 1998.

On the background of the literature studied and the knowledge of Sweden being a small open economy used to international trade, with high taxes and strict labour market legislation, no really huge productivity premiums from offshoring could be expected. The reason behind offshoring also needs to be considered, like Mattila and Strandell (2006) point out. Their conclusion is that Swedish firms to a larger extent move their production or source out mainly in order to enlarge the market rather than trying to reduce their costs, at least in the short run. This could slightly differ from results for larger economies like the UK where the cost effectiveness may be a stronger issue. Moreover using the most recent firm data in the analyses can give an indication of the directions of the effects on productivity while the investments in IT human capital and IT maturity of firms may play a role as to tell whether IT has actually facilitated offshoring and hence strengthened productivity. Just like Sweden

193


being a country used to foreign trade the same could be applied on IT. If the level of IT use is high the possible gains in productivity may already have been taken.

11.2.4 Data sources The statistics available for this study originate mainly from the Structural Business Statistics with economic and descriptive data on all firms active in Sweden. Information on level of education comes from the Education Register including all individuals 16-75 years of age and the international trade in services and goods is captured by the International Trade Statistics. The trade in goods data are based on a cut-off survey coming close to a total while the international trade in services is sample based. Deflators used come either from the lowest available industry level or from the National Accounts. The firm use of IT will be derived from the EU-harmonised ICT-usage and e-commerce survey. All data are available at Statistics Sweden. The international trade in services data are used under kind permission from the Riksbank (central bank of Sweden). Due to a change of methods in the collection of data there is a time series break in the micro level international trade in services series between 2002 and 2003 and therefore only data from 2003 onwards will be included in the regression analysis.

Theoretically the input-output tables would have been an alternative data source, but these are not revised often enough to be a serious alternative here.

11.2.5 Methodology Based on a Cobb Douglas production function including capital, skills, intermediates, ownership, and international engagement and controlling for changes over time and differences between industries, estimations were made on pooled unbalanced panels of firms for the period 2000-2005.2 The offshoring variable used is total imports over total purchases of the input in question. This is in line with the measure used by Feenstra and Hansen (1996).

The production function can be expressed as:

1−++==

KSU

IMIMGKSU Sititititititit IMSIMGKSUAQ αααµααααα (1)

where Q is output, total factor productivity, inputs of unskilled labour, skilled labour, capital, intermediate goods and intermediate services of the i:th firm in period t, and where

ititititititit IMSIMGKSUA ,,,,,,

µ is an indicator of returns to scale, where 0>µ indicates increasing returns to scale. The parameters are the same for all firms.3

Dividing both sides of Equation 1 by total employment ititit SUL += labour productivity can

be written as it

it

it LQq = µ

itLαααααititititititit imsimgksuAq IMSIMGKSU= (2)

where u, s, k, img and ims are shares of unskilled and skilled workers (skill intensity) respectively, capital intensity, the intensity with which materials and services are used in the production and L is supposed to capture scale effects.

2 Firm level IT maturity variables based on the e-commerce survey are available from 2001 onwards. 3 If value added is chosen to illustrate the output, the explanatory intermediate variables have to be dropped since the value added per definition excludes those inputs.

194


Total factor productivity can be written as a function of the different components of the knowledge capital stock:

),,,,( ,GOFFS

itMNEit

EXPit

SPit

LEitit FA κκκκκ= (3)

On the sources of knowledge coming from learning ( LE

itκ ) spillovers in general, domestic and international ( SP

itκ ), there is no information. This leads to the assumption that these components of knowledge are the same for all firms. In addition there is information on exporting activities ( EXP

itκ ), and whether or not the firm is a multinational firm (foreign or domestically owned) ( ). Moreover there is information on the intensity with which firms import their intermediate goods and services, labelled offshoring ( ). Since

(share of skilled and unskilled employees) only one of these shares needs to be included in the equation.

MNEitκ

GOFFSit

,κ

1=+ itit us

Firm size, Lit is measured as employment of the i:th firm. By substituting multinational firm and offshoring of services or intermediate goods for Ait in equation 2 the following specification is estimated, with dummies for industry, ,iλ and time, ,tλ added:

( )

1 ,lnlnlnlnln

lnln

187654

3210

<++++++++

+++=

βελλβββββ

ββββ

ittiititititit

Aitititit

Limsimgks

expoffIMNEqit

444444 3444444 21 (4)

where offI denotes offshoring of the intermediate input in question (services or goods),

is a dummy variable for multinational firms (it

itMNE 1=itMNE if firm is multinational) and is an intensity variable illustrating international experience as total exports over total

salesitexp

4. A time dummy controlling for business cycle effects as well as age of the firm and age squared (non-linearity) are also included. Equation 4 is thus an augmented production function where the coefficient estimates on the capture their contribution to the TFP. To this certain variables reflecting the investments in IT or IT maturity can be added, either in the form of dummies or as intensity variables.

itA

11.2.6 Results There have been no dramatic alterations of the international trade flows the last ten years. The major trade partners of Sweden are still other high wage countries like the EU and North America.5 Almost 15 per cent of the firms are active in offshoring of intermediate goods while the exact share of services offshorers is not possible to calculate since the survey is now sample based. However, before 2003 the share amounted to around four per cent. In terms of value added the shares are much higher and amount to circa 60 per cent for the offshorers of intermediate goods and to circa 40 per cent for the services ones.

4 The effect captured by the MNE dummy should reflect superior technology but could also follow by foreigners having better access to markets abroad, resulting in higher productivity (with constant technology). 5 High wage countries include EU15, Norway, Iceland, Switzerland, Canada, the United States, New Zealand, Australia and Japan.

195


Table 11.2.1. Facts about average offshorers in 2005 Unweighted means

Variable All firms Offshorers of services

Offshorers of intermediate goods

Number of full time employees 11.04 337.8 43.8Labour productivity (value added per employee), TSEK 516.0 1253.9 617.8Gross production per employee, TSEK 1210.0 4528.2 1647.6Intermediates per employee, TSEK 711.1 3360.5 1057.0Capital intensity (capital per employee), TSEK 805.5 1426.2 473.7Offshoring intensity (imports/total purchases), per cent 0.08 77.9 2.4Export intensity (exports/value of production), per cent 0.008 0.43 0.04Share of MNEs, per cent 6.2 69.5 24.7 Share of Swedish MNEs, per cent 2.9 28.3 10.3Share of exporters, per cent 11.7 91.4 56.7Share of employees with post upper secondary education, per cent* 15.0 14.5 14.6Share of employees with post upper secondary IT intensive education, per cent** 3.5 4.2

3.9

Share of imports from low wage countries, per cent - 16.0 22.1Note: TSEK means thousand SEK. *The share of employees with post upper secondary education among offshorers of services has changed markedly between the years 2004 and 2005. From being quite stable at the level around 20 per cent it now amounts to slightly under 15 per cent. This could of course follow from structural changes in the industry or firms offshoring, but could as well be a sampling phenomenon. **IT intensive educations here are equalised with post upper secondary educations in mathematics, physics and information technology. Source: Statistics Sweden and the Riksbank The descriptive results show that offshorers in general are firms with higher labour productivity and higher capital intensity. These firms are also more often exporters as well as affiliated to multinational corporations than firms in general. This is particularly true for the services offshorers. Similar results are presented by both Görg et al (2008) and Criscuolo et al (2008). According to the e-commerce survey the Swedish firms are reaching a saturation point in that most firms have PCs and Internet via fast broadband, but the share of employees with direct access is somewhat lower. Less than every second manufacturing employee has gained this access and two thirds of the services firm’s employees. However, the number of firms buying and selling over the Internet is still low as well as the use of firm IT systems, which can follow from either a slow uptake of this in firms or from firms being security conscious and not daring to use outwardly directed systems.

Following the data limitations, the offshoring of intermediate goods and services were treated in separate regressions. The offshoring regressions demonstrate that there are productivity bonuses arising from both offshoring of services and intermediate goods. However, the economic values are quite small, particularly for the offshoring of intermediate goods. A one per cent increase in the offshoring of services intensity led to a 0.4 per cent boost in productivity for both manufacturers and services firms. Even if the changes are marginal, a downward shift in the effects from the offshoring of services may be detectable when using the most recent firm data including the year 2005. Otherwise the significances and the estimates show patterns expected and well recognised from earlier studies. Due to different levels of detail in the data sets only the offshoring of intermediate goods could be separated with respect to low or high wage countries. This showed that the effects on productivity from offshoring to low wage countries could hardly be separated from zero, implying that there are mainly effects from offshoring of intermediate goods to high wage countries. Criscuolo

196


(2006) found similar results when separating the UK offshoring of services with respect to country group.

Since IT developments have been suggested as one of the facilitators of globalisation there might be a risk that something important has been omitted when not taking this into account. Both Abramovsky and Griffith (2005) and Bartel et al (2005) saw a connection between IT and offshoring of business services. Therefore a variable measuring the IT intensity or maturity of the firm should be considered. In theory this could easily be carried out by separating the capital variable in two, and allowing the one represent investments in IT and the other all other investments. The skills variable can be treated in the same way, taking as its point of origin a certain level of education and sorting out those high in IT.

Unfortunately, no microdata on IT investments are available over time for the Swedish economy, but there is good information on type and level of education, so the latter can easily be included in the regressions.

Table 11.2.2. Effects on firm productivity from offshoring of services OLS estimations

Note: Heteroskedasticity-consistent standard errors are shown in brackets, ***, **, *, meaning significant at the one-, five-, ten-percent level respectively. The estimations consist of unbalanced panels including all firms in the sample with at least one employee from the international trade in services sample survey. Unreported time and industry dummies are included in order to control for differences in labour productivity and over the years. Lowest available industry level or National Accounts deflators have been used.

Variables Dependent variable: Log of labour productivity

Offshoring of services 2003-2005

Offshoring of services 2003-2005 including IT skills variable

Industry Manufacturing firms

Services firms

Manufacturing firms

Services firms

Log of offshoring intensity, services 0.004 0.004 0.004 0.003 (0.002)* (0.001)*** (0.002)* (0.001)*** Log of employment 0.018 -0.054 0.018 -0.055 (0.009)** (0.006)*** (0.009)** (0.006)*** Log of capital intensity 0.087 0.106 0.088 0.105 (0.009)*** (0.006)*** (0.009)*** (0.006)*** Log of skill intensity 0.007 0.023 (0.005) (0.004)*** Log of skill intensity, IT 0.001 0.005 (0.009) (0.005) Log of skill intensity, non-IT 0.007 0.023 (0.008) (0.005)*** Swedish MNEs 0.072 0.189 0.073 0.183 (0.031)** (0.026)*** (0.031)** (0.026)*** Foreign MNEs 0.065 0.137 0.064 0.135 (0.032)** (0.024)*** (0.032)** (0.024)*** Exporter 0.005 0.030 0.000 0.027 (0.065) (0.034) (0.064) (0.034) Age 0.006 0.021 0.005 0.020 (0.008) (0.006)*** (0.008) (0.006)*** Age squared 0.000 -0.001 0.000 -0.001 (0.000) (.0000)*** (0.000) (0.000)*** Number of observations 1975 4624 1978 4630 R2 0.40 0.28 0.40 0.28

Source: Statistics Sweden and the Riksbank.

197


Table 11.2.3. Effects on firm productivity from offshoring of intermediate goods OLS estimations

Note: Heteroskedasticity-consistent standard errors are shown in brackets, ***, **, *, meaning significant at the one-, five-, ten-percent level respectively. The estimations consist of unbalanced panels including all firms in the Business Register with at least one employee except those based on the E-commerce survey where firms smaller than ten employees are left out. Unreported time and industry dummies are included in order to control for differences in labour productivity and over the years. Lowest available industry level or National Accounts deflators have been used. As IT maturity 1 the share of employees with Internet access via fast broadband has been used and for the IT maturity 2, the firm shares of selling and buying on the Internet is summed up.


Offshoring of intermediate goods 2000-

2005

Offshoring of intermediate goods

including IT skills 2000-2005

Offshoring of intermediate goods

including IT skills and IT maturity 2001-2005

Offshor-ers in EC

survey

Industry Manufac-turing

Services Manufac-turing

Services Manufac-turing

Services All

Log of offshoring intensity, intermediate goods to low wage countries 0.001 0.001 0.001 0.001 0.001 0.002 0.002 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.001)** (0.001)*** (0.000)*** Log of offshoring intensity , intermediate goods to high wage countries 0.003 0.003 0.003 0.003 0.001 0.001 0.005 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.001) (0.001)** (0.001)*** Log of employment 0.039 0.019 0.039 0.016 0.001 -0.040 -0.017 (0.001)*** (0.001)*** (0.001)*** (0.001)*** (0.005) (0.005)*** (0.005)*** Log of capital intensity 0.105 0.122 0.105 0.121 0.099 0.092 0.102 (0.001)*** (0.001)*** (0.001)*** (0.001)*** (0.005)*** (0.004)*** (0.004)*** Log of skill intensity 0.006 0.017 (0.001)*** (0.000)*** Log of skill intensity, IT 0.004 0.015 0.003 0.004 0.000 (0.001)*** (0.001)*** (0.006) (0.004) (0.005) Log of skill intensity, non-IT 0.004 0.012 0.004 0.012 0.017 (0.001)*** (0.000)*** (0.005) (0.004)*** (0.005)*** Log of firm IT maturity 1 0.001 0.001 0.000 (0.000)*** (0.000)*** (0.000)*** Log of firm IT maturity 2 0.006 -0.007 -0.001 (0.004) (0.003)** (0.004) Swedish MNEs 0.039 0.197 0.038 0.193 0.073 0.144 0.09 (0.006)*** (0.005)*** (0.006)*** (0.005)*** (0.017)*** (0.018)*** (0.016)*** Foreign MNEs 0.072 0.227 0.071 0.224 0.059 0.082 0.048 (0.007)*** (0.005)*** (0.007)*** (0.005)*** (0.018)*** (0.017)*** (0.016)*** Exporter 0.051 0.105 0.051 0.103 -0.021 0.088 0.05 (0.004)*** (0.003)*** (0.004)*** (0.003)*** (0.019) (0.016)*** (0.019)*** Age 0.025 0.038 0.025 0.038 0.012 0.034 0.026 (0.001)*** (0.001)*** (0.001)*** (0.001)*** (0.004)*** (0.004)*** (0.004)*** Age squared -0.001 -0.002 -0.001 -0.002 -0.001 -0.002 -0.001 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Number of observations 236923 578600 236922 578599 5399 6836 6759 R2 0.19 0.23 0.19 0.23 0.35 0.43 0.32

Source: Statistics Sweden

198


Employees high in IT skills are considered to be those with post upper secondary education in mathematics, physics or information technology. The general skills variable show positive estimates in most offshoring equations, implying that education is important for productivity. Splitting the skills variable in two made it clear that among the offshorers of services the effects on productivity mainly came from general high skills in the services firms while all firms offshoring intermediate goods gained from all types of skills, but more strongly so for the services firms even there. Since the skills variable partly lost its significance when separated in two, this suggests that the type of investments in human capital is of importance for firm productivity

Besides the more obvious factors like investments in skills and capital, there are several alternative paths that could be explored trying to describe the IT maturity of a firm. The ICT usage and e-commerce survey which includes several aspects of the firm use of computers and the Internet is one. As an addition to the offshoring model a variable describing the firm share of employees with access to Internet via fast broadband was used. This followed from several attempts to find variables that not only worked logically but also in practice, like the firm share of PC users. Even the possible systematic way of using IT was considered. That is why the firm share of buying and selling activities on the Internet were added up to a second IT maturity variable.

As from 2003 when the international trade in services became a sample survey, this led to severe matching problems, mainly because some sample surveys are more or less mutually exclusive in order not to put too heavy a burden on firms. When trying to match the e-commerce and the international trade in services survey around 80 per cent of the observations were lost. Consequently there is no real meaning in pursuing the effects on productivity from offshoring of services and other variables for this tiny group of firms. This is a clear drawback since particularly the services offshoring could well be expected to have strong links to IT, just like Abramovsky and Griffith (2005) suggest.

Fortunately, the trade in goods survey comes closer to a census, briefly described in Hagsten et al (2007) and the E-commerce survey thus will not suffer more than marginal losses of observations in the matching process. However, other problems can arise partly due to the sample design and partly due to the rate of non-responses. A high rate of non-response may lead to a loss in representativity and can distort regression results.

The regressions including the IT maturity variables reveal that the offshoring of intermediate goods to high wage countries now receives smaller productivity boosts than before or no significant result at all while the offshoring to low wage countries remains steady on an already low level. Several of the estimates in these regressions differ more markedly from the ones carried out on the total number of firms, emphasising the fact that these results may be sliding away from representativity. However, the features come closer to the effects on services firm productivity in the international trade in services data set in that the services firms seem to be the only ones on average to gain in productivity from improvements of the firm share of high skilled employees, but only from general non-IT skills.

The IT maturity variables were found to be partly significant, but the estimates for the broadband access were minimal, possibly implying that the gains from this have already been taken. The services firms were the only ones affected by an increase in buying and selling over Internet, but negatively, something that could be interpreted as this being an area still to explore even for firms high in human capital. It could also follow from an uncertainty over the degree of security when the firm connects itself outside the intra system. The somewhat small or not significant estimates may imply, just like Crespi et al (2007) suggest, that

199


European firms as contrary to American ones have not made enough or the right organisational investments in order to gain fully from the investments in ICT.

The question can also be posed whether there is some kind of interaction between the IT and offshoring variables. To investigate this both variables for offshoring of intermediate goods were interacted with the IT maturity 1 variable. This led to no significant estimates suggesting that IT in this form does not affect productivity over offshoring.

When splitting the E-commerce sample in offshorers versus non-offshorers, no real differences were found between the two groups in the estimation of the IT variables. In other words, IT seems to affect firms independently of their offshoring activities. However, the non-offshorers gained in productivity from an increase in the share of IT intensive employees while the offshorers received a somewhat stronger positive productivity bonus from an increase in the share of general high skilled employees, revealing that these two groups of firms differ in characteristics. The decreasing return to scale is quite strong in the offshorer group, just like the services firms in the samples. This is not surprising since there is an overrepresentation of large firms in the samples as well as there are many large services firms.

In running the equations on subgroups based on skills it appears that for firms high in IT human capital, offshoring of services does not affect productivity at all, while firms high in human capital independent of type received positive estimates.6 Among the firms offshoring intermediate goods to low wage countries only those low in human capital (not reported here) gained in productivity. All firms high in human capital received productivity bonuses from the offshoring of intermediate goods to high wage countries.

The regressions on the international trade in services dataset gave no significant estimates from an increase in the share of employees high in IT skills while all firms except those high in IT human capital were positively affected by an increased share of general non-IT high skills. Estimations on the full set of firms give somewhat different results showing that firms high in human capital are affected only by improvements in IT skills, and particularly so with those firms high in IT human capital.7 Criscuolo (2006) finds no evidence of IT intensive firms (IT capital) having higher marginal returns to offshoring. Unfortunately data on this physical capital are not available in Sweden over time.

Quite striking is the fact that after the matching of the Business Register and the E-commerce survey, firms high in IT human capital generally won more in productivity than other firms did from an increase in the offshoring to low wage countries, while the opposite is true for the offshoring by the same firms to high wages countries. Firms high in general human capital gained slightly in productivity from an increase in the offshoring of intermediate goods to low wage countries but were not affected by offshoring to the high wage countries. This is at odds with earlier results and should be interpreted carefully. Those firms high in human capital lost clearly in productivity from an increase in the buying and selling over the Internet. The firms high in IT human capital did not seem to nurture any relationship between productivity and the IT maturity variables at all.

6 These results can of course partly be driven by the definition chosen for a firm high in skills (or human capital). A firm is considered to be high in skills if at least 10 per cent of their employees hold post upper secondary degrees. 7 Results for firms low in human capital are not fully reported here but are available on request.

200


Table 11.2.4. Effects on firm productivity from offshoring of intermediate goods or services sorted by type of human capital OLS estimations

Note: Heteroskedasticity-consistent standard errors are shown in brackets, ***, **, *, meaning significant at the one-, five-, ten-percent level respectively. The estimations consist of unbalanced panels including all firms with at least one employee except those based on the IT usage and e-commerce survey where firms smaller than ten employees are left out. Unreported time and industry dummies are included in order to control for differences in labour productivity and over the years. Lowest available industry level or National Accounts deflators have been used. As IT maturity 1 the share of employees with Internet via broadband access has been used and for the IT maturity 2, the firm share of buying and selling on the Internet is summed up. HCIT means that at least ten per cent of the firm employees hold post upper secondary IT intensive degrees and HChigh is the share of employees with higher degrees all types.


International Trade in Services

2000-2005

Business register 2000-2005

E-commerce survey 2001-2005

Sub-groups HCIThigh HChigh HCIThigh HChigh HCIThigh HChigh Log of offshoring intensity, services 0.004 0.004

(0.003) (0.002)** Log of offshoring intensity, intermediate goods to low wage countries 0.000 0.000 0.005 0.002 (0.000) (0.000) (0.002)*** (0.001)** Log of offshoring intensity , intermediate goods to high wage countries 0.001 0.002 -0.003 0.001 (0.000)*** (0.000)*** (0.002)* (0.001) Log of employment -0.004 -0.038 0.039 0.028 -0.012 -0.009 (0.013) (0.008)*** (0.003)*** (0.002)*** (0.014) (0.008) Log of capital intensity 0.091 0.111 0.106 0.117 0.108 0.108 (0.012)*** (0.005)*** (0.002)*** (0.001)*** (0.012)*** (0.007)*** Log of skill intensity, IT 0.005 0.008 0.012 0.008 -0.040 0.012 (0.034) (0.008) (0.005)** (0.001)*** (0.033) (0.007) Log of skill intensity, non-IT -0.012 0.024 0.003 0.001 0.023 0.042 (0.014) (0.011)** (0.002) (0.001) (0.016) (0.011)*** Log of firm IT maturity 1 0.000 0.001 (0.000) (0.000)** Log of firm IT maturity 2 -0.007 -0.012 (0.01) (0.006)* Swedish MNEs 0.119 0.152 0.142 0.170 0.174 0.152 (0.049)** (0.032)*** (0.012)*** (0.008)*** (0.046)*** (0.027)*** Foreign MNEs 0.149 0.118 0.260 0.214 0.147 0.098 (0.049)*** (0.030)*** (0.012)*** (0.007)*** (0.05)*** (0.029)*** Exporter 0.011 0.045 0.071 0.104 0.068 0.047 (0.084) (0.051) (0.009)*** (0.005)*** (0.05) (0.029) Age 0.002 0.004 0.039 0.037 0.025 0.036 (0.012) (0.008)*** (0.002)*** (0.001)*** (0.011)** (0.006)*** Age squared 0.000 0.000 -0.002 -0.002 -0.001 -0.001 (0.001) (0.000) (0.000)*** (0.000)*** (0.001)* (0.000)*** Number of observations 1182 2928 62302 173963 1177 3079 R2 0.27 0.29 0.20 0.21 0.33 0.37

Source Statistics Sweden

201


The robustness of the results can of course always be discussed. Some prefer to use dynamic specifications in exploring the effects on productivity from offshoring. This has been tried in earlier work on Swedish offshoring by Hagsten et al (2007) where it was difficult to find suitable instruments. However, the main reason why Ordinary Least Squares are preferred is the short period of time studied. The equations have also been run in the fixed effect form of the Ordinary Least Squares and showed similar patterns and significances as with time dummies for all regressions except those including the IT maturity variables, implying that the detailed industry dummies to a certain degree manage to control also for firm specifics. In the Fixed Effects estimations on the matched Business Register and E-commerce survey the major variables lost their significance giving fuel to the argument that the high rate of non-responses plays tricks with the results. Sometimes it is discussed whether gross production or value added gives the most decent picture of production. Here it is believed that value added to a larger extent covers the sought after processes within the firms. In using value added instead of gross production some purely statistical problem can also be avoided in that the intermediates do not have to be included as explanatory variables.

11.2.7 Conclusions The general estimates of productivity effects from offshoring coincide with the expectations and earlier results, allowing small productivity gains from both offshoring of services and intermediate goods, with a somewhat stronger boost from the offshoring of services. When the year 2005 was added to the panel of firms the effects on productivity from offshoring of services declined slightly. This could imply that the possibility to gain from offshoring is fading but could also be a pure random act.

It is somewhat difficult to untangle exactly how different aspects of IT work their way through the offshoring regressions. Offshoring does not improve productivity for firms high in IT human capital while firms high in all human capital seem to get a boost in productivity. However, firms high in IT human capital in the E-commerce data set get one of the strongest bonuses on productivity from offshoring to low wage countries. This data set, like most sample surveys, has an overrepresentation of large firms revealing not only a tendency to decreasing return to scale but also a stronger relationship between IT and offshoring.

The split of the skills into two variables showed that investments in all human capital will not necessarily improve firm productivity, but rather certain types. In this case the general non-IT higher skills seemed to be most important for productivity, and the services firms gained the most. As an exception to this firms high in IT human capital in the full dataset gained strongly from an increase of IT skills. Similar results were found by Ilmakunnas and Maliranta (2005) for Finnish firms. This also coincides with what Romer (1990) concludes in that not only human capital is important for growth but also what type of education. It should not be forgotten that effects from improved skills on productivity may be delayed in time and are therefore not properly captured in these estimations without lagged variables.

When including the IT maturity variables, small but positive effects or maybe no effects were expected, just like for the offshoring estimates. This was also confirmed for the IT maturity 1 variable illustrating the firm employee share of access to the Internet via fast broadband. However, the IT maturity 2 variable showing the firm share of buying and selling over the Internet did not seem to be related to firm productivity in general but for the services firms and firms high in human capital. They clearly lost in productivity from an increase of e-commerce over the Internet. Since IT generally facilitates different systems this is a bit odd. Maybe this type of system is the upcoming challenge for firms to construct and use efficiently as well as securely or it is something that only the services offshorer, which could

202


203

not be taken into account here, would have gained from. A possible interaction between offshoring and IT maturity was also explored. This gave no proof of IT maturity affecting firm productivity over its offshoring.

It should be emphasized once more that the absent opportunity to study the IT maturity variables together with the international trade in services data is very unfortunate since both literature and the characteristics of some services (weightless) suggest that the links between productivity and offshoring over IT may be far stronger or show another pattern than for the quite weak links found here for the offshoring of intermediate goods.



11.3 Performance, IT maturity and offshoring behaviour of Italian manufacturing corporations - In the dire straits of globalisation

Andrea de Panizza1 Abstract: This study examines productivity and profitability of Italian manufacturing corporations in relation to IT usage and offshoring of intermediate goods. The information set is based on a balanced panel of enterprises' economic accounts and foreign trade statistics for the years 2000-2004, linked to 2002 and 2004 surveys on ICT usage. The analytical framework is similar to one previously developed for Sweden, allowing for (partial) comparability. Offshoring is positively related to productivity, although the significance of intensity variables depends on employment size and industry. The same occurs for some variables of IT maturity (workers using PCs and a composite indicator), and for human resources as proxied by cost of labour (i.e. wage levels). These variables also show a positive impact on profitability, although limited to productions which are easy to outsource. Offshoring decisions and IT maturity, instead, do not present any strong mutual relation. The key issue of the direction of causality between IT maturity, offshoring and productivity is also tentatively addressed: lagged offshoring appears to weakly impact productivity, while lagged IT maturity does not, and a reverse causality from productivity to IT maturity is revealed. This first evidence, albeit limited, challenges some commonplaces, suggesting the coexistence of different business models. A richer information set should allow for a more appropriate treatment of these issues, as well as for extending the analysis to other, crucial determinants of performance.

1 IPTS (Institute for Prospective Technological Studies of the European Commission's Joint Research Centre), and ISTAT (National Statistics Institute of Italy); the views expressed are those of the Author, and do not necessarily coincide with those of parent Institutions.

204


11.3.1 Setting the scene

This paper investigates relationships linking IT maturity and offshoring with productivity and profitability in Italian manufacturing firms. An empirical analysis is performed on micro-data for the years 2000-2004, which correspond to the deployment of the first globalisation-induced crisis hitting the Italian economy.

Indeed, with respect to the other large EU countries, Italy is characterised by strong specialisation in labour intensive/low technology manufacturing, more exposed to price competition from emerging economies. With specific reference to these ‘traditional’ industries, offshoring can be viewed as a channel for the survival of enterprises via cheaper labour (letting home only a few functions), while the motivations of markets conquest and acquisition of technologies are deemed more important in other sectors. In all cases, though, the impact of offshoring on productivity is likely to be mediated by and to go along with other factors, such as labour market arrangements.

In practice, after the business cycle peak of year 2000, the Italian economy underperformed versus both historical records and nearly all EU countries. Unlike in previous crises, employment proved resilient, resulting in a prolonged stagnation of labour productivity, which fell for the first time in manufacturing (Figure 11.3.1).

Figure 11.3.1 – Italy: dynamics of employment, value added and productivity. Yearly % change, 1993-2007

- 2.0

- 1.0

-

1.0

2.0

3.0

Manuf. TOTAL Manuf. TOTAL Manuf. TOTAL Manuf. TOTAL

1996/92 2000/96 2004/2000 2007/04

- 2.0

- 1.0

-

1.0

2.0

3.0

Employment (FTE) Apparent labour productivity value added at factors cost

Enterprises managed to partly compensate for the drop in profits (associated to the decrease in value added), by keeping wages low (see e.g. Tronti, 2007).2 The period 2000-2004, however, was also marked by a strong restructuring and selection process in Italian manufacturing industries, with a decrease of about 6.6 percent in the number of firms, contrasted to a growth of 0.6 percent for the rest of the EU25.3

2 Profitability rates, however, fell with respect to capital, labour and labour costs alike (de Panizza, Calza and Rossetti, forthcoming). 3 The comparatively large number of micro-enterprises in Italy is mirrored in an average employment size much lower than the EU and, correspondingly, swells the country share in the population of EU manufacturing firms up to an astonishing 25 percent. The shrinking in the population of Italian firms, thus, resulted in a reduction of 1.2 percent for the whole of the EU25, which hinders an otherwise positive variation.

205


Here, we address the role of internationalisation and IT usage on the way firm performed in productivity and explore their impact on profitability which, in turn, corresponds to higher chances of survival and development.

The analytical framework is developed starting from that used with respect to Sweden by Hagsten et al (2008), and it is rooted in previous and ongoing work on offshoring.4 In a nutshell, the presence and intensity of offshoring is recorded by tracking the purchase of intermediate goods from abroad, and it is inserted within a Cobb-Douglas type log-linearized production function as an explanatory variable of (apparent labour) productivity, measured as value added per person employed.

The main features of the dataset and descriptive results are presented in Section 11.3.2, key analytical results are discussed in Section 11.3.3, and some concluding remarks are sketched in Section 11.3.4.

11.3.2 Features of the dataset and descriptive analysis

The dataset used in regression analysis consists of 4745 records referring to 3633 enterprises. This results from the merging of 2002 and 2004 ICT usage surveys in enterprises (each with about 10thousand respondents) with a balanced panel reporting information from 2000 to 2004 on economic accounts plus offshoring for about 45 thousand manufacturing corporations (other firms are excluded, due to the lack of economic accounts). Issues of causal direction and of changes along time are addressed by means of a derived (core) dataset, which includes 1144 enterprises responding to both ICT surveys, 1551 hit in 2002 only and 938 in 2004 only. Lacking direct information on the quality of human resources, average labour cost per person employed was used as a proxy. The analysis is limited to the purchase of intermediate goods in manufacturing, as no comparable data on services were available.

Due to the mix of these features the dataset, and all the more so its core subset, are clearly biased towards the upper end of the employment size distribution of firms, and (partly in relation to the above) distorted with respect to industrial composition. The smallest firms are outside the observation field of ICT surveys, and corporations are typically more robust and 'modern' than unincorporated firms of the same size. Hence, full representativeness of the industrial system could not be achieved, even by weighting regressions, which we do not do here. The issue, however, is not to be regarded as negative. Although including only the 0.65 percent of firms, the dataset covers about 15% of employment and more than one fifth of turnover in manufacturing. Basically, firms portrayed in the dataset are those driving the whole of manufacturing. Indeed, with respect to the universe of manufacturing enterprises with 10+ persons employed these firms are relatively more productive (across all industries), and score higher in both IT maturity and offshoring intensity. On this latter aspect, we can contrast them only with the original panel in which, however, about 30 per cent of firms were to some extent engaged in offshoring (Figure 11.3.2).

4 First proposed by Hagsten and Karpaty (2006), partly followed in de Panizza, Calza and Rossetti (2007), and ibid. (forthcoming).

206


Figure 11.3.2 - Features of the merged dataset: A. Coverage and economic ratios

0%

5%

10%

15%

20%

25%

2000 2004 2000 2004

panel+ICT core (ICT 02 & 04)

p.emp. (rhs) App.Lab.Prod.

SAMPLE COVERAGE (% of Universe)

1.1

1.2

1.3

2000 2004 2000 2004

panel+ICT core (ICT 02 & 04)

0

10

20

30

40

App.Lab.Prod. Lab.cost p p.e. p.emp. (rhs)

RATIOS vs. UNIVERSE

Legend: persons employed, apparent labour productivity and labour costs per person employed

B. IT Maturity and Offshoring behaviours

0

10

20

30

40

50

60

70

80

2002 2004 2002 2004 2002 2004

Universe panel+ICT core

pcpct dslIT MATURITY

0

5

10

15

20

25

30

2002 2004 2002 2004 2002 2004

PANEL panel+ICT core

high income

low income

OFFSHORING

Legend: percentage of employees using a Pc, availability of xDsl connection;

Impacts of employment size and of industry specificities are addressed by including a variable for employment, and by treating separately four groups of industries according to a Pavitt-like taxonomy.5 The latter allows broadly distinguishing industries according to their patterns with respect to productivity, market dynamics and some behavioural features. In addition, NACE four digits dummies are used within each group, as well as dummies for geographical location and multinationality of enterprises.

As expected, offshoring intensity is clearly dependent on employment size of firms and on their main industry, so that large firms and those operating in hi-tech industries and in sectors characterised by economies of scale rank comparatively high. When we consider

5 Following Pavitt (1984), economic sectors can be grouped into four clusters according to their features with respect to innovation & knowledge. Hereunder a slight departure from the original taxonomy is adopted, with an eye to market dynamics, as follows: "Traditional" industries (in Pavitt's phrasing, Supplier dominated), including food and beverages, textiles and apparel, footwear, paper and printing, and wood industry, plus, in this paper, also furniture, metal works, and non metal mineral products; "specialised suppliers" including machinery and equipment plus part of the electronic industry; science based, including IT, pharmaceuticals, and aeronautics; plus "scale intensive" productions. The latter present some similarities with traditional industries, in as cost reduction is a key objective of innovation, which also has a relatively low degree of appropriability. These similarities (and, by contrast, those between the two other groups), are quite evident also in the results of the analysis hereunder.

207


localisation, it is also evident that firms in traditional low-tech sectors exposed to competition of emerging economies are leading in offshoring to low income countries (Figure 11.3.3).

Figure 11.3.3 –Industry and employment size vs. intensity of offshoring to high and low income countries

0

5

10

15

20

25

30

35

Trad

itiona

l

Spec

. sup

pl.

Scie

nce

base

d

Ec.o

f sca

le

TOTA

L

Trad

itiona

l

Spec

. sup

pl.

Scie

nce

base

d

Ec.o

f sca

le

TOTA

L

Trad

itiona

l

Spec

. sup

pl.

Scie

nce

base

d

Ec.o

f sca

le

TOTA

L

<50 50 -249 250+

high

low

60

65

Trad

itiona

l

Spec

.sup

pl.

Sc. B

ased

Ec.o

f sca

le

TOTA

L

All f irms(unw eighted)

With respect to IT maturity, instead, things are not straightforward, as larger firms are likely to score higher when variables such as DSL are considered, but also to have comparatively larger shares of blue-collars, so that no differences can be traced with respect to the percentage of workers using PCs (see above, Figure 11.3.2B). With respect to this latter aspect, instead, a clear sectoral hierarchy emerges, with high tech industries far at the top and traditional ones detached at the bottom (Figure 11.3.4)

Figure 11.3.4 – Impact of Industry on IT maturity: percentage of workers using PC, year 2004

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

TRAD SPECIALISEDSUPPLIER

SC.BASED SCALE EC. TOTAL

pcpct

Variables available from ICT usage surveys present obvious overlapping, and only some of them are able to discriminate among enterprises and prima facie present a direct relationship with performance (Table 11.3.1). Amongst these, the percentage of workers using PCs proved clearly superior (a finding in line with a previous work on macroeconomic performance at the international level, see de Panizza and Visaggio, 2007), followed by xDSL and intranet usage. These facts respond to the logical hierarchy among usage

208


indicators, partly visible in cross-correlations. A composite indicator of IT maturity obtained by summing all variables6 shows the highest correlation with productivity. However, the additional explanatory power of this and other composite indicators tested (in particular, focusing on PC usage and DSL) resulted really minimal, so that, also for presentation purposes, individual variables are used in regression analysis hereunder. These stylised facts are exemplified in Table 11.3.1 for the whole dataset, without considering sector/size specificities (see above) and time differences (for instance, in 2004 pc usage proved more significant than the composite for some size classes and industries). It is also worth noting that profitability (defined as gross profits per unit of capital) is mildly related to productivity, while IT maturity variables do not seem to have any impact on it.

Table 11.3.1 – simple correlations among productivity, profitability and IT maturity variables

logp

rod

Prof

itabi

lity

IT m

atur

ity

pc u

sage

%

Inte

rnet

Inte

rnet

%

xDSL

intra

net

Intra

net %

web

Epur

chas

es

Epur

chas

es%

Esal

es

Esal

es%

Ln(productivity) 1.00 Ln(profitability) 0.29 1.00 IT (composite) 0.39 0.02 1.00 pc usage % 0.34 0.08 0.69 1.00 Internet 0.17 -0.05 0.36 0.17 1.00 Internet% 0.24 0.08 0.63 0.70 0.23 1.00 xDSL 0.28 -0.01 0.72 0.27 0.24 0.26 1.00 Intranet 0.25 -0.01 0.53 0.27 0.22 0.22 0.35 1.00 Intranet% 0.33 0.05 0.72 0.70 0.17 0.53 0.35 0.74 1.00 web site 0.17 -0.04 0.63 0.19 0.36 0.20 0.25 0.24 0.22 1.00 Epurchases 0.11 0.00 0.29 0.16 0.10 0.15 0.22 0.20 0.21 0.16 1.00 Epurchases % 0.05 0.01 0.16 0.06 0.03 0.05 0.10 0.09 0.09 0.06 0.31 1.00 Esales 0.13 0.01 0.25 0.10 0.09 0.09 0.19 0.15 0.15 0.13 0.21 0.18 1.00 Esales% 0.11 0.02 0.23 0.09 0.05 0.06 0.15 0.13 0.13 0.06 0.14 0.18 0.60 1.00

11.3.3 Linking behaviour with performance: results of regression analysis

Italian foreign trade stagnated for most of the period under exam, with respect to both exports and imports. Considering the 45 thousand firms in the base panel, their percentage involved in imports of intermediates grew marginally, but showed some changes in the pattern of import intensity, with a rebalancing from high to low income countries and some related sectoral differences.

Delocalisation became more important for corporations operating in traditional industries (apparel, shoes, iron works, tiles, and furniture). Firms in other sectors either slightly reduced their internationalisation overall or, in industries with positive returns to scale, they increased it marginally and, again, mainly in emerging economies.

Initially, results obtained with a basic model (K, L, human resources, offshoring & IT maturity) were overall quite similar to those obtained for Sweden, both in terms of the model overall explanatory power (about 30% of variability) and to some of its key components. In particular, and as expected, human resources and capital intensity were the most relevant factors, although the influence of capital intensity on productivity is not significant in both

6 IT_Maturity = (Pc*pcpct+inter*interpct+dsl+web+intra*intrapct+epurch*epurchpct+esales*esalespct).

209


science based and scale intensive industries. Offshoring too proved weak, in particular when to low income countries.

A marked improvement on the overall explanatory power (to values approaching 60%) was obtained by excluding from the analysis the few cases for which (log) productivity was not positive, but with no significant changes on the significance of individual variables (Table 11.3.2).7

Table 11.3.2 - impacts on Productivity from production mix, offshoring and IT maturity (years 2002 & 2004):

Dependent variable: Ln of labour productivity TOTAL TRADITIONAL SPEC.SUPPLIERS SC.BASED SCALE

INTENSIVE Business size (Ln of Employment) -0.04 (0.008)*** -0.03 (0.01)*** -0.02 (0.013) -0.09 (0.037)** -0.05 (0.012)***Human Resources (Ln of Labour cost p.p.e.) 0.89 (0.082)*** 1.05 (0.05)*** 0.75 (0.057)*** 1.04 (0.113)*** 0.95 (0.044)***Capital intensity (Ln of Capital p.p.e.) 0.09 (0.007)*** 0.10 (0.009)*** 0.10 (0.015)*** 0.04 (0.023)* 0.08 (0.012)***Vert.Integr. (Share of VA on turnover) 0.19 (0.028)*** 0.24 (0.041)*** 0.07 (0.039)* 0.31 (0.16)* 0.21 (0.042)***

% employees w/pc 0.16 (0.032)*** 0.21 (0.05)*** -0.03 (0.075) -0.15 (0.149) 0.14 (0.044)***% emp. w/internet 0.02 (0.04) 0.04 (0.074) 0.14 (0.083)* 0.15 (0.134) -0.07 (0.065) Dsl (y/n) 0.01 (0.012) -0.01 (0.019) 0.03 (0.028) 0.01 (0.065) 0.01 (0.019) E-sales (y/n) 0.02 (0.017) 0.03 (0.027) 0.09 (0.04)** 0.12 (0.128) -0.01 (0.025) E-purchases (y/n) 0.00 (0.015) -0.04 (0.024) 0.02 (0.025) 0.10 (0.057)* 0.02 (0.027) E-sales (% turnv.) 0.02 (0.062) -0.05 (0.14) 0.28 (0.16)* 0.03 (0.308) 0.01 (0.074) IT

MA

TU

RIT

Y

E-purch (% turnv) 0.14 (0.081)* 0.18 (0.13) -0.36 (0.122)*** 0.19 (0.317) 0.24 (0.107)** Low income 0.02 (0.004)*** 0.03 (0.007)*** -0.02 (0.01) -0.00 (0.017) 0.03 (0.008)***OFFSH-

ORING High income 0.03 (0.004)*** 0.02 (0.006)*** 0.02 (0.01)** 0.03 (0.017)* 0.04 (0.007)***North-West 0.01 (0.015) 0.02 (0.023) -0.00 (0.032) -0.06 (0.075) -0.01 (0.028) North-East 0.03 (0.014)** 0.02 (0.021) 0.07 (0.033)** 0.11 (0.064)* 0.00 (0.027)

GEO LOC.

South -0.07 (0.02)*** -0.10 (0.029)*** -0.13 (0.051)** -0.12 (0.097) -- -- MNC (y/n) 0.05 (0.014)*** 0.03 (0.022) 0.01 (0.03) 0.10 (0.063) 0.05 (0.019)***Year =2002 0.04 (0.011)*** 0.06 (0.017)*** 0.02 (0.024) -0.02 (0.045) 0.06 (0.017)***No. 4745 2225 723 256 1541 R2 0.647 0.655 0.695 0.674 0.616 Note: Heteroskedasticity-consistent standard errors in brackets; ***, **, * = variables significant at the 1, 5 and 10 percent levels, respectively. Unreported four-digit industry dummies always included. All firms have at least 10 persons employed and apparent labour productivity > 0 & <500k€

Business size usually is positively related to productivity, but here the impact results significantly negative, overall and for most industries, once controls for other aspects are introduced. This is not surprising, in as smaller firms in the sample are likely to be 'the best' in their class. Our proxy for Human resources (labour cost per person employed), instead, has a strong positive impact across all industries, and Capital intensity too drives productivity, although the evidence appears weaker for the group of science based industries.

Both types of offshoring overall were significant at the 1% level. However, the positive impact on productivity of offshoring to low income countries is significant only for traditional and scale intensive productions – thus confirming common sense evidence – while that of offshoring to high income countries is significant in all industries, though only to a limited extent for science based ones.

Amongst IT maturity variables, as anticipated, only the percentage of employees with a Pc is significant overall (at the 1% level). The latter, however, does not show any clear impact for science based and specialised supply industries. E-purchases variables show a mixed 7 Apparent labour productivity can be negative, due to its definition.

210


impact across industries. The intensity variable (% of turnover) presents a positive association with productivity in industries where economies of scale are relevant (ability to trade inputs globally) and negative in specialised supply industries, who are also the only group for which E-sales show a positive and significant impact. These results are in line with the evidence suggested in a previous study for manufacturing of machinery and equipment (Becchetti, de Panizza and Oropallo, 2007), that in this specific industry sub-contractors are not necessarily on the lowest ladder of the value chain. xDsl, finally, now does not seem to be relevant for any type of business, as this variable does not discriminate in our sample, where pioneers are overrepresented with respect to the whole economy.

The introduction of a variable for vertical integration (share of value added on turnover, i.e. the complement of overall purchases of intermediate goods & services) also proved to impact productivity, and contributed to differentiate results for offshoring to low income countries. Finally, the introduction of dummies for multinationals, for (NUTS1) geographical location of enterprises and for year of reference improved results only marginally with respect to basic regressions (not reported). Nonetheless, all these variables were overall significant, with MNCs and enterprises located in northern regions clearly more productive than enterprises with only local branches and/or located in the South, and framework conditions easier in 2002 than in 2004.

A highly tentative exercise to address the issue of causality was carried out on the core sub sample of corporations responding to both 2002 and 2004 ICT surveys and, where possible, on the larger sample of respondents to the 2004 surveys for which we have economic accounts for the previous years. Some basic regressions were performed for 2004, with the aim of checking whether lagged variables had any impact. An important limitation to the interpretation of results arises from the fact that, in the period under exam, only IT maturity (pc usage) moved fast, while offshoring progressed little and, in aggregate terms, productivity in 2004 slowed down to the same level of 2002.

A first test, addressing the IT maturity - productivity nexus, showed that, controlling for other variables, lagged (2002) PC usage and composite IT usage proved to be irrelevant in explaining productivity, while lagged (2002) productivity was significant, although weakly, in determining PC usage (table 11.3.3).

The (absence of) influence of lagged IT maturity on productivity was confirmed also in more detailed sector/size analysis. This result, to be regarded as very preliminary, suggests that the relationship between (historical) IT maturity and productivity is mediated by other aspects, while that from (historical) productivity to IT maturity points at the coexistence of different business models, with those grounded on productivity corresponding to a more dynamic behaviour with respect to IT (productive) usage.

A twin exercise addressing the issue of the direction of causality between offshoring and productivity, instead, reveals that lagged offshoring is weakly significant for productivity (12-14%), while the opposite does not hold (unreported). In other words, results confirm that enterprises transforming themselves through offshoring later on improved their probability to rank high in productivity, while of course it is not at all obvious that firms at the top of productivity would later on, due to this, become offshorers.

211


Table 11.3.3: Test on the direction of causality between Productivity and IT Maturity plus Offshoring Pc Usage intensity (% of workers) & lagged productivity EXPLANATORY VARIABLES

CORE SAMPLE LARGER SAMPLE Ln Productivity 0.06 (0.03)** 0.07 (0.02)*** Ln Productivity 2002 0.02 (0.03) 0.02 (0.01)* Ln employment 0.01 (0.01) 0.00 (0.01) Ln Capital 0.00 (0.01) 0.00 (0) Ln Labour cost 0.05 (0.04) 0.06 (0.04) Ln Vertical integration

-0.07 (0.02)*** -

0.08 (0.01)*** Ln Offshoring (total) 0.00 (0.01) 0.00 (0) Multinationality 0.03 (0.01)** 0 (0.01)*** No 1134 2060 R squared 0.489 0.468 Productivity & Lagged IT maturity & Lagged Offshoring IT Maturity

0.27 (0.07)*** 0.22 (0.08)*** 0.31 (0.06)***

IT Maturity 2002 -0.05 (0.06) … … Ln employment -0.02 (0.01)** -0.03 (0.01)** -0.02 (0.01) Ln Capital

0.08 (0.01)*** 0.08 (0.01)*** 0.09 (0.01)***

Ln Labour cost 0.76

(0.08)*** 0.76 (0.08)*** 0.80 (0.09)***

Ln Vertical integration 0.19

(0.04)*** 0.19 (0.04)*** 0.23 (0.04)***

Ln Offshoring (total) 0.05

(0.01)*** 0.04 (0.01)*** 0.03 (0.01)**

Ln Offshoring (total) 2002 0.02 (0.02) 0.02 (0.01) o Multinationality 0.02 (0.03) 0.02 (0.03) 0.02 (0.02) No 1134 1134 2060 R squared 0.728 0.730 0.686

Note: only enterprises with productivity [0, 500thousand €]. Sectoral & geographical dummies always included, as well as other IT variables

It is worth adding that, in both exercises, the significance of lagged (T-2) offshoring (vs. productivity) and productivity (vs. IT maturity) was higher in larger samples, where smaller firms weight more, and was further improved when current variables were excluded, using T-1 values instead.

A similar question was addressed with respect to offshoring: can we infer an influence of IT maturity in offshoring decisions (or vice versa)? Regression results reported in Table 4 show no clear impact, in both directions. Among other variables, human resources, multinationality and (negative) vertical integration are relevant in both cases, while business size is significant for offshoring decisions, but does not show any significant impact on IT maturity.

212


Table 11.3.4: Cross relationships between offshoring & IT maturity in 2004

(Including lagged variables) OFFSHORING Betas St.err. IT MATURITY Betas St.err. p.empl. using a pc (%) 0.12 (0.29) Total Offshoring 0.00 (0.01) p.empl. using a pc (%)2002 0.29 (0.22) Total offshoring 2003 0.00 (0.01) Total offshoring 2002 0.01 (0.01) Business size (Ln of Employment)

0.38 (0.04)*** >> 0.00 (0.01)

Human Resources (Ln of Labour cost p. p.e.)

0.34 (0.13)*** >> 0.10 (0.04)***

Capital intensity (Ln of Capital p.p.e.)

0.15 (0.05)*** >> 0.01 (0.01)

Vert.Integr. (Share of VA on turnover)

-0.56 (0.1)*** >> -0.06 (0.02)***

Multinationality 0.20 (0.1)** >> 0.04 (0.01)*** Localisation / North 0.12 (0.17) >> -0.01 (0.02) Localisation / North East 0.19 (0.17) >> 0.04 (0.02)* Localisation / Centre -0.05 (0.19) >> 0.02 (0.03) No 1134 Rsq. 0.523 >> 0.481

Note: dependent variables = total offshoring; % of persons employed using a pc

Finally, it is worth discussing whether IT maturity and offshoring do have any impact on the profitability of firms which is what counts for the entrepreneur and makes the firm survive. To this end, profitability (gross profit per unit of capital) was estimated by means of the previous set of predictors, with mixed results (Table 11.3.5). Table 5: impacts on Profitability from production mix, offshoring and IT maturity (years 2002 & 2004):

TOTAL TRADITIONAL SPEC.SUPPLIERS SC.BASED SCALE INTENSIVE

Business size (Ln of Employment) -0.04 (0.01)*** -0.05 (0.01)*** -0.02 (0.02) -0.07 (0.03)** -0.04 (0.01)***Human Resources (Ln of Labour cost p. p.e.) 0.32 (0.09)*** 0.50 (0.05)*** 0.13 (0.05)** 0.49 (0.19)** 0.45 (0.09)***Capital intensity (Ln of Capital p.p.e.) -0.34 (0.02)*** -0.30 (0.02)*** -0.32 (0.02)*** -0.36 (0.05)*** -0.42 (0.06)***Vert.Integr. (Share of VA on turnover) 0.04 (0.02) 0.04 (0.03) 0.05 (0.06) 0.05 (0.11) 0.07 (0.04)*

% emp. w/pc 0.17 (0.04)*** 0.24 (0.05)*** -0.04 (0.12) -0.28 (0.24) 0.13 (0.07)* % emp. w/intern. -0.09 (0.06) -0.08 (0.09) 0.10 (0.14) 0.05 (0.21) -0.14 (0.09) Dsl (y/n) 0.01 (0.02) -0.02 (0.02) 0.05 (0.04) 0.02 (0.08) 0.03 (0.03) Esales (y/n) 0.01 (0.02) 0.02 (0.04) 0.07 (0.06) 0.09 (0.21) -0.01 (0.03) Epurch (y/n) 0.00 (0.02) -0.02 (0.02) 0.01 (0.04) -0.04 (0.11) 0.01 (0.03) Esales (% turnv.) 0.09 (0.1) 0.20 (0.18) 0.41 (0.23)* 0.00 (0.44) -0.01 (0.13) IT

MA

TU

RIT

Y

Epurch (% turnv) 0.12 (0.13) 0.18 (0.17) -0.43 (0.2)** 0.95 (0.89) 0.11 (0.12) Low income 0.02 (0.01)*** 0.01 (0.01)* 0.02 (0.02) -0.02 (0.04) 0.04 (0.01)***OFFSH-

ORING High income 0.06 (0.02)*** 0.04 (0.03) 0.03 (0.04) 0.13 (0.09) 0.06 (0.03)** North-West 0.01 (0.02) 0.03 (0.03) 0.00 (0.05) -0.14 (0.12) -0.09 (0.08) North-East 0.03 (0.02)* 0.03 (0.02) 0.07 (0.05) 0.12 (0.1) -0.08 (0.07)

GEO LOC.

South 0.02 (0.04) 0.01 (0.04) -0.05 (0.06) -0.08 (0.12) -- -- MNC (y/n) 0.02 (0.01)** 0.02 (0.01)** -0.02 (0.02) 0.00 (0.04) 0.02 (0.01)** Year =2002 0.03 (0.02)** 0.04 (0.02)** 0.04 (0.03) -0.10 (0.08) 0.06 (0.03)** No. 4692 2193 715 251 1533 R2 0.469 0.440 0.437 0.458 0.560

213


214

Offshoring has a strong impact on profitability. However, this is due only to scale intensive productions, and to a limited extent to traditional industries for offshoring to low income countries. When it comes to IT maturity, again, the indicator for the intensity of PC usage is the only whose impact is overall significant, due to traditional and to a lesser extent to scale intensive industries, in a similar fashion to what we found for productivity. Amongst other variables, it is worth noting that while the proxy for human resources present a positive and significant correlation with profitability throughout different industries, the opposite happens for capital intensity. Multinationality seems to add a plus to profitability, but only in traditional and scale intensive sectors, while geographical location and vertical integration do not show any significant impact.

11.3.4 Concluding remarks

The analysis proposed in this paper for Italian manufacturing corporations provides new evidence confirming that skills and capital intensity are key determinants of productivity. Offshoring and IT maturity too result to have both had a positive impact on productivity. In this case, though, the significance of (intensity) variables used depended crucially on industry specificities. IT maturity and offshoring had as well a positive and significant impact on profitability, but only for productions which could be more easily delocalised.

Labour skills appear to have improved profitability (while capital intensity as such, for the years considered, had a negative impact), and at the same time to have played a role in offshoring and IT usage decisions. These latter two, instead, did not show any clear mutual relationship. Finally, with respect to the key issue of directions of causality, lagged offshoring appears to weakly impact productivity, while lagged IT maturity does not, and a reverse causality from productivity to IT maturity is revealed.

The latter evidence, albeit limited, challenges some commonplaces, suggesting the coexistence (and, for the time being, the viability) of different business models, rather than a strong, univocal relationship framing a single techno-economic paradigm. A richer information set (including innovation, R&D and skills surveys, but also data for service firms) should allow for a more appropriate treatment of these issues, as well as for extending the analysis beyond IT and the offshoring of intermediate goods in manufacturing corporations to other, crucial determinants of performance.



11.4 Offshoring, Productivity and IT-maturity: The French evidence Yoann Barbesol, INSEE Simon Quantin, INSEE

11.4.1 Background Within the frame of the Eurostat project working on productivity effects from several dimensions of IT, this paper aims to estimate the impact on productivity:

• from offshoring combined with IT-maturity: does IT-maturity affect directly offshorers’ productivity or indirectly, through offshoring ?

• from IT-maturity on offshoring firms compared to non-offshorers; We first analyse the impact on productivity from offshoring. Since IT development seems to be one of the facilitators of globalisation, there is a strong possibility that estimations from offshoring equations may be biased if the firm IT maturity is not controlled for. Hence a variable measuring the IT-maturity (intensity) of the firm will be added: the e-commerce survey will help us to include several aspects on how firms use computers and the Internet.

11.4.1.1 Offshoring / outsourcing: definitions The term “offshoring” is often associated with “outsourcing” but refers to a different concept. Whereas outsourcing refers to the relocation of jobs and processes to external providers regardless of the provider’s location, offshoring refers to the relocation of jobs and processes to any foreign country without distinguishing whether the provider is external or affiliated with the firm. Outsourcing may therefore include job relocations both within and between countries, whereas offshoring refers only to international relocations. The term offshore outsourcing therefore only covers the relocation of jobs or processes to an external and internationally located provider

Figure 11.4.1

Between countriesWithin countries

International Insourcing

Domestic supplyWithin Firms(insourcing)

Offshoring

International outsourcing

Domestic outsourcingBetween firms

(outsourcing)

Sourcing

InternationalNational

Between countriesWithin countries

International Insourcing

Domestic supplyWithin Firms(insourcing)

Offshoring

International outsourcing

Domestic outsourcingBetween firms

(outsourcing)

Sourcing

InternationalNational

Source: STI Working Paper 2006/1

215


11.4.1.2 Expectations In theory, offshoring provides:

• Better access to knowledge. This effect might be observed by using skill variables as the share of employee with a post-secondary education. Due to data limitations, we will not add these variables, contrary to Sweden.

• Lower production costs (change of relative prices of inputs) • International market closer to perfect competition • Large fixed costs impede low productivity firms to engage in offshoring

11.4.2 Empirical approach

11.4.2.1 Estimating the firm-level productivity The impact of offshoring is commonly analysed by estimating labour productivity through a production function framework. The standard approach used to estimate the impact of offshoring / outsourcing on labour productivity at the level of a firm is to estimate a production function given by:

),( itititit LKFAY =

Where Y refers to the output, A the technology factor, K is capital and L is employment. i refers to the firm and t to time. The functional form is assumed to be a Cobb-Douglas function. In this case, labour productivity can be written:

µαitititit LkAy =

where µ and α are indicators of return to scale ( 0>µ indicates increasing return to scale), k is capital intensity. Note that the parameters µ and α are the same for all firms, and if value-added is chosen to illustrate the output, the explanatory intermediate variable is omitted since the value-added definition excludes this input.

It is common to let outsourcing / offshoring take effect through the technology factor of the production function. This basically means that outsourcing/offshoring may alter the firm’s production technology by allowing it to shift the intercept of the log-linear equation production function. Thus, the base regression equation can be written:

ititititit offLky εββββ ++++= )ln()ln()ln()ln( 3210

where off refers to the measure of offshoring. It could be useful to split offshoring measure between materials and services offshoring. The remaining production technology factors are picked up by the constant 0β and the error term itε .

The effect of offshoring on productivity is determined by the level of significance of the estimated coefficients on off.

This estimation method is likely to be faced with some constraints. Problems of endogeneity can be expected since the results might be driven by unobserved covariates that are correlated with both productivity and offshoring/ outsourcing. For example, plants with a high productivity level may also be more skill-intensive and for that reason more likely to engage in outsourcing of low-skill intensive processes in order to focus on core competences. This may result in biased and inconsistent regressions. A way of reducing these biases is to use lagged variables of the outsourcing variable or IV estimation, which is widely used in the literature. Moreover, looking at growth rates by first-differencing is known to exacerbate potential problems of measurement errors in the data (Griliches and Hausman, 1986).

To link the offshoring equation to IT use, we add two variables:

• an IT-maturity variable based on the e-commerce survey information. • a broadband variable, a dummy for firms using broadband access.

216


As the offshoring behaviour differs a bit according to type of firm, international experience (like firms which export) and ownership status (multinational, foreign owned), such variables are also included.

11.4.2.2 Measuring offshore outsourcing and IT-maturity Offshore outsourcing is measured as the ratio of imported inputs over total purchases. Due to data limitations, only the offshoring of intermediate goods can be treated.

Since the IT-maturity of firms can be expressed in several ways, a further variable was tested. The degree to which a firm sells and/or buys over the Internet should logically be of importance for the IT maturity level of the firm. A second IT-maturity variable has also been tested by Sweden, which is the share of employees with Internet via broadband access.

11.4.3 Description of the average French offshorer in 2004. These results show that offshorers in general are: firms with high labour productivity; exports intensive firms; firms affiliated to multinational corporations. These firms are also over-represented in the E-commerce survey. However, E-commerce survey firms are larger, more export intensive, and firms from this sample are more homogenous, which reduces the threat of composition bias.

Table 11.4.1: Facts about French offshorers in 2004

Variable All firms Offshorers of intermediate

goods

EC-Survey firms Offshorers of intermediate goods in EC

survey

Number of firms 447,020 31,494 6,313 3,176 Mean number of employees 16,6 111,1 453,3 647,0

Mean of labour productivity (value added per employee)

47,9 76,1 62,6 68,8

Mean of offshoring intensity (imports / total purchases) in percent

3,5% 50,0% 15,6% 31,0%

Mean of export intensity (exports / value of production) in percent

1,7% 18,0% 11,8% 23,0%

Share of imports from low wage countries

0,6% 8,2% 2,9% 5,9%

Share of French multinational firms 1,7% 9,1% 19,0% 24,3% Share of foreign owned firms 1,5% 13,0% 20,0% 32,7%

Note: Retail sectors were excluded Source: INSEE

11.4.4 Offshoring of intermediate goods impact on productivity In this section, we analyze whether offshoring of intermediate goods affects productivity. We run two regressions;

• The first one uses exhaustive information (Table 11.4.2), • The second uses only the EC-survey firms (Table 11.4.3) leads to the same conclusions:

offshoring of intermediate goods is correlated with higher productivity. It is worth stressing that results are left quite unchanged whether we use the E-Commerce survey or exhaustive information except for the “log of employment” estimate.

The “log of employment” estimate in these regressions differs markedly from the ones carried out on the total number of firms, emphasizing the fact that the results with EC-survey firms might not necessarily represent all firms in general. Moreover, remember that this parameter is an estimate of returns to scale. According to our specification, EC-survey firms are decreasing returns to scale firms.

217


However offshoring intensity is insignificant1 and other estimates remain nearly the same. Finally, no bias due to omitted variable is observed. Table 11.4.2: Effects of firm productivity performance from offshoring of intermediate goods Dependent variable: log of labour

productivity Estimate

parameters Robust

standard errors Estimate

Parameters Robust

standard errors

All firms All firms

Intercept -1.5310 (0.02)*** -1.5311 (0.02)*** Log of employment 0.0078 (0.00)*** 0.0077 (0.00)*** Log of capital intensity (capital per employee)

0.1780 (0.00)*** 0.1779 (0.00)***

French multinational 0.1989 (0.01)*** 0.1986 (0.01)*** Foreign owner 0.1872 (0.01)*** 0.1837 (0.01)*** Offshoring intensity 0.0166 (0.00)*** Age 0.0179 (0.00)*** 0.0179 (0.00)*** Age squared -0.0004 (0.00)*** -0.0004 (0.00)*** Number of observations 579,363 579,363 Note: Between estimations. Unreported industry dummies are included in order to control for differences in labour productivity and over the years Table 11.4.3: Effects on EC-survey firm productivity performance from offshoring of intermediate goods Dependent variable: log of labour

productivity Estimate

parameters Robust

standard errors Estimate

Parameters Robust

standard errors

All firms from EC-survey All firms from EC-survey

Intercept -0.8723 (0.08)*** -0.8730 (0.08)*** Log of employment -0.0174 (0.01)*** -0.0175 (0.01)*** Log of capital intensity (capital per employee)

0.1843 (0.01)*** 0.1839 (0.01)***

French multinational 0.1138 (0.02)*** 0.1142 (0.02)*** Foreign owner 0.1505 (0.02)*** 0.1484 (0.02)*** Offshoring intensity 0.0150 (0.01) Age 0.0024 (0.00) 0.0024 (0.00) Age squared -0.0001 (0.00)*** -0.0001 (0.00)*** Number of observations 10,150 10,150 Note: Between estimations. Unreported industry dummies are included in order to control for differences in labour productivity and over the years

11.4.5 Offshorers productivity and IT-maturity Elsewhere in this Project Report we show that IT maturity impacts firms’ productivity. In the following section we try to identify if that effect is a direct one or operates through the channel of offshoring.

1 Offshoring intensity parameter estimate (1.5%) is close to the previous estimation (1.7 %). The fact that it is insignificant is probably due to a sampling discrepancy between the two estimations. Offshorers stand for 50% of the firms in the E-Commerce survey (10% in the Business register).

218


11.4.5.1 IT-maturity impacts offshorers’ productivity… In this section, we just add an IT-maturity variable to the previous specification (dummy if firm sells and/or buys over the Internet). Results are shown in table 11.4.4. Results are left unchanged whether we proxy IT-maturity with a dummy variable or with the continuous variable ECPCT (share of firm purchases + share of firm sales over the Internet);

Table 11.4.4: Effects on offshorers productivity performance from IT-maturity

Dependent variable: log of labour productivity

Estimate parameters

Robust standard

errors

Estimate parameters

Robust standard

errors

Estimate parameters

Robust standard

errors

Intercept 0.119 (0.15) 0.121 (0.14) -0.004 (0.11) Log of employment

-0.033 (0.01)*** -0.033 (0.01)*** -0.008 (0.01)

Log of capital intensity (capital per employee)

0.216 (0.01)*** 0.216 (0.01)*** 0.182 (0.001)***

French multinational

0.107 (0.02)*** 0.107 (0.02)*** 0.093 (0.02)***

Foreign owner 0.118 (0.02)*** 0.118 (0.02)*** 0.110 (0.02)*** International experience

0.053 (0.02)** 0.053 (0.02)** 0.073 (0.02)***

Log of offshoring intensity of intermediate goods

-0.013 (0.00)*** -0.012 (0.00)*** -0.014 (0.00)***

Log of offshoring intensity of intermediate goods * Purchases with internet (dummy)

-0.002 (0.01) 0.002 (0.01)

IT-maturity 0.050 (0.01)*** 0.048 (0.02)*** 0.034 (0.01)*** Broadband 0.045 (0.01)*** Age 0.001 (0.00) 0.001 (0.00) 0.001 (0.00) Age squared 0.000 (0.00) 0.000 (0.00) 0.000 (0.00) Number of observations

5,461 5,461 3,176

Note: Heteroskedasticity -consistent standard errors are shown. Unreported industry dummies are included in order to control for differences on labour productivity and over the years

IT-maturity is significantly correlated with offshorers labour productivity (table 11.4.4 - column (1)). Table 11.4.62 shows results when distinguishing manufacturing firms and services firms. When splitting our sample into these 2 groups, manufacturing firms and services firms, we detect effects on productivity from IT-maturity only for services firms. However these results should be cautiously looked at, as services are under-represented (only 165 observations in the EC-survey).

It is now worth examining through which channel IT-maturity impacts their productivity; directly or by boosting international trade, since IT development appears clearly as one of the facilitators of globalisation

2 Tables 11.4.6 and 11.4.7 are at the end of this section of Chapter 11.

219


11.4.5.2 … but IT-maturity impact on offshorers productivity does not operate through offshoring We introduce a crossed variable (log of offshoring * purchases over Internet dummy) to capture the effects of IT-maturity on labour productivity that operate through offshoring. We find no evidence of such a mechanism (table 11.4.4 - column (2)).

To extend the scope of the study, we introduce a dummy for broadband, as employees working in firms equipped with such technology may be more productive. The results are shown in the column (3) of table 11.4.4. The broadband variable impacts significantly labour productivity. However we point out that the point of estimate for the log of employment (which stands for the indicator of return to scale) becomes insignificant and the coefficient of the capital variable falls from 21.6% to 18.2%. The panel attrition (42% of the firms are deleted owing to the addition of the broadband dummy) may explain these variations.

11.4.6 IT-maturity impact on productivity As we have shown that the IT-maturity effect on labour productivity does not pass through offshoring, we want to check that IT impact does not differ between offshorers and non-offshorers.

In the specification, we add the following crossed variable; IT-maturity*(offshoring dummy). IT-maturity effect is measured by the continuous variable, share of firm purchases + share of firm sales over the Internet. We find evidence that IT impact is the same between offshorers and non-offshorers, as shown below3 (cf table 11.4.5).

Table 11.4.5: IT-maturity impact on productivity does not differ between offshorers and non-offshorers Dependent variable: log of labour

productivity Estimate parameters Robust standard errors

All EC-survey firms Intercept -0.470 (0.12)*** Log of employment -0.025 (0.01)*** Log of capital intensity (capital per employee)

0.172 (0.01)***

French multinational 0.128 (0.02)*** Foreign owner 0.148 (0.02)*** Age 0.001 (0.00) Age squared 0.000 (0.00)** IT- maturity 0.106 (0.04)*** IT- maturity * offshoring dummy

-0.029 (0.05)

Number of observations 9,657 Note: Between estimations. Unreported industry dummies are included in order to control for differences in labour productivity and over the years However, these results are based on a debatable assumption that the technology used by offshoring firms and non-offshoring ones is similar. We show in Table 11.4.7 that these doubts are relevant. However, controlling for labour quality should be a way to improve the comparison between IT impact on offshorers and IT impact on non-offshorers.

3 If we use a dummy variable instead, all estimate parameters are left unchanged except for IT-maturity whose parameters falls to 5%, which is consistent with the results shown in table 11.4.4.

220


221

Table 11.4.6: Effects on firm productivity performance from IT maturity Dependant variable: Log of labour productivity

Estimate parameters

Robust standard errors

Estimate parameters


Manufacturing firms Services firms Intercept -0.135 (0.02)*** -0.108 (0.01)*** Log of offshoring intensity of intermediate goods from high wage countries

0.025 (0.00)*** -0.67 (0.00)***

Log of offshoring intensity of intermediate goods from low wage countries

0.005 (0.00)*** -0.003 (0.00)

Log of employment 0.006 (0.00) -0.049 (0.00)*** Log of capital intensity (capital per employee)

0.234 (0.01)*** 0.176 (0.00)***

French multinational 0.043 (0.01)*** 0.336 (0.01)*** Foreign owner 0.031 (0.01)*** 0.421 (0.01)*** International experience of firm (exports over total sales)

-0.072 (0.02)*** -0.040 (0.02)***

Age 0.003 (0.00)*** 0.006 (0.00)*** Age squared 0.000 (0.00)*** 0.000 (0.00)*** Log of firm IT-maturity

0.001 (0.00) 0.067 (0.00)***

Number of observations

1,957 165

Note: Heteroskedasticity -consistent standard errors are shown. Unreported industry dummies are included in order to control for differences on labour productivity and over the years Table 11.4.7: IT-maturity impact on productivity differs between offshorers and non-offshorers Dependant variable: log of labour productivity

Estimate parameters


Estimate parameters


Offshorers Non-offshorers Intercept 0.242 (0.14) -0.490 (0.11)*** Log of employment -0.030 (0.01)*** -0.035 (0.01)*** Log of capital intensity (capital per employee)

0.217 (0.01)*** 0.171 (0.01)***

French multinational

0.109 (0.02)*** 0.145 (0.03)***

Foreign owner 0.110 (0.02)*** 0.225 (0.04)*** Age 0.001 (0.00) 0.002 (0.00) Age squared 0.000 (0.00) 0.000 (0.00) IT- maturity 0.070 (0.03)** 0.126 (0.04) Number of observations

5,460 4,237

Note: Between estimations. Unreported industry dummies are included in order to control for differences in labour productivity and over the years


CHAPTER 12: ICT, INNOVATION AND PRODUCTIVITY George van Leeuwen, Statistics Netherlands Shikeb Farooqui, UK Office for National Statistics Summary: It seems that broadband connectivity and electronic selling are important drivers of innovativeness and that augmenting innovation models with ICT use variables adds additional power for explaining differences in innovativeness. However, the picture is less clear for the ‘direct’ ICT productivity impact and conditional on performing R&D. After controlling for the more ‘usual suspects’ like innovation (R&D) expenditures, the (implicit) direct ICT contribution to productivity appears to be only weakly significant. By contrast, this contribution is very sizable and significant if a less restrictive definition of innovativeness is adopted by using also the data of firms that relied on using a general purpose technology such as ICT and – at the same time – did not apply technological innovations (generating new or improved products or production methods). This last result has been obtained by ignoring R&D expenditures. Therefore, the question whether ICT use or R&D are more important for explaining differences in productivity remains open to debate. Arguably, there is too much focus in CIS (and CIS based research) on technological innovations (in particular the creation of new and improved products) to be able to disentangle productivity into contributions of ICT use and R&D driven innovation. Taking also into account the measurement problems for technological innovations, it might be considered to downplay the role of innovative sales and to upgrade the role of ICT use in order to seek a new balance between the two in innovation research. Innovation expenditures can be underestimated in CIS, not only because much R&D is not reported, but also due to the neglect of ICT investment. In Chapter 10 it has been shown that the ICT use variables collected in the EC surveys are fairly good predictors of ICT capital stocks. Therefore, augmenting the model of Griffith et al. (2007) with ICT indicators could be a next step forward to understanding the relative importance of ICT use and R&D driven innovation. This route seems not to be impossible, at least for the Netherlands, if non-technological innovators are not excluded from the empirical analysis from the outset.

12.1. Introduction A recurrent conclusion of research carried out using firm-level data is that using R&D or ICT

‘stocks’ that are built from investment in R&D and computers respectively, may not be sufficient to

explain heterogeneity in firm performance (e.g. productivity). Additional information on how these

stocks are used seems to be of increasing importance. This is also reflected in the development of

statistical systems, notably the extension of R&D surveys to Innovation surveys and the adaptation of

Automation Surveys to – what is now called – Ecommerce (EC) surveys with a stronger focus on how

firms use their ICT capital stocks.

Since much of the empirical literature is focused on either innovation in the R&D view or ICT for

explaining differences in productivity (growth), one cannot escape from the impression that research

on these topics seems to belong to separate branches of economics. However, there are several

reasons to believe that ICT use and innovation can be strongly related. First, innovation can no longer

be conceived as a pure R&D story only. Innovation imitation may be an equally important driver for

innovativeness as is performing own R&D (see e.g. Griffith et al. (2004) and Van der Wiel et al.

(2008)). Second, one can imagine that the innovation enabling features of ICT in particular

materialize in the throughput stage of the innovation process. Here, ICT use may play an important

222


role in 1) rolling out (new) products (e.g. Ecommerce), 2) in capturing and processing knowledge

developed elsewhere or 3) in managing knowledge flows within and between firms (e.g. by exploiting

ICT network externalities).

Identifying whether ICT use has a separate role for explaining differences in productivity

conditional on being innovative in the traditional (R&D) view raises several problems. To achieve

this, one has to make a distinction between the direct and indirect productivity contributions of

innovation as well as ICT variables and their interactions. Applying a reduced-form single-equation

model for disentangling the different productivity contributions of ICT use conditional on innovation

seems not appropriate for this purpose, because such an approach remains silent on the importance of

different forms of ICT use for innovation, i.e. the indirect effect of ICT use on productivity.

Today, the state of the art in empirical innovation (R&D) research concerns the application of the

so-called CDM model on CIS data (see e.g. Crépon et al., 1998). This model explicitly accounts for

innovation as a separate process by adopting the construct of an innovation production function and

by linking this function to overall performance, e.g. productivity (growth). The bottom line of the

CDM model is that it replaces the traditional reduced-form approach to the assessment of R&D as a

driver of productivity (growth) by a structural modelling approach that acknowledges the joint

dependence of innovation inputs, innovation outputs and productivity on all innovation characteristics

made available in CIS (see e.g. Van Leeuwen and Klomp, 2006).

In this chapter the CDM model is extended with ICT variables in an attempt 1) to identify the role

of ICT (use) in the innovation process, 2) to investigate the relative importance of ICT use for

different stages of the innovation process and 3) to assess whether ICT use has explanatory power

over and above (CIS) innovation variables for productivity (growth). In this exercise, special

attention is given to broadband connectivity and Ecommerce as potentially important ICT drivers of

productivity that can be channelled via innovation.

The plan of the chapter is as follows. Before presenting the model (in section 12.5) some data

issues are discussed in section 12.3. This discussion offers a background for understanding the

empirical route chosen. This will be illustrated further with the help of simple descriptive measures in

section 12.4. Estimation results obtained with the help of data for the Netherlands are presented in

sections 12.6–12.8, whereas section 12.9 includes a comparison with a recent Swedish study.1

12.2. Data issues: censoring and selectivity due to data linking To pursue a research strategy as explained above is far from easy, not least because one has to deal

with the applied censoring in the innovation survey and the missing data patterns that are due to the

merging of three different surveys.

The censoring problem emerges due to the fact that firms may have been engaged in innovation

but nevertheless did not succeed in producing technological innovations (new or improved products) 1 The Swedish study of Hagén et al. (2008) uses two ICT use variables in a structural innovation model.

223


or implementing process innovations (see Box 1). For these firms it makes sense to have ‘measured’

zero sales of new and/or improved products. The censoring problem is most manifest for the input

side of the innovation process, as data on innovation costs (e.g. R&D expenditures) or on the

innovation throughput stage (use of information sources, innovation cooperation etc.) are missing for

a considerable number of firms in this case.2 However, this does not exclude the possibility that these

firms were innovative in a broader sense, for instance by relying on a general purpose technology

such as ICT.

Another complication concerns the merging of different data sets. This may result in many holes

in the merged data, even for the firms for which we have data on innovation output, innovation

(R&D) expenditures and details on the throughput stage of the innovation process. The small overlap

between Production Statistics, Ecommerce and Innovation surveys causes a myriad of missing data

patterns that raises additional problems of selectivity or endogenous attrition for the data used in the

estimation.3

Box 12.1: Censoring in CIS The censoring problem can be clarified by looking at the filtering questions that guide the respondents in filling in the CIS questionnaire. Take, for instance, the basic filtering questions 2.1 – 2.4 in CIS4 (covering the period 2002 – 2004):

2.1) During 2002 – 2004 did your enterprise introduce a) new or improved products? b) new or improved services? If ‘yes’: report share of innovative sales (new and/or improved).in total sales (turnin). If not: go to (question) 3.1 3.1) During 2002 – 2004 did your enterprise introduce a) new or improved methods of producing goods? b) new or improved logistics, distribution methods? If yes: report on innovation expenditure, innovation cooperation etc. (questions 5.1 and following). If not: go to (question) 4.1. 4.1) During 2002 – 2004 did your enterprise have any innovation activities that were abandoned or are still ongoing? If not: go to ‘hampering factors for innovation’ (question 8.2).

The routing of questions stresses the importance of so-called technological innovations (i.e. the creation of new and/or improved products). Moreover, there is no explicit reference to ICT use. Nevertheless, in some cases, the wording of the question more implicitly refers to ICT (for instance Ecommerce as an example of new or improved logistics and distribution methods).

As the linking of data from Ecommerce surveys (EC) and Innovation surveys (CIS) to

performance measures is at the heart of this study, we first look at the overlap between the three

surveys presented in Table 12.1. First, it can be noted that, in general, the linking of data on ICT use

to firm performance leads to smaller data sets than when linking data on firm performance and

2 Firms may have stated to have abandoned or ongoing innovation project, but only very few of these firms are reporting R&D expenditures. 3 Measures of firm performance (e.g. productivity) are derived from the SBS surveys, innovation variables are surveyed in CIS and the data on ICT use are sourced from the Ecommerce Surveys.

224


innovation. This result primarily reflects the smaller sample size of Ecommerce surveys.

Furthermore, by extending the merging to three data sources and by focussing on technological

innovators, one exacerbates the selectivity problem considerably, not at least due to the sampling of

many small firms, the small overlap of samples and endogenous ‘attrition’ as a result of not being

successful in innovation in the traditional view. Consequently, the final result of merging the three

surveys is a database that seems to be biased towards the selection of larger manufacturing firms.4

Table 12.1: Linking Ecommerce (EC), Innovation (CIS) and Production (PS) surveys EC and CIS with PS EC/CIS/PS EC/SBS CIS/PS All1 Innovative2

Breakdown by year3 N N 2002 5063 7320 2731 630 2004 3549 5786 1659 516 All 8612 13106 4390 1146 Breakdown by sector % % Manufacturing 0.31 0.29 0.34 0.61 Construction 0.10 0.12 0.10 0.04 Distributed services 0.25 0.36 0.26 0.14 Other business services 0.33 0.23 0.30 0.21 Breakdown by size class employed persons < 20 0.32 0.20 0.20 0.08 20 – 50 0.22 0.26 0.21 0.14 50 – 100 0.14 0.24 0.18 0.16 100 – 250 0.17 0.19 0.23 0.32 250 – 500 0.09 0.06 0.11 0.18 > 500 0.07 0.04 0.08 0.12

1 Including firms that did not succeed in producing technological innovations (product and/or process); 2 Firms that realized technological innovations (strict innovators); 3 For CIS 2002 refers to the period 2000 – 2002 and 2004 to the period 2002 – 2004. Source: Statistics Netherlands

The usual way to proceed is to use the full sample (i.e. the 1149 firms of the last column of Table

12.1) and to ‘control’ for selectivity in the estimation. However, our primary goal is to identify direct

and indirect productivity contributions of two important ICT indicators (broadband intensity use and

Ecommerce). There may still be a productivity contribution of ICT for the firms that did not realize

technological innovations (had zero innovative sales) or firms that are not engaged in R&D. Ignoring

these firms could bias the estimate for the contribution of innovation to productivity. For this reason

we will compare the results of two models: 1) a model that uses all linked EC, CIS and PS data,

4 The bias towards manufacturing also mirrors the well known fact that ‘technological innovations’ are more important (and better measurable) in manufacturing than in the service sector.

225


including the firms with zero innovative sales and 2) a model that can be applied to a subset of firms

that reported innovative sales as well as innovation expenditures and data on innovation throughput.

The last model is used to determine the contribution of ICT use to technological innovation after

controlling for – amongst others – innovation costs as an input into innovation.

12.3. ICT use, innovation and productivity: indications from descriptive statistics To get a first impression on the relation between ICT use, innovation and productivity, one can

look at some simple descriptive statistics obtained after merging the three surveys and after allocating

firms according to their level of broadband intensity use (DSLpct)5. Figure 12.1 shows the average

knowledge production (measured as the share of innovative sales in total sales (Turnin)) for four

classes of broadband intensity, and after imputing a value of zero for firms that did not produce

technological innovations.7

Figure 12.1: Broadband intensity and innovation output

0102030405060708090

100

0 – 25 25 – 50 50 – 75 75 – 100

Class of DSLpct

DSLpct

Turnin

At first sight, the figure suggests weak evidence for a positive correlation between broadband

penetration (DSLpct) and the share of innovative sales in total sales. The only thing one can say is

that the firms with the highest levels of broadband intensity seem to have significantly higher shares

of innovative sales compared to firms that do not use DSL at all or only at a modest scale. However,

this is not very surprising as many ‘non-technological’ innovators can be found in the lowest class of

DSLpct.8

5 Broadband intensity (DSLpct) is used as the measure of ICT intensity as this measure refers to the level of IT intensiveness as well as to the use of communication infrastructure (DSL). It is constructed by multiplying the share of employees that use computers (PCpct) with the binary DSL indicator (DSL). 7 It should be remembered that we do not have available other inputs into innovation for the ‘non-technological’ innovators. 8 The number of firms in the DSLpct classes of figure 1 is (from low to high) 2165, 570, 414 and 1241.

226


Figure 12.2: Broadband intensity, Innovation investment and innovation output

0102030405060708090

100

0 – 25 25 – 50 50 – 75 75 – 100

Class of DSLpct

DSLpctRyTurnin

0

20

40

60

80

100

120

0 – 25 25 – 50 50 – 75 75 – 100

Class of DSLpct

All

Innov.

Figure 12.3: Broadband intensity and labour productivity (x 1000 Euro)

Figure 12.2 repeats the same exercise for the strict innovators and after including also their inputs

into innovation (innovation costs as a percentage of total sales (Ry)).9 Again the correlation between

broadband connectivity and innovation seems to be weak in the averages. Moreover, the relatively

better innovation performance of the firms belonging to the highest DSLpct class cannot be attributed

227

9The distribution of firms across classes of DSLpct in figure 2 is (from low to high) 391, 219, 174 and 471.


to being more broadband enabled only, because the same firms also spend (relatively) more on inputs

into innovation.

Next we look at the productivity distributions in figure 12.3. Here, it can be seen that the

productivity differences between classes of DSLpct are more pronounced than when focussing on

innovation. Although productivity is not monotonically increasing in DSLpct, one cannot escape

from the impression that broadband enabled workers have higher levels of productivity on average.

To sum up: it is difficult to obtain a clear picture of the relation between ICT use, innovation and

productivity from simple descriptive statistics. There may be a positive correlation between ICT use

and innovation, but this may be due that the fact that many firms that produced no technological

innovations were also not broadband enabled at all or at (relatively) low levels. Furthermore, within

the class of firms that had the highest level of broadband connectivity, average productivity was

higher for non-technological innovators! Thus, technological innovation in the traditional view and

using a general purpose technology such as ICT may impact productivity in different ways as well as

may complement each other.

12.4. Model The simple fact that ICT is an enabler of innovation but also an instance of innovation makes it

difficult to identify the role of ICT use in the causal chain that runs from ICT use to productivity. The

example of broadband connectivity discussed in the preceding section illustrates that the potential

benefits of ICT use can be captured at different stages within the value chain and that innovation and

productivity can be jointly dependent on the same ICT use variables. Formalizing this joint

dependence calls for the application of a modelling approach that takes into account the direct and

indirect contributions of ICT usage to productivity. This has been implemented by using a structural-

equations model (SEM, see Box 12.2).

The model is an extension of the well-known seminal CDM innovation model (see Crépon et al.

1998). This extension explicitly models the role of ICT use for innovativeness 1) by including

broadband intensity use and Ecommerce variables into the traditional innovation production function,

2) by including a separate equation for the share of electronic sales in total sales 3) by linking this

equation to knowledge production (the innovation output equation).

228


Box 12.2: Augmenting the CDM innovation model with ICT The ICT augmented CDM model consist of a system of four equations: (1) VA Prod = f{Capital intensity (C), Labour (C), Epurchpct (C), DSLpct (C), Innovative sales per employee (C), Process innovation (D)} (2) Innovative sales per employee = f{Size control (C), Innovation expenditure per employee (C), Esalespct (C), DSLpct (C), Innovation Cooperation (D), Process innovation (D)} (3) Esalespct = f{Size control (C), DSLPct (C), Stage of ICT maturity (D)} (4) Innovation expenditure per employee = f{Size control (C), R&D firm (D), Innovation Cooperation (D)}. (C) in (1) – (4) refers to variables measured on a continuous scale and (D) to dummy variables. The jointly endogenous variables of the system are underlined (the definition of all variables is given in Appendix A). Equation (1) is the familiar Cobb-Douglas Production Function for value added per employee in logarithms. Logarithmic transforms are also applied for innovation output (innovative sales per employee, equation (2)) and innovation inputs (innovation expenditure per employee, equation (4)). The endogenous variable Esalespct (equation (3)) is measured as electronic sales scaled by total sales. As three out of four endogenous variables are measured in logarithms, this implies that the four-equation system can only be estimated for innovators in a strict sense, i.e. the firms belonging to the merged (PS/EC/CIS) data with both innovative sales and innovation expenditures > 0. As many firms stated not to have realized innovative sales, the model can also be applied to a substantial larger data set after imputing a value of zero for innovative sales and after removing equation (4) from the system. In this case the innovation output variable is replaced by the share of innovative sales in total sales. Expressing the most important variables of the system in per employee terms allows a comparison of direct and indirect contributions of ICT use to productivity. For instance, using (1) and (2) the contribution of broadband connectivity (DSLpct) to productivity (log(Y/L)) can be decomposed as

DSLpctLY

DSLpctLInnosales

LInnosalesLY

DSLpctLY

∂∂

+∂

∂∂

∂=

∂∂ )/log()/log(.

)/log()/log()/log( ,

with the direct effect given by DSLpct

LY∂∂ )/log( ,

and the indirect contribution by DSLpct

LInnosalesLInnosales

LY∂

∂∂

∂ )/log(.)/log(

)/log( .

The rationale for doing this can be understood as follows. Beginning with broadband connectivity,

it is assumed that fast internet connections may contribute to the outcome of innovation efforts of

other innovation inputs (such as e.g. spending on R&D) by enhancing the management of knowledge

flows in the innovation throughput stage. In this way broadband connectivity represents ICT network

externalities that support the use of information sources in knowledge production. Thus, this variable

can be used as an alternative for binary indicators derived from the use of data on information sources

as collected in CIS.

Furthermore, one can imagine that a firm’s innovativeness may be affected positively by ICT

applications such as electronic sales. Firms with higher intensities of electronic sales (measured as

229


the share of electronic sales in total sales) may be better placed in the market. Electronic selling can

be seen as a way to roll out new or improved products or to complement production with better

marketing and distribution systems, and these factors can increase the probability of innovation

success as measured as the sales of new and/or improved products. In turn, the alignment of the ICT

infrastructure to better communicate with customers may improve the conditions for increasing

electronic sales. In the empirical implementation we combine several variables collected in the

Ecommerce survey to construct a measure for ICT maturity seen from a seller’s perspective (see also

Appendix A). This measure will be used to explain differences in electronic sales in addition to the

broadband intensity use indicator (DSLpct).

12.5. Results This section summarizes the results of the simultaneous-equation models using data for the

Netherlands. The augmented CDM model of Box 2, which uses the data of strict innovators, will be

labelled model I. Model II differs from model I by excluding the innovation input equation and by

using the share of innovative sales in total sales instead of (the logarithm of) innovative sales per

employee as the measure of innovation output. In order to ‘control’ for selectivity and joint

endogeneity, the following route has been chosen:

• Estimate a Heckman selection model on each equation separately;

• Evaluate the equation specific selection bias and capture in a new “mills bias” variable;

• Re-estimate all four equations jointly, in a system, adding the “mills bias” variable to the model specifications.

This sequence generates a lot of estimation results. To save space we will not discuss the

estimates of the Heckman models extensively (see Appendix B for the innovation input and output

equation of model I). An interesting conclusion that can be drawn from this exercise is that

broadband connectivity is an important determinant of the propensity to innovate, irrespective of

whether one looks at the input or at the output side of knowledge production.

Table 12.2 collects the results of the system approach for the most important variables. Notice that

model I resembles the core model of the OECD micro data benchmark project (see OECD, 2008). In

this model innovation inputs, innovation output and productivity are expressed in per employee terms.

Thus the estimates can be interpreted as elasticities. This nice feature does not carry over to model II

as innovation output is measured as the share of innovative sales in total sales. This difference should

be taken into account for the interpretation of the estimates. Nevertheless a comparison of the two

models remains useful. This is so because model II uses considerably more data than model I.

230


Table 12 2; Estimates of structural model for ICT, innovation and productivity1 Model I (N = 1036) Model II (N = 3953) Est. T Est. T(1) Productivity Innovative sales per employee 0.14 1.9 Share of innovative sales in total sales 0.72 1.6Capital intensity 0.21 8.1 0.23 24.6Labour 0.06 4.8 0.04 0.4DSLpct 0.10 1.8 0.33 7.2Epurchpct -0.07 -0.6 -0.14 -2.1Process innovation 0.01 0.1Constant 2.82 11.7 2.59 2.3R2 0.31 0.18 (2) Innovation output2 Labour -0.01 -0.2 0.04 1.4Innovation expenditure per employee 0.44 6.0 Esalespct 2.84 6.1 0.58 10.3DSLpct 0.29 2.3 0.07 7.6Cooperation 0.03 0.4 Process innovation -0.06 -0.8 Constant 2.79 13.3 -0.40 -1.1R2 0.08 0.09 (3) Esalespct Labour 0.02 4.1 0.02 1.1DSLpct -0.02 -1.3 0.01 2.1_Ifase_1 -0.11 -4.3 -0.05 -7.2_Ifase_2 -0.10 -7.1 -0.06 -10.7_Ifase_3 0.01 0.5 0.03 4.0_Ifase_4 0.04 1.8 0.06 5.9_Ifase_5 0.11 4.9 0.13 11.7Constant 0.03 1.0 -0.13 0.6R2 0.15 0.16 (4) Innovation expenditure per employee Labour -0.23 -6.5 R&D permanent 0.88 9.5 Cooperation 0.56 6.3 Constant 1.32 7.1 R2 0.18

1 All equations include industry and time dummies and Heckman correction terms to control for selectivity; 2 In model I innovation output is measured as the (logarithm) of innovative sales per employee and in model II by the share of innovative sales in total sales.

12.6. Results for the core variables of the CDM model A useful starting point is to begin the discussion at the bottom of Table 12.2 with results for inputs

into innovation, i.e. the estimates of the equation for innovation expenditure per employee. Here, not

very surprisingly, performing R&D on a permanent basis is the most important explanatory variable.

231


The estimate of ‘innovation cooperation’ is also positive and very significant, a result which

corroborates earlier results concerning the testing of the absorptive capacity hypothesis (see e.g.

Cohen and Levinthal, 1989). Furthermore, the negative estimate for the coefficient of labour points to

decreasing innovation expenditures per employee if firm size increases.

Looking next at the results for the knowledge production function (the innovation output equation

(2)), one can see similar patterns as reported in e.g. Van Leeuwen and Klomp (2006) and Polder

(2008)10. Innovation cooperation is not significant here as its impact on innovation has already been

captured by innovation expenditures. Moreover, the return to innovation expenditures (in terms of

innovation output) is almost equal to the corresponding estimate of Polder (2008). Comparable

results are also found for the estimates of labour and process innovation (both insignificant). Thus,

conditional on being innovative, size no longer matters for explaining differences in knowledge

output.

The final part of the model consist of the link between productivity and innovation output. Again,

the results are very comparable in magnitude to those reported earlier, although the significance is

slightly weaker. Thus, taken on the whole, the results of model I of Table 12.2 are fairly well in line

with the Dutch contribution of the OECD benchmark project reported in Polder (2008).

12.7. The contribution of ICT to innovation Returning to the main research issues, we can look at the estimates of ICT variables for answering

the question whether ICT use has an impact on productivity over and above other innovation inputs.

Starting with the estimates for broadband connectivity (DSLpct), the results of model I show that the

‘direct’ productivity contribution of broadband connectivity is positive, but only weakly significant.

However, it can also be verified that its productivity impact is also channelled through innovation: the

estimate of DSLpct in the innovation output equation of Model I is rather sizable and significant.

This result suggests that if broadband penetration increases by 10%, then innovative sales per

employee would increase by approximately 3% (ceteris paribus).11

Furthermore, it can be seen that electronic sales also adds explanatory power for innovation output.

It should be noted that the vast majority of firms do not apply Ecommerce activities at all. Thus, the

average share of electronic sales in the sample is still rather small in the Netherlands.12 Nevertheless,

the estimates reported in Table 12.2 imply that enhancing innovation activities by electronic selling

increases the probability of innovation success significantly and that achieving higher levels of ICT

maturity increases E- sales. In particular, electronic links with ICT systems of customers appears to

contribute to E-selling (see the estimate for fase_5 of the E-sales equation of the system).

10 This study concerns the Dutch contribution to the OECD micro data benchmark project and which uses CIS4 data (covering 2002 – 2004) for manufacturing,. By contrast, this Chapter uses two CIS waves and covers all industries. 11 The mean value of DSLpct was about 0.5 in 2004. 12 The share of electronic sales in total sales is 8% for the data used in model I and 5% for the data used in model II.

232


The importance of ICT use for explaining differences in innovation success is confirmed in model

II. Contrary to model I, the emphasis in this model lies on the relation between the two ICT use

variables and innovation outputs and not on the role of innovation expenditures for being successful in

innovation. After all, defining innovativeness as being engaged in R&D (by far the most important

part of innovation expenditures) may be too restrictive if innovation imitation becomes more

important. Ignoring innovation expenditures might offer an explanation for the weak significance of

the productivity contribution of innovation output found for model II. By contrast, the results of this

model also show that the direct contribution to productivity of broadband connectivity is considerably

higher than in model I, whereas the contribution to innovation output is lower, although remaining

significant.13 According to this model, an increase of broadband enabled workers of 10% would

increase the share of innovation sales in total sales by ~0.7%. Taking into account the average value

for this innovation output indicator (in the estimation sample close to 10%), this is a rather substantial

outcome. Moreover, the contribution of electronic sales is even higher. Doubling the share of

electronic sales (from 5% to 10% of total sales) would increase innovation output by about 3%!

12.8. A comparison with Swedish results It is interesting to compare the results for the Netherlands with those found in the Swedish study of

Hagén et al. (2008) which, to our knowledge, is the first to include ICT in the CDM innovation

model. There are some differences in implementation that are worth mentioning:

• Broadband connectivity has been implemented with the help of a binary indicator in the Swedish study and this indicator is only used in the innovation output equation of the model;

• The Swedish study controls for skill biases of productivity estimates by including a Human Capital indicator;

• Ecommerce variables are used for the construction of an ICT integration indicator and this indicator is included in the innovation input equation;

• Due to very small overlap between EC and CIS surveys, the Swedish study could only use 209 (mostly larger) firms.

Table 12.3 shows results for the innovation output and productivity equations. The general picture

is that many impacts are much higher for Sweden than for the Netherlands. The only exception is the

return to innovation investment in term of innovation output, which appears to be only weakly

significant in the Swedish case. However, taking into account the high Swedish elasticity of

innovation output in the productivity equation, the returns to innovation investment in terms of output

gains are comparable.

13 It should be noted that the value of DSLpct in the productivity equation of model II might be overstated as we do not control for ICT use related skills (see Chapter 8 of this report).

233


Table 12.3: Comparing Swedish and Dutch results Sweden (N = 209) Netherlands (N = 1036) (1) Productivity Est. Sign.1 Est. Sign.1

Innovative sales per employee 0.44 a 0.14 bCapital intensity 0.22 a 0.21 aLabour 0.30 a 0.06 aBroadband connectivity na2 0.10 bHuman capital indicator 0.75 a na2 (2) Innovative sales per employee Labour 0.58 a -0.01 Innovation expenditure per employee 0.17 c 0.44 aBroadband connectivity 0.63 a 0.29 aCooperation na2 0.03 Process innovation (improved production methods) 0.55 a -0.06

1a: significant at the 1% level; b: significant at the 5% level; c: significant at the 10% level. 2na: variable not used in the estimations.

Another striking difference is the implied scale estimate (the coefficient of labour in the

productivity equation). The remarkably high estimate found for Sweden suggests that, after

controlling for skill differences, much of the productivity differences comes from increasing returns to

scale in labour and capital. Moreover, the Swedish estimate of labour in the innovation output

equation is also very sizable, whereas it is zero for the Netherlands. These differences might be due

to the selectivity of the Swedish sample. Finally, the difference for broadband connectivity may be

due to the fact that Swedish application uses a binary indicator instead of the continuous measure

DSLpct. Thus, cutting through these results, it can be concluded that direction of impacts is similar in

both countries and that differences in magnitude are related to selectivity and differences in

implementation of impact variables.

234


Appendix A: Description of variables1

Variables Description Source2

I) Endogenous variables Productivity (L) Valued added (factor costs) per employed person PS

Innovation output 1 (L) Innovative sales per employee = (turnin*total sales)/employed persons CIS/PS

Innovation output 2 Share of innovative sales in total sales (turnin)3 CISInputs into innovation (L) Total innovation expenditures per employed person CIS/PSEsalespct Share of electronic sales in total sales EC II) Exogenous variables Capital intensity (L) Depreciation costs per employed person PSLabour (L) Employed persons (in FTE) PS Epurchpct Share of electronic purchases in total purchases ECDSLpct Broadband intensity use indicator = DSL*PCpct ECPCpct Share of employed persons that use computers ECDSL Binary indicator for using broadband connections EC_Ifase_0 No external ICT connections (reference) EC_Ifase_1 External connections no Website EC_Ifase_2 External connections, only Website EC_Ifase_3 External connections, Website and Ecommerce EC_Ifase_4 As _Ifase_3 + electronic after sales services EC

_Ifase_5 As _Ifase_4 + ICT links with ICT systems of suppliers and customers EC

Process innovation Binary indicator for implementing improved production methods . CISCooperation Binary indicator for innovation cooperation CISR&D permanent Binary indicator for performing R&D on a permanent basis CIS III) Other Industry dummies Manufacturing (SIC2 15 - 37, reference) PS Construction (SIC2 45) PS Distributive services (SIC2 50 - 62, SIC2 60-64) PS Other services (SCI2 73, 74, 93) PSTime dummies Dummy for CIS edition 2000 - 2002 (reference) or 2002 - 2004 CIS

1 (L) refers to logarithmic transformation of variables; 2 PS: Production Survey; EC: Ecommerce Survey; CIS: Innovation Survey; 3 A value of zero has been imputed for firms that did not produce product or process innovations.

235


Appendix B: Estimates of Heckman models for innovation inputs and innovation output1

Innov. expenditure per employee Innov. sales per employee Outcome equations1 Est. T Est. T Labour -0.49 -10.8 -0.06 -0.5Esalespct 0.25 1.2Epurchpct 0.58 2.0DSLpct 0.43 1.8Process innovation 0.03 0.5RD permanent 0.88 9.7 0.17 2.1Cooperation 0.50 5.8 0.25 3.4Industry dummy yes yes Time dummy yes yes Stage of ICT maturity yes yes _cons 4.21 12.4 3.50 2.7N 1036 1036 Probit part1 Labour 0.22 14.8 0.22 14.5Esalespct 0.14 1.4 0.16 1.3Epurchpct -0.11 -0.8 -0.11 -0.7DSLpct 0.49 9.9 0.46 8.1Industry dummy yes yes Time dummy yes yes Stage of ICT maturity yes yes _cons -1.61 -19.8 -1.64 -19.5Rho -0.81 -10.6 0.01 0.0Sigma 1.86 1.13 Lambda -1.51 0.01 N 6707 6707 Log L -4225.6 -4068.8 P> Chi2 LR 0.00 0.99 1 The estimates of the equations of interest (for inputs into innovation and innovation output) are estimated simultaneously with the selection equations (the Probit part of the Heckman model). The latter equations determine the importance of ICT use variables for being selected in the model that explains differences in innovation inputs and innovation output.

236


Appendix C: A Note on Causality

12C.1. Introduction Recent studies have focused on the complementary aspect of IT investment and innovative

activity. Firms that invest in IT but also undertake organisational restructuring or invest in innovation

reap greater benefits than firms that only implement changes along one dimension14.

Although this line of thinking is now well established in the micro-econometrics literature on ICT

and innovation an issue that has been left largely un-tackled is the nature of causality. The preceding

analysis shows how e-sales is not only a specific form of ICT use, but that this ICT use also manifests

itself in the form of an innovation input. Closer inspection of Equation (3) in Table 12.2 shows that

firms are likely to have higher e-sales intensity if they are in the third or higher stage of ICT maturity,

which then translates into higher innovation output.

Also shown is both the direct and indirect impact of DSLPCT on innovation output. Indirectly

DSLPCT impacts the innovation input e-sales, which reinforces the direct impact. Is it these

underlying investments in ICT, captured for example by DSLPCT and ICT maturity indicators, that

allow firms to successfully innovate in today’s competitive markets. Or is it because firms are

innovating in other dimensions, and this guides their decisions to improve and innovate in their choice

of ICT investment and use?

12C.2. Question The modelling strategy presented in this Chapter implicitly assumes a direction of causality that

runs from ICT investment to productivity through innovation. ICT investment reinforces the

innovation process, by acting as an innovation input, but retains its explanatory power in the

productivity equation, over and above that explained by the innovation output.

In this appendix, using DSLPCT as our chosen measure of ICT investment, we refocus the

direction of causality and ask the question: “Do innovative firms invest more than average in

DSLPCT investment, and if so, do certain types of innovative activity warrant more intensive

DSLPCT investment?”

12C.3. Methodology We use responses from the UK CIS4 survey. All questions in the CIS4 survey refer to innovative

activities over the period 2002-2004. We therefore merge the CIS4 survey with all survey responses

for the period 2002-2005 from the E-commerce survey. This allows us to analyse a larger merged

sample with a longitudinal element – as opposed to the standard cross-sectional studies.

14 Crespi, Criscuolo & Haskel (2007) provide a treatment of this using UK data. US work along the same lines includes Bresnahan, Brynjolfsson & Hitt (2002), Ichniowski & Shaw (2003) and Black and Lynch (2003).

237


Based on this merged sample we calculate the average DSLPCT level for a firm between 2001-05,

and regress it against CIS indicators. We control for the initial firm specific “stock” of DSLPCT on

the right hand side of the regression i.e. the DSLPCT response returned by a firm in the E-commerce

survey the first time it was sampled.

The idea is simple: controlling for the initial DSLPCT a higher average DSLPCT indicates growth

in the % of employees using high-speed internet. Then we want to see whether a firm changed its

DSLPCT stock in conjunction with any innovative activity.

The basic regression specification is the following:

AVDSLPCT(2001-2005)=β0DSLPCT(INITIAL) + β’C (CIS INDICATORS) + γ’(EC CONTROLS) + ε

Where (CIS INDICATORS) is a vector of chosen indicators from the CIS and β’C is the vector of

coefficients, and (EC CONTROLS) is the vector of controls taken from the E-commerce survey and γ’

is the vector of control coefficients.

EC controls are retained in this arrangement to capture any firm level propensity to invest and

experiment in ICT use not directly correlated with innovation.

Initially we considered almost all responses from the four main sections of the CIS survey:

(1) The nature of product and process innovation i.e. whether it was new to enterprise or new to market, whether it was carried out alone or in conjunction with other enterprises, whether it was abandoned or not.

(2) R&D and knowledge acquisition: whether this was internal to the firm or sourced from external sources, including the type of external sources e.g. clients, competitors or consultants

(3) The geographical source of cooperation i.e. whether it is regional, within national borders or international.

(4) Advanced management practices, including, organisational change, marketing and corporate strategy.

(5) Skills indicators from the CIS survey

Many of the responses are highly correlated and it is not meaningful to include all as explanatory

variables in the regression specification, therefore, we introduced some aggregated variables into the

model.

For example, we capture product innovation on a scale of 1-3. Innovation that is new to the market

is ranked highest at 3. Firms that did not product innovate ranked lowest. Innovation that is new to

the enterprise but not new to the market lies in between. Service innovation is retained separately.

The section on information sharing is very detailed and covers not only the different institutions

with who information is shared but also their individual geographic locations. We condense all this

information into one firm specific variable that does not distinguish between the types of institution

but does retain the geographical element, the variable “international communication” is categorised

according to whether there is any information sharing at all, and whether this information sharing is

national or international. Firms that have international knowledge sharing partners are ranked highest

on a scale of 1-3 and firms that share no information lowest.

All other variables, apart from the obvious % variables, are binary (0,1) indicators.

238


239

12C.4. Results The analysis uses a total of 2501 observations from the CIS4 and E-commerce overlap.

Conditional on the initial stock of DSLPCT, and the main predictors of IT capital (as identified in

Chapter 10, notably: Intranet, Extranet, WAN, e-trade intensity and business integration) firms that

are more innovative have higher stocks of DSLPCT.

Column 1 of the regression table presents results based on simple OLS analysis. Column 2 adds

firm ownership i.e. multinational status as an extra control. Most of the results are robust to the

inclusion of this control.

It should not come as a surprise that the two measures of skills are significantly correlated with

DSLPCT growth, although the coefficients are of small magnitude. Also firms that report

implementing organisational changes during 2002-2004 have higher stocks of DSLPCT, on average

by 2%.

Service innovation is significantly correlated and of reasonable magnitude – firms that were active

in service innovation during 2002-2004 increased their DSLpct stock on average by 3%. Also, the

nature of the innovation, i.e. whether it is new to the enterprise or market is related to DSLPCT

investment.

Another strong result to come out of the analysis is firms that collaborate with external resources

increase their stock of DSLPCT:

• Firms that actively acquired R&D from external sources, during 2002-04, increased their stock of DSLpct by 2%.

• Firms that actively sourced information from consultants, during 2002-04, increased their stock of DSLpct by 1.7%.

• Firms that developed products mainly through collaboration, during 2002-04, increased their stock of DSLpct by 5%.

• Firms actively undertaking international communication, during 2002-04, increased their stock of DSLpct by 2.8%.

Surprisingly, firms that source their innovative information from suppliers and competitors do not

invest as much in DSLPCT. But this probably hints at the fact that for these cases the medium of

information acquisition is different. It might also reflect that fact these firms are not the most

innovative overall and operate at a lower level on the technology frontier.

12C.5. Discussion The analysis shows there is evidence from UK firm-level data that success of firms in gathering

new ideas for innovation from outside the value chain (i.e. not from customers or suppliers) is

associated with a higher proportion of high speed internet enabled employees.

Greater connectivity, through a high-speed internet enabled workforce, allows firms to exploit

knowledge network externalities originating outside the firm. The decision to build and maintain

knowledge networks may be a key driver explaining the need to invest in certain types of ICT, such as


240

DSLpct. An increased knowledge of the market facilitated through information sharing with

consultants and international partners may impact innovation inputs such as e-sales, but also have a

direct effect on innovation output.

To this extent DSLpct should be considered a reasonable proxy for knowledge creation,

management and sharing. A possible extension to the CDM model presented in the main body of this

Chapter, would be to endogenise the investment decision in DSLpct by including an additional

equation in the model that captures the relationship between DSLpct and knowledge flows within the

firm.

Table 12C.1. DSLPCT DSLPCT Acquisition of R&D 0.021 0.021 (0.007)*** (0.007)*** Prodinv 0.006 0.004 (0.003)* (0.003) Information sourced from suppliers -0.025 -0.025 (0.004)*** (0.004)*** Information sourced from competitors -0.008 -0.007 (0.004)** (0.004)* Information sourced from consultants 0.018 0.017 (0.004)*** (0.004)*** Implementing organisational changes 0.023 0.023 (0.007)*** (0.007)*** Service Innovation 0.030 0.030 (0.007)*** (0.007)*** Product developed mainly by collaboration with other enterprises 0.048 0.050 (0.011)*** (0.012)*** % employees with science and engineering degrees 0.001 0.001 (0.000)*** (0.000)*** % employees with other degrees 0.001 0.001 (0.000)*** (0.000)*** International Communication 0.013 0.014 (0.004)*** (0.005)*** EC controls + + ()*** ()*** Ownership Controls NO YES Observations 2501 2501 R-squared 0.78 0.79 Standard errors in parentheses * significant at 10%; ** significant at 5%; *** significant at 1%


Chapter 13

IT Outsourcing in Finnish Business

Mika Maliranta (ETLA, The Research Institute of the Finnish Economy) Petri Rouvinen (ETLA, The Research Institute of the Finnish Economy) Aarno Airaksinen (Statistics Finland) ABSTRACT: This paper reviews the characteristics and magnitude of information technology (IT) outsourcing as well as studies its labour productivity effects with a representative sample of Finnish businesses. Depending on the IT task in question, on average from one-third to two-thirds of IT has been outsourced. Of the ten categories considered, the development of non-Internet business-to-business applications (e.g., EDI) is the leading activity in this respect. The various dimensions of IT outsourcing are all highly positively correlated. After controlling for industry and regional effects as well as characteristics of firms and their employees, it is found that an externally-supported computer user is about 20% more productive than an otherwise similar worker without a computer. Which corresponds to about 5% output elasticity of outsourced IT; the effect of internally-supported computer use is not statistically significantly different for zero, and it is also several times smaller in magnitude. While the issues of causality, timing, self-selection, and unobserved firm heterogeneity are not fully addressed, the findings nevertheless suggest that IT outsourcing may have significant economic consequences.

241


13.1 Introduction Despite the fact that firms increasingly outsource some or all of their (routine) information

technology (IT) tasks, its economic consequences remain ill-understood. In this paper we

employ Statistics Finland’s IT and e-commerce survey to empirically study the firm-level

productivity effects of IT outsourcing. We find that IT outsourcing enhances an

organization’s IT use and thus also boosts its labour productivity.

In the new millennium the dominant fad in management strategy has been to

concentrate on one’s core competences. Faced with intensifying global competition, firms

are increasingly forced to cut production costs. With the prolonged economic upswing in

Finland, firms feel that they are unable to attract sufficiently qualified employees. All of

these observations have made IT outsourcing attractive to Finnish firms. Furthermore,

changes in both domestic and international supply of IT services have made them more

accessible and affordable. Thus, it is no wonder that in a survey by the leading Finnish IT

trade journal Tietokone (January 2008 issue, page 9), more than half of the top management

in Finnish firms – and more than one-third of IT management – agrees that it is their firm’s

priority to outsource as much IT as possible.

In their review Mahnke, Overby, and Vang (2005) come to the conclusion that three

main theoretical perspectives underlie empirical work on IT outsourcing: (1) transaction cost

economics as well as (2) capacity-based and (3) relational views. The literature suggests that

the primary reasons for outsourcing are to lower cost, to gain flexibility, to optimize financial

structure, to seek outside IT expertise, to gain strategic advantage (to focus on core

competences, to facilitate mergers and acquisitions, to reduce time-to-market, to circumvent

problems in attracting IT talent), and to respond to political pressures (dissatisfaction with

internal IT, pressure from vendors, desire to imitate competitors etc). The risks identified in

the literature include the loss of absorptive capacity and/or control, declining innovativeness,

deteriorating performance, increasing transaction or hidden costs, leaking out of business

secrets, as well as immediate and/or subsequent de-motivation of the employees.1

1 Heshmati (2003) provides a survey of the more general literature on outsourcing and its relationship to efficiency and productivity. Siegel and Griliches (1992), Fixler and Siegel (1999), Görg and Hanley (2005), and Görg, Hanley, and Strobl (2008) are among the papers considering the relationship of productivity and services outsourcing in general. Abraham and Taylor (1996) suggest that a firm’s decision to contract out business support is influenced by the associated savings, the volatility of output demand, and skill availability of the outside contractor. In the model by Grossman and Helpman (2002) the equilibrium level of outsourcing is determined by firms’ trade-off between the relatively high governance costs of integrated firms as opposed to the search costs for partners by outsourcing firms. Domberger et al. (1995) study competitive tendering of

242


Thouin, Hoffman, and Ford (2008) study the effects of IT budget, outsourcing, and

(internal) personnel on financial performance of integrated health care delivery systems.

They find that higher IT expenditures and outsourcing intensity are associated with better

profitability. Florin, Bradford, and Pagach (2005) study investors’ reactions to IT

outsourcing. IT outsourcing announcements are positively associated with short-term

abnormal returns but long-term returns become negative due to organizational restructuring

efforts. Domberger, Fernandez, and Fiebig (2000) suggest that there might be learning in IT

outsourcing: first-term contracts tend to be more expensive than subsequent contracts. They

also find that competitive bidding – as opposed to directly negotiated contracts – does not

lead to lower prices but it is associated with better performance, which they attribute to the

clients better-defined expectations and requirements. Mahnke, Özcan, and Overby (2006)

suggest that timing might be important for IT outsourcing also from another perspective: the

governance choices with respect to IT are influenced by firms’ attempts to position

themselves as early- or late-movers in varying technological regimes. Bertschek and Müller

(2006) use a semi-parametric endogenous switching model to study IT outsourcing. While IT

outsourcing does not seem to make firms too different in observed dimensions (partial

production elasticities of key inputs), firms without IT outsourcing produce more efficiently

than those with IT outsourcing, which they attribute to coordination costs.2 Knittel and

Stango (2008) examine the effect of IT outsourcing with a panel of US credit unions. They

find that IT outsourcing has significant productivity benefits primarily towards the end of

their 1992–2005 observation period. The effect is present only when studied within-firm and

switching to outsourcing is endogenous. In cross-section they find that less productive firms

are more likely to outsource. Clayton (2005), along with eight associated researchers, touches

upon the issue of IT outsourcing. Findings suggest that IT outsourcing is complementary to

IT capital. They do not, however, find that IT outsourcing would be a significant determinant

of productivity when included in a regression with IT investment.3

service provision with data on cleaning contracts: they find that competition reduced prices while the quality was maintained or even enhanced. 2 Abramovsky and Griffith (2006) suggest that information and communication technology (ICT) in itself has consequences on a firm's choice to outsource (regardless of type): they find that more ICT-intensive firms outsource more and that they are also more likely to off-shore. Bardhan et al. (2006) come to a similar conclusion; they also find that outsourcing of production processes is associated with lower cost and higher quality. Bartel, Lach, and Sicherman (2005) present a model formalizing these observations. 3 In what follows, we do not measure IT investment per se but rather IT use directly, which is a function of prior and current IT investment as well as rented, leased, or similarly acquired equipment (as a part of the outsourcing contract the provider typically exploits its own IT capital for which the customer is not billed separately). Furthermore, in our regression setting we do not consider outsourced IT services that are best characterized as investment as opposed to purchases for immediate consumption.

243


13.2 Data Our data on IT originates from Statistics Finland’s Information technology use and electronic

commerce in enterprises 2005 -survey4 conducted in the spring of 2005 with some questions

referring to the time of the survey and some to the statistical year of 2004. The survey’s

question 7.1 is: To what extent are the following information technology functions performed

by your firm’s own (hired) labour / outside labour?5 Answers are requested in the following

ten categories:

(a) Design/development of Internet homepages,6 (b) Maintenance of Internet homepages,7 (c) Internet marketplace for private/retail customers,8 (d) Internet or extranet marketplace for businesses,9 (e) Other business to business commerce application (for example EDI),10 (f) Development and maintenance of applications,11 (g) Development of other information technology systems,12 (h) Operation/maintenance of servers,13 (i) Operation/maintenance of a PC environment,14 and (j) User support,15

with the following six mutually exclusive answering alternatives given for each: Completely

performed by external labour,16 Mostly performed by external labour,17 Equally performed

by external and own labour,18 Mostly performed by own labour,19 Completely performed by

own labour,20 and I am unable to say / Irrelevant.21

We will discuss IT outsourcing in light of the answers provided to these questions. It

should be noted that some categories refer to the external purchases of investment inputs –

expected to deliver a flow of service used internally and/or contribute to goods and/or

services sold externally in the future – while some refer to the external purchases of services

4 In Finnish: Tietotekniikka ja sähköinen kauppa yrityksissä 2005. 5 In Finnish: Missä määrin seuraavat tietotekniikkatoiminnot tehdään yrityksenne omalla työvoimalla / ulkopuolisella työvoimalla? 6 In Finnish: Internet-kotisivujen suunnittelu/kehittäminen. 7 In Finnish: Internet-kotisivujen ylläpito. 8 In Finnish: Internet-kauppapaikka yksityis/vähittäisasiakkaille. 9 In Finnish: Internet tai extranet kauppapaikka yrityksille. 10 In Finnish: Muu yritystenvälisen liiketoiminnan sovellus (esim. EDI). 11 In Finnish: Sovellusten kehittäminen ja ylläpito. 12 In Finnish: Muu tietotekniikkasysteemien kehittäminen. 13 In Finnish: Palvelinten käyttö/ylläpito. 14 In Finnish: PC-ympäristön käyttö/ylläpito. 15 In Finnish: Käyttäjätuki. 16 In Finnish: Kokonaan ulkopuolisella työvoimalla. 17 In Finnish: Pääosin ulkopuolisella työvoimalla. 18 In Finnish: Yhtä paljon ulkopuolisella ja omalla työvoimalla. 19 In Finnish: Pääosin omalla työvoimalla. 20 In Finnish: Kokonaan omalla työvoimalla. 21 In Finnish: En osaa sanoa / Ei relevantti.

244


for immediate consumption. Of the above categories (a) and (g) as well as to a lesser extent

(c), (d), and (e) may be considered investment; categories (b), (h), (i), and (j) may be

considered consumption; category (f) involves both. The usefulness of (f) is hindered by the

fact that it is unclear. For example what are the respective roles of packaged software and

firm-specific solutions?

The structure of the IT outsourcing section suggests a continuous coding of the

relevant questions: outsourcing goes from 0 (completely internal) to 100% (completely

external) with a clearly identified mid-point of 50% (equally). While the mostly internal and

external alternatives are less clear cut, it is not unreasonable to code these respectively as

25% and 75% outsourcing. The analysis below is based on this choice of coding.

For the non-IT variables we use values referring to the statistical year of 2005.22 We

include firms with at least ten employees in manufacturing (D, Nace 2002 industry codes 15–

37); electricity, gas and water supply (E, code 40); construction (F, code 45); and private

services (F–K, codes 50–74) in which at least some employees use computers at their work.

Furthermore, the firm’s capital stock, labour input, and value added must be observed. As

also “importance” (employment) weighted figures are considered, we exclude 18 firms with

over 2,000 employees. After these restrictions we are left with a usable cross-section of

1,839 firms.

IT outsourcing is widespread (see Table 13.1): depending on the category (and

conditional on considering the question relevant and being able to answer), on average the

sample firms have outsourced 40–66% (weighted: 37–64%) of the IT activity in question.

Most firms (56–74%; weighted: 50–83%) bi-source (Du, Lu, and Tao, 2006), i.e., both

purchase externally and provide internally the service in question; 6–29% (weighted: 6–36%)

of the sample firms perform the IT activity in question completely internally and 14–31%

(weighted: 10–28%) completely externally. On average 17–93% of the sample firms

consider the various IT outsourcing questions relevant (depends on the employed IT system)

and are able to answer.

22 The details on firm-level employment (educational structure etc.) refer to the end of year 2004.

245


Table 13.1: Unweighted and weighted distributions of the IT outsourcing responses.

Sh. of those Conditional of being able to answer andUnweighted answering / considering relevant, the extent of

considering outsourcing (Mean and X% outsourced):relevant (%) Mean 0% 25% 50% 75% 100%

(a) Design/development of Internet homepages 17 18 32 86 56 15 18

(b) Maintenance of Internet homepages 26 10 19 86 46 25 19

(c) Internet marketplace for retail customers 22 13 24 17 52 19 22

(d) Internet marketplace for business customers 19 16 29 22 52 18 18

(e) Other business to business commerce appl. 13 14 34 28 66 8 31

14 16 39 (f) Development and maintenance of applications 84 66 6 25

(g) Development of other IT systems 26 20 28 86 56 8 18

(h) Operation/maintenance of servers 26 13 21 91 50 19 21

(i) Operation/maintenance of a PC environment 28 12 19 93 40 28 13

(j) User support 24 14 19 93 41 29 14

Weighted Asnw./relev. Mean 0% 25% 50% 75% 100%

19 19 32 (a) Design/development of Internet homepages 86 54 14 16

31 9 19 (b) Maintenance of Internet homepages 86 43 24 17

26 10 25 (c) Internet marketplace for retail customers 17 49 20 19

20 22 30 (d) Internet marketplace for business customers 22 51 16 13

16 15 33 (e) Other business to business commerce appl. 28 64 8 28

14 18 42 (f) Development and maintenance of applications 84 63 6 19

30 25 29 (g) Development of other IT systems 86 51 7 10

(h) Operation/maintenance of servers 29 11 19 91 46 22 19

29 8 18 (i) Operation/maintenance of a PC environment 93 39 31 15

25 10 15 (j) User support 93 37 36 14

The most frequently answered question concerns the outsourcing of user support (see

Table 13.1): 1,705 of the 1,839 firms (93%) consider the question relevant and are able to

provide an answer. Given that by construction at least some employees of the sample firms

use computers at their work, obviously all should have considered the question relevant if

understood as intended. The other questions may be conditional on the specific features of a

firm’s IT infrastructure. Thus, while it seems that relatively few answers are provided

regarding the outsourcing of online (c) retail or (d) business-to-business sales, some eighty

per cent of the firms having online sales do provide an answer.

246


247

Table 13.2. Unweighted and weighted correlation coefficients of the IT outsourcing variables.

Note: * indicates statistical significance at 1% level. The number of observations below the coefficient (in small print).

Various types of IT outsourcing are highly correlated (see Table 13.2). The

correlations among the Internet and e-commerce outsourcing measures (a, b, c, and d) is .56–

.87 (weighted: .51–.87). In application (f) and system development (g) the correlation is .69

(weighted: .65). As for the computer network as well as its terminals and users (h, i, and j),

the correlation is .75–.84 (weighted: .73–.88). Other business-to-business commerce

application (e) is highly correlated with both e-commerce measures (c and d). Internet or

extranet marketplace for businesses (d) is correlated with both development measures (f and

g).

Unweighted (a) (b) (c) (d) (e) (f) (g) (h) (i) (j)

(a) Design/development of Internet homepages 1.001,587

(b) Maintenance of Internet homepages .65 * 1.001,569 1,581

(c) Internet marketplace for retail customers .62 * .62 * 1.00317 318 320

(d) Internet marketplace for business customers .56 * .58 * .87 * 1.00383 387 240 396

(e) Other business to business commerce appl. .39 * .34 * .65 * .72 * 1.00461 465 158 217 506

(f) Development and maintenance of applications .40 * .36 * .55 * .59 * .62 * 1.001,399 1,397 304 382 501 1,550

(g) Development of other IT systems .39 * .36 * .41 * .48 * .44 * .69 * 1.00 1,420 1,415 309 384 491 1,484 1,581

(h) Operation/maintenance of servers .31 * .31 * .31 * .35 * .33 * .46 * .66 *1,547

1.00 1,505 1,504 317 395 503 1,514 1,674

(i) Operation/maintenance of a PC environment 0.30 * 0.30 * 0.31 * 0.32 * 0.27 * 0.41 * .59 *1,552

.82 *1,643

1.001,516 1,514 313 391 502 1,518 1,704

(j) User support 0.30 * 0.30 * 0.33 * 0.32 * 0.24 * 0.41 * .59 * .75 * .84 * 1.001,518 1,516 315 390 501 1,511 1,549 1,637 1,663 1,705

Weighted (a) (b) (c) (d) (e) (f) (g) (h) (i) (j)

(a) Design/development of Internet homepages 1.001,587

(b) Maintenance of Internet homepages .58 * 1.001,569 1,581

(c) Internet marketplace for retail customers .59 * .61 * 1.00317 318 320

(d) Internet marketplace for business customers .56 * .51 * .87 * 1.00383 387 240 396

(e) Other business to business commerce appl. .42 * .40 * .67 * .64 * 1.00461 465 158 217 506

(f) Development and maintenance of applications .43 * .38 * .57 * .58 * .67 * 1.001,399 1,397 304 382 501 1,550

(g) Development of other IT systems .41 * .36 * .34 * .42 * .44 * .65 * 1.00 1,420 1,415 309 384 491 1,484 1,581

(h) Operation/maintenance of servers .28 * .27 * .17 * .34 * .32 * .39 * .60 *1,547

1.00 1,505 1,504 317 395 503 1,514 1,674

(i) Operation/maintenance of a PC environment 0.21 * 0.20 * 0.18 * 0.28 * 0.23 * 0.30 * .50 *1,552

.79 *1,643

1.001,516 1,514 313 391 502 1,518 1,704

(j) User support 0.22 * 0.22 * 0.18 * 0.24 * 0.23 * 0.29 * .50 *1,549

.73 *1,637

.88 * 1.001,518 1,516 315 390 501 1,511 1,663 1,705


248

Table 13.3. Unweighted and weighted descriptive statistics.

Note: Underlining indicates that the pair wise correlation in question is statistically significance at 1% level.

Table 13.3 reports the descriptive statistics of the variables used in the analysis

(besides the eighteen industry and twenty regional dummies) along with their correlations

with the IT measures.

Unweighted

.44

Obs.

.65

Mean

1

St. dev.

IT: (h) Operation/maintenance of servers

Min.

1,674

Max.

.460

IT

.363

(a)

0

(b)

1

(c)

-.12

(d)

.28

(e)

.27

(f)

.17

(g)

.34

(h)

.32

(i)

.39

(j)

.60

IT: Share of employees using a computer at work

1

1,839

IT: (i) Operation/maintenance of a PC environment

.666

1,704

.352

.389

.010

.363

1

0

1

1

IT: (a) Design/development of Internet homepages

-.09

1,587

.21

.555

.20

.332

.18

0

.28

1

.23

-.17

.30

1

.50

IT: (b) Maintenance of Internet homepages

.79

1,581

1

.458

IT: (j) User support

.370

1,705

0

.368

1

.364

-.20

0

.65

1

1

-.07

IT: (c) Internet marketplace for retail customers

.22

320

.22

.518

.18

.362

.24

0

.23

1

.29

-.11

.50

.62

.73

.62

.88

1

1

IT: (d) Internet marketplace for business customer

CD: ln(real value added / labor input)

sIT: (e) Other business to business commerce appl.

-.10

396

1,839

.520

10.96

.346

0.57

0

5.51

1

14.22

-.11

.24

.56

-.07

.58

-.05

.87

-.14

1

-.11

506

.01

.664

-.02

.317

.03

0

.12

1

.09

-.07

CD: ln(real capital stock / labor input)

.39

1,839

.34

10.48

.65

-1.77

.72

3.35

1

17.88

IT: (f) Development and maintenance of pplications

.12

a

IT: (g) Development of other IT systems

5.64

1,550

.03

.655

.00

.295

.10

0

.11

1

-.02

-.16

.10

.40

-.01

.36

-.01

.55

.04

.59

.02

.62

CD: ln(labor input)

1

1,839

1,581

1.16

.559

2.30

.308

7.55

0

.16

1

-.06

-.18

-.05

.39

-.18

.36

-.08

.41

-.07

.48

-.04

.44

-.06

.69

-.02

1

.03

IT: (h) Operation/maintenance of servers

.01

1,674

Firm: Young (established in/after year 2000)

.498

1,839

.362

.180

0

.384

1

0

-.18

1

.31

.02

.31

.00

.31

.06

.35

.18

.33

.03

.46

.12

.66

.04

1

.05

IT: (i) Operation/maintenance of a PC environment

.04

1,704

.04

.403

.03

.351

Firm: Old (established in/before year 1985)

0

1,839

1

.371

-.15

.483

.30

0

.30

1

.31

.00

.32

-.01

.27

-.05

.41

-.08

.59

-.06

.82

-.12

1

-.02

IT: (j) User support

.01

1,705

-.03

.411

-.04

.357

-.07

0

Fim: Multiple establishments (dummy)

1

1,839

-.18

.753

.30

.431

.30

0

.33

1

.32

.15

.24

-.09

.41

-.07

.59

-.07

.75

.00

.84

-.04

1

-.06

CD: ln(real value added / labor input)

-.05

1,839

-.06

10.86

-.04

0.57

-.01

5.51

Ed.: Share of employees with bachelor or eq. techn. e

14.22

1,770

.20

.056

-.03

.058

-.01

0

-.07

.800

-.04

.17

-.05

-.16

-.04

-.10

-.06

.02

-.03

-.01

.01

-.12

-.02

-.06

CD: ln(real capital stock / labor input)

-.12

1,839

-.07

10.10

-.01

1.78

-.02

-3.35

d.

Ed.: Share of employees with Master or eq. techn. ed.

17.88

1,770

.05

.065

.09

.087

.05

0

.09

.732

.11

.21

.03

-.16

.05

-.14

.00

-.20

-.01

-.10

.01

-.10

-.02

-.11

CD: ln(labor input)

-.13

1,839

-.07

4.06

.02

1.21

.02

2.30

Ed.: Share of employees with PhD or eq. techn. ed.

7.55

1,770

.14

.040

-.01

.071

-.05

0

.06

.538

.03

.29

-.03

-.18

-.08

-.14

-.16

-.22

-.14

-.17

-.08

-.15

-.19

-.18

Firm: Young (established in/after year 2000)

-.20

1,839

-.07

.147

.03

.355

.01

0

Ed.: Share of employees with bachelor or eq. non-t. e

1

1,770

.05

.099

.03

.082

.05

0

.12

.818

.01

.35

.09

-.07

.02

-.12

.01

.03

.03

-.07

.04

-.13

.01

-.11

Firm: Old (established in/before year 1985)

-.05

1,839

-.05

.363

-.12

.481

-.11

0

d.

Ed.: Share of employees with Master or eq. non-t. .

1

1,770

-.04

.039

.03

.047

.01

0

.01

.538

.07

.28

-.03

-.13

.04

-.07

.03

-.06

.00

-.18

-.01

-.12

-.01

-.17

Fim: Multiple establishments (dummy)

-.08

1,839

-.06

.468

-.05

.499

-.06

0

ed

Ed.: Share of employees with PhD or eq. non-t. ed.

1

1,770

.17

.033

-.05

.072

-.06

0

.07

.755

.06

.25

.02

-.09

-.03

-.08

-.08

-.18

-.08

-.08

-.06

-.15

-.11

-.17

Ed.: Share of employees with bachelor or eq. techn. e

-.14

d.

Ed.: Share of employees with Master or eq. techn. ed.

-.04

1,770

-.07

.049

-.07

.063

-.09

0

Labor: Young (share of those under 3

.800

1,770

.14

.357

-.10

.173

-.07

0

-.06

1

-.03

-.06

-.09

-.05

-.10

.01

-.14

-.08

-.08

-.17

-.06

-.01

-.09

-.11

1,770

-.07

.058

-.14

.096

-.09

0

5)

Labor: Old (share of those at least 45)

.732

1,770

.25

.374

-.12

.153

-.11

0

-.17

1

-.15

-.01

-.09

.12

-.14

.10

-.17

.14

-.14

.25

-.11

.14

-.14

.18

Ed.: Share of employees with PhD or eq. techn. ed.

.13

1,770

.15

.033

.18

.075

.16

0

Labor: Women (share of female employees)

.538

1,770

.30

.349

-.12

.243

-.09

0

-.11

1

-.15

.13

-.12

.00

-.19

.00

-.19

.14

-.13

.00

-.07

.04

-.11

-.01

Ed.: Share of employees with bachelor or eq. non-t. e

.04

d.

Ed.: Share of employees with Master or eq. non-t. . 1,770

-.01

.108

-.11

.105

-.09

0 .818 .41 -.05 -.07 .02 -.02 .02 .00 -.01 -.03 -.06 -.05ed

Ed.: Share of employees with PhD or eq. non-t. ed. 1,770 .041 .059 0 .538 .33 -.10 -.09 -.03 -.10 .00 -.06 -.04 -.06 -.03 -.051,770 .036 .081 0 .755 .31 -.06 -.07 -.06 -.04 -.01 -.10 -.09 -.07 -.04 -.06

Labor: Young (share of those under 3 5)

Labor: Old (share of those at least 45) 1,770 .368 .192 0 1 -.01 .00 -.03 -.02 -.09 -.04 -.09 -.03 .00 -.01 .021,770 .360 .176 0 1 -.09 .07 .09 .07 .16 .08 .13 .08 .06 .06 .04

Labor: Women (share of female employees) 1,770 .341 .262 0 1 .24 -.03 -.06 .05 -.04 .14 .07 .09 .05 .03 .03

Weighted Obs. Mean St. dev. Min. Max. IT (a) (b) (c) (d) (e) (f) (g) (h) (i) (j)

IT: Share of employees using a computer at work 1,839 .734 .305 .010 1 1IT: (a) Design/development of Internet homepages 1,587 .541 .325 0 1 -.17 1IT: (b) Maintenance of Internet homepages 1,581 .434 .362 0 1 -.22 .58 1IT: (c) Internet marketplace for retail customers 320 .492 .360 0 1 -.08 .59 .61 1 IT: (d) Internet marketplace for business customer sIT: (e) Other business to business commerce appl.

396 .509 .319 0 1 -.04 .56 .51 .87 1 506 .645 .317 0 1 -.14 .42 .40 .67 .64 1

IT: (f) Development and maintenance of pplications a

IT: (g) Development of other IT systems 1,550 .634 .286 0 1 -.18 .43 .38 .57 .58 .67 1 1,581 .513 .282 0 1 -.18 .41 .36 .34 .42


13.3 Model While the outsourcing questions are of interest in their own right, the key question is their

relation to firm performance. In what follows, we will briefly consider the productivity

effects of outsourcing in a setting previously employed by Maliranta and Rouvinen

(2004;2006).23

We wish to study the productivity effects of using a computer at work as well as IT

outsourcing at the level of an individual worker. These IT characteristics are, however, only

observed at the firm level, that is, the data are grouped in a certain way. Assuming that

workers are – after controlling for observable individual qualities – reasonably similar, a

firm-level model revealing these effects can nevertheless be devised and estimated.

A Cobb-Douglas production function of firm i can be written as

iK L Zi i i i iY A K L eεβ β= Z β (1)

where Y is net output, is disembodied technology, is capital, is labour, is a vector

of the other relevant firm and individual qualities and

A K L Z

ε is a stochastic error term. Workers

may have different marginal productivities depending on whether or not they use a computer

at work. Let ITL be the number of workers using a computer at work. Adding this to (1)

yields

,1L

iK ZIT ii i i i IT i

i

LY A K L e

L

β

εβ θ

= + Z β , (2)

where ITθ is a parameter capturing the possible additional productivity effects associated with

computer use. Slight manipulation yields a labour productivity specification

( ),ln ln ln 1 1 ln lnIT ii ii K L IT K L i

i i i

LY KA LL L L

ln iβ β θ β β

= + + + + + − + + Z iZβ ε

) i

, (3)

where accounts for deviations from constant returns to scale.

Approximating

( 1 lnK L Lβ β+ −

( )( ),IT i iL Lln with 1 ITθ+ ( ),IT IT i iL Lθ yields

( ),ln ln ln 1 ln lnIT ii ii K L IT K L i

i i i

LY KA LL L L iβ β θ β β ε

≈ + + + + − + +

Z iZβ . (4)

A worker using a computer is supported either by an in-house IT staff or by an

outsider contractor to whom the task has been outsourced. Outsourcing could be

incorporated in suggesting that it makes all factors proportionately more productive. This A

23 See also Olsen (2006).

249


could be the case for instance if IT outsourcing provides strategic advantages.24

Alternatively, and in our opinion more appropriately, one could think that IT outsourcing

might enhance the computer use of those exposed to it, in which case (4) could be re-written

as follows:

( )

( )

,,

ln ln ln

1 ln ln

i ii K

i i

ITout iITin iL ITin L ITout

i i

K L i i

Y KAL L

LLL L

L

β

β θ β θ

β β ε

≈ +

+ +

+ + − + +Z iZβ

, (5)

where represent the computer users supported by internal and by external staff

with

ITinL IToutL

ITinθ and IToutθ as the respective parameters to be estimated.

and , where is the fraction of outsourced user support

(derived from the answers to question (j) discussed in the previous section; naturally the

subsequent analysis is limited to firms that have provided an answer to this question).

= ×ITL OutsourcedIToutL

= × )ITL tsourced−(1 OuITin L Outsourced

25

In what follows, we estimate a model incorporating IT outsourcing with the variables

discussed in the previous section as well as with a stochastic error term.26

24 While we do not consider this the most appropriate way to proceed, we nevertheless estimate a model where the term is included in . Outsourcede A25 As the discussion in the previous section suggest, the IT outsourcing measures largely convey the same information. Introducing them jointly in a regression setting might cause severe multicollinearity problems. Furthermore, immediate consumption and investment call for rather different modelling approaches; externally purchased IT investment inputs are perhaps best likened to R&D (for discussion on externally purchased R&D inputs see Ulset, 1996). We experimented with combining answers to (h), (i), and (j) using Cronbach’s Alpha method. The reliability coefficient of .91, i.e., well above the usually employed cut-off of .70 suggesting that pooling the answers is indeed appropriate. Employing the pooled IT outsourcing measure yields both qualitatively and quantitatively nearly identical results. 26 For instance Greenan and Mairesse (2000) have considered a model similar to ours even though without the IT outsourcing extension. Below disturbances are assumed to be uncorrelated across observations but arbitrary differences in their variances (heteroscedasticity) is allowed.

250


13.4 Results Before proceeding, we eliminate outliers by using the standardized or Pearson residuals: a

preliminary regression is performed and 14 observations with the standardized residuals over

four standard deviations away from the mean are dropped, which with normally distributed

errors would be roughly equivalent to eliminating three out of 100,000 observations.

Detailed employee characteristics (age, education etc. for each individual) are missing for 69

firms; the missing information is coded as being zero but we also include a separate dummy

variable indicating where such replacements have been made.

Table 13.4. Estimation results (dependent variable: ln(value added/labour input)). Employment weighted Unweighted

(4.) (5.) (1.) (2.) (3.) (6.)

IT: Sh. of employees using a computer at work ** .139 .097 * **.063 * .076IT: Outsourcing sh. (outsourced user support (j.)) .139 *** .085 ***IT: Share of computer users × Outsourcing sh. *** .203 ***.135IT: Share of computer users × (1 - Outsourcing sh.) .018.024

.078 *** .087 *** CD: ln(real capital stock / labor input) .087 ***.076 *** .079 *** .079 ***CD: ln(labor input) *** *** .035 .038 .037 ****** *** ***.035 .041 .041Firm: Young (established in/after year 2000) .049 .048 .053.032 .027 .028Firm: Old (established in/before year 1985) -.002 .002 -.005 .001 .001 .004

-.084 *** -.079 ** Fim: Multiple establishments (dummy) -.082 **-.058 ** -.067 *** -.067 ***Ed.: Sh. of employees with Bachelor or eq. techn. ed. *** *** 1.093 1.185 1.202 ****** ** **.522 .496 .495Ed.: Sh. of employees with Master or eq. techn. ed. .532 ** .562 ** .766 *** .782 *** .787 *** .547 **Ed.: Sh. of employees with PhD or eq. techn. ed. *** *** .999 .925 .929 *****.422 .410 * .412 *Ed.: Sh. of employees with Bachelor or eq. non-t. ed. .320 .405 + .417*** ** ** +.360 .372 .376Ed.: Sh. of employees with Master or eq. non-t. ed. .662 ** .558 * .701 *** .803 *** .790 *** .508 *Ed.: Sh. of employees with PhD or eq. non-t. ed. *** *** 1.682 1.763 1.783 ****** *** ***1.192 1.193 1.195

-.235 -.175 Labor: Young (Sh. of those under 35) -.187-.370 *** -.437 *** -.436 ***Labor: Old (Sh. of those at least 45) .064 .032 -.135 -.188 + -.193 + .011Labor: Women (Sh. of female employees) -.389 *** -.359 *** -.361 *** -.415 *** -.413 *** -.358 ***Labor: Missing -.006 .022 -.150 -.208 -.210 .006* *Also including a constant term as well as industry (18 categories) and regional (20 categories) dummies.

Are the two IT intensity outsourcing measures equal (= H0)? F(1, 2207) × 12.56 ***F(1, 2207) ***6.90

Observations: 1,825 1,691 1,825 1,691 1,691 1,691R-squared: 0.50 0.52 0.520.37 0.38 0.38

Note: Estimated with White (1980) heteroskedasticity consistent ordinary least squares in Stata for Windows version 9.2. ***, **, *, and + respectively indicate significance at 1, 5, 10, and 15 per cent level.

Table 13.4 provides both unweighted (Columns 1–3) and employment weighted

(Columns 4–6) heteroskedasticity-consistent (robust) ordinary least squares production

function estimates with the share of computer using employees (IT intensity) as the only IT

measure (Columns 1 and 4), the share of computer users and the IT outsourcing intensity as

the measures (Columns 2 and 5), as well as the estimates, where the IT intensity is split into

251


two by interacting it with the share of outsourcing and one minus the share of outsourcing

(Columns 3 and 6). As the effect of outsourcing should in our opinion be a function of the

firm’s IT intensity and we wish to make interpretations regarding an individual worker, the

rightmost Column 6 in Table 13.4 provides our preferred estimates.27

Before proceeding, a word of caution: The results in Table 13.4 do not take into

account the output elasticity of labour, that is, L ITβ θ (.203 for the Share of Computer users ×

Outsourcing share in Column 6) rather than ITθ (correspondingly .214) is reported.

Furthermore, Table 13.4 does not report output elasticities of ICT, which are more common

in this literature (correspondingly the IT output elasticity for the computer users with

outsourced support is about five per cent). Maliranta and Rouvinen (2006) discuss these

issues in some detail.

After controlling for industry and regional effects as well as labour and other firm-

level characteristics, the “excess” productivity associated with using a computer at work is

little over ten per cent (Column 3), which is inline with our earlier results (Maliranta and

Rouvinen, 2004). Regardless of how IT outsourcing is introduced to the model, it seems to

have a considerable impact: If it is assumed to enhance all factors proportionally, a

completely IT outsourcing firm, as opposed to one with completely internal IT support, seem

to have some ten per cent higher labour productivity (Column 2). If assumed to enhance the

labour input associated with using a computer at work, the difference between internally and

externally supported labour is qualitatively large and statistically significant (Column 6): in

fact it seems that a computer is enhancing a worker’s productivity only if s/he is externally

supported. A worker using an externally supported computer at work is about twenty per

cent more productive than a similar worker without a computer.

27 In one wishes to make interpretations at the level of a firm, Column 3 is preferred.

252


13.5 Discussion In section 13.3 we neither explicitly model firms’ motives and behaviour nor identify the

actual channels of influence when it comes to IT outsourcing. The results in section 13.4 are

partial correlations. They nevertheless suggest IT outsourcing might have considerable

effects. Earlier literature suggests that its effects might be conditional on timing (first-term

vs. subsequent contracts; associated re-organization and coordination costs). Outsourcing

might also have unintended long-run effects (forgone learning associated with internal

provision of IT services; “hollowing out” of the firm etc.). One might also want to take into

account unobserved firm heterogeneity as well as (self-) selection into IT outsourcing before

drawing definite conclusions about its effects.

Even though we control for a number of firm characteristics, in this chapter we have

not considered how the estimated effects might vary by firm characteristics. For example, to

the extent that our findings reflect economies of scale, it might be that smaller firms have

more to gain from IT outsourcing.28

Most likely outsourcing will move at least some IT investment outside the firm, in

which case outsourcing studies using IT capital to measure intensity of use might be biased,

as the benefits of the IT-related productivity gains might be captured by the outsourcing

measure (Knittel and Stango, 2008). This is not, however, an issue in our case as we directly

measure IT use.

28 We thank Tony Clayton for pointing this out.

253


254

13.6 Conclusion Our findings suggest that – despite being little studied and understood – IT outsourcing has

potentially considerable economic consequences. IT outsourcing seems to enhance the

involved organization’s IT use and thus also boost its labour productivity.29 In the longer run

a firm has to consider, from its own point of view, whether IT is just a support

department/function, or is it in fact the glue that binds together everything else?

29 Some caveats of this finding are discussed in Section 13.5.


CHAPTER 14

ICT Impacts: From micro to macro

Eric J. Bartelsman

Vrije Universiteit Amsterdam, Tinbergen Institute, and IZA

14.1 Abstract

This Chapter provides some insight into how the aggregate impact of ICT depends on the effects of ICT at the firm-level as well as on the market interactions between ICT using firms. The Chapter concludes with practical tips for practitioners and with suggestions for future research.

14.2 Introduction

What is the impact of ICT on the economy? This simple question from EU policy makers was the logical follow-up to the observation that production of ICT goods and services in Europe lagged that in the U.S., and varied greatly between smaller northern EU countries and the large mainland countries. Subsequent research, using the growth accounting methodology, was able to parse output growth into contributions from ICT production, and from the use of services provided by ICT-related capital stock (O’Mahony, Timmer, and van Ark, 2008). In recent years, the contribution from ICT use was higher in the U.S. than in the EU. Nonetheless, a large portion of the difference in output growth between the EU and US in recent years, is the results of the acceleration of productivity (TFP, total factor productivity) in the US, and thus a measure of our ignorance according to the growth accounting methodology.

The questions: “why is ICT production and ICT use low in the EU” and “what is the impact of ICT on TFP” thus remains on the table. In the growth accounting framework the answers are: “we do not know” and “no impact”. In part, the inability of the growth accounting framework to provide answers relates to the mismatch in level of analysis between the theory and practice of growth accounting. The theory starts from profit maximizing firms operating in competitive markets but in practice, it is applied to aggregate (or industry-level) data. The mismatch is justified by appealing to a ‘representative firm’. The evidence is now overwhelming (e.g. Dunne, Jensen and Roberts, 2008), that markets are populated by firms that differ in many dimensions, even in industries producing homogeneous products with common technology. Further, these heterogeneous firms interact strategically with other firms in an environment where various deviations from the assumptions of growth accounting (imperfect competition, adjustment frictions, fixed entry costs, non-constant returns to scale) prevent the most productive firm from capturing the whole market.

255


With firm-level data and a model of how individual firms make decisions and of how these firms interact in a market, a much richer set of possible impacts of ICT on the (aggregate) economy may be studied. The widespread use of ICT may affect many aspects, for example it may change the strategic interactions between firms, it may lower the costs with which supply meets demand, it may speed up the process of firms to adopt new technology and it may increase the speed with which the market selects more productive firms and forces exit of less productive firms. Recent research has been developing such models of dynamic industries with heterogeneous firms (e.g. Melitz (2003), Restuccia and Rogerson (2007), Klette and Kortum (2004))

The main ingredients of such models, applied to the question of ICT, would start with firm level decisions to adopt ICT. This decision would depend not only on relative prices of ICT versus other asset types, but also on the expected value of future profits, given market interactions. Next, the impact of ICT use at the firm level on employment, output and productivity needs to be modelled. Finally, the market interactions and resource reallocation across firms (including firm entry and exit) determine the macro outcomes. These ingredients all need empirical counterparts, indicators built up from firm-level decisions, market interactions and aggregate outcomes. The purpose of this Chapter is to give a sketch of possible micro-macro models, along with a description of the necessary indicators needed to provide empirical support for the model.

The organization of this Chapter is as follows. We start with a description of the type of data used for economic policy analysis and the underlying theoretical model underlying the analysis. We then provide more detail on the type of model that can help explain how ICT impacts the economy. Following this, the indicators collected in the ICT Impacts project are described. We continue with an application that looks at impacts of ICT on productivity and the decision to adopt ICT, using a combination of micro based indicators and macro (industry-level) data. We conclude with some observations on the type of data that are not yet collected but would help in tracking the impacts of ICT on market interactions.

14.3 Micro and macro indicators for policy

Policy makers are the main customer of indicators published by national statistical institutes (NSIs). The indicators and policy requirements evolve in an interplay with economic theory that describes causes and effects of policies and provides implications that are testable with appropriate indicators. In this section, we argue that the questions concerning ICT impacts require theories, policies and indicators that traverse from the micro level (firm and household) to the macro level (industries, countries) and back.

Most economic indicators currently supplied by national statistical institutes (NSIs) pertain to the macro economy, or are aggregates at the industry or sectoral level. The most famous of these, GDP, summarizes movements in the economy that correspond to concepts from established macroeconomic models taught in every undergraduate curriculum (see e.g. Mankiw (2006)). Outside these well-known benchmark indicators, NSIs provide summary statistics on an increasingly broad range of topics. It seems that no EU Ministerial meeting

256


goes by without demands for new indicators on the hot-topic of the day. Unfortunately, producing an indicator that has no theoretical basis not only wastes scarce resources of NSIs, but may also produce incomplete or worse, incorrect, guidance to policy makers.

Official macroeconomic statistics have been well grounded in a theoretical framework since the times of Keynes and Stone. The movements of time series on GDP and the components of final demand - consumption, investment, government spending, and net-exports - provide policy makers with information on the main economic problem in Keynes’ times, namely demand shortfall. The machinery of national accounts is geared towards providing the best possible indicators of these time series, based on information gathered and integrated using the supply-use framework.

Currently, a mismatch between economic theory, economic policy, and statistical indicators has become evident. Economists have been critical of the underlying macro model as a guide to policy, because changes in policy would alter the relationship between final demand, e.g. consumption or investment with output and interest rates (the Lucas critique). More importantly, policy makers themselves have shifted from demand management to structural policy, namely policy aimed at creating an economic environment that is conducive to efficient resource allocation decisions among firms and households. Finally, the statistics needed to uncover the drivers of expenditure or other economic decisions must include information on the observed behaviour of individual households and firms over time. Policy makers concerned with structural policy require data not only describing actions of buyers and sellers, but also the functioning of markets.

Ideally policy makers would have information sources available to conduct policy evaluation. In other words, to give an answer to the question ‘what would have happened to the economy had it not undergone the treatment applied through the change in policy?’ Of course, randomized experiments generally are not an option for policy makers. Instead, identification of the effects of a change in policy, say a subsidy for ICT-education, would need some way to control for other things that vary in the time before and after the policy (say a shift in quality of ICT hardware), or to split individuals into those that undergo the treatment versus those that don’t (for example based on amount of ICT hardware available at the firm).

The mismatch of theory, policy, and indicators does not imply a complete redesign of the statistical system, but points to a different way of using the available information. Even though the theoretical framework underlying national accounts was not designed to provide information to evaluate structural policies directly, the data underlying national accounts may be used to provide some guidance. In Figure 14.1, the National Accounts data are shown to be built up from micro-level surveys. One of the statistical sources is the structural business survey. In this survey, firms are queried annually on their gross output, their use of intermediates, and the disbursement of value added among labour and capital. Further, information is available on employment at firms and through household surveys and government registers, also on jobs and workers. Two different strategies have been followed to help identify the effect of structural policies with use of existing data. One is to use micro

257


data at the national level. The other is to use aggregate (or industry) data pooled across countries.

The latter strategy was one of the driving forces behind EU/FP6 support to build the EUKLEMS database. By pooling across countries, identification of policies may be achieved by relying on the difference in timing and magnitude of the ‘treatment’, while using the industry dimension to control for other possible variation. Much analytical work has been done with cross-country, industry, and time datasets, to study the effects of policies ranging from corporate taxes, R&D subsidies, environmental policies, etc1.

Figure 14.1

(augmented) EU KLEMS

Longitudinal Micro Data

National Accounts Industry Data

Single country

Multiple countries

Macro and Sectoral

Timeseries

Surveys, Business Registers

n.a.

For analysis of the impact of ICT, much work already has been done using the EUKLEMS database. The growth accounting method, derived from microeconomic theory of firm-level optimization, is applied as the organizing framework. Table 14.1 shows some recent data on the contributions to output growth from the EUKLEMS project. The table shows that output growth in the market sector in the US, from 1995 through 2005 was about 1.5 percentage point higher per year than in the EU. The lag in the EU not only occurred in the ICT producing sector, where the EU has relatively little activity in the very high growth industries, but was broad based in sectors where ICT adoption in the U.S. has increased significantly, such as the financial and distribution sectors. The contribution to output growth from the services provided by ICT capital was lower in the EU than in the US, owing to a lower share of expenditures on ICT capital and a slightly lower rate of growth in ICT capital

1 Analytical work using EUKLEMS is available on www.euklems.net. Earlier work, using the OECD STAN database, may be found easily by searching for Stan database via scholar.google.com

258



stocks. Most of the difference in output growth, however, is from lower growth of TFP, or that portion of output growth not explained by decisions of firms about the use of productive inputs. While these facts do aid in revealing symptoms of differences in structural policy across the EU and the US, they do not really explain why the differences in outcomes exist, nor what policies may improve the economic situation in the EU. The data do not allow for unearthing the determinants of firm-level decisions, nor for the effect of market selection in determining aggregate outcomes.

Table 14.1 1995-2005 EU US

VA% --Kict --TFP VA% --Kict --TFP

Market 2.1 .4 1.0 3.7 .6 1.7

EleCom 3.8 .8 2.8 10.5 .8 8.7

MexElec 1.2 .2 1.7 1.8 .2 2.2

DISTR 2.6 .3 1.5 4.1 .5 2.1

FinBu 3.5 .9 -1.0 4.3 .7 .4

Source: O’Mahony, Timmer, van Ark (2008).

The other strategy for identifying policy and for learning about drivers of economic behaviour is to look directly at micro data, preferably longitudinally linked. In other chapters of the present report, we have shown results of firm-level analysis of the impact of ICT on output, productivity and employment.2 Further, we have analyzed the relationships between ICT use, off-shoring, outsourcing and innovation.3 While it is possible to identify the drivers of individual decisions, it actually may not be easy to assess the impact of policy changes. For example, it would be interesting to find out to what extent low internet prices, brought about by good telecommunications policy, affect the firm-level use and impact of ICT. Unfortunately, the policy generally affects all firms in the same manner, and identification thus relies on controlling for all other effects before and after the policy change, or requires distinguishing between firms according to how much they are potentially affected by policy.

Ideally, researchers would have access to cross-country micro data to conduct evaluation of policy changes. As seen in the bottom left corner of Figure 14.1, these data are not available. Confidentiality of the firm-level data underlying the national accounts at the NSIs prevents pooling these data together. The advantage of having such a dataset is obvious: not only can one compare the effects of policy changes across countries, but one also has the firm-level data to understand and track how the policy affects actual decision making. Given the impossibility of conducting research with cross-country micro data, the ICT Impacts Project

2 See Chapters 6 and 8. 3 See Chapters 11 and 12.

259


augmented the cross-country industry datasets with indicators drawn from the firm-level datasets in the individual countries. With appropriately chosen indicators, it becomes feasible to track the effects of policy not only on market outcomes, but also on decisions at the firm level and interactions in the market.

14.4 Theory for micro and macro indicators

As argued above, a sound theoretical basis is needed to design indicators that are useful for policy makers. The basic idea that indicators from single country longitudinal micro data can be aggregated to the industry level and merged into cross country datasets does not provide much guidance as to which indicators need to be collected. One of the goals of the ICT Impacts Project was to search for indicators that could aid in answering questions concerning the effect of ICT at the firm level, its effects on markets, and its effect at the macro (or industry) level. In particular, one wants to know why ICT adoption is lower in the EU than in the US, and what the impact of ICT is on TFP.

Underlying the search for these indicators is a framework for the mechanisms tying micro behaviour, market selection, and aggregate outcomes. Figure 14.2 shows a schematic, with firm decisions, market interactions, and macro outcomes. In Chapter 2 we described the microeconomic behaviour of firms deciding on capital and labour inputs, as well as on appropriate human capital and knowledge capital. These choices lead to improvements in their technology or product quality. In Figure 14.2, we see that firms make these choices conditional on their capabilities, the prices of inputs, and their expectations of their interactions with other firms and customers. Indicators are needed that provide information on the choices made by firms, but also on the constraints and opportunities provided by markets. In the case under study, ICT, we should be mindful that the technology has different paths for affecting firms and markets.

Figure 14.2 Firm choices

• Knowledge capital

• ICT adoption

• Organisation

• Technology

• Input choices

Market interactions

• Competition type

• Market shares

Aggregate outcomes

Reviewing the bountiful literature on ICT impacts (e.g. Brynjolfsson and Kahin (2000); Soete and ter Weel (2005)), the following impacts need to be addressed. First, one needs to look at the ICT producing sector, which contributes directly to output and employment. Our statistical system currently provides this information at micro and industry level. Next, ICT may be considered in the same way that other capital (equipment and structures) enters into

260


production. The growth accounting method computes the contribution of ICT capital in the same way it computes the contribution of machine tools. Some technical considerations here are the rapid change in the deflator for ICT investment, and the high rate of depreciation. At present, NSIs vary in their ability to provide reliable indicators of investment in ICT by asset type and industry, along with the appropriate deflators. In particular, measuring the investment in own-account software has proven to be difficult. However, the EUKLEMS database provides a best attempt at creating harmonized indicators of the service flows from ICT-capital. Finally, ICT may have special properties that provide additional impacts. In the early literature, one often used term for such properties was ‘spillovers’, namely effects that go above and beyond the normal returns to capital. Recent literature has been more specific about the precise mechanisms through which ICT may affect TFP, as presented below. Generally, the statistical system does not provide direct evidence on these impacts.

In traditional macroeconomic growth models, TFP is boosted through innovative activities undertaken by firms, namely research and development. R&D generates ‘blueprints’ that boost the productivity with which a firm can employ traditional factor inputs, capital and labour. Further, R&D provides spillovers of two sorts: one is intertemporal, as the knowledge built up in the past may provide benefits by improving the efficiency of researchers in the present (“standing on the shoulders of giants”). The other effect is that knowledge that is built up may provide benefits to others that have not paid for it (knowledge spillovers).

ICT spending may share certain characteristics with R&D. To a certain extent, firms that are early adopters of new ICT technologies are investing in innovation, similar to conducting R&D. These firms indeed are searching for new and better ways to bring goods and services to their customers, and new and better ways to organize their supply chain and business processes. Once a configuration of software, networks, and business processes are seen to be successful, they are like a blueprint and can be used at any scale of production, similar to the outcome of R&D. However, increasing the scale of operations does require complementary inputs, hardware and trained workers. Next, the spillovers likely are very different between R&D and innovative adoption of ICT: while the blueprint from ICT use may be applied to any scale of production within the firm, it may not be of direct use to competitors. Further, the depreciation of ICT-ideas likely is different from traditional R&D.

A major difference between innovative ICT adoption and traditional R&D is the overall effect of ICT on transactions costs in a market. By allowing fast, high-bandwidth communication, and codification of many aspects of transactions, many markets are able to function with fewer frictions. Not only does this lower overall macroeconomic costs, but it may fundamentally change the market interactions that impact on individual firm choices and thus on aggregate outcomes. This effect may be the most difficult to capture in the present statistical system, where annual data are collected (annual sales, stock of employment or computers at a fixed point in time). Ideally, transactions data, or even bids and offers, in real markets at high frequency would be needed to measure what the impact of ICT would be on transactions and transaction costs.

261


The indicators that we are able to collect attempt to capture characteristics of firm behaviour, market selection and outcomes that are expected to be affected by ICT. First, innovative use of ICT, as opposed to investment in ICT capital, often requires adopting the leading edge of technology. At present, this appears to include wireless applications, broadband use, and electronic sales and purchases. At the micro level, access to these technologies by workers in a firm indeed is seen to be correlated with the availability of high skilled workers at the firm and to sales of innovative products. Next, innovative use of ICT is expected to be risky, in the sense that firms should start small, fail often, but grow rapidly if successful (see Bartelsman, Perotti, and Scarpetta,(2008)). To capture these features, information is collected on the size and productivity distribution of entrants, and incumbents.

The features of market selection are more difficult to capture with a statistical system that only collects realized market outcomes at a low frequency. The indicators we do collect reflect both the working of the market, but also the choices of customers and suppliers, making it difficult to disentangle the effects of ICT on the market from the effect of the market on the adoption of ICT. For example, the use of ecommerce may lower transactions costs and increase the transparency of a market, thereby generating higher fluctuations in market shares across firms in a market. However, as firms are better able to win market share with an innovative product or service, they may be more willing to undertake the innovative activity needed to have a chance at producing an innovative product. Overall, we collect a host of indicators related to ICT use, firm demographics and dynamics, and reallocation of resources. Table 14.2 provides a summary, and Chapter 4 describes the variables collected in each of the harmonized cross-country datasets.

Table 14.2 Domain Indicator

ICT Indicators Pct of workers with access to Internet Pct of workers with access to broadband Pct of sales / purchases via electronic means Firm Demographics Entry and Exit rates Job creation / destruction rates … above indicators by size class Firm Dynamics Distribution of firm growth Distribution of productivity Reallocation Market share churning Olley-Pakes cross-term Firm growth by productivity quartile Productivity decompositions

262


14.5 An application: Adoption and impact of ICT

Deriving a full model of firm behaviour and market outcomes, designing and collecting appropriate indicators from the available data, and testing the model using the indicators will require a considerable amount of research and clearly falls outside the scope of the present ICT Impacts Project. However, a quick glance at the importance of interactions between micro decisions and market selection is given in Figure 14.3. Here, the indicator showing the percentage of workers with access to broadband internet (horizontal axis) is plotted against the variability of market shares, as measured by the inter-quartile range of the distribution of firm-level output growth in an industry (vertical axis). Each datapoint in the figure is at the aggregate level, showing broadband and market share variability for a particular country, industry and year. The figure shows a positive correlation, with higher broadband penetration being associated with higher market share variability. In our reading of relevant theoretical models, the causality goes both ways: 1) the potential to gain market share (or the risk to lose share) provides an incentive for firms to adopt the latest technology, but also 2) the adoption of the latest broadband technology - which also requires complementary changes in supply chains, business processes and customer relations - may fail and lead the firm to lose market share, or may succeed and lead to growth.

While Figure 14.3 whets the appetite for more research in this area, the remainder of this section will continue with the question posed in the introduction: what is the impact of ICT use on TFP, and what are the drivers of the decision to use ICT.

Figure 14.3

Firm-growth Distrbution vs Broadband Use

0

0.5

1

1.5

2

2.5

3

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1Percent of w orkers w ith broadband access

Inte

rqua

rtile

rang

e of

firm

gro

wth

di

strib

utio

n

ICT Impact

Using the EUKLEMS database, augmented with indicators from our harmonized cross-country dataset, we attempt to shed some light on the impact of ICT. To start, we run a production function regression on a panel dataset (countries, industries, time), with the log of output regressed on the log of traditional inputs and one of the ICT indicators. This

263


regression roughly mimics at the aggregate (industry) level the regressions done using the micro data in each country separately.

Table 14.3 (at the end of the Chapter) summarizes results from 8 separate regressions, with output regressed on each of the four different ICT indicators and either country and time fixed effects, or industry and time fixed effects. The regressions include data from 8 countries at the 30-industry level of detail, for the period 2001-2005.4 In all specifications, the coefficients on labour (hours worked), non-ICT capital services, and ICT-capital services are in the range seen in the literature. The effects of the ICT variables on value added, controlling for capital (ICT and non-ICT) and labour and fixed effects, are all significantly positive. This means that the correlation between ICT and TFP is positive, and points to either a spillover, or to other, omitted, complementary inputs.

In the chapter on ICT and productivity, the regressions were run on firm-level data, one country at a time. To make a comparison, we run the same specification as above, one country at a time. Table 14.4 provides the coefficient on the ICT variable (percent of workers with broadband access, or DSLpct), in a regression of (log) value added on the labour and capital services variables and the ICT variable, as well as time fixed effects. As seen, the results vary across countries, with a significant effect of DSL in Finland, U.K, Netherlands, and Sweden, and insignificant (sometimes negative) effects elsewhere. The coefficients on the labour and non-ICT capital are as expected, but ICT capital results are sometimes small and insignificant.

Similarly, table 14.5 presents the basic regression, but now run separately for each of 4 industry groups. Here, the results are strongly positive in the major market sectors except in the non-ICT manufacturing sector. Unfortunately the short time period of available data makes it impossible to get estimates for countries and broad sectors separately, in order to compare directly with the results from the firm-level regressions in Chapter 6.

Broadband adoption

While the above regressions provide some feeling that, on average, the access of workers to DSL correlates significantly with TFP, we worry about reverse causality, namely that good firms adopt broadband. Further, policy makers are interested in knowing if some policy levers may increase the adoption of ICT. To this end, we consider the adoption decision. Often, technology adoption is modelled by looking at costs and benefits. Another way this is described in the literature is by looking at the ability of firms to adopt and their desire to adopt. Typically, the variables hypothesized to affect adoption of a technology would include its price, the ability and readiness of the firm to take on the new technology (in our case, skilled ICT workers, and availability of ICT capital), as well as some measure of expected benefits.5 Given the non-rival, but likely appropriable, nature of firm-level ICT projects, firms that can successfully replicate, or scale-up, their operations upon success in

4 The exact time period and industries differ by country. 5 Unfortunately, no time series data are available across countries for the price of DSL access. The OECD does have data for 2007, but this provides no information separate from the country fixed-effects.

264


the market, have a higher potential benefit to the ICT investment (see e.g. Brynjolfsson et al (2007), Bartelsman et al. (2008)). To proxy this, we use the width of the distribution of firm-level output growth.6

Table 14.6 present the results from this regression for Austria, where all four variables are significantly positive. When no fixed effects are included in the regression 82 percent of the variance across sectors and time in broadband access is explained; including fixed effects the regression fit is 87 percent. The results in other countries are not as ‘perfect’, but in all countries two or more of the hypothesized variables are significantly positive, and more than 50 percent of the variance in DSLpct is explained. When all countries are pooled, all four variables have a positive and significant effect.

Adoption and Impact

The next step is to combine the adoption and production equation in a system to be estimated simultaneously. The system uses lagged exogenous variable as instruments. Table 14.7 shows results for the coefficients in four versions of the two equations, with the (log) of value added as the dependent variable of the production equation, and the percent of broadband enabled workers or internet-enabled workers as the dependent variable in the two sets of adoption equations. For each ICT variable, the system is run either with country and time fixed effects, or with industry and time fixed effects. As shown, the broadband or internet access variable always has a significant positive effect on productivity. The ICT capital variable becomes insignificant (and sometimes negative), while the hours and non-ICT capital variables are as expected. The explanatory variables in the adoption equations also show up significantly positive in all specifications. Table 14.8 changes the specification, with the production equation run in (log) first-differences, and the ICT variable being broadband access. Because estimation of capital output elasticities is particularly sensitive to the fact that capital is subject to steep adjustment costs, a combined ‘input growth’ variable is constructed from the separate traditional inputs, using expenditure shares as weights. With this specification, the percentage of broadband enabled workers is seen to boost output growth by one-tenth to one-quarter of a percentage point, while the explanatory variables in the adoption equation continue to be significantly positive.

14.6 Conclusions

In the above results, we have shown a positive correlation between productivity and the percentage of workers in an industry with access to (broadband) internet. Further, the results continue to hold, after controlling for the adoption equation, although the effect on productivity becomes lower. In the adoption equation, it is seen that countries or industries with a high proportion of skilled workers are more intense users of new technology. Further, the amount of variability in firm-level output growth across firms in an industry is seen to boost the intensity of firm use of internet. This result is consistent with findings in Bartelsman et al (2008).

6 In Chapter 5, we show that the width of the distribution of firm-level growth rates is similar for the full PS sample, and the linked PSEC sample.

265


More generally, the empirical application is an example of how to collect firm-level indicators and merge them into aggregate data in order to track firm-level decisions, market selection and aggregate outcomes. While this style of theoretical and empirical work is rather new, it points in a clear direction for types of indicators that NSIs should consider. To start, by linking special surveys on firm level behaviour into the business register, one can create industry-level indicators of this behaviour. Next, one needs to provide access to researchers to generate industry-level indicators that capture the dynamics and reallocation of resources across firms in an industry. The exact specification of these will depend on the model setup of the researcher, but the indicators should be quite easy to compute if the data infrastructure is in place. Examples of these indicators are given in Table 14.2.

The main ingredient missing from the statistical system is direct evidence on how transactions are conducted in a market place. The underlying ‘Keynesian’ design of National Accounts has placed little importance on these intermediate transactions. The treatment of large parts of the economy as ‘margins’ (transport, trade, banking’), rather than as creators of value and locations of technological progress is evidence of this. Instead, especially in the light of understanding the role of ICT, information needs to be collected on how (potential) buyers and sellers conduct all the stages leading up to, and following a transaction.

266


Table 14.3: ICT Impacts – cross-country regressions Indicator: DSL% PC% Inter% Ecomm% DSL% PC% Inter% Ecomm%

ICT 0.14 0.31 0.36 0.01 0.28 0.26 0.37 0.02 (0.07) (0.09) (0.09) (0.00) (0.09) (0.09) (0.09) (0.01) hrs 0.59 0.59 0.60 0.59 0.52 0.52 0.53 0.49 (0.02) (0.02) (0.02) (0.02) (0.02) (0.02) (0.02) (0.02) KNIT 0.23 0.24 0.24 0.23 0.46 0.46 0.45 0.47 (0.02) (0.02) (0.02) (0.02) (0.01) (0.01) (0.01) (0.01) KICT 0.12 0.12 0.12 0.13 0.04 0.04 0.03 0.06 (0.02) (0.01) (0.02) (0.02) (0.02) (0.01) (0.02) (0.01) R2 0.98 0.98 0.98 0.98 0.96 0.96 0.96 0.96 Deg.F. 572 572 572 572 590 590 590 590 Dummies i,t i,t i,t i,t c,t c,t c,t c,t Standard errors in parentheses

Model: dummieskakahrsaICTaav ICTNITijt +++++= 43210

Where: Vijt (log) real value added, by industry, country, time

Kict ICT capital service

Knit Non-IT capital service

Hrs Hours

ICT Impacts project ICT Indicator

PC% Pct of workers with PCs

DSL% Pct of workers with broadband (DSL) access

Inter% Pct of workers with internet access

Ecomm% Pct of sales via net

Dummies industry(i), country(i), or time(t)

267


Table 14.4: ICT Impacts – single country regressions

Country DSL% (t-stat)

AUT -.12 (.9)

FIN .46 (3.5)

FRA .45 (1.6)

GBR .39 (3.9)

GER -.22 (.5)

ITA -.69 (1.4)

NLD .37 (2.5)

SWE .47 (2.9)

Model: dummieskakahrsaICTaav ICTNITjt +++++= 43210

Table 14.5: ICT Impacts – cross-country sector regressions

Sector DSL% (t-stat)

Elecom 1.24 (9.6)

MexElec -.00 (.1)

DISTR .38 (6.0)

FINBU .14 (3.7)

Model dummieskakahrsaICTaav ICTNITijt +++++= 43210

268


Table 14.6: ICT Adoption Model - Austria

W Kit% HiSkl Churn dummy R-sq D.F.

.21 .88 .57 .19 t .87 92

(6.6) (17.1) (4.9) (5.5)

T-statistics in parentheses

Model: 0 1 1 2 1 3 1 4% %ITijtDSL b b w b Cap b HiSkl b Churn dummies− − −= + + + + +

Where:

W Average wage

Cap%it ICT-capital as share of capital.

HiSkl High skilled worker share

DSL% ICT-indicator

Churn Interquartile range of firm-level growth rate distribution

269


Table 14.7: System Model

Coef Variable DSL% Internet%

a1 ICT-indicator: 1.24 .90 1.20 1.05

a2 Kn .35 .27 .34 .27

a3 Kit -.07 .05 -.08 .05

a4 Hrs .72 .68 .72 .68

b1 w(-1) .24 .02 .30 .01

b2 Cap%it .31 .20 .32 .17

b3 HiSkl .18 .38 .19 .33

b4 Churn .30 .15 .28 .14

dummies c,t i,t c,t i,t

D.F. 659 646 649 646

Model: 0 1 2 3 4

0 1 1 2 1 3 1 4

: %

: % %

IT Nijt

ITijt

a v a a DSL a k a k a hrs dummies

b DSL b b w b Cap b HiSkl b Churn dummies− − −

= + + + + +

= + + + + +

Where:

V (log) real value added

Kit ICT capital service

Kn Non-IT capital service

Hrs Hours

w Average wage

Cap%it ICT-capital as share of cap.

HiSkl High skilled worker share

DSL% Broadband penetration

Churn Interquartile range of firm-level growth rate distribution

Kit generally insignificant; All other coefs significant at 1%-level

270


271

Table 14.8: System Model – Alternative Specification Coef Variable VA growth a1 DSL% 0.12 0.27 a2 Growth K, Kit, Hrs 0.67 1.03 b1 w(-1) 0.26 0.16 b2 Cap%it 0.21 0.12 b3 HiSkl 0.18 0.44 b4 Churn 0.78 0.44 Dummies c , t i , t D.F 501 484 Churn: interquartile range of firm-level VA growth distribution

Eurostat Agreement No. 49102.2005.017-2006.128 References

References Abraham K G. and Taylor S K. (1996) Firms' Use of Outside Contractors: Theory and Evidence. Journal of Labor Economics, 14(3), 394.

Abramovsky L. and Griffith R. (2005) Outsourcing and Offshoring of business Services: How important is ICT? The Institute for Fiscal Studies Working paper 05/22.

Acemoglu D. Aghion P. and Zilibotti F. (2006) Distance to Frontier, Selection, and Economic Growth. Journal of the European Economic Association 4, no. 1: 37-74.

Acemoglu D. Aghion P. Griffith R. and Zilibotti F. (2004) Vertical integration and Technology: Theory and Evidence. The Institute for Fiscal Studies Working Paper 04/34.

Ackerberg D A. and Caves K. (2007) Structural Identification of Production Functions. mimeo.

Airaksinen A. (2004) Impacts of ICT usage on business organisation and business processes – Final report. NESIS work package 5.4

Allen A. (2001) Historical Revisions to Computer Producer Prices. Office for National Statistics, UK. Available on-line at www.statistics.gov.uk.

Allen A. and Ball A. (2003) The Introduction of Hedonic Regression Techniques for the quality adjustment of computing equipment in the Producer Prices Index (PPI) and Harmonised Index of Consumer Prices (HICP). Office for National Statistics, UK. Available on-line at www.statistics.gov.uk.

Amiti M. and Shang-Jin W. (2006) Service Offshoring and Productivity: Evidence from the United States. National Bureau of Economic Research, Working Paper 11926, Cambridge, Massachusetts.

Antonucci, T. and Pianta M. (2002) Employment Effects of Product and Process Innovation in Europe. International Review of Applied Economics, Vol. 16, No. 3, pp. 295 – 307.

Airaksinen A. (2004) Impacts of ICT usage on business organisation and business processes – Final report. NESIS work package 5.4

Atrostic B K. and Nguyen S V. (2002) Computer Networks and US Manufacturing Plant Productivity: New Evidence from the CNUS Data. CES Working Paper 02-01, Center for Economic Studies.

Aubert P. and Sillard P. (2005) Offshoring and Job Losses in French Industry. Paper prepared to the OECD workshop on globalisation in Paris, November 2005,

Baily M N. (1986) Productivity Growth and Materials Use in U.S. Manufacturing. The Quarterly Journal of Economics February, pp 185-195.

Ball A. Hughes J. Smith K and Waldron K. (2004) Changes to methodology employed in the CPI and RPI from February 2004. Office for National Statistics, UK. Available on-line at www.statistics.gov.uk.

Bardhan I. Whitaker J. and Mithas S. (2006) Information Technology, Production Process Outsourcing, and Manufacturing Plant Performance. Journal of Management Information Systems, 23(2), 13-40.

Bartel A. Lach S. and Sicherman N. (2005) Outsourcing and Technological Change. Centre for Economic Policy Research Centre discussion Paper number 5082, May.

Bartelsman E J. and Doms M. (2000) Understanding Productivity: Lessons from Longitudinal Microdata. Journal of Economic Literature, 38(3), p.569.

272


Bartelsman E J. and Barnes M. (2001) Comparative Analysis of Firm-Level Data: A Low Marginal Cost. Approach Paper for Workshop on Firm-Level Statistics, 26-27 November 2001, OECD.

Bartelsman E J. (2004) The Analysis of Microdata from an International Perspective. OECD Statistics Directorate, (STD/CSTAT 12).

Bartelsman E J. Bassanini A. Haltiwanger J. Jarmin R S. Scarpetta S. and Schank T. (2004) The spread of ICT and productivity growth. Is Europe really lagging behind in the New Economy? The ICT Revolution: Productivity Differences and the Digital Divide.

Bartelsman E J. Haltiwanger J. and Scarpetta S. (2004) Microeconomic evidence of creative destruction in industrial and developing countries.

Bartelsman E J. and Hinloopen J. (2004) Unleashing Animal Spirits: Investment in ICT and Economic Growth. The economics of the digital economy, Cheltenham: Edward Elgar.

Bartelsman E J. and Scarpetta S. (2004) Experimentation Within and Between Firms: Any Role for Policy and Institutions? Presented at the 2004 American Economic Association Meeting.

Bartelsman E J. Perotti E. and Scarpetta S. (2008) Barriers to Exit, Experimentation, and Comparative Advantage. Draft.

Becchetti L. De Panizza A. and Oropallo F. (2007) Role of Industrial District Externalities in Export and Value-added Performance: Evidence from the Population of Italian Firms. Regional Studies, Vol. 41.5, pp. 601–621, July 2007.

Bertschek I. and Müller M. (2006) Productivity Effects of IT-Outsourcing: Semiparametric Evidence for German Companies. In S. Sperlich, W. Härdle and G. Aydinli (Eds.), The Art of Semiparametrics (pp. 130-154). Heidelberg and New York: Physica-Verlag, Springer.

Black S E. and Lynch L M. (2003) Measuring Organisational Capital in the New Economy, in Corrado, C., Haltiwanger, J., and Sichel, D., eds Measuring Capital in the New Economy, Chicago, University of Chicago Press.

Bloom N. Sadun R. and Van Reenen J. (2007) Americans Do I.T. Better: US Multinationals and the Productivity Miracle, Centre for Economic Performance Discussion Paper No 788, April.

Booz Allen Hamilton. (2002) International e-Economy Benchmarking; the world’s most effective policies for the e-economy.

Bresnahan T. F. and Trajtenberg M. (1995) General purpose technologies 'Engines of growth'?. Journal of Econometrics, Elsevier, vol. 65(1), pages 83-108, January.

Bresnahan T. F. Brynjolfsson E. and Hitt L M. (2002) Information Technology, Workplace Organisation, and the Demand for Skilled Labour: Firm-Level Evidence. The Quarterly Journal of Economics, vol. 117(1), 339-376.

Brynjolfsson E. and Hitt L M. (2001) Computing Productivity: Firm-Level Evidence. MIT Sloan Working Paper No. 4210-01.

Brynjolfsson E. and Kahin B. (2000) Understanding the Digital Economy; data, tools and research. MIT Press, Cambridge.

Brynjolfsson E. McAfee A. Zhu F. and Sorell M. (2006) Scale without Mass: Business Process Replication and Industry Dynamics. Harvard Business School Technology & Operations Mgt. Unit Research Paper No. 07-016.

Calza M G. de Panizza A. and Rossetti S. (forthcoming) Internationalisation, economic structure and performance of Italian manufacturing firms.

273


Calza, M G. de Panizza A .and Rossetti S. (2007) Dinamica dell’internazionalizzazione delle imprese manifatturiere italiane nel periodo 2000-2004. Rapporto ICE 2006 - 2007 "L'Italia nell'economia internazionale", pp. 194-201.

Cap Gemini, Ernst & Young, (2002). Web-based Survey on Electronic Public Services. For DG Information Society e-Europe 2002.

Chamberlin G. Chesson A. Clayton T. and Farooqui S. (2006) Survey-based measures of software investment in the UK. Economic Trends, No. 627, February 2006

Chesher A. and Nesheim L. (2006) Review of the Literature on the Statistical Properties of Linked Datasets. DTI Economics Paper number 3.

Chesson A. Clayton T. and Farooqui S. (2006) Survey Based Estimates of Software Investment. UK Office for National Statistics.

Clayton T. and Criscuolo C. (2002) Electronic Commerce and Business Change. Economic Trends, 583, pp 62-69.

Clayton T. Farooqui S. Gales G. Leaver M. and Sadun R. (2005) IT investment, ICT Use and UK firm Productivity. Office for National Statistics / London School of Economics.

Cohen W M. and Levinthal D A. (1989) Innovation and Learning: The Two Faces of R&D. The Economic Journal, Vol. 99, pp. 569 – 596.

Corrado C. Hulten C. and Sichel D. (2006) Intangible Capital and Economic Growth. NBER working paper 11942.

Crépon B. Duguet E. and Mairesse J. (1998) Research, Innovation and Productivity: An Econometric Analysis at the Firm Level. The Economics of Innovation and New Technology, Vol. 7, pp. 115 – 158.

Crespi G. Criscuolo C. and Haskel J. (2007) Information Technology, Organisational Change and Productivity. Centre of Economic Policy Research (6105) Discussion Paper Series dated 14 Jan 2007.

Crespi G. Criscuolo C. Haskel J and Slaughter J. (2007) Productivity growth, knowledge flows and spillovers. Centre for Research into Business Activity

Criscuolo C. and Leaver M. (2006) Offshore Outsourcing and Productivity. OECD and Office of National Statistics

Criscuolo C. Hagsten E. Hanley A. Karpaty P. and Svanberg S. (2008) Offshoring and Productivity in the Case of Ireland, Sweden and the United Kingdom. OECD, forthcoming

Criscuolo C. (2006) Does offshoring matter for productivity? Paper presented for the OECD Working Party on Statistics in Paris, November 2006.

Criscuolo C. Waldron K. (2003) E-commerce and productivity, Economic Trends. No. 600, November 2003.

Daveri F. and Jona-Lasinio C. (2007) Off-Shoring and Productivity Growth in the Italian Manufacturing Industries. paper presented at the CESifo Economic Studies Conference on Productivity and Growth, Munich 22-23 June.

Domberger S. Fernandez P. and Fiebig D G. (2000) Modelling the Price, Performance and Contract Characteristics of IT Outsourcing. Journal of Information Technology, 15(2), 107-118.

Domberger S. Hall C. and Ah Lik Li E. (1995) The determinants of price and quality in competitively tendered contracts. Economic Journal, 105(433), 1454-1470.

Doms M. Dunne T. and Troske K R. (1997) Workers, Wages and Technology. The Quarterly Journal of Economics, February, pp 253-290.

274


Draca M. Sadun R. and Van Reenen J. (2006) Productivity and ICT: A Review of the Evidence. CEP Discussion Paper 0749, Centre of Economic Performance, London School of Economics.

Du J. Lu Y. and Tao Z. (2006) Why Do Firms Conduct Bi-sourcing? Economics Letters, 92(2), 245–249.

Dunne T J. Jensen J. and Roberts M J. (2008) Producer Dynamics: New Evidence from Micro Data. University of Chicago Press.

Ekholm K. and Hakkala K. (2006) The Effects of Offshoring on Labour Demand, Evidence from Sweden., CEPR Working paper number 5648.

Elbers C. Lanjouw J O. and Lanjouw P. (2003) Micro-Level Estimation of Poverty and Inequality. Econometrica 71, no. 1: 355-364.

Farooqui, S. (2005) Information and Communication Technology use and productivity, In: Economic Trends 625, Office of National Statistics (ONS), London.

Farooqui S. and Sadun R. (2006) Broadband Availability, Use, and Impact on Returns to ICT in UK Firms. OECD paper DSTI/ICCP/IIS 2006(9).

Fazio G. Lam K. and Ritchie F. (2005) Sample bias in microeconometric analyses of official microdata. DTI Economics Papers.

Feenstra R C. and Hanson G H. (1996) Globalization, Outsourcing and Wage Inequality. The American Economic Review, volume 86, number 2, pp 240-245.

Fitzgerald D. and Haller S. (2008) Exchange Rates and Producer Prices: Evidence From Micro-Data. Presented at the ‘Comparative Analysis of Enterprise Data’ conference, Budapest, May 2008

Fixler D J. and Siegel D. (1999) Outsourcing and Productivity Growth in Services. Structural Change and Economic Dynamics. 10(2), 177–194.

Florin J. Bradford M. and Pagach D. (2005) Information technology Outsourcing and Organizational Restructuring: An Explanation of Their Effects on Firm Value. Journal of High Technology Management Research, 16(2), 241-253.

Foster, L. Haltiwanger, J. and Krizan C J. (2001) Aggregate Productivity Growth: Lessons from Microeconomic Evidence New Developments in Productivity Analysis, editors: Edward Dean, Michael Harper and Charles Hulten, Chicago: University of Chicago Press, 2001.

Galindo-Rueda F. and Haskel J. (2005) Skills, Workforce Characteristics and Firm-Level Productivity: Evidence from the Matched ABI/Employer Skills Survey. Institute for the Study of Labour, Discussion Paper Series IZA DP number 1542, March.

Goodridge P. Clayton, T. (2004) E-business and labour productivity in manufacturing and services. Economic Trends, No. 609, August 2004

Görg H. and Hanley A. (2005) International outsourcing and productivity: evidence from the Irish electronics industry. North American Journal of Economics & Finance, 16(2), 255–269.

Görg H. Hanley A. and Strobl E. (2008) Productivity effects of international outsourcing: Evidence from plant level data. Canadian Journal of Economics, volume 41 number 2, forthcoming.

de Graaf X. (2007) ICT Maturity and Firm Productivity. in T. Wolters (ed.), 'Measuring the New Economy. Statistics between Hard-Boiled Indicators and Intangible Phenomena', Elsevier, Amsterdam etc.; pp. 143-176.

de Graaf X. Keller W. van Leeuwen G. van der Pijll E. and Wolters T.(2007) E-Business, ICT and Productivity. Research Paper Division of Business Statistics, Statistics Netherlands, The Hague.

275


Greenan N. and Mairesse J. (2000) Computers and Productivity in France: Some Evidence. Economics of Innovation and New Technology, 9(3), 275-315.

Griffith R. Huergo E. Mairesse J. and Peters B. (2007) Innovation and Productivity Across Four European Countries. Oxford Review of Economic Policy, Vol. 22(4), pp. 483 – 498.

Griffith R. Redding S. and Van Reenen J. (2004) Mapping The Two Faces of R&D: Productivity Growth in a Panel of OECD Industries. The Review of Economics and Statistics, Vol. 86(4), pp. 883 – 895.

Grossman G M. and Helpman E. (2005) Outsourcing in a Global Economy. Review of Economic Studies, volume 72, pp 135-159.

Grossman G M. and Helpman E. (2002) Integration Versus Outsourcing in Industry Equilibrium. Quarterly Journal of Economics, 117(1), 85-120.

Gunnarsson G. Mellander E. and Savvidou E. (2001) Is Human Capital the Key to the It Productivity Paradox? The Research Institute of Industrial Economics, Working Paper number 551.

Gust C. and Marquez J. (2004) International comparisons of productivity growth: the role of information technology and regulatory practices. Labour Economics, 11(1), p.33-58.

Hagen H O. Ahlstrand C. Daniels M. Nilsson E. and Adermon A. (2007) Innovation matters. In: Yearbook on Productivity, Statistics Sweden.

Hagen H O. (2007) IT use and broadband in the Swedish business sector. Presented at the OECD workshop on the Economic and Social Impacts of Broadband Communications: from ICT Measurement to Policy Implications. 22 May 2007.

Hagsten E. and Svanberg S. (2008) Offshoring, IT Maturity and Productivity: Evidence from Swedish firm level data. Preliminary draft (april)

Hagsten E. Karpaty P. and Svanberg S. (2007) Effects on productivity from Swedish offshoring. Statistics Sweden and Örebro University.

Hall B H. and Mairesse J. (1995) Exploring the relationship between R&D and productivity in French manufacturing firms. Journal of Econometrics, Vol. 65(1), pp. 263 – 293.

Handel M. (2003) Implications of Information Technology for Employment, Skills, and Wages: A Review of Recent Research. Final Report, SRI International.

Harrison R. Jaumandreu J. Mairesse J. and Peters B. (2005) Does innovation stimulate employment? A firm-level analysis using comparable micro data on four European countries. MPRA Paper No. 1245.

Hempell T. (2002) Does Experience Matter? Productivity effects of ICT in the German service sector. Discussion paper 02-43, Centre for European Economic Research, Manheim.

Heshmati A. (2003) Productivity Growth, Efficiency and Outsourcing in Manufacturing and Service Industries. Journal of Economic Surveys, 17(1), 79-112.

Holdway M. (1999) Quality-Adjusting Computer Prices in the Producer Price Index: An Overview. Bureau of Labor Statistics, February.

Ilmakunnas P. and Maliranta M. (2005) Technology, Labour Characteristics and Wage-Productivity Gaps. Oxford Bulletin of Economics and Statistics, 67, 5(2005), pp 623-644.

Jaumandreu J. Mairesse J. and Janz N. (2005) Innovation and Employment in European Firms: Microeconometric Evidence. Final Report Project, DG Research, EC.

Jorgenson D W. Griliches Z. (1967) The explanation of productivity change. The Review of Economic Studies, Vol. 34, No. 3 (Jul., 1967), pp. 249-283.

276


Judson, D H. (2007) Information integration for constructing social statistics: history, theory and ideas towards a research programme. Journal of the Royal Statistical Society: Series A (Statistics in Society) 170, no. 2 (March): 483-501.

Knittel C R. and Stango V. (2008) The Productivity Effects of IT Outsourcing. Mimeo.

Koeniger W. and Prat J. (2007) Employment Protection, Product Market Regulation and Firm Selection. The Economic Journal, 117(521), p.F302-F332.

Klette T J. and Kortum S. (2004) Innovating firms and Aggregate Innovation. Journal of Political Economy. 2004, vol. 112, no. 5.

Krueger A B. (1993) How computers have changed the wage structure: Evidence from microdata. 1984 – 1989’, Quarterly Journal of Economics, Vol. 108, pp. 33 – 60.

Lehr B. and Lichtenberg F. (1999) Information Technology and its impact on Productivity: Firm-Level Evidence from Government and Private Data Sources, 1977-1993. Canadian Journal of Economics, Vol.32 n°2 , Special Issue on Service Sector Productivity and the Productivity Paradox (Apr., 1999), pp. 335-362.

Lewbel A. (2005) Modelling Heterogeneity. Boston College Working Paper.

Mahnke V. Overby M L. and Vang J. (2005) Strategic Outsourcing of IT Services: Theoretical Stocktaking and Empirical Challenges. Industry and Innovation, 12(2), 205-253.

Mahnke V. Özcan S. and Overby M L. (2006) Outsourcing Innovative Capabilities for IT-Enabled Services. Industry & Innovation, 13(2), 189–207.

Maliranta M. and Rouvinen P. (2003) Productivity Effects of ICT in Finnish Business. Helsinki: ETLA, The Research Institute of the Finnish Economy, Discussion Papers, No. 852.

Maliranta M. and Rouvinen P. (2004) ICT and Business Productivity – Finnish Micro-Level Evidence. In OECD (2004).

Maliranta M. and Rouvinen P. (2006) Informational Mobility and Productivity: Finnish Evidence. Economics of Innovation and New Technology, 15(6), 605–616.

Mankiw N G. (2006) The Macroeconomist as Scientist and Engineer. Journal of Economic Perspectives, 20(4), pp29-46;

Marrano G M. Haskel J. and Wallis G. (2007) Whatever happened to the Knowledge Economy, ICT, Intangible Investment and Britain’s Productivity Record Revisited. Queen Mary University of London working paper 603.

Martin R. (2005) Providing evidence base from business micro data: theory and evidence. PhD Thesis, London School of Economics

Matteucci N. and Sterlacchini A. (2003) ICT and employment growth in Italian industries, Working Papers 193, Universita' Politecnica delle Marche (I), Dipartimento di Economia.

Mattila L. and Strandell A. (2006) Defining and Measuring relocation and outsourcing of production. Institute of Growth Policy Studies A2006:005.

Melitz M J. (2003) The Impact of Trade on Intra-Industry Reallocations and Aggregate Industry Productivity. Econometrica. October 2003. Vol 71 Issue 6 Pp 1695 - 1725

Melitz M J. (2007) Volatility, Labor Market Flexibility, and the Pattern of Comparative Advantage. NBER Working Paper.

Merkouris T. (2004) Combining Independent Regression Estimators from Multiple Surveys. Journal of the American Statistical Association 99, no. 468: 1131-1140.

Motohashi K. (2007) Network use and productivity: a look into the black box. Presentation for OECD Joint WPIIS-WPIE Workshop, May 2007

277


Muysken, J. Schim van der Loeff S. and Cheshko V. (2006) Beyond unobserved heterogeneity in computer wage premiums. Working Paper 2006–006, UNU-MERIT, Maastricht, Netherlands.

OECD (2004) The Economic Impact of ICT, Measurement, Evidence and Implications. Editor: Pilat D, OECD, Paris.

OECD (2008) STI-Outlook, OECD, Paris (forthcoming).

Oliner S D. and Sichell D E. (2000) The Resurgence of Growth in the Late 1990’s: Is Information Technology the Story? The Federal Reserve Board San Francisco.

Olsen K B. (2006) Productivity impacts of offshoring and outsourcing, a review. STI Working Paper 2006/1.

Olsen K B. (2006) Productivity Impacts of Offshoring and Outsourcing: A Review. OECD Directorate for Science. Technology and Industry (STI) Working Papers, 2006(1).

O’Mahony M. Timmer M. Van Ark B. (2008) The EU KLEMS Growth and Productivity Accounts: An Overview. University of Groningen & University of Birmingham; downloadable at www.euklems.net.

ONS (2007) The ONS productivity handbook: a statistical handbook. Edited by: Camus D.

Pavitt K. (1984) Sectoral Patterns of Technical Change: Towards a Taxonomy and a Theory. Research Policy, 13: 343-73.

Pianta M. (2005) Innovation and Employment. The Oxford Handbook of Innovation, ed: Fagerberg et al., Oxford University Press, pp. 568 – 598.

Picart C. (2006) Gazelles in France. INSEE Working Paper.

Polder J M. (2008) Innovation and Productivity: Firm-Level Evidence and Aggregate Implications. Research Paper Division of Business Statistics, Statistics Netherlands (forthcoming).

Renssen R H. and Nieuwenbroek N J. (1997) Aligning Estimates for Common Variables in Two or More Sample Surveys. Journal of the American Statistical Association 92, no. 437.

Renssen R H. Kroese A H. and Willeboordse A J. (2001) Aligning estimates by repeated weighting. Unpublished report (BPA H 491-01-TMO), Heerlen, The Netherlands: Statistics Netherlands.

Restuccia D, and Rogerson R. (2007) Policy Distortions and Aggregate Productivity with Heterogeneous Plants.

Ridder G. and Moffitt R. (2003) The econometrics of data combination. In: Handbook of Econometrics. November 2007

Ritchie F J. (2004) Business Data Linking: Recent UK Experience. Austrian Journal of Statistics.

Romer P M. (1990) Endogenous Technological Change. Journal of Political Economy, volume 98, number 5, pp 71-102.

Särndal C E. (2007) The calibration approach in survey theory and practice. Survey Methodology: 99.

Schreyer P. (2000) The contribution of information and communication technology to output growth: a study of the G7 countries. STI Working Papers 2000/2.

Siegel D. and Griliches Z. (1992) Purchased Services, Outsourcing, Computers, and Productivity in Manufacturing. In Z. Griliches (Ed.), Output measurement in the service sectors (sivut 429–458). Chicago and London: University of Chicago Press.

Soete L. and Weel B. (2005) The Economics of the Digital Society. Edward Elgar.

278



279

Solow R M.(1987) We’d better watch out. NY Timers Book Review July 12th

Spiezia V. and Vivarelli M. (2002) Innovation and Employment: A Critical Survey. Productivity, Inequality, and the Digital Economy, eds.: Greenan et al., MIT Press, pp. 101 – 132.

Statistics Canada (2006) Efficient Small-Domain Estimation by Combining Information from Multiple Surveys through Regression. Working paper, HSMD-2006-007E.

Syverson C. (2004) Market Structure and Productivity: A concrete example Journal of the Political Economy. 2004, vol. 112, no. 6

Thouin M F. Hoffman J J. and Ford E W. (2008) The Effect of Information technology Investment on Firm-Level Performance in the Health Care Industry. Health Care Management Review, 33(1), 60-68.

Tronti L. (2007) “The July Protocol and Economic Growth: The Chance Missed", Social Pacts, Employment and Growth. A Reappraisal of Ezio Tarantelli’s Thought. Acocella N. and Leoni R. (eds.), Physika-Verlag, Heidelberg.

Ulset S. (1996) R&D Outsourcing and Contractual Governance: An Empirical Study of Commercial R&D Projects. Journal of Economic Behavior and Organization, 30(1), 63-82.

Van Ark B. O’Mahony M. and Timmer M P. (2008) The Productivity Gap Between Europe and the United States: Trends and Causes. Journal of Economic Perspectives, Vol. 22(1), pp. 25 – 44.

Van der Wiel H. and Van Leeuwen G. (2004) ICT and Productivity, In: Fostering Productivity, Patterns, Determinants and Policy Implications. Contributions to Economic Analysis no. 263, Elsevier, the Netherlands, pp. 93 – 112.

Van der Wiel H. Creusen H C. Van Leeuwen G. and Van der Pijll E C. (2008) Cross the border and look around. Research Document, Netherlands Bureau for Economic Policy Analysis (CPB), (forthcoming).

Van Leeuwen G. and Klomp L. (2006) On the contribution of innovation to productivity growth. Special Edition of The Economics of Innovation and New Technology, Vol. 15 (4/5), pp. 367 – 390.

Van Leeuwen G. and Van der Wiel H. (2003) ICT, Innovation and Productivity. CAED Conference 2003, London.

Van Leeuwen, G. and Van der Wiel H. (2003) Do ICT spillovers matter? Evidence from Dutch firm-level data. Discussion Paper No. 26, CPB Netherlands Bureau for Economic Policy Analysis, The Hague, Netherlands.

Vaze P. (2001) ICT Deflation and Growth: A Sensitivity Analysis. Office for National Statistics, UK. Available on-line at www.statistics.gov.uk.

Wallis G. (2006) ICT Deflation and Productivity Measurement. Office for National Statistics, UK. Available on-line at www.statistics.gov.uk.

White H. (1980) A Heteroscedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroscedasticity. Econometrica, 48(4), 817-838.

Wu C. (2004) Combining information from multiple surveys through the empirical likelihood method. The Canadian Journal of Statistics 32, no. 1: 15-26.

Zieschang K D. (1990) Sample Weighting Methods and Estimation of Totals in the Consumer Expenditure Survey. Journal of the American Statistical Association 85, no. 412: 986-1001.

Eurostat Agreement No. 49102.2005.017-2006.128 Appendix I

APPENDIX I

User guide

Peter G. Stam UK Office for National Statistics Objective 1. This document is intended to act as an uncomplicated guide for countries who wish to become involved in the EU ICT Impacts project. It includes a brief background to the project, guidance in relation to the vital firm level data preparation and metadata stages. Definitions and explanations of variables, weighting and deflators used in the project are followed by information on the appropriate aggregation of data before and after running the code. Advice then follows about the actual checking of results, disclosure and common analysis.

2. This note has been requested by Eurostat as an element of the above project and draws closely upon the ‘Summary for Phase 2 Countries’ (dated 31st May 2007).

Background and overview 3. The project has been designed to demonstrate a common approach so that it can easily be extended to other countries.

4. The “EU ICT impact assessment by linking data from different sources” project was launched in April 2006. Eurostat has prioritised making better use of data collected, particularly on the Information Economy and competitiveness. Linking data sources should permit National Statistical Institutes (NSIs) “to identify how ICT adoption affects business behaviour and performance” using data from existing sources linked at firm level. Also to develop processes and datasets, this will allow the impact of technology to be assessed over time and across the EU.

5. An early cross country comparative study of this work has already been led by OECD, based on testing different models for behaviour and performance on different national datasets. The UK Office for National Statistics has tested a variety of metrics (IT investment, IT use, e-transactions and communications) on UK data, and broadly confirmed the OECD conclusions. Linked surveys across countries allow research into the effects of technology inputs on behaviour and performance. This project is intended to be a more comprehensive approach to metrics for impact assessment, involving more countries.

6. A number of member states have already undertaken firm level data analysis on:

- the policy and economic environmental factors that influence ICT adoption - effects of ICT investment and use on firm productivity - effects of ICT use on market behaviour (e.g. pricing) - how other inputs (e.g. skills, communications infrastructure) affect the productivity of

ICT 7. As the project has progressed it has become clear that both firm level analysis (to validate impact indicators) and industry level analysis (to develop stable indicators and investigate industry / country level effects) need to be developed.

8. The project has links with the EU KLEMS (European Capital Labour Energy Materials and Services) project “Productivity in the European Union: A Comparative Industry Approach” (where one of the inputs is ICT capital). Further information on this project can be found at http://www.euklems.net/.

280


9. A useful summary of the methodology used in the ICT project can be found in Bartelsman and Barnes (2001).

Expectations 10. By its conclusion this project will have:

▫ Created a framework for analysis; using common surveys and code ▫ Reviewed analytical conclusions ▫ Made recommendations on indicators of ICT impact and on methodology to apply

them ▫ Made recommendations on surveys to support indicators of ICT impact,

innovation and competitiveness Metadata 11. This project combines data from several sources so it is crucial that metadata (that is, data about the underlying data) is clearly defined and checked before it is submitted. Linking data from different countries obviously has inherent metadata hazards, such as units (currencies) and linguistic meanings on source surveys etc. Thorough metadata checking must be performed and the data standardised before submission to running through the core code. By doing this, unnecessary output errors will be minimised later. The checks have been broken down by source variables (that is the E-Commerce variables and the Production Survey variables) and, results stored by the project leaders in the UK.

12. While good metadata checking did occur at the beginning of this project, variables have been defined and extended. For this reason the U.K took the lead in extending a second round of variable metadata checks through simple survey-specific questionnaires. Upon return of these questionnaires there was a process of checking that cross country data was still compatible, and advising individual countries what changes were required in order to assemble data which could be put side by side with any other project member’s data. The objective of these metadata checks is to ensure that any analysis produced from the data is not misinformed or incorrectly specified and is an essential action in order to participate in the project.

13. The metadata framework permits production of indicators which are completely comparable across countries, and auditable back to source data, but also allows us to generate data which can be used for analysis across countries and industries of differences in behaviour and performance associated with ICT use.

Firm level data preparation 14. In order for the core code to work analysts in each of the NSIs must review the native micro data sets and write (and document) routines to transform the native data into a set of input data files that can be read by the project’s code.

15. This will involve:

▫ Rationalising native data sets, for example eliminating industry sectors from the Production Survey [PS] (otherwise known as the ‘Structural Business Survey’) datasets that are not covered by the eCommerce Survey [EC] (otherwise known as the Community Survey on ICT Usage and e-Commerce in Enterprises) to speed up the runtime of the code.

281


▫ Checking naming conventions and ensuring that variables that might take different names in native datasets in different years of the surveys are transformed to a common name in the input data files.

▫ Checking the domain and order of the native data, and making any necessary transformations to the domain and ordering conventions assumed by the common code.

▫ Reviewing the treatment of missing data, outliers etc. 16. Similar to the cleaning of data it is crucial that joining countries properly treat ‘zero’ and ‘missing’ values. This project distinguishes between the two as such:

Missing value = “.” Zero Value = “0”

17. Before running this code a country must assemble:

▫ Business_Register (BR) A register with variables: unique_firm_id, year, industry_code num_employees and others. Users must be able to link records between the eCommerce (EC) and Production surveys (PS),. This is a single file covering the whole period for which EC and PS data are available.

▫ Production_Survey (PS), (otherwise known as the ‘Structural Business Survey’) With variables: unique_firm_id, industry_code, num_employees, survey_weight and other_vars (see below). There is one PS file for each year of data.

▫ eCommerce_Survey (EC), (otherwise known as the ‘Community Survey on ICT usage and e-commerce in enterprises)

With variables: unique_firm_id, industry_code, IT_use and other_vars (see below). There is one EC file for each year of data.

▫ Industry_Concordance Linking all occurring industry code values in a participating country to the industry classification used in the ICT project (EUK). This is an important file, requiring individual action from each user of the code.

18. It is normal practice to include at least one year of PS data prior to the earliest available EC data. This is to facilitate computation of lagged variables.

19. The survey files must be named with the following convention: prefixYYYY, (for example PS1999 PS2000 PS2001; Inv1999 Inv2000; eCom1999 eCom2001). This project examines data from these surveys over the period 2000 to 2005, where these surveys overlap. The year (YYYY) refers to the year that the data pertain to, not the year the survey was conducted. It is not necessary for surveys to be available for all years.

20. Cleaning of data early on in the process makes it relatively uncomplicated to find errors, as opposed to cleaning after merging the data.

Variables 21. Table I.1 displays a list of the core variables. It was adapted from the ‘Summary for Phase 2 Countries’. Table I.1 is, however, subject to change as further themes are investigated by project members.

282


Table I.1: Summary of key variables Variable Name Description Domain E-Commerce Survey Variables

DSL Firm has broadband Boolean (1=Yes, 0=No) DSLpct % of workers with access to broadband Percentage (Range 0–1) Epurch Firm orders through Internet (or EDI) Boolean Epurchpct % of orders through Internet (or EDI) Percentage Esales Firm sells through Internet (or EDI) Boolean Esalespct % of sales through Internet (or EDI) Percentage Inter Firm has Internet Boolean Interpct % of workers with access to Internet Percentage Intra Firm has intranet Boolean Intrapct % of workers with access to intranet Percentage PC Firm uses computers Boolean PCpct % of workers using computers Percentage Web Firm has website Boolean Production Survey Variables

Country Country code Text string Year Year to which data pertain Numerical string Euk Industry classification Text string Sz_Cls Size class (based on employment) Numerical class (0 – 7) Frgn_Own Multinational dummy Boolean (1=Yes, 0=No) SRC Tabulation type (for sorting data) Numerical class (1 – 10) NQ Nominal sales Level NV Nominal value added Level NM Nominal material inputs Level Pay Payroll Level E Employment Level Wage Derived Pay/E Level LnWage Log of Wage Level K Capital stock Level Productivity Variables

(Computed by code)

LPQ Labour productivity based on real sales Level LPV Labour productivity based on value added Level TFP TFP (Value added with capital and labour) Level MFP MFP (Gross output with capital, labour and

materials) Level

Additional Variables

Emp_BR Number of employees given in BR dataset Level Export Export dummy, exporter = 1 non-exporter = 0 Boolean (1=Yes, 0=No) Age Age of firm in years Level Frgn_own Foreign owner (Business Register) Boolean He High-growth enterprises) (Computed by code) Boolean Gzl Gazelle (Computed by code) Boolean

283


Variable Name Description Domain Skills Hkpct Number of employees with upper secondary

education Level

Hkitpct Number of employees with IT upper secondary education

Level

Hknitpct Number of employees with non-IT upper secondary education

Level

ICT Integration Intlink_1 Internal link to systems for re-ordering

replacement supplies Boolean (1=Yes, 0=No)

Intlink_2 Internal link to invoicing and payment systems

Boolean

Intlink_3 Internal link to systems for managing production, logistics or service operations

Boolean

Extlink_s External links to suppliers’ business systems Boolean Extlink_c External links to customers’ business

systems Boolean

Intlink_1 Internal link to systems for re-ordering replacement supplies

Boolean

Intlink_2 Internal link to invoicing and payment systems

Boolean

22. It is possible that some of the variables may be available on different native datasets, other than on structural business surveys. In this case the user must pre merge the required variables into the PS files (which are used as inputs to the programme code).

23. It is not essential that all of these variables are held in order to effectively participate in this project. Variable names are as they are used in the code for compatibility. Countries do not need to change their variable names, as the code provides opportunity to assign ‘native’ names to ‘code’ names. Likewise it is not necessary to hold all of the survey years as the code is robust in its handling of missing data.

24. All E-Commerce (EC) variables should be scaled from 0 - 1. This is obvious for the Boolean variables (PC, Web etc), which should only take values of 0 or 1 (or missing). However, the dimensions of the percentage variables on the EC survey (PCpct, Epurchpct etc) must be checked and transformed to a scale of 0 - 1 . Any values outside of the range 0 - 1 should be investigated and corrected. The advantage of this approach is that all EC summary statistics will have the same dimensions and there will be no need for the code to do any re-scaling. This approach must be upheld throughout the duration of the project in order to ensure consistency within the data; all EC summary statistics will then have the same dimensions.

Weighting 25. The code uses a variable from the Production Survey (PS) called "wgt_PS" which represents the weight of that firm within the PS sample frame. This variable should be ≥ 1, that is standing for the number of firms represented by that record in the dataset. Where native data has probabilities (wgt_PS ≤ 1), the inverse must be taken before the project code is run. Any values outside the appropriate range need to be investigated and corrected or replaced with missing values. The weight variable, however, is not essential to the running of the code.

284


Deflators 26. Data in this project shall be deflated; either using National Accounts data or by using (where available) price deflators derived from the EU KLEMS project (http://www.euklems.net/).

Appropriate aggregation of data 27. The project utilises two methods of aggregating the data into industry classifications. Joining countries must be able to conform to the EUK industry classifications (see Chapter 4 above) and be able to classify data at firm level. This could comprise the creation of a mapping of country industry codes to EU-KLEMS industry codes. Using a standardised industry coding approach allows industry analysis to be conducted at cross country level and also assists in the consistent aggregation of the data.

28. The code will then use this aggregation to create an alternative classification structure (which it calls ALT). This structure has been created to allow relevant analysis with easy appropriate aggregation. Further details on this classification is provided in APPENDIX A

29. When reviewing outputs users can select which classification to view by using the SRC variable as summarised in Table I.2: Table I.2:

SRC Tabulation by: 1 Euk, Year 2 Euk, Year, Sz_Cls 3 Euk, Year, Mnc 4 Euk, Year, Frgn_Own 4.2 Euk, Year, Hge 4.4 Euk, Year, Gzl 5 ALT, Year 6 ALT, Year, Sz_Cls 7 ALT, Year, Mnc 8 ALT, Year, Frgn_Own 9 ALT, Year, Hge 10 ALT, Year, Gzl

30. As shown in Table I.2 the variables Mnc, Frgn_Own, Hge, Gzl can be used as aggregation tools. These are flag variables (1=Yes 0=No) and are created automatically by the code.

31. Sz_Cls (size class) is also a variable which the code automatically generates using employment data. The structure of which is explained in Table I.3:

Table I.3: Sz_Cls Employment 0 emp=0 1 0≤emp<10 2 10≤emp<20 3 20≤emp<50 4 50≤emp<100 5 100≤emp<250 6 250≤emp<500 7 Emp≥500

285



Running the code 32. Once a country has organised the data they must ensure that they have software compatible with the project code. This is either Statistical Analysis System (SAS) http://www.sas.com or STATA http://www.stata.com. The U.K. Office for National Statistics is custodian of the code. For access and advice please contact either:

Peter Stam [email protected] or Mark Franklin [email protected] Office for National Statistics Government Buildings Cardiff Road Newport NP10 8XG United Kingdom 33. In order to begin analysis a country must open the code file run_ONSPROD which will then call in the other modules and place output datasets within the specified output folder. Users must then

- Set the running parameters in this file (Not changing any other files)

- Save as and run (execute) this file. 34. Countries are required to state for the code which variables they hold. If any variables are not held then leaving the relevant flag in the code blank will ensure that the code understands the variable is missing and will treat it appropriately.

35. The user must specify input and output folders and ensure that the survey files mentioned previously are placed in the input directory. This takes place in the section titled ‘USER EDITS to STATA/SAS Program’.

36. Modules can be activated or constrained by editing the user edits in the run_ONSPROD file. Each part of the code (coverage, summary stats, correlations etc) can be switched on or off to save time or to assist in the diagnosis of run time errors. To use this function a country must set the parameters for the run of the code. For example by specifying TRUE for the Docov option in the run_onsprod file, the code will create the COVERAGE dataset. See the table below for further explanation of module options.

37. The STATA format is numerical (0 = restrict, 1 = run), the SAS format follows a ‘TRUE’ / ‘FALSE’ system.

38. Listing of the modules follows in Table I.4; more detail on the output datasets that these modules create is available later in this guide:

286

http://www.sas.com/

http://www.stata.com/




Table I.4: Summary of module running

39. Further detail on output datasets is found in Table I.6

Module name Module description DoReWgt A module to compute weighting procedures DoCov A module to create COVERAGE output dataset. DoIndd A module to compute industry dynamic statistics DoStat A module to compute summary statistic datasets. DoCorr A module to compute ST and CR output datasets. DoReg A module to compute regressions on the data DoTabs A module to provide users with some summary tables from output datasets.

40. Table 1.5 (at the end of this Appendix) is a screenshot on how the SAS code can be set up to run to the users specification and how to set up with individual data. In this example the code will run fully for data pertaining to 2000 – 2004; but it will not create any summary tables (as the ‘DoTabs’ module has been turned off)

41. The code will then be executed and will call in the specified programmes. Run time will vary according to:

▫ The size of input datasets ▫ The number of years of survey data ▫ The number of computational modules.

With large datasets the full suite of project code can take several hours to compute.

42. Table I.6 displays the outputs that the code produces for each country.

Table I.5: Summary of outputs produced by the code Name Description Coverage Information on linked BR, PS and EC Demogr Firm demographics data from BR Inddyn Industry-specific dynamics ECstat Summary statistics from EC file PSstat Summary statistics from BS file PSECstat Summary statistics from linked EC and BS files PSst File with moments of distributional variables in PS PScr File with moments of joint distribution of two variables in PS PSECst File with moments of distributional variables in merged PS / EC PSECcr File with moments of joint distribution of two variables in merged PS / EC Regw Production function regressions using wages RegSkills Production function regressions using skills (where available) Availability File with flag for variable availability for each country Features File with flag for availability of ‘analysis feature’ for each country Checking of results 43. It is the responsibility of countries to check their own data to verify that the code has executed without errors and that the output datasets do not contain any errors.

Disclosure 44. This code runs on firm-level data and would normally be subject to stringent disclosure controls. Despite the fact that the outputs of the code are aggregated at industry level, some

287


cases are still subject to disclosure restrictions. An example of this is where the underlying number of firms is below a threshold level.

45. An objective of this project is to conduct cross-country comparisons and there has been extensive discussion over the subject of disclosure. A ‘rule of thumb’ has been adopted that each project member willing to share outputs with other project members for research activities can do so on similar terms and control access to firm-level data.

46. There is no presumption that new participants would benefit from this arrangement, although they are free to negotiate access to other countries’ results. It is also likely that some high-level outputs will be published (subject to full disclosure approval be all project members). Obviously new participants would have no external restrictions on their own use of the outputs for their country. However failure to agree to this arrangement reduces the size of the cross-country database, and means that joining countries will not be included in any cross-country analysis (potentially damaging the project’s ability to conduct valuable macro analysis).

Analysis 47. In the original brief, Eurostat discussed that national and international research on ICT adoption and its drivers is well established, but there is scope to extend it as longer datasets for surveys become available in more countries.

48. Further discussion between project members has refined the analysis outputs as outlined below.

Core analysis:

Productivity and core ICT metrics Labour Productivity Gross output MFP Net Output MFP

ICT Characteristics: Fast growth Firms Employment effects and ICT adoption

There is wider interest from Eurostat in:

▫ Human capital (e.g. skills) ▫ Physical capital (e.g. ICT soft and hardware, non-ICT hardware) ▫ Materials and services, including communications ▫ Knowledge capital ▫ Can we measure the forms of capital across countries, and how firm investment

behaves? ▫ What is the impact of ICT on:

Output and quality/price. The impact on output is the productivity agenda, the impact on quality/price is the profitability agenda. How does ICT feed into productivity growth: does it need other inputs, does it have TFP effects or is it just a part of capital deepening? How does ICT affect quality and prices; e.g. do online sellers reduce prices relative to the competition?

Other inputs. How does ICT affect demands for unskilled/routine jobs? How does ICT affect organisational structures? How does ICT affect the process of innovation that gives rise to changes in technological capital?

49. Country level analysis needs to ensure limited analytical resources are not duplicated. It is important that analysis is as identical as possible across countries. For that reason, measures of impact this project is interested in for analysis have been decided. See table I.7.

288


Table I.6: Levels Growth rates

Productivity measured as - Gross sales / employee (vs

industry) - V.A. / employee (vs industry)

Productivity measured in terms of Multi-Factor Productivity (for countries which have firm level data on investment, IT investment, labour and other inputs)

Productivity change measured as - Gross sales / employee (vs

industry) - V.A. / employee (vs industry)

Growth rates measured for:

- Gross sales (deflated by industry PPI)

- Market share of IT using firms in industry

- Value added (double deflated where possible)

- Employment, or share of employment in industry

Sub themes 50. In addition to the common code and the analysis that flows from it, Phase 1 saw the commencement of additional analyses into certain sub-themes. If joining countries feel able to assist, the themes currently underway [and their leaders] are:

▫ ICT characteristics of fast growing firms [France] ▫ Employment effects of ICT adoptions [Slovenia] ▫ ICT Investment [Netherlands] ▫ Skills [Sweden] ▫ Offshoring and outsourcing [Sweden] ▫ Innovation [Netherlands]

Resource Requirements 51. Time requirements for effective participation in this project will obviously vary between countries, depending on the resources available to individual NSIs. In order to gain some idea on these requirements, information on the time budgeted for phase 2 countries to join the project is provided below.

▫ Data sourcing and cleaning – Between 20 and 30 days. This will depend upon the data quality and ease of access to the micro-data

▫ Application of common analysis – 20 days ▫ Reviewing, reworking and developing – 19 days ▫ Conclusions and recommendations – 6 days

52. It must be stressed that these time allocations are estimates only. In particular large variability has been found in the time it takes NSIs to carry out data cleaning. Key factors include quality of the data, quantity of the data and ease of access to the data. Analysis is also a very open ended activity which makes estimation of time needed unreliable.

Further Information For further information please contact: Peter Stam: Office for National Statistics, UK E-mail: [email protected] Telephone: +44 (0) 1633 455982

289



290

Table I.7: Screenshot of STATA code front screen for code run set-up

***Calls main macro, defined in ONSprod.sas program; ***The parameters given below are for TEST data; ***USER should call macro with parameters relevant for own case; ***See descriptions of macro parameters at bottom of this file; %ONSPROD( _ _ _ ccc = TST, |

Code run restrictions typestr = LPV LPQ TFP MFP,| startyr = 2000, | User specifies parameters here

endyr = 2004, |<-- survlib = dirIN, | unit = firm_id, | year = yyyy, _ _ _| BusReg = br, | ind_BR = indn, | emp_BR = empment, | frgn_own = foreign, | mnc = multifl, | age = age, | pre_PS = survey, | ind_PS = indn, | nv = nvadd, | Code naming maps. nq = ngo, | e = emp, | Left: pay = payroll, | nm = nmat, |<-- Right:

Code variable name

k = kstock, | hkpct = hk_emppct, |

Variable name to be mapped (specific to country) hkitpct = it_hkemppct, |

hknitpct = nonit_hkemppct, | export = exportflag, | wgt_PS = , | pre_EC = ecommerce_sur | ind_EC = indn, | PC = PC, | PCpct = PCpct, | Web = Web, | Epurch = Epurch, | Epurchpct = Epurchpct, | Esales = Esales, | Esalespct = Esalespct, | Inter = inter, | Interpct = interpct, | Intra = intra, | Intrapct = intrapct, | DSL = DSL, | pfile = TSTdefl, | concfile = conc, | hierfile = althier, _ _ _| keepBR = false, | DoReWgt = true, | DoCov = true, | Module parameters. DoIndd = true, | DoStat = true, |<-- Users can ‘turn on /

off’ run modules here

DoCorr = true, | DoReg = true, | DoTabs = false, _ _ _| debug = 0 );

49102.2005.017-2006.128 - ICT IMPACTS - FINAL REPORT V2

Documents

uk statistics finland

official national statistics

ict investment

ict outsourcing

identical firm level

data peter stam

micro data analysis

identical data sources