-
the key factors behind such a bproductivity gapQ has long been
recognized in the lack ofability of the Italian labour force to
adapt to the everchanging needs of the global market.
Labour Economics 12 (2005) 557576
www.elsevier.com/locate/econbase1. Introduction
Between 1995 and 2002, the annual growth rate of hourly labour
productivity in
manufacturing has been 4.5% in the US, 4.6% in France, 2.4% in
Germany and only 0.9%
in Italy (OECD, 2002). The Governor of the Bank of Italy,
Antonio Fazio, in his bFinalConsiderationsQ of this years Report to
the annual General Meeting, has urged immediatepolicy responses to
stop the loss of competitiveness suffered by the Italian system.
One ofTraining, productivity and wages in Italy
Gabriella Conti*
Department of Economics and Institute for Social and Economic
Research, University of Essex,
Wivenhoe Park, Colchester CO43SQ, United Kingdom
Abstract
This paper presents for the first time panel evidence on the
productivity and wage effects of
training in Italy. It is based on an original dataset which has
been created aggregating individual-level
data on training with firm-level data on productivity and wages
into an industry panel covering all
sectors of the Italian economy for the years 19961999. I use
several modelling specifications and a
variety of panel data techniques to argue that training
significantly boosts productivity. However, no
such effect is uncovered for wages. This seems to suggest that
firms do actually reap more of the
returns.
D 2005 Elsevier B.V. All rights reserved.
JEL classification: C23; J24; J31
Keywords: Training; Productivity; Wages0927-5371/$ -
doi:10.1016/j.
* Tel.: +44 1
E-mail add206 874875.see front matter D 2005 Elsevier B.V. All
rights reserved.
labeco.2005.05.007
ress: [email protected].
-
However, notwithstanding the complete redesign of the training
system carried out in the
recent years, Italys performance appears highly unsatisfactory
also on this ground. The
available evidence clearly shows that Italy has one of the
lowest values of training
incidence in the European Union, together with other
Mediterranean countries, such as
Portugal and Greece (Brunello, 2002).
These facts appear rather puzzling if analysed in the light of
rigorous and sound
economic theory. The recent approach to training in imperfect
labour markets (Stevens,
1994a,b,c and 1996; Acemoglu and Pischke, 1998 and 1999a,b)
points to some forms of
labour market imperfections as driving a wedge between increases
in wages and increases
in productivity, allowing the firms to recoup some of the costs
of training, and fostering
their incentives to invest in it. This is consistent with some
empirical findings which show
a lower level of on-the-job training in the US compared to
Germany and Japan (Lynch,
1994). However, this is not true for Italy. Italy is on the top
of ranking of regulated labour
markets, with a strong role of unions in wage bargaining, and
high hiring and firing costs.
However, in sharp contrasts with the predictions of the theory,
Italy is trapped in a low-
training equilibrium.
The research presented in this paper is an attempt to shed some
light on this puzzle,
testing for training effects on labour productivity and wages.
Many studies have tried to
establish this link in an international context. However, no
such work has been done for
Italy. Moreover, the available literature has achieved
controversial results, which seem
to depend strongly on the training measure used, the modelling
specifications and the
estimation techniques adopted, and the controls included in the
empirical model.
Overall, the majority of the studies have found a positive
impact of training on
productivity, although often not significant. In addition, some
forms of training (general
training and off-the-job training) seem to have a greater
effect. This study takes stock of
the available knowledge, and tries to overcome the limitations
of the previous studies in
several ways.
First of all, due to the lack of longitudinal data, many studies
have failed to control
for unobserved heterogeneity (Black and Lynch, 1995 and 1996),
and potential
endogeneity of training (Bartel, 1994; Bishop, 1994; Barrett and
OConnell, 2001);
longitudinal data with repeated training information have become
available only recently
(Black and Lynch, 2001; Dearden et al., 2000; Ballot et al.,
2001; Zwick, 2002). This
paper deals with both the issues of unobserved heterogeneity and
endogeneity of
training, by using a variety of panel data techniques on an
original dataset which
contains longitudinal information on training and measures of
corporate productivity
covering all sectors of the Italian economy for the years
19961999. Coherently with the
previous literature, I show that failing to take into account
these issues leads to severe
biases in the estimates.
Secondly, most of the available studies have used a flow measure
of training, due to
the lack of an appropriate measure of the stock of human
capital. However, measuring
training participation only over a relatively short period of
time fails to take into account
the role played by skills accumulated during the working life.
The richness of the
database allows me to overcome also this limitation. So, I
construct a stock measure of
G. Conti / Labour Economics 12 (2005) 557576558training, using a
question which has been consistently asked over time in the
Italian
Labour Force Survey.
-
used for the empirical estimation. A simple empirical framework
to analyze the impact oftraining on productivity and wages is
specified in Section 3. Section 4 is devoted to the
presentation and the discussion of the results. A section of
concluding remarks and
directions for future research closes the paper.
2. The data
The empirical analysis is based on a original panel which has
been created merging two
different complementary datasets. In doing so, I have followed
the methodology adopted
by Dearden et al. (2000). I have used the 19961999 waves of the
Italian Labour Force
Survey (April quarter), assembled with accounting data on firms
for the corresponding
years, drawn from the AIDA Database. The reason for this choice
relies on the fact that no
one Italian dataset contains the information on training and
measures of corporate
performance required for this kind of analysis.
The first database used is the Italian Labour Force Survey. This
is a household-level
survey carried out with a detailed questionnaire every quarter
(in the months of January,
April, July and October) since 1959. Around 75,000 households
are interviewed every
three months, for a total of approximately 200,000 individuals.
From the LFS I gather
information on training, personal characteristics of the
individuals (sex, age, region of
residence), measures of their skills (educational qualification
and occupation), job
characteristics (hours of work) and workplace characteristics
(industry sector). The
sample includes all men and women aged between 15 and 64
inclusive who were
employed at the time of the survey (including the
self-employed).Thirdly, many studies have used very parsimonious
specifications both in terms of the
modelling strategy adopted and in terms of the controls for
firms and workers
characteristics included in the empirical model. In this paper,
I use several different
specifications of the baseline model (an augmented Cobb-Douglas
production function), in
order to relax the assumption of constant returns to scale and
to evaluate the effect of
training on the growth of labour productivity and wages. In
addition, the richness of the
database allows me to control for a plethora of firms and
workers characteristics,
including another important intangible investment of the firm,
namely R&D, and other
measures of workers skills, such as education. I show that the
results are sensitive to the
modelling specification adopted, and that the inclusion of
several controls significantly
reduces the magnitude and the level of significance of the
estimated returns.
Finally, following Dearden et al. (2000), the production
function estimates are explicitly
compared with the wage equations. This is important to examine
how the benefits from
training are shared between the firms and the workers. The
recent models of training in
imperfect labour markets predict that the benefits from training
do not fully accrue to the
workers. I show that firms do indeed reap most of the returns:
the main finding I obtain is
that training significantly boosts productivity, while no
significant effect is uncovered for
wages. This result is robust to alternative specifications.
The outline of the paper is as follows. In Section 2 I describe
the data and the variables
G. Conti / Labour Economics 12 (2005) 557576 559Two main
training questions are asked in the LFS. The first refers to
courses undertaken
by the individual in the month before the interview, and has
changed slightly in 1998;
-
however, it was possible to build up a consistent series because
the basic structure of the
question was unchanged. The question in 19961997 was: bDuring
the month before theinterview, have you attended any of the
following courses?Q1; then, in 19981999 it wasrephrased as: bDuring
the month before the reference week, have you attended any of
thefollowing courses or have you taken part in any of the following
on-the-job training
activities?Q.2 The main drawback with this training measure is
that it is a flow variable. Inorder to derive a measure of the
workers stock of post-schooling human capital, I have
used the answer to the second major question available in the
LFS, which refers to training
received by the individual during his whole life. In 19961997
the question was: bWhat isthe highest level of vocational training
achieved during your life?Q; in 19981999 it wasrephrased as:
bDuring your life have you concluded a vocational training course
or haveyou taken part in an on-the job training activity?Q. I have
used this question, combinedwith the previous information on
current training, to calculate the stock of post-schooling
human capital in each year.3
Most of the literature uses the current level of training to
measure the stock of human
G. Conti / Labour Economics 12 (2005) 557576560capital in the
firm. However, this procedure is correct only if one assumes
fully
depreciation of the skills after one period. Moreover, it is not
consistent with the
underlying theory.4 Henceforth, in this paper I have followed a
different procedure.
Following Boon and van der Eijken (1998), I have constructed the
stock of human capital
as the sum of the proportion of workers trained at time t in
industry i (the flow) and the
stock of the previous year,5 taking into account depreciation,
according to the Perpetual
Inventory Method:
TRAINit FLOWit 1 d STOCKi;t1 1
where y measures the depreciation rate of the human capital.
Since an exact value fory is essentially unknown, I have
experimented with various rates of depreciation. Theresults were
not sensitive to different values, within a plausible range 5%35%,
so I
1 The information collected in this question refers exclusively
to courses which are connected with the current
job, or relevant for a job that the interviewed might be able to
do in the future.2 In the 19961997 questionnaire, a list of 10
courses was available; in 1998, this number was increased to 14
courses, and in 1999 two more options were added. These
additions have been necessary in order to take into
account the greater range of options available, following the
introduction of the reform in the training system after
the Treu Law 175/98.3 Along with the type of courses attended,
the LFS also contains additional information related to
training
undertaken in the month preceding the interview, such as the
scope of the course, its overall duration, the number
of weekly hours of training (only available for the years
19961997). However, the focus on the stock of
accumulated post-schooling human capital, combined with the
great amount of non-response, hampered the use
of these additional questions.4 Some papers calculate the stock
of training by cumulating past flows: however, this methodology
suffers from
the weakness of not having suitable initial values.5 The
availability of LFS data for 1995 (April quarter) allowed me to
obtain a measure of the stock of human
capital with reliable starting values also for 1996. Moreover, I
have used the answer to the question at t1 inorder to avoid
double-counting: although the question refers to the training
already concluded, and not stillongoing, there is a small chance
that the worker would have completed the course in the beginning of
the
preceding month, hence referring to the same episode in both
replies.
-
have chosen a value of y=0.15, taking the estimates derived in
Groot (1998) as abenchmark.6 Moreover, I have controlled for
turnover in all the estimated specifications, in
order to take into account the loss in human capital arising
from separations. It is crucial to
account for reallocation of workers across industries, since the
stock measure is derived
from information on past courses, but the exact reference period
is not known. Henceforth,
it could be well the case that the training course was attended
while the worker was
G. Conti / Labour Economics 12 (2005) 557576 561employed in a
different sector. Since I dont follow the same workers over time, I
cannot
derive a proper measure of turnover at individual level;
however, including controls for
inflow and outflow rates across different sectors will control
for much of this problem.
Moreover, each year at least 40% of the workers interviewed
reported that they had
completed a formal course in a school with duration of at least
one year and release of final
certificate: henceforth, it is reasonable to believe that the
training undertaken mostly
provides portable skills, which are likely to have a
long-lasting impact on productivity.
Finally, given that we are most likely to see workers moving
across sectors in case of wage
and productivity gains, we should see the effect of the initial
training even if this training is
not directly relevant in the current job.
The second source used in this paper is AIDA (Analisi
Informatizzata delle Aziende).
This is a private database,7 which provides accounting
information from the balance sheets
of all Italian companies with an annual turnover higher than one
million Euros.8 The
original sample contained 189,059 firms, covering a ten-year
period from 1993 to 2002.
However, an accurate work of data cleaning has reduced the
sample size. In the first place,
observations for years before 1996 have been excluded, due to
severe reduction in the
number of firms with reported information compared to the
following years: preserving
those observations would have severely affected the
representativeness of the sample, due
to the fact that only information for a smaller subset of firms
is available for the early
years. Secondly, the years subsequent to 1999 have been
excluded, due to the impossibility
of deriving a consistent series for the training variable in the
Labour Force Survey. Then,
40,141 firms (332,936 observations) have been excluded due to
incomplete balance sheets,
and further 16,543 firms (45,673 observations) for lack of
consistency between specific
budget items. Henceforth, the final sample is an unbalanced
panel of 132,039 firms,
covering all sectors of the Italian economy for the years
19961999. From the AIDA
database I have derived information on value added, wages,
capital stocks, R&D
expenditure and employment. Real values have been obtained by
deflating the nominal
measures with two digit producer price indices for the different
years provided by ISTAT
(the Central Statistics Institute).
The data drawn from the two datasets have then been aggregated
into proportions
(for the variables training, male, age, education and
occupation, taken from the LFS)
6 Groot (1998) develops a model to estimate the rate of
depreciation of human capital. He estimates the model
on data for Great Britain and the Netherlands, and finds that
the rate of depreciation of education is 1117% per
year. Since there is no reason to believe that economic and
technological changes happen at a faster rate in Italy
than in UK or the Netherlands, I have used an average value of
15% in my estimations.7 It is provided by Bureau van Dijk.8 It is
important to note that the absence of any dimensional limit
constitutes one of the main strenghts of thedatabase used, given
the structural composition of the Italian industry, mainly formed
of small and medium
enterprises.
-
Table 1
Training incidence by sector
Rank ATECO2002 Description Flow(%) Stock(%)
1 11 Education, Health and related Social Services 10.74
43.38
2 8 Finance, Banking and Real Estate 8.12 37.29
3 2 Energy, Mining and Quarrying 5.96 35.46
4 9 Business Services and other Professional Activities 5.79
32.39
5 12 Community, Social and Personal Services 4.44 27.26
6 7 Transports and Communication 3.72 23.49
G. Conti / Labour Economics 12 (2005) 557576562and averages (for
the variables value added, wages, capital stock, R&D
expenditure,
hours worked and employees) at industry9 level, and then merged.
The rationale behind
this choice relies on the different level of aggregation
available in the two datasets: while
the AIDA database contains data disaggregated at the firm level
(5digit ATECO2002),
the Labour Force Survey only provides divisional information at
a higher level of
aggregation (12 sectors). Aggregating the data also at regional
level has, henceforth,
two advantages: on the one side, it increases the level of
disaggregation in a
geographical dimension, on the other, it allows to take into
account the high
productivity differentials and the marked disparities in
industry agglomeration and
labour market outcomes existing in the Italian regions. As
argued in Dearden et al.
(2000) aggregation allows to capture the within-industry
spillovers that would be left
out in case of a firm- or individual-level analysis,10 although
the advantages arising
from this methodology have to be weighed against the possible
problems due to
aggregation bias (see Grunfeld and Griliches, 1960).
Ideally, the aggregation process would have left me with data on
228 industries over 4
7 3 Manufacturing 2.65 22.15
8 5 Wholesale and Retail Trade 2.09 17.96
9 6 Hotels and Restaurants 1.98 19.77
10 4 Construction 1.79 17.97
11 1 Agriculture 1.54 11.91years, for a total of 912 data
points. After cleaning the AIDA database, I was left with 228
industries, for a total of 866 observations. However, I was
worried about the quality of the
data in some cells with a very low number of firms. The majority
of these industries were
operating in the public sector, and a closer inspection of the
data revealed that the series
were quite unreliable,11 so I decided, somewhat reluctantly, to
drop the cells containing
less than 20 firms, otherwise the measurement error in the micro
data would have been
9 Here industry is defined as a cluster of firms located in the
same region and operating in the same sector of
activity. The sectors are coded according to the ATECO2002
classification, using a 2digit level of disaggregation
into 12 sectos. The number of regions amounts to 19, due to the
lack of information on one of the smaller regions
in the North of Italy (Val dAosta) in the Labour Force Survey.
Dropping it corresponded to a loss of only 318
observations in the AIDA database; moreover, further 18
observations were excluded as consisting of firms
located outside Italy.10 The new growth theory (see Aghion and
Howitt, 1998) has stressed the role played by human capital
externalities in fostering long-term economic performance.11
Measuring productivity and wages in the public sector is a
well-known difficult problem.
-
Table 2
Summary statistics (pooled sample)
Variable Mean Std. Dev. Min. Max.
Proportions
Stock of training 0.243 0.115 0.028 0.680
Flow of training 0.039 0.036 0 0.214
G. Conti / Labour Economics 12 (2005) 557576 563Male employees
0.665 0.181 0.251 1
1524 0.091 0.044 0.005 0.267
2534 0.280 0.068 0.054 0.567
3544 0.290 0.051 0.156 0.517
4554 0.234 0.057 0.097 0.500
5564 0.114 0.058 0 0.468
Degree/postdegree 0.096 0.117 0 0.468
Upper-secondary 0.284 0.135 0.041 0.711
Vocational 0.072 0.046 0 0.285
Compulsory education 0.548 0.220 0.089 0.941
Levels
Log real value added per employee 10.701 0.367 9.152
14.219exacerbated and worsened attenuation bias.12 The final sample
consists of 176 industry
groupings observed over a maximum period of 4 years, for a total
of 633 data points used
in the empirical estimates.
The basic characteristics of the matched sample are described in
the following two
tables. Table 1 illustrates the incidence of training across
sectors,13 ranking each of them
both by its propensity to train,14 and by the stock of
accumulated post-schooling human
capital of their workforce.15 It can be readily seen that
high-training industries are those
providing services of different nature; this seems pretty
obvious, given that this kind of
businesses crucially rely on the role of human resources. The
high ranking of Finance,
Banking and Real Estate also comes at no surprise, given the
intensive use of computers
12 These accounted for 26.9% of the original sample, but only
for 16.9% of the total employment.13 The absence of sector 10,
Public Administration, Defence and Social Insurance, is a
consequence of the
sample selection procedure outlined above.14 Here I refer to
propensity to train as the proportion of workers who have attended
a course in the four weeks
preceding the interview, according to the question asked in the
LFS.15 The second variable is the derived measure of training,
obtained following the procedure outlined above.
Log real wage per employee 9.894 0.272 8.527 13.742
Log capital-labour ratio 10.823 0.712 8.435 14.208
Log R&D expenditure per employee 5.063 1.447 0 9.054
Average hours worked 40.5 3.5 29.4 48.9
Average firm size 138.9 626.1 7.6 8137.9
Growth rates
Labour productivity 0.028 0.177 1.474 1.218Wages 0.033 0.153
1.495 1.249Inflow rate 0.061 0.163 0 1.351
Outflow rate 0.109 0.203 0 1.638
-
Table 3
Summary statistics (pooled sample)
Variable High training Low training
Proportions
Stock of training 0.331 0.155
Flow of training 0.061 0.016
G. Conti / Labour Economics 12 (2005) 557576564Male employees
0.609 0.721
Age 1524 0.083 0.098
Age 2534 0.289 0.271
Age 3544 0.306 0.274
Age 4554 0.236 0.233
Age 5564 0.085 0.124
Degree/postdegree 0.155 0.036
Upper-secondary 0.347 0.222
Vocational 0.084 0.060
Compulsory education 0.413 0.682
Levels
Log real value added per employee 10.771 10.632and IT equipments
in these sectors. By the same token, also the high training
incidence
observed in the Energy, Mining and Quarrying sector is to be
expected, since these
industries use specialized equipment and require stringent
safety measures. Finally, the
low ranking of industries in Transport and Communication and in
Manufacturing sectors
are not surprising, if we take into account the peculiar
industrial structure in Italy, mainly
characterized by small and medium enterprises (SMEs),
specialized in products with a low
technological content (such as clothing, furnishing and
electrical appliances), and
employing low-skilled labour. The skill content of the workforce
also holds as an
explanation for the low ranking of firms in the Trade, Tourism,
Construction and
Agricultural Sector.16
Log real wage per employee 9.949 9.838
Log capital-labour ratio 10.809 10.836
Log R&D expenditure per employee 4.976 5.150
Average hours worked 38.71 42.37
Average firm size 226.74 49.33
Growth rates
Labour productivity 0.043 0.012
Wages 0.042 0.024
Inflow rate 0.068 0.053
Outflow rate 0.099 0.121
16 Note that the ranking of the sectors does not change
according to whether we consider the flow or the stock
measure of training. However, in case of the Trade sector, only
19.81% of the workforce has already attended a
training course during the lifetime, while the proportion is
21.68% for the Tourism sector. The differences
between the two measures, however, can be easily reconciled if
one thinks that usually those employed in hotels
and restaurant are specifically trained for this job at the
beginning of the holiday season, so it is understandable to
find a lower proportion of these workers being trained in March,
which is the month the question in the LFS refers
to.
-
Summary statistics for the variables used in this paper are
provided in Table 2. It is
worth noting that, after having accounted for depreciation,
there is quite enough
variation in the stock of training of the workforce across the
industries, ranging from a
minimum of 2.8% to a maximum of 68%. On the contrary, 2917
industries report 0
training propensity: the lack of sufficient variation in this
measure hinders the possibility
of using it to identify the effect of training. The table also
shows that the sample consists
mainly of middle-aged male employees working on average 40.5
hours a week in a
medium-sized firm,18 among whom more than a half has attained
only a compulsory level
of education.
Finally, in Table 3 I split the sample into dhigh-trainingT and
dlow-trainingTindustries, according to the stock of human capital
embodied in their workforce.19 High
training industries are mainly composed by larger firms, who
employ more middle-aged
female workers with a higher level of education, who work fewer
hours, are more
productive and get paid higher wages as expected. Moreover, they
also experience a higher
it is assumed that the production function for the economy is
represented by a standard
Cobb-Douglas:
G. Conti / Labour Economics 12 (2005) 557576 565Q ALaKb 2where Q
is value added,23 L is effective labour, K is capital, and A is a
Hicks-neutral
technology parameter. Following Dearden et al. (2000), under the
assumption that
17 This amounts to 16.5% of the sample.18 It is worth stressing
that the dimensionality of the firm in the database is
representative of the Italian economy,
mainly formed by SMEs.19 Here the criterion is the median
training stock, which amounts to 0.217. However, very similar
numbers and
the same relative proportions are obtained if the sample is
split according to the median training intensity.20 It is also
worth noting that productivity and wages are closely linked in
high-training industries, whereas
workers in low-training industries experience greater increases
in wages.21 These corresponds to the mean inter-industry inflow and
outflow rates, and have been calculated as the
absolute change in industry-level employment between t and t1
divided by the average employment in the twoperiods. They provide a
measure of workers reallocation across sectors. In order to derive
them, data on
employment in 1995 have been used; given the relevant amount of
missing values for those years, inflow and
outflow dummies have been included in all the estimations in
order to preserve the sample size.22 Since some industries in the
sample do not invest in R&D at all, in order not to furtherly
reduce the sample
size, I have used a small value (1 euro) for their R&D
stock. Hence, a dummy variable is added in al the estimated
models, which equals 1 for those industries not engaging in
R&D activities, and 0 otherwise.23rate of labour productivity
and wage growth,20 and have a higher inflow and a lower
outflow rate.21 The fact that high-training industries are less
capital intensive and engage
less in R&D22 can be easily explained by noting that the
majority of these industries
operates in the service sector.
3. The model
Following a modelling strategy consolidated in the literature
(see Dearden et al., 2000),Griliches and Ringstad (1971) list
numerous justifications for the value-added specification of the
production
function.
-
G. Conti / Labour Economics 12 (2005) 557576566training has a
positive effect on workers productivity, effective labour can be
written
as:
L NU cNT 3
where NU are untrained workers and NT are trained workers (and
we expect gN1).Substitution of Eq. (3) into (2) yields:
Q A NU cNT aKb 4which, after some manipulations, can be
rewritten as:
Q A 1 c 1 TRAIN aNaKb 5
where TRAIN =NT/N represents the proportion of trained workers
in an industry. The
production function can be rewritten in logarithmic form
as:24
lnQ lnA a c 1 TRAIN alnN blnK 6
Finally, under the assumption of constant returns to scale, Eq.
(6) can be respecified in
per-capita terms as:
lnQ
N
lnA 1 b c 1 TRAIN bln K
N
7
where the dependent variable, labour productivity, is measured
as the natural logarithm of
real value added per employee from the balance sheets, TRAIN is
the proportion of trained
workers in an industry, and ln KN
is measured as the natural logarithm of the real value of
tangible fixed assets from the balance sheets (plant and
machinery, land and buildings,
tools and equipment).
Following Dearden et al. (2000), I have firstly estimated the
above production function,
in order to assess the effect of training on the average level
of productivity for the
economy in the years 19961999; in a second step, I have
estimated a wage equation
keeping the same explanatory variables, in order to compare the
gains from training
accruing to firms and to workers. Both equations can be
expressed in terms of the
following general specification:
yit a b1TRAIN b2Xit eit 8
where yit is the outcome of interest (labour productivity or
wages), Xit the vector of
explanatory variables, and qit = fi+uit, i.e. the error term is
composed of a time-invariantindustry-specific effect, and a
time-varying white noise.
Eq. (8), however, may suffer from the major weakness that some
of the regressors could
be correlated with the error term due to the presence of
industry-specific time-invariant24 Here I use the approximation
ln(1+x)=x, assuming (g1)TRAIN is small.
-
factors.25 To deal with this potential source of bias, a
first-difference version of Eq. (8) has
been estimated:
yi;t yi;t1 b1 TRAINit TRAINi;t1 b2 Xit Xi;t1 eit ei;t1 9
This equation relates productivity growth to the change in the
proportion of trained
workers. As argued in Barrett et al. (2001), the assumption
underlying this model points to
the change in the stock of human capital, rather than the flow,
as the main factor fostering
G. Conti / Labour Economics 12 (2005) 557576 567long-term
economic performance.
In addition, in order to avoid omitted variable bias (and hence
overestimate the true
returns to training), in the empirical estimation I have
included several controls, taking into
account observed heterogeneity both in the workers dimension (by
adding proxies for
human capital such as age and education), and in the firms
dimension (by including per-
capita expenditure in R&D as a proxy for the rate of
innovation); I have also controlled for
gender, working hours, and inflow and outflow rates. Time
dummies have been included
to control for time-varying effects, such as the impact of
technological progress or some
other unobserved factor linked to the business cycle. Finally,
several estimation techniques
have been applied: firstly, the model has been estimated using
standard linear techniques
and the within-group estimator. However, for short panels the
consistency of the latter
estimator requires the regressors to be strictly exogenous,
which is not a suitable
assumption in the present case, because transitory shocks on
productivity could be
correlated with training26 (as well as the other inputs),
resulting in an underestimation of
the true returns. Hence, I have drawn on more recent advances in
the Generalized Method
of Moments techniques to deal with this limitations. The GMM
handles not only
unobserved heterogeneity, but also potential endogeneity of
training. In the original First-
Difference GMM estimator developed by Arellano and Bond (1991),
and then extended by
Arellano and Bover (1995), the variables are first-differenced,
in order to eliminate time-
invariant industry-specific effects, and the predetermined and
endogenous variables in first
differences are instrumented with suitable lags of their own
levels, in order to correct for
simultaneity. However, it is well known that the original
Arellano-Bond estimator has poor
finite sample properties when the lagged levels of a series are
weak instruments for the
first differences, especially for variables which are close to a
random walk.27 Blundell and
Bond (1998 and 2000) described how to increase efficiency by
taking into account
additional nonlinear moment conditions, which corresponds to
adding T2 equations in
levels to the system,28 in which pre-determined and endogenous
variables in levels are
instrumented with suitable lags of their own differences. The
so-called extended System-
GMM estimator, as any valid instrumental variable strategy,
handles not only endogeneity,
25 For example, technological change may occur at a faster rate
in some industries, having an impact on both the
regressors and the dependent variable: in this case,
cross-section estimates are inconsistent.26 For example, firms may
choose to train the workforce in periods in which the demand is
low.27 Griliches and Mairesse (1997) have noted that this a severe
problem especially in the context of the
production functions. If the variables evolve in a random walk
like fashion, the past levels have no power as
instruments for the current growth rates, unless one assumes the
existence of lags in adjustments to shocks, in
which case, nonentheless, the power of the internal instruments
is rather low.28 The additional equations come from the moment
restriction: E (qitDqi,t1)=0, where i indexes the industry,and t
=3, 4,. . .,T is the total number of periods in which the industry
is present.
-
but it should also correct for bias arising from transitory
measurement error both in the
dependent variable and the regressors. I have implemented the
two-step version of the
heterogeneity both in an observed and in an unobserved
dimension. However, endogeneity
G. Conti / Labour Economics 12 (2005) 557576568and serial
correlation must be taken into account in the context of production
functions:
shocks in productivity might be well correlated with training,
since firms can adjust their
29 In Monte Carlo simulations it has often been found that the
asymptotic standard errors of the efficient two-stepextended
GMM-SYS estimator using the finite-sample correction for the
two-step
covariance matrix developed by Windmeijer (2000).29
4. The results
Table 4 presents the results for the productivity regressions.
In Model 1, training is
measured in levels, as the proportion of workers who have
accumulated post-schooling
skills during their working life, using the TRAIN variable
derived following the
methodology outlined above. Firstly I have estimated the OLS as
a reference. Training
has a positive and significant effect on labour productivity in
the basic specification which
includes only capital, R&D and hours worked as controls;
however, this impact is clearly
overstated, since the coefficient becomes negative and
significant after controlling for
inflow and outflow rates, workers observed characteristics (sex
and age) and skills
(education). This could be well a signal of endogeneity. In
fact, fixed effect estimates
recover a positive impact of training on labour productivity,
which remains significant
with a high point estimate also after conditioning upon the full
set of controls. Turning to
the other variables, the coefficient on the capital-labor ratio
is highly significant, and its
magnitude confirms the existence of constant returns to scale,
since the share of the wage
bill in value added is about 0.46. Productivity appears to fall
in hours worked, but the
effect is well determined only in the baseline model. Lastly,
R&D expenditure has a strong
and significant impact on productivity only in the simplest
specification. However, the
negative sign of the coefficient seems to be mainly driven by
endogeneity: when
reestimating all the equations using the lagged value, the
coefficient reverts to a positive
sign, and the effect of training is reinforced (the coefficient
is 0.449 with a level of
significance of 1%). The estimation results for the
first-difference version of this model
(the last four columns in Table 4) confirm the robustness of the
main finding: the change in
the stock of accumulated human capital has a positive effect on
labour productivity
growth, with a level of significance stable at 5% across all
different specifications. This
confirms the fact that training has also a long-lasting effect
on industry productivity.
Turning to the other variables, the capital-labour ratio
maintains the high level of
significance achieved in the equations in levels, while R&D
and hours worked are poorly
estimated. On the other side, both measures of turnover exhibit
positive and highly
significant coefficients: this seems to suggest that a higher
speed of reallocation of workers
across sectors leads to a better match, which fosters
productivity.
Until now the estimation techniques adopted have taken into
account industriesGMM estimator are severely downward biased in
small samples. Windmeijer (2000) has developed a variance
correction to increase the accuracy of the inference in two-step
GMM estimations, and overcome this limitation.
-
Table 4
Production function estimates
Model 1 OLS(1) OLS(2) FE(1) FE(2) Model 2 OLS(1) OLS(2) FE(1)
FE(2)
Train(%) 0.234 (0.130) 0.424 (0.178) 0.322 (0.167) 0.349 (0.172)
DTrain(%) 0.314 (0.132) 0.343 (0.144) 0.383 (0.159) 0.376
(0.159)Ln(K/N) 0.333 (0.017) 0.330 (0.018) 0.431 (0.027) 0.431
(0.028) DLn(K/N) 0.314 (0.025) 0.317 (0.026) 0.399 (0.032) 0.392
(0.032)Ln(R&D/N) 0.029 (0.009) 0.015 (0.009) 0.016 (0.009)
0.017 (0.010) DLn(R&D/N) 0.006 (0.008) 0.005 (0.008) 0.003
(0.009) 0.002 (0.009)Ln(Hours/N) 0.919 (0.169) 0.365 (0.318) 0.236
(0.346) 0.315 (0.364) DLn(Hours/N) 0.124 (0.281) 0.102 (0.299)
0.451 (0.338) 0.354 (0.343)Inflow(%) 0.044 (0.073) 0.045 (0.059)
0.018 (0.047) 0.007 (0.048) DInflow(%) 0.079 (0.034) 0.087 (0.035)
0.133 (0.038) 0.127 (0.038)Outflow(%) 0.078 (0.062) 0.066 (0.049)
0.027 (0.039) 0.024 (0.039) DOutflow(%) 0.040 (0.031) 0.045 (0.031)
0.071 (0.032) 0.065 (0.035)Male 0.381 (0.158) 0.152 (0.251) DMale
0.441 (0.202) 0.558 (0.224)Age 2534 0.171 (0.408) 0.393 (0.389)
DAge 2534 0.005 (0.318) 0.028 (0.363)Age 3544 0.404 (0.396) 0.404
(0.381) DAge 3544 0.343 (0.303) 0.337 (0.329)Age 4554 0.470 (0.048)
0.127 (0.396) DAge 4554 0.202 (0.318) 0.374 (0.351)Age 5564 0.193
(0.429) 0.395 (0.458) DAge 5564 0.176 (0.363) 0.270
(0.413)Degree/post 0.391 (0.327) 0.022 (0.433) DDegree/post 0.527
(0.344) 1.013 (0.393)Uppersec 0.238 (0.217) 0.294 (0.238) DUppersec
0.466 (0.195) 0.588 (0.217)Vocational 0.168 (0.435) 0.060 (0.408)
DVocational 0.515 (0.329) 0.495 (0.372)R2 0.428 0.669 0.407 0.413
R2 0.283 0.292 0.411 0.451
NT 633 633 633 633 NT 456 456 456 456
Dependent variable: log(value added per worker) in Model 1,
change in log(value added per worker) in Model 2.
All models include year dummies, R&D dummies, inflow and
outflow dummies. Models OLS(2) also include region and sector
dummies.
G.Conti/LabourEconomics
12(2005)557576
569
-
inputs to changes in demand, as outlined above. Henceforth, the
GMM-estimator has been
G. Conti / Labour Economics 12 (2005) 557576570implemented to
overcome the limitations of the techniques previously adopted. The
results
are presented in Table 5. Each regression includes all the
variables used in the previous
estimations, but only the results on the variables more closely
related to the production
process are reported.30 In the first column, the
First-Difference GMM is firstly estimated
as a benchmark. The coefficient on training is positive, with a
magnitude that resembles
the one obtained in the previous specification, but fails to
achieve significance at a
conventional level. Furthermore, the coefficient on capital is
unreasonably low: this is in
line with what shown by Blundell and Bond (1998): the lagged
levels of a series provide
weak instruments for the first differences, and produce
implausibly low estimates
because measurement error in the explanatory variables bias the
coefficients towards
zero. Consequently, I have implemented the full two-step GMM
system estimator, using
the finite-sample correction for the two-step covariance matrix
proposed by Windmeijer
(2000). Training recovers significance at a conventional level,
and the coefficient on
capital is doubled. In the next four columns, I have
respectively added a lag for the
dependent variable, the training variable, and for capital,
hours worked, R&D and the
turnover rates. The coefficient on training varies somehow
across the different
specifications, but it is always strongly significant and with a
higher point estimate
than the one estimated when treating it as exogenous. The other
variables included also
exhibit highly significant values, and they have the expected
sign; in particular, it should
be noted that, while the current levels of R&D fail to
achieve significance, its lagged
levels are positive and significant at 10%. All the
specifications easily pass all the
diagnostics tests.
In order to check the robustness of the results obtained in the
dynamic specification, I
included employment (and eventually its lag) to test for
non-constant returns: only in one
case the coefficient achieved a level of significance of 10%,
but in that case the diagnostics
exhibited clear evidence of misspecification; furthermore, when
included in the full
dynamic model (last column), employment and its lag were jointly
insignificant with a p-
value of 0.24.
Now Ill turn to discuss the wage equation results. In order to
ease comparability, I have
used exactly the same models as in the productivity regression.
There is no significant
effect of training on wages in the static model of Table 6; in
particular, when controlling
for skills using linear estimation techniques, the coefficient
becomes negative and
significant, suggesting some form of endogenous effect similar
to the one known as
dAshenfelters dipT. All the other variables are conventionally
signed. Turning to theresults for the model in first differences, a
positive and significant impact of the increase in
the stock of trained workforce on wage growth is uncovered,
which seems suggestive of
the existence of a long-run effect of accumulated skills on the
earnings of the individuals.
Finally, Table 7 presents the estimation results for the GMM
estimation of the effect of
training on wages. When taking endogeneity into account, the
estimated coefficients fail to
achieve significance at a conventional level in all the
specifications adopted. Turning to
the effect of the other variables, industries with a high
capital-labour ratio and engaging30 Full results are available from
the author upon request.
-
Table 5
Production function estimates
Model 1 GMM-FD GMM-SYS(1) GMM-SYS(2) GMM-SYS(3) GMM-SYS(4)
GMM-SYS(5)
Ln(vadd/N)t1 0.516 (0.088) 0.505 (0.084) 0.527 (0.082) 0.502
(0.079)Train(%) 0.313 (0.210) 0.334 (0.203) 0.317 (0.168) 0.383
(0.147) 0.451 (0.177) 0.408 (0.183)
Train(%)t1 0.412 (0.279) 0.431 (0.293) 0.349 (0.282)Ln(K/N)
0.153 (0.068) 0.317 (0.056) 0.246 (0.050) 0.254 (0.044) 0.348
(0.057) 0.336 (0.055)
Ln(K/N)t1 0.134 (0.057) 0.131 (0.055)Ln(R&D/N) 0.006 (0.029)
0.006 (0.023) 0.013 (0.018) 0.013 (0.016) 0.026 (0.016) 0.021
(0.017)Ln(R&D/N)t1 0.037 (0.015) 0.035 (0.014)Ln(Hours/N) 0.359
(0.583) 0.942 (0.558) 0.432 (0.398) 0.526 (0.420) 0.199 (0.424)
0.293 (0.398)Ln(Hours/N)t1 0.710 (0.455) 0.732 (0.464)Inflow(%)
0.062 (0.055) 0.006 (0.087) 0.166 (0.134) 0.166 (0.145) 0.139
(0.137) 0.132 (0.135)Inflow(%)t1 0.018 (0.075) 0.039
(0.085)Outflow(%) 0.095 (0.061) 0.060 (0.065) 0.359 (0.133) 0.344
(0.121) 0.303 (0.117) 0.299 (0.113)Outflow(%)t1 0.059 (0.047) 0.054
(0.040)Hansen test 0.315 0.306 0.292 0.250 0.236 0.166
AR(1) test 0.001 0.001 0.003 0.003 0.004 0.005
AR(2) test 0.441 0.955
NT 456 456 456 456 456 456
Dependent variable: log(value added per worker).
All models include year dummies, R&D dummies, inflow and
outflow dummies.
All the variables are treated as endogenous (except the
dummies).
Model GMM-SYS(5) includes lags for all the variables.
p-values are reported for AR(1), AR(2) and Hansen tests. A full
stop in the AR(2) test box indicates that no output has been
reported: the residuals and the L(2) residuals
have no obs in common, so the AR(2) is trivially zero.
G.Conti/LabourEconomics
12(2005)557576
571
-
Table 6
Wage equation estimates
Model 1 OLS(1) OLS(2) FE(1) FE(2) Model 2 OLS(1) OLS(2) FE(1)
FE(2)
Train(%) 0.059 (0.112) 0.459 (0.167) 0.203 (0.172) 0.215 (0.176)
DTrain(%) 0.253 (0.123) 0.305 (0.134) 0.319 (0.157) 0.287
(0.156)Ln(K/N) 0.171 (0.014) 0.177 (0.017) 0.342 (0.028) 0.339
(0.029) DLn(K/N) 0.184 (0.023) 0.187 (0.024) 0.234 (0.032) 0.225
(0.031)Ln(R&D/N) 0.017 (0.008) 0.001 (0.008) 0.003 (0.010)
0.003 (0.010) DLn(R&D/N) 0.005 (0.008) 0.004 (0.008) 0.003
(0.009) 0.004 (0.009)Ln(Hours/N) 0.959 (0.145) 0.537 (0.299) 0.245
(0.355) 0.273 (0.373) DLn(Hours/N) 0.164 (0.262) 0.136 (0.280)
0.435 (0.334) 0.343 (0.336)Inflow(%) 0.021 (0.062) 0.080 (0.056)
0.085 (0.048) 0.073 (0.135) DInflow(%) 0.030 (0.032) 0.025 (0.033)
0.004 (0.037) 0.001 (0.037)Outflow(%) 0.070 (0.053) 0.033 (0.047)
0.017 (0.040) 0.010 (0.041) DOutflow(%) 0.044 (0.029) 0.049 (0.029)
0.066 (0.034) 0.071 (0.034)Male 0.499 (0.148) 0.030 (0.258) DMale
0.341 (0.189) 0.480 (0.219)Age 2534 0.061 (0.383) 0.429 (0.399)
DAge 2534 0.018 (0.297) 0.169 (0.356)Age 3544 0.676 (0.372) 0.549
(0.391) DAge 3544 0.386 (0.283) 0.363 (0.323)Age 4554 0.570 (0.383)
0.615 (0.406) DAge 4554 0.497 (0.297) 0.567 (0.344)Age 5564 0.261
(0.403) 0.351 (0.469) DAge 5564 0.126 (0.339) 0.006
(0.405)Degree/post 0.329 (0.307) 0.001 (0.444) DDegree/post 0.522
(0.322) 0.845 (0.386)Uppersec 0.194 (0.204) 0.207 (0.245) DUppersec
0.514 (0.182) 0.697 (0.213)Vocational 0.352 (0.408) 0.179 (0.418)
DVocational 0.709 (0.308) 0.787 (0.365)R2 0.236 0.468 0.331 0.337
R2 0.168 0.175 0.246 0.306
NT 633 633 633 633 NT 456 456 456 456
Dependent variable: log(wage per worker) in Model 1, change in
log(wage per worker) in Model 2.
All models include year dummies, R&D dummies, inflow and
outflow dummies. Models OLS(2) also include region and sector
dummies.
G.Conti/LabourEconomics
12(2005)557576
572
-
Table 7
Wage equation estimates
Model 1 GMM-FD GMM-SYS(1) GMM-SYS(2) GMM-SYS(3) GMM-SYS(4)
GMM-SYS(5)
Ln(W/N)t1 0.331 (0.118) 0.341 (0.108) 0.299 (0.117) 0.314
(0.111)Train(%) 0.199 (0.197) 0.021 (0.167) 0.091 (0.128) 0.158
(0.124) 0.181 (0.154) 0.123 (0.160)
Train(%)t1 0.434 (0.206) 0.375 (0.228) 0.412 (0.260)Ln(K/N)
0.092 (0.061) 0.154 (0.043) 0.153 (0.045) 0.166 (0.041) 0.194
(0.052) 0.185 (0.060)
Ln(K/N)t1 0.053 (0.029) 0.046 (0.030)Ln(R&D/N) 0.031 (0.027)
0.002 (0.019) 0.005 (0.017) 0.009 (0.015) 0.019 (0.021) 0.017
(0.022)Ln(R&D/N)t1 0.031 (0.011) 0.023 (0.010)Ln(Hours/N) 0.144
(0.562) 1.022 (0.409) 0.738 (0.297) 0.825 (0.286) 0.588 (0.315)
0.607 (0.454)Ln(Hours/N)t1 0.664 (0.369) 0.685 (0.408)Inflow(%)
0.092 (0.053) 0.053 (0.101) 0.022 (0.134) 0.007 (0.116) 0.032
(0.134) 0.016 (0.123)Inflow(%)t1 0.003 (0.053) 0.008
(0.056)Outflow(%) 0.062 (0.063) 0.027 (0.067) 0.248 (0.114) 0.244
(0.113) 0.248 (0.100) 0.264 (0.112)Outflow(%)t1 0.012 (0.029) 0.016
(0.031)Hansen test 0.373 0.259 0.463 0.638 0.484 0.382
AR(1) test 0.002 0.001 0.004 0.004 0.006 0.007
AR(2) test 0.814 0.609 . . . .
NT 456 456 456 456 456 456
Dependent variable: log(wage per worker).
All models include year dummies, R&D dummies, inflow and
outflow dummies.
All the variables are treated as endogenous (except the
dummies).
Model GMM-SYS(5) includes lags for all the variables.
p-values are reported for AR(1), AR(2) and Hansen tests. A full
stop in the AR(2) test box indicates that no output has been
reported: the residuals and the L(2) residuals
have no obs in common, so the AR(2) is trivially zero.
G.Conti/LabourEconomics
12(2005)557576
573
-
more in R&D pay higher wages, while longer working hours are
associated with a lower
pay. Finally, a higher degree of workers reallocation across
industries is associated with
higher wages, which seems to be suggestive of the fact that
turnover leads to a better
matching.
To summarize the results, training seems to have a positive and
strongly significant
effect on productivity. This effect disappears only when
controlling for observed
heterogeneity with the standard linear techniques, but the sign
and the size of the
greater than the effect on wages (for example, in the
full-dynamic specification, the
G. Conti / Labour Economics 12 (2005) 557576574coefficient is
0.408 in the productivity regression and 0.123 in the wage
equation): this is
consistent with a human capital model in which some of the costs
of training are borne
by the employees. Using the results obtained in the full-dynamic
model, this implies that
raising the stock of trained workers in an industry by one
percentage point leads to a
0.4% increase in productivity and to a 0.1% increase in wages.
This effect is much
smaller if compared to the results obtained in other papers in
this literature.32
Nonetheless, it is quite substantial on its own. There could be
several reasons behind
this: on the one hand, estimation at aggregate level captures
the within-industry spillovers
arising from training externalities, which are missed in case of
a firm-level analysis.33 On
the other, it fails to account for non-random selection of
workers in the training pool. On
a more general ground, it could reflect the existence of other
unobserved industry-specific
factors that have not been controlled for (such as the existence
of other human resource
management practices, as argued in Ichniowski et al., 1997).
Above all, also allowing for
these caveats, the key qualitative result still holds: firms do
seem to actually reap more of
the returns.
5. Concluding remarks
This paper has examined for the first time the productivity and
wage effects of training
in Italy. It is based on an original dataset, which has been
created aggregating individual-
level data on training from the Labour Force Survey with
firm-level data on productivity
31 This finding also emerges in a recent paper by Arulampalam et
al. (2004). They show that training has no
significant effect on wages at all the quantiles of the
conditional wage distribution.32 For example, Dearden et al. (2000)
found that increasing the proportion of workers being trained in
an
industry by 5% leads to a 4% increase in productivity and to a
1.5% increase in wages: this amounts to,
respectively, a 2% increase in productivity and to a 0.6%
increase in wages in the Italian case.33 Under this respect, the
availability of a linked employer-employee database with training
information, stillcoefficient clearly reflects endogenous effects.
Most importantly, it persists when
unobserved heterogeneity is taken into account, and it is robust
to the first-difference
specification and to the GMM estimation. On the other side, the
effect of training on
wages is much less robust.31 It achieves significance at a
conventional level only in the
first-difference specification, but it does not pass the GMM
estimation. Moreover,
whatever specification is considered, the estimated impact on
productivity is alwayslacking for Italy, will provide useful
results in terms of comparability between private and social
returns to post-
schooling human capital.
-
I am deeply indebted to my supervisor, Amanda Gosling, for her
generous and
Acemoglu, D., Pischke, J.-S., 1998. Why do firms train? Theory
and evidence. Quarterly Journal of Economics113 (1), 79119 (also
available as NBER Working Paper, n.5605).
Acemoglu, D., Pischke, J.-S., 1999a. Beyond Becker: Training in
imperfect labour markets. Economic Journal
109 (453), F112F142.
Acemoglu, D., Pischke, J.-S., 1999b. The structure of wages and
investment in general training. Journal of
Political Economy 107 (3), 539572 (also available as NBER
Working Paper, n.6357).
Aghion, P., Howitt, P., 1998. Endogenous Growth Theory. MIT
Press, Cambridge, Mass.
Arellano, M., Bond, S., 1991. Some tests of specification for
panel data: Monte Carlo evidence and an application
to employment equations. Review of Economic Studies 58 (2),
277297.
Arellano, M., Bover, O., 1995. Another look at the
instrumental-variable estimation of error-components model.
Journal of Econometrics 68 (1), 2952.
Arulampalam, W., Booth, A.L. Bryan, M.L., 2004. Training in
Europe. ISER working paper 0401.
Ballot, G., Fakhfakh, F., Taymaz, E., 2001. Firms human capital,
R&D and performance: a study on French and
Swedish Firms. Labour Economics 8, 443462.
Barrett, A., OConnell, P., 2001. Does training generally work?
The returns to in-company training. Industrial and
Labour Relations Review 54 (3), 647662 (also available as IZA
Discussion Paper, n.51).invaluable support and guidance throughout
the realisation of this research. I am also
grateful to two anonymous referees and to seminar participants
at the 16th annual
conference of the European Association of Labour Economists
2004, for helpful
comments and suggestions which greatly improved the paper. Many
people have helped
me in accessing the data. I owe special thanks to Sergio
Destefanis, Tullio Jappelli and
Mario Padula for supplying the AIDA data, and to Marco Musella
and Francesco Pastore
for supplying the Labour Force Survey data. Financial support
from the Economic and
Social Research Council, award no. PTA030200300812, is
gratefully acknowledged.
The usual disclaimer applies.
Referencesand wages from AIDA into an industry-level panel,
covering the years from 1996 to 1999.
Given the availability of longitudinal information I have been
able to control for
unobserved heterogeneity and potential endogeneity of training.
The richness of the
database has also allowed me to construct a stock measure of
training, and to control for
several workers and firms characteristics. I have then allowed
for flexibility in the
modelling specification, and estimated all the models in the
dual form of a production
function and a wage equation, in order to assess how the
benefits from training are shared
between the firms and the workers.
The main finding is that training has a positive and significant
effect on productivity.
This finding is robust to several estimation strategies,
including System-GMM. However,
the effect uncovered for wages is much less robust, and smaller
in size. This proves clear
evidence of the fact that firms do actually reap more of the
returns.
Acknowledgements
G. Conti / Labour Economics 12 (2005) 557576 575Bartel, A.P.,
1994. Productivity gains from the implementation of employee
training programs. Industrial
Relations 33 (4), 411425 (also available as NBER Working Paper,
n.3893).
-
Bishop, J.H., 1994. The impact of previous training on
productivityand wages. In: Lynch, L.M. (Ed.), Training
and the Private Sector, International Comparisons, NBER Series
in Comparative Labour Markets. University
of Chicago Press, Chicago, pp. 161199.
Black, S.E., Lynch, L.M., 1995. Beyond the incidence of
training: evidence from a national employers survey,
NBER Working Paper, n.5321.
Black, S.E., Lynch, L.M., 1996. Human capital investments and
productivity. American Economic Review
(Papers and Proceedings) 86 (2), 263267.
G. Conti / Labour Economics 12 (2005) 557576576Black, S.E.,
Lynch, L.M., 2001. How to compete: the impact of workplace
practices and information technology
on productivity. The Review of Economics and Statistics 83,
434445.
Blundell, R., Bond, S., 1998. Initial conditions and moment
restrictions in dynamic panel data models. Journal of
Econometrics 87 (1), 115145.
Blundell, R., Bond, S., 2000. GMM estimation with persistent
panel data: an application to production function.
Econometric Reviews 19 (3), 321340.
Boon, M., van der Eijken, B., 1998. Employee training and
productivity in Dutch manufacturing firms. Statistics
Netherlands 13.
Brunello, G., 2002. Is training more frequent when wage
compression is higher? Evidence form 11 European
countries, CESifo Working Paper, n.637 (4).
Dearden, L., Reed, H., Van Reenen, J., 2000. Who gains when
workers train? Training and corporate productivity
in a panel of british industries, IFS Working Paper,
n.00/01.
Griliches, Z., Mairesse, J., 1997. Production functions: the
search for identification. In: Strom, S. (Ed.), Essays in
Honour of Rayner Frisch, Econometric Society Monograph Series.
Cambridge University Press.
Griliches, Z., Ringstad, V., 1971. Economies of scale and the
form of the production function. North Holland,
Amsterdam.
Groot, W., 1998. Empirical estimates of the rate of depreciation
of education. Applied Economics Letters 5 (8),
535538.
Grunfeld, D., Griliches, Z., 1960. Is aggregation necessarily
bad? The Review of Economics and Statistics XLII
(1), 113.
Ichniowski, C., Shaw, K., Prennushi, P., 1997. The effects of
human resource management practices on
productivity: a study of steel finishinglines. American Economic
Review 87 (3), 291313.
Lynch, L.M., 1994. Introduction. In: Lynch, L.M. (Ed.), Training
and the Private Sector, International
Comparisons, NBER Series in Comparative Labour Markets.
University of Chicago Press, Chicago, pp. 1
24.
OECD, 2002. Economic Outlook, Paris.
Stevens, M., 1994a. Labour contracts and efficiency in
on-the-job training. Economic Journal 10 (423), 408419.
Stevens, M., 1994b. A theoretical model of on-the-job training
with imperfect competition. Oxford Economic
Papers 46 (4), 537562.
Stevens, M., 1994c. An investment model for the supply of
training by employers. Economic Journal 104 (424),
556570.
Stevens, M., 1996. Transferable training and poaching
externality. In: Booth, A.L., Snower, D.J. (Eds.), Acquiring
Skills, Market Failures, Their Symptoms and Policy Responses.
Cambridge University Press, Cambridge.
Windmeijer, F., 2000. A finite sample correction for the
variance of linear two-step GMM estimators, IFS
Working Paper, n.00/01.
Zwick, T., 2002. Continuous training and firm productivity in
Germany, ZEW Discussion Paper, n.0250.
Training, productivity and wages in ItalyIntroductionThe dataThe
modelThe resultsConcluding remarksAcknowledgementsReferences