Page 1
Unclassified STD/CSSP/WPTGS(2015)24 Organisation de Coopération et de Développement Économiques Organisation for Economic Co-operation and Development 23-Mar-2015
___________________________________________________________________________________________
_____________ English - Or. English STATISTICS DIRECTORATE
COMMITTEE ON STATISTICS AND STATISTICAL POLICY
Working Party on International Trade in Goods and Trade in Services Statistics
THE STABILITY OF TIVA ESTIMATES: TESTING DIFFERENT METHODS OF RECONCILING
CONTRADICTORY OFFICIAL DATA
24-26 March 2015
Conference Centre, OECD Headquarters, Paris
The presence of trade asymmetries and national data discrepancies means that when building an Inter Country
Input Output table (ICIO), decisions must be taken on the correct statistical approach to combining and
harmonizing the conflicting data. In the current creation of the ICIO, two important steps (among many) are
taken that are addressed in this paper. First, the national SUT or IO data are harmonized with the most recent
available information from the National Accounts. Secondly, when combining national data into an international
IO, the inevitable discrepancies are currently distributed proportionally across countries and industries (using
RAS).
This paper analyses the robustness of TiVA indicators by exploring the effect of both of these steps on two focal
indicators: Vertical Specialization and bilateral exports. We show that the main TiVA results are robust to
adjustments to national accounts, and that also when the RAS procedure is adapted in such a way that the ICIO
that is created fully reflects the national statistics of one single country, while permitting other countries’ trade
flows to change to respond to the needs of global balancing, the focal indicators generally do not change
significantly. Improvements in the symmetry of bilateral trade statistics, and further investigations into the best
way of balancing them when discrepancies continue, would however still improve the TiVA estimates.
Agenda item 5.4
Contact persons: Guannan Miao ([email protected] ) or Fabienne Fortanier
([email protected] )
JT03372904
Complete document available on OLIS in its original format
This document and any map included herein are without prejudice to the status of or sovereignty over any territory, to the delimitation of
international frontiers and boundaries and to the name of any territory, city or area.
ST
D/C
SS
P/W
PT
GS
(20
15
)24
Un
classified
En
glish
- Or. E
ng
lish
Page 2
STD/CSSP/WPTGS(2015)24
2
THE STABILITY OF TIVA ESTIMATES: TESTING DIFFERENT METHODS OF
RECONCILING CONTRADICTORY OFFICIAL DATA
Guannan Miao and Fabienne Fortanier1
1. Introduction: the comparability of TiVA estimates with national data
1. The release of Trade in Value Added (TiVA) indicators by OECD and WTO in 2013 has
been successful in many ways. By decomposing exports into its various value added components
(domestic and foreign, direct and indirect) which are directly comparable to GDP, the indicators have
helped our collective understanding of Global Value Chains. They have highlighted the importance of
imports for export success, of the large (but indirect) role of services in exports, and have changed the
way bilateral trade balances are perceived. These conclusions and their implications continue to be
widely discussed among even the highest levels of policy makers. However, despite this success (or
perhaps even because of it), an often-raised and important concern is that the data presented in TiVA
are not well comparable with official national data. In particular, gross trade data in the TiVA
database are sometimes quite different from the trade statistics reported by the country, particularly
when broken down by partner country. This lack of comparability potentially reduces the usefulness
of TiVA indicators for national policy making, but also begs the question if – when the national data
were fully considered – the TiVA indicators might be very different from the ones currently published.
2. This paper aims to address this concern head-on. We start by introducing the two root causes
of the differences between the data in TiVA and those published by NSOs. These include first of all,
the presence of trade asymmetries, or differences between the national data as reported by one
country compared to the mirror flows reported by the trading partner. Secondly, there are substantial
discrepancies in the national data sources that are used to create the Inter-Country Input-Output
(ICIO) table that underpins TiVA. National Supply and Use Tables (SUTs) or Input-Output (IO)
tables, Systems of National Accounts (SNA), and bilateral trade statistics (all official statistics
published by NSOs) may give quite different pictures regarding e.g. value added by industry, trade
flows, and on how imports are attributed to end use categories.
3. The presence of trade asymmetries and national data discrepancies means that when
building an ICIO, decisions must be taken on the correct statistical approach to combining and
harmonizing the conflicting data. In the current creation of the ICIO, two important steps (among
many) are taken that are addressed in this paper. First, the national SUT or IO data are harmonized
with the most recent available information from National Accounts statistics on main aggregates of
GDP. The advantage is that the subsequent results can be readily compared with national GDP.
1 The authors thank Nadim Ahmad for his encouragement and insightful comments.
Page 3
STD/CSSP/WPTGS(2015)24
3
However, since these main National Accounts aggregates are updated and revised more frequently
than the underlying SUTs and/or IOs, discrepancies between both may occur. Secondly, when
combining national data into an international IO, the inevitable discrepancies are currently distributed
proportionally across countries and industries using a method called RAS, which is traditionally used
to compiling Social Accounting Matrices (SAMs) and IO tables. However, this by definition means
that the final balanced2 ICIO is a compromise of 57 national perspectives, and countries may be
interested in understanding how this result may differ from their own national statistics.
4. In this paper we test the impact of these two decisions. In our first test, we study the effect
of the adjustment of national SUT/IO data to national accounts aggregates. In the second test, we
compare the current balanced TiVA results with the results that are obtained when constraining the
RAS procedure in such a way that the ICIO that is created fully reflects the national statistics of one
single country, while permitting other countries’ trade flows to change to respond to the needs of
global balancing (i.e. essentially creating 57 – the number of countries in TiVA – different ICIO
tables). We focus specifically on the variation in the outcomes from these 57 national perspectives to
obtain a better understanding of the robustness of TiVA indicators.
5. The remainder of this paper is organized as follows. First, section 2 explains in more detail
why there are differences between the data presented in the OECD-WTO TiVA database compared to
national statistics. We also illustrate this with two country studies, for United Kingdom and Greece.
Section 3 presents the methodology that we use to for the two tests, and also provides a brief
introduction to the RAS procedure, and explains why we use the degree of Vertical Specialization
(VS) and bilateral exports as our main target variables for comparing results. The results are
subsequently presented in section 4, first focusing on the effect of benchmarking SUTs to National
accounts, and subsequently on different national constraints can impact the balancing results. Section
5 concludes.
2. Prevalence of discrepancies in national data and international trade asymmetries
6. One of the key questions that users of TiVA indicators ask is why the data reported in TiVA
can be quite different from the data reported by national statistical offices. This problem holds
particularly true for the bilateral trade data reported in TiVA. As illustrated in table 1, with case
studies for United Kingdom and Greece, the values that are reported by NSOs on their bilateral trade
position differs from the final values that are included in TiVA and that used for the calculations of
the indicators on for example the import share of exports or domestic value added in final demand.
This difference makes it difficult to compare TiVA with other national data and therefore hampers
(policy) analysis. But it also raises the question what would be the TiVA indicators if the reported
data were taken into account, which comes down to a request for more insights into the robustness
and stability of TiVA indicators.
2 Balanced meaning that all matrix cells add up precisely to the row and column totals.
Page 4
STD/CSSP/WPTGS(2015)24
4
Table 1. Examples of differences between gross merchandise trade values included in TiVA and as
reported by countries (by partner country, 2009, in million USD)
Reporter Partner
Exports Imports Trade Balance
Merchandise
trade TiVA
Merchandise
trade TiVA
Merchandise
trade TiVA
UK Germany 38,181 40,721 62,055 55,658 -23,874 -14,937
Netherlands 27,319 15,812 33,546 28,383 -6,226 -12,571
China 6,583 7,150 42,759 38,719 -36,176 -31,569
Canada 5,595 7,975 7,739 5,991 -2,144 1,984
Norway 4,111 3,432 23,047 26,122 -18,936 -22,690
USA 51,893 40,281 45,463 36,955 6,430 3,326
Greece Italy 2,163 2,110 7,569 9,455 -5,406 -7,346
Russia 325 317 3,421 4,666 -3,096 -4,349
Germany 2,183 2,028 8,149 9,020 -5,966 -6,992
Belgium 283 173 2,634 1,938 -2,351 -1,765
France 727 919 3,617 4,367 -2,890 -3,448
USA 935 1,042 1,884 2,547 -949 -1,505
Sources: OECD-WTO TiVA and UN Comtrade database. Note that TiVA merchandise trade covers ISIC sectors A to E.
7. There are two main factors that drive these differences. First, there are substantive
asymmetries in international merchandise trade and services trade data, meaning that the data
published by a reporting country on its bilateral trade flows may be vastly different from the data
published by its partner countries on the same trade flows. There are many explanations for these
asymmetries; the most important ones include differences across countries in the reporting definitions
(e.g. country of consignment versus country of origin), definitions of geographical areas,
(mis)classification of products; differences in valuation (FOB for exports and CIF for imports);
different ways of reporting confidential data; and for some countries, the sheer size of re-exports.
8. Trade asymmetries are problematic for TiVA because the ICIO can only include one
number to describe the exports of one country and industry and imports of its partner country and
industry. The larger the asymmetries, the more difficult the choice of the ‘right’ number becomes and
the more the final figure will deviate from what is reported by at least one (if not both) trading
partners involved3. Therefore, even though countries often describe trade asymmetries as “when you
fix one problem, another ten come up” or “too tough to chew”, dealing with trade asymmetries has
been and will continue to be a part of the core work for the national statistical offices. International
organisations and relevant task forces can facilitate and drive the reconciliation processes, and thereby
help to minimize asymmetries, but a bottom-up approach is clearly preferable (see also
STD/CSSP/WPTGS(2014)20).
9. Second, even within an individual country, discrepancies exist between the various national
data sources that are used in TiVA. Here we use United Kingdom and Greece’s 2005 data to
demonstrate the magnitude of these data conflicts, which can be partially explained by the difference
in definition, but also be caused by data cleaning processes and imputation methods, the difference
between survey and administrative data, or timing of revisions.
3 At the moment, TiVA uses import shares by partner country to describe bilateral trade relationships.
Page 5
STD/CSSP/WPTGS(2015)24
5
2.1 SUTs and National Accounts
10. Our first example compares Supply and Use tables with data from National Accounts, for
the year 2005. Table 2 shows, for United Kingdom, in particular that while the differences are small
for final consumption expenditure and gross fixed capital formation, this is not the case for value
added by industry, where discrepancies are varying from 4% to almost 20% across sectors. At the
aggregate level, SNA provides a total value added 0.8% higher than the equivalent data given in the
SUT. Finally, when we check the breakdown of trade by goods and services, we observe bigger gaps
between two data sources than at the total trade level. As for Greece, final consumption expenditure
reported in SUT is 8% higher than reported in SNA while gross capital formation is 7% lower than
SNA. Value added data by sector sees differences of up to 20% (positive or negative). The biggest
discrepancy can be seen in the area of trade: for both imports and exports, SNA flows are
substantially higher than reported in the Greek SUT. This revision could possibly reflect that non-
resident expenditures were initially classified as part of household final consumption expenditure in
the SUT, but were subsequently reclassified as exports (and similarly for resident expenditures
abroad).
Table 2. Data comparison for selected variables of SNA and SUT, 2005, in millions of national currency
United Kingdom Greece
SNA
(1)
SUT
(2)
Ratio
(1)/(2)
SNA
(1)
SUT
(2)
Ratio
(1)/(2)
Final consumption expenditure (FCE) 1,084,211
169,662
Household FCE* 815,938 814,964 1.001 134,725 149,143 0.903
General government FCE 268,273 268,088 1.001 34,937 33,225 1.052
Gross capital formation 213,938
41,322
Gross fixed capital formation** 209,689 209,381 1.001 40,020 38,873 1.030
Changes in inventories 4,249 4,472 0.950 1,302 -402 -3.240
Exports of goods and services 340,424 330,794 1.029 44,807 32,861 1.364
Exports of goods 217,476 213,536 1.018 20,490 16,337 1.254
Exports of services 122,948 117,258 1.049 24,317 16,524 1.472
Imports of goods and services 375,862 373,641 1.006 62,741 58,881 1.066
Imports of goods 281,850 293,862 0.959 51,875 50,556 1.026
Imports of services 94,012 79,779 1.178 10,866 8,325 1.305
Gross Domestic Product (GDP) 1,262,710
193,050
Total Value Added*** 1,125,300 1,116,648 1.008 172,595 174,624 0.988
Agriculture, hunting, forestry, fishing (A-B) 7,035 7,530 0.934 8,428 8,403 1.003
Mining, Manufacturing, Utilities (C-E) 184,756 192,249 0.961 21,957 22,494 0.976
Manufacturing (D) 133,390 148,111 0.901 16,428 16,987 0.967
Construction (F) 82,112 69,868 1.175 12,050 10,949 1.101
Wholesale, retail, restaurants and hotels (G-H) 193,664 162,712 1.190 38,094 41,720 0.913
Transport, storage and communication (I) 95,232 80,889 1.177 14,781 17,729 0.834
Other Activities (J-P) 562,501 603,400 0.932 77,284 73,327 1.054
Sources: SNA data are from UN; SUT are from Eurostat. Final consumption expenditure and Gross capital formation are in
purchasers’ price. Exports of goods and services include re-exports.
* Household FCE includes non-profit institutions service households and
** Gross fixed capital formation includes acquisitions less disposals of valuables.
*** Letters in parenthesis refer to ISIC Rev.3 industry classifications
Page 6
STD/CSSP/WPTGS(2015)24
6
2.2 SUTs and merchandise trade statistics
11. A second set of national data sources that can be compared include merchandise trade
statistics and SUTs, for the exports (including re-exports) and imports of goods. Table 3 first
compares the export figures. At the aggregate level for United Kingdom, both sources provide a rather
similar picture for the reference year (2005). Merchandise exports totalled 208 billion pounds, or only
3 percent smaller than what is reported in the SUT. However, the level of discrepancy rises when
breaking down the trade statistics by product category (using the CPA classification), with differences
up to ± 40% (See Table 3). For Greece, important differences can be observed even at aggregated
level, with merchandise trade data that are 15% lower than the data reported in the SUTs. Also at the
product level, differences are substantial.
Table 3. SNA and Merchandise trade data, exports in millions of national currency (2005)
United Kingdom Greece
CPA
Merchandise
Trade (MT) SUT
Ratio
(MT/SUT)
Merchandise
Trade (MT) SUT
Ratio
(MT/SUT)
01 1,422 1,270 1.120 1,219 1,380 0.883
02 52 52 1.000 6 8 0.750
05 369 400 0.923 293 346 0.847
10 52 44 1.182 5 8 0.625
11 12,015 11,830 1.016 - - -
13 17 -
32 32 1.000
14 7,156 5,088 1.406 116 117 0.991
15 9,325 10,234 0.911 1,841 2,040 0.902
16 647 684 0.946 142 142 1.000
17 2,867 3,183 0.901 624 651 0.959
18 2,306 3,843 0.600 1,276 1,744 0.732
19 897 1,441 0.622 59 65 0.908
20 338 354 0.955 66 67 0.985
21 2,206 2,425 0.910 112 129 0.868
22 4,220 3,379 1.249 128 101 1.267
23 8,718 10,463 0.833 1,345 2,363 0.569
24 35,023 33,574 1.043 1,972 2,036 0.969
25 5,578 5,029 1.109 460 465 0.989
26 2,071 1,983 1.044 357 374 0.955
27 9,749 10,936 0.891 1,440 1,661 0.867
28 4,277 4,202 1.018 382 378 1.011
29 19,546 18,630 1.049 569 575 0.990
30 10,091 9,808 1.029 60 19 3.158
31 6,982 6,739 1.036 391 454 0.861
32 11,814 19,010 0.621 252 306 0.824
33 8,223 8,537 0.963 129 138 0.935
34 23,539 21,736 1.083 199 211 0.943
35 13,380 13,773 0.971 336 361 0.931
36 4,536 4,717 0.962 140 154 0.909
37 3 -
1 - -
40 101 166 0.608 8 15 0.533
Total 207,521 213,530 0.972 13,959 16,337 0.854
Data sources: Merchandise trade data as reported to UN Comtrade, SUT as reported to Eurostat.
Page 7
STD/CSSP/WPTGS(2015)24
7
12. On the import side, similar discrepancies can be observed. What is even more important, is
that since imports are also classified by End-Use categories in SUTs, the merchandise data must be
well-aligned with these categories as well. In table 4a (for the UK) and 4b (for Greece), a comparison
is made between imports by product and main End-Use category as reported in the SUT with similar
data derived from merchandise trade statistics.
Table 4a. Differences between merchandise trade statistics and SUTs for imports by product and end-use
category, in millions national currency, 2005, United Kingdom
CPA
code
Merchandise
Trade
End use category as % of total
merchandise imports
SUT
End use category as % of
total SUT imports
Intermediate
Final
demand Dual*
Intermediate
Final
demand
01 6,526 24% 76% - 6,767 17% 83%
02 149 100% - - 135 43% 57%
05 287 4% 96% - 280 50% 50%
10 1,909 100% - - 1,624 98% 2%
11 13,683 100% - - 13,136 98% 2%
13 932 100% - - - - -
14 4,707 9% - 91% 1,449 99% 1%
15 19,815 15% 85% - 23,584 36% 64%
16 239 3% 97% - 1,613 - 100%
17 6,225 36% 64% - 6,275 27% 73%
18 9,990 - 100% - 11,280 13% 87%
19 3,886 3% 97% - 4,094 39% 61%
20 3,086 95% 5% - 2,807 73% 27%
21 5,700 88% 12% - 5,420 85% 15%
22 3,131 47% 53% - 2,330 23% 77%
23 8,165 100% - - 12,085 42% 58%
24 32,272 67% 10% 23% 29,548 84% 16%
25 7,429 81% 19% - 6,874 61% 39%
26 3,035 85% 15% - 2,916 64% 36%
27 9,974 100% - - 7,245 100% -
28 6,115 70% 30% - 5,834 61% 39%
29 20,394 38% 61% 2% 20,106 37% 63%
30 15,769 34% 3% 63% 9,850 13% 87%
31 8,518 65% 35% - 8,197 57% 43%
32 17,558 29% 47% 24% 15,082 50% 50%
33 8,228 28% 72% - 8,694 64% 36%
34 37,275 31% 10% 60% 31,429 43% 57%
35 10,986 28% 12% 59% 13,210 60% 40%
36 10,133 12% 80% 8% 8,914 8% 92%
37 5 100% - - - - -
40 442 100% - - 466 69% 31%
Total 276,562 46% 33% 20% 261,247 51% 49%
Source: Merchandise trade data as reported to UN Comtrade converted to End-Use using an OECD-adapted version of the
UN conversion table (cf BTDIxE). SUT data are sourced from Eurostat.
* Dual: certain products have not been converted to an individual end use category but have been treated separately as ‘dual
use’ products, including e.g. personal computers, cell phones, and cars.
Page 8
STD/CSSP/WPTGS(2015)24
8
Table 4b. Differences between merchandise trade statistics and SUTs for imports by product and end-
use category*, in millions national currency, 2005, Greece
CPA
code
Merchandise
Trade
End use category as % of total
merchandise imports
SUT
End use category as % of
total SUT imports
Intermediate
Final
demand Dual*
Intermediate
Final
demand
01 1,228 63% 37% - 1,378 54% 46%
02 30 100% 0% - -98 65% 35%
05 82 1% 99% - 92 13% 87%
10 43 100% 0% - 48 100% 0%
13 15 100% 0% - 15 100% 0%
14 87 97% 0% 3% 159 100% 0%
15 3,913 20% 80% - 3,271 23% 77%
16 193 13% 87% - 414 1% 99%
17 1,056 57% 43% - 1,533 50% 50%
18 1,408 8% 92% - 1,688 4% 96%
19 634 5% 95% - 1,103 5% 95%
20 469 94% 6% - 509 94% 6%
21 981 77% 23% - 1,488 75% 25%
22 338 60% 40% - 292 21% 79%
23 1,740 100% 0% - 1,824 63% 37%
24 6,975 51% 12% 36% 6,567 64% 36%
25 1,002 81% 19% - 981 55% 45%
26 676 81% 19% - 762 71% 29%
27 2,365 100% 0% - 2,504 100% 0%
28 869 58% 42% - 896 61% 39%
29 3,062 27% 73% - 4,537 40% 60%
30 923 36% 8% 56% 882 8% 92%
31 918 61% 39% - 904 82% 18%
32 1,481 15% 54% 31% 1,467 22% 78%
33 1,076 19% 81% - 1,271 55% 45%
34 4,176 17% 16% 67% 4,853 16% 84%
35 2,642 8% 92% - 2,653 4% 96%
36 1,081 13% 86% 1% 1,708 11% 89%
37 0 100% 0% - - - -
40 131 100% 0% - 132 70% 30%
Total 39,592 42% 42% 16% 43,831 42% 58%
Source: Merchandise trade data as reported to UN Comtrade converted to End-Use using an OECD-adapted version of the
UN conversion table (cf BTDIxE). SUT data are sourced from Eurostat.
* Dual: certain products have not been converted to an individual end use category but have been treated separately as ‘dual
use’ products, including e.g. personal computers, cell phones, and cars.
13. To link merchandise import data to SUTs, they have to be converted from their reported HS
product category, into to Industries and End-Use classifications. This was achieved using the
approach of OECD’s BTDIxE database, which involves applying a conversion key (HS-ISIC-End Use)
to all trade flows at the HS 6-digit level. An End-Use category for each HS product is also assigned at
the 6-digit level, distinguishing between Intermediate goods and Final demand.4 Confidential data and
4 We report the results for “Final Demand” after combining consumptions and capital goods as defined
in BTDIxE. When separating ICIO, however, the imports shares are calculated independently for
these two categories. The artificial boundaries between Consumptions and Capitals sometimes can be
Page 9
STD/CSSP/WPTGS(2015)24
9
goods with multiple uses, such as cars and medicines related products are kept as a separate End Use
category (Dual). When comparing the data as presented in tables 4a and 4b, important differences can
be observed for both total merchandise imports (in 2005, merchandise imports were 6% higher for the
United Kingdom, and 11 % lower for Greece, as compared to SUTs). These differences become even
larger when looking at some of the breakdowns by CPA and End-Use.
2.3 SUTs and services trade statistics
14. So far we have only discussed discrepancies in NSO data with respect to merchandise trade.
Similar or even larger differences exist in the area of trade in services. This can be explained by the
fact that trade in services data are more difficult to collect (for example they are compiled from a
variety of different data sources, and are very often based on (small) survey samples rather than a
complete observation of all in and outgoing flows as in merchandise trade), and that they are even
more difficult to attribute to individual industries. We compare SUT data on total trade in services for
2005 of Greece and the UK with their official trade in services statistics from the Balance of
Payments (BOP) in table 5. This comparison shows that for the UK, the figures for exports and
imports are rather well-aligned across the two sources. However, this is clearly much less so the case
of Greece, where Trade in Services data record flows that are almost twice as high as the Greek SUT
table.
Table 5. Trade in Services, SUT and EBOPS S200, National Currency 2005
United Kingdom Greece
BOP total
services (S200) SUT
Ratio
(BOP/SUT)
BOP total
services (S200) SUT
Ratio
(BOP/SUT)
Export
s 114,219 117,258 0.97
31,095 16,524 1.88
Import
s 89,622 79,779 1.12
13,381 8,325 1.61
3. Methodology
15. The question of how to reconcile various data sources has long been addressed in the
context of the construction of national SUTs and IOs, which are built by combining information from
customs, various business surveys and other administrative sources. But with the increased attention
for TiVA and the underlying ICIO, the issue of reconciling not only national but also international
data sources has become even more prominent. As illustrated in section 2, official data may present
different values for the same variables, which means that in order to reconcile national and
international information, many methodological choices have to be made. It is exactly some of these
steps that this paper aims to address in more detail and to test for their impacts on the final results.
This section describes our methodology for testing the stability of TiVA estimates under various data
reconciliation options.
very blurry in both datasets; forcing data to separate introduced more variances than what is reported
in Table 4.
Page 10
STD/CSSP/WPTGS(2015)24
10
16. The first test deals with the national reconciliation: we compare the effect of reconciling
SUT data to the most recent national account aggregates on the main TiVA indicators, as compared to
not benchmarking the data to national accounts statistics. The second test deals with the international
reconciliation, and compares the results of the balancing the ICIO as currently done in TiVA (i.e., to
adjust the national trade data of all countries in roughly equal ways to come to a global picture) with
the balanced matrices from a single country’s perspective; thereby essentially creating as many
versions of the TiVA indicators as there are countries in the dataset. But before explaining the set-up
of our tests in more detail, we first explain the RAS technique in more detail (in section 3.1) and
subsequently introduce the central variable of interest that we selected for this study to comparing the
results: Vertical Specialisation (VS) (section 3.2).
3.1 What is RAS?
17. In compiling and updating national SAM and IO tables, RAS has become a well-established
technique. During its history of more than 50 years, it has not lost any of its popularity compared to
other alternatives such as optimisation models. RAS offers an algorithm that is easy to program and
understand, and this method has been adopted in the OECD-WTO TiVA project to balance national
IO tables and the ICIO in an international context.
18. Generally speaking, RAS is a bi-proportional scaling method which iteratively adjusts an
old matrix A0 with row sums u0 and column sums v0, to a new matrix A that satisfies a new set of
given row sums and column sums, u and v respectively. The adjustments are made by simply
applying the proportional distribution of the rows (columns) to the new row (column) total, and
repeating this process until the adjustment is complete. Annex 2 gives a simple numerical example of
RAS.
19. The RAS method was set in concrete for updating IO tables by Stone in the early 1960s
(Stone, 1961; Stone and Brown, 1962). Theoretically, a SAM or IO should always balance, meaning
that the row sum should equal the column sum, but empirically they never do so in the first instance.
This is due to the fact that creating an SAM or IO means that different data sources (different surveys
and administrative data) have to be combined which are never fully consistent. Thus, the RAS
technique is useful when for example one data source is preferred to describes the detailed input-
output industry structure of an economy, but another data source is considered better (or more timely)
with respect to the total output (or input) that is produced by industry. Similarly, RAS can be used to
harmonize SUTs (with structural information) to the most recent National Accounts data (‘new’
constraints for the row and column totals). And in an international IO setting, to rebalance after
splitting import data by partners, RAS also play a critical role.
20. Academics have continued to expand the usability of RAS procedure by developing a
number of additional features and successfully tackling some of the challenges. RAS has thereby
developed into a family of related procedures, and is now capable to:
Incorporate constrains with row and column sums (as explained above, in RAS);
Some elements of the matrix should not be changed even if the columns and rows sums do
and to incorporate sub-constrains on subsets of matrix elements (as in Modified RAS
(MRAS) or Three Stage RAS (TRAS));
Page 11
STD/CSSP/WPTGS(2015)24
11
Consider the reliability of the initial estimates and of the external constraints (extensions of
MRAS);
Handel negative values and preserve the sign of matrix elements (this is called Generalised
RAS (GRAS));
And deal conflicting external data (as in Konfliktfreies RAS (KRAS)).
21. Box 1 gives more details on each of these different RAS procedures.5 In this paper, we use
the standard TRAS and MRAS procedure. Negative values in TiVA are very few and generally only
exist in inventory adjustments. They are dealt with before the RAS procedure which is common
practice in the construction of IOs. Similarly, we do not have conflicting external data – constraints
are imposed separately and their effects on our TiVA estimates are even the subject of our current
investigation.
BOX 1. Different RAS techniques
MRAS. The modified RAS procedure (MRAS) is useful when some of the matrix elements of A are known in
addition to its row and column sums. In other words, MRAS utilises additional information compared to RAS, and
holds part of matrix A0 constant when applying the scaling procedure. It does so by first replacing the known
elements by zero and subtracting these values from the (new) row and column sum totals (also called
constraints), and subsequently subjecting the new net of A0 to the standard RAS procedure, adding back the
known elements afterwards.
This also applies to some aggregates of elements of matrix A. For example, in an inter-regional input-output
system, national aggregates may constitute partial information as constraints; and for more disaggregated
national table, the known aggregated terms can serve as constraints. These are realised by Oosterhaven (1986)
using national constraint to construct interregional and inter-industry tables; Batten (1985) is proposing further
constraints for intermediate and final demand data in a national table; Jackson (1993) use partition coefficients
for groups of cells of a disaggregated base year matrix to disaggregate cells in an updated but aggregated
matrix.
TRAS. The concept of three stage RAS (TRAS) is similar to MRAS, and was first proposed by Gilchrist
(1999), which was based Bacharach’s two stage RAS (1970) and Miller’s research (1985). It is an RAS algorithm
to account for generalised information on various sub-aggregates of the cells in a target matrix. This information
includes, but is not restricted to row and column sums, and the contents of particular cells with Canadian data as
an example.
A variation of MRAS method takes into account the uncertainty of the preliminary estimates. Allen (1974),
Lecomber (1975) and Allen (1975) addressed this concern. It has accomplished by introducing a matrix E contain
“reliability information” about the elements in A0. Lahr (2001), on the other hand, considers the uncertainties of
external constraints in treating the tolerances of the RAS criteria as functions of the varying reliabilities of row and
column sums.
GRAS and KRAS. Junius (2003) argues when negative entries are present, the stand RAS approach of
applying the RAS-algorithm to the matrix A0 easily leads to a new matrix A with a structure that may strongly
deviate from the structure of the old matrix, in particular, in the rows and columns with relatively large negative
entries. He also suggests in practice if the negative entries are treated outside the RAS-Procedure (i.e. removing
5 See also Lenzen et al. (2009) and Lahr et al. (2004), who provide very detailed literature reviews on
the RAS techniques.
Page 12
STD/CSSP/WPTGS(2015)24
12
before RAS and substitute back into to the matrix after) indicates that the negative entries in the adapted RAS
procedure do not contribute to the minimization, hence a loss of information. His generalised approach takes both
negative and positive numbers into account. Lenzen and his research colleagues (2009) again proposed a
generalised iterative scaling method which is a step ahead of GRAS method, Konfliktfreies RAS (KRAS), which is
able to balance and reconcile IO tables and Social Accounting Matrixes under conflicting external information and
inconsistent constraints. Additional to traditional and earlier RAS procedure described, KRAS can handle
constraints on arbitrarily sized and shaped subsets of matrix elements, and include reliability of the initial estimate
and the external constraints therefore find a compromised solution between inconsistent constraints.
22. It is important to note that in certain circumstances, RAS may go awry (i.e., not result in a
balanced matrix). Miller and Blair (2009) explain that this problem of non-convergence may
especially occur when the matrix contains too many zeroes (e.g. in highly disaggregated matrices),
because RAS only makes adjustments to non-zero elements. Too few of such non-zero cells may
make it impossible to find an adequate solution.
3.2 Selection of target indicators
23. To evaluate the impact of RAS and the different conflict national data on the TiVA results,
we do not need to assess the impact on a full set of (currently) 38 different TiVA indicators. Instead,
we focus on two indicators that is among the most important and that also drives many of the other
indicators.
24. First of all, we select the so-called degree of Vertical Specialization (VS).6 This indictor was
introduced by Hummels et al. (2001), and measures the value of imported inputs used directly and
indirectly in production of an exported good. In addition to being one of the central indicators in
TiVA, the VS indicator has the advantage that it can be calculated for national IO tables, as well as
from an international IO table. This facilitates therefore a direct comparison of the results across the
four scenarios that we present below. In matrix notation, it is expressed as
𝑉𝑆 = 𝐴𝑚[𝐼 − 𝐴𝐷]−1𝐸
25. In this formula, 𝐴𝑚 is the n x n imported coefficient matrix, I is the identity matrix, and 𝐴𝐷
is the n x n domestic coefficient matrix. 𝐸 is a n x 1 vector of exports. VS is also an n x 1 vector is
total exports of a country. The size of the matrices is determined by the number of industries (n) used
in the calculations.
26. The second indicator we choose to focus on is bilateral exports, which is especially relevant
for the test involving international reconciliation. This indicator is key in addressing concerns
regarding discrepancies between the data published in TIVA and gross trade statistics as nationally
reported.
6 The VS indicator is currently not published in TiVA but it is very similar to foreign (imported) value-
added content of exports. The difference between the two variables is that VS can be calculated both
from national Leontief inverse matrices and the global Leontief inverse matrix.
Page 13
STD/CSSP/WPTGS(2015)24
13
27. In additional to VS (as a share of exports) and total exports, we also report value added
(VA), gross output (GO), and total exports (TEXP) in each of our results, to provide better insights
into what element(s) cause(s) the differences at 18 industry level.
3.3 Testing different options of national and international data reconciliation
National data reconciliation
28. In our first test, we explore to what extent our key target variable, Vertical Specialization, is
different when it is calculated from a country’s national SUT and IO tables directly, versus after the
SUT and IOs have been adjusted to the main GDP main aggregates published in the national accounts
statistics.
29. The UK and Greece are chosen for this study because their SUT data differ with respect to
their alignment with national accounts: the data for the UK are relatively well-aligned, whereas the
Greek data show larger differences between these two sources. The comparison is done for the year
2005, because this is the latest benchmark year available in TiVA database as well as in national
country’s SUTs. This means that the national data are as robust as possible as they are compiled from
the most comprehensive survey information. Before performing the calculations, the national
published SUTs and IOs (sourced from Eurostat, and available with a breakdown into 59 industries)
are first aggregated to the 18 TiVA industries (see also Annex 1).
30. The national accounts constraints that are subsequently introduced include all the variables
listed in table 2 (see above), whereas the fundamental structure of the national IO is maintained. The
constraints for value added are given at the A7 industry level. Changes are hence proportionally
attributed to each of the underlying industries. Output is modified to using a constant value added-to-
output ratio. Import data (with a breakdown in goods and services) are used as additional constraints
for the rows, and final demand and exports (again, of goods and services) to adapt the column sums7.
International data reconciliation
31. Our second test assesses what changes are imposed on a country when its data are balanced
in an ICIO framework. We compare how the results from the current balancing procedure used in the
TiVA ICIO compare with those that iteratively keep one country’s data fixed (NB: in both cases, data
have been constrained to the National Accounts). Exports and VS shares are then derived from the
different balanced ICIOs, where the 𝐴𝑑 matrix involved country’s domestic IO coefficients and 𝐴𝑚 is
calculated by adding up the imports coefficients from all partners.
32. In essence, we run 57 separate RAS procedures, each time with a new starting point by
allowing one and only one country to keep its domestic constraints including total imports, import
7 Note that final consumption expenditure and gross capital formation should be included at basic
prices, but that benchmarking to national accounts imposes purchasers’ price. This will introduce a
very small bias: upwards for final demand and downwards for intermediate use.
Page 14
STD/CSSP/WPTGS(2015)24
14
partner shares by industry, and aggregated export statistics, whereas other countries’ data are adjusted
in RAS process to come to a balanced ICIO8.
4. Results
4.1 Test 1: the effect of constraining SUTs to national accounts on Vertical Specialisation
33. The results of our first test is reported in tables 5a and 5b, taking again the examples of the
UK and Greece. The tables display value-added and gross outputs by industry– both as presented in
the original SUTs and after adjustments to the National Accounts. Total export data are presented
excluding re-exports (and hence are slightly smaller than those data in Table 2 above, after also
adjusting for exchange rates). The tables show that for the UK, the alignment with national accounts
data resulted in an increase in value-added and gross outputs, whereas also total exports increased
(from 531 million to 544 million USD). For Greece, value added and output decreased, while total
exports increased 35%, from 40 million to 54 million (32 and 43 million euros).
Table 5a Vertical specialisation calculated from original SUT and SUT after aligns with SNA data, United
Kingdom, 2005, USD millions
Industry
Original SUT SUT benchmarked to national accounts VS share
Difference
(2)-(1) VA GO VS TEXP (1) VS
share VA GO VS TEXP
(1) VS
share
01T05 13,691 36,614.66 435.55 3,051 0.14 12,791 34,208 555 4,012 0.14 0.00
10T14 49,915 71,866 1,965 22,800 0.09 58,089 83,634 2,768 32,768 0.08 0.00
15T16 40,033 119,653 2,571 16,091 0.16 36,054 107,761 3,112 19,376 0.16 0.00
17T19 7,780 19,953 2,698 10,513 0.26 7,007 17,970 2,950 11,306 0.26 0.00
20T22 40,433 94,146 2,597 11,666 0.22 36,414 84,788 3,290 14,701 0.22 0.00
23T26 58,111 189,573 29,553 73,699 0.40 52,335 170,731 32,753 82,412 0.40 0.00
27T28 29,260 76,775 5,757 18,047 0.32 26,352 69,144 6,945 21,697 0.32 0.00
29 22,264 57,953 8,218 29,060 0.28 20,051 52,193 9,082 31,506 0.29 0.01
30T33 29,986 78,051 11,326 32,846 0.34 27,005 70,294 12,948 36,787 0.35 0.01
34T35 29,484 108,306 27,771 58,210 0.48 26,553 97,541 29,877 61,944 0.48 0.01
36T37 11,944 30,609 1,496 6,178 0.24 10,757 27,567 1,815 7,429 0.24 0.00
40T41 30,336 108,742 88 689 0.13 35,304 126,550 136 1,126 0.12 -0.01
45 127,033 327,205 168 1,581 0.11 149,295 384,546 148 1,510 0.10 -0.01
50T55 295,841 574,240 7,409 60,797 0.12 352,118 683,477 7,600 61,097 0.12 0.00
60T64 147,071 328,012 4,798 35,431 0.14 173,149 386,172 4,725 35,018 0.13 0.00
65T67 144,642 294,943 3,981 53,599 0.07 134,838 274,951 3,461 43,939 0.08 0.00
70T74 472,940 731,615 5,073 75,359 0.07 440,883 682,025 4,413 60,499 0.07 0.01
75T95 479,512 856,777 2,006 21,353 0.09 447,011 798,704 1,583 16,760 0.09 0.00
Total 2,030,275 4,105,032 117,911 530,970 0.22 2,046,006 4,152,255 128,159 543,885 0.24 0.01
34. When turning to the calculated VS shares, some differences can be observed when
comparing the calculations from the NA benchmarked and original SUTs. For United Kingdom,
overall VS share for the entire economy increased 1 percent after alignment with National Accounts.
8 I.e., in each run of the RAS procedure, different subsets of matrix A0 are fixed
Page 15
STD/CSSP/WPTGS(2015)24
15
Also at the industry level, the VS shares based on the adjusted SUT are either 1 percent higher or
lower, as compared to the original data. As for Greece, the overall impact of correcting SUT to SNA
controlled totals is an upward adjustment of 2 percent. But when examining the details by industry,
important differences can be observed: the highest VS share increase can be seen in Transport, storage
and communication (ISIC 60T64), at 3 percent. This may be explained by the fact that on one hand,
the value-added of Transport, storage and communication decreased nearly 20 percent, while on the
other hand both imports and exports in services are heavily adjusted to meet the new totals in SNA.
Table 5b Vertical specialisation calculated from original SUT and SUT after aligns with SNA data, Greece,
2005, USD millions
Industry
Original SUT SUT benchmarked to national accounts VS share
Difference
(2)-(1) VA GO VS TEXP (1) VS
share VA GO VS TEXP
(1) VS
share
01T05 10,450 16,773 241 1,623 0.15 10,482 16,823 461 2,989 0.15 0.01
10T14 998 1,863 8 152 0.05
1,001 1,870 15 273 0.05 0.00
15T16 5,741 19,410 455 2,136 0.21 5,552 18,772 843 3,770 0.22 0.01
17T19 2,186 5,187 516 1,808 0.29
2,114 5,016 788 2,639 0.30 0.01
20T22 2,000 4,970 58 228 0.26 1,934 4,806 112 426 0.26 0.01
23T26 5,117 20,699 2,452 4,519 0.54
4,948 20,018 3,946 7,166 0.55 0.01
27T28 2,694 11,039 906 2,241 0.40 2,605 10,676 1,460 3,481 0.42 0.02
29 891 2,192 251 704 0.36
862 2,120 383 1,037 0.37 0.01
30T33 897 2,335 187 564 0.33 867 2,258 291 847 0.34 0.01
34T35 815 2,045 238 706 0.34
788 1,978 358 1,035 0.35 0.01
36T37 786 2,031 42 147 0.29 760 1,964 78 266 0.29 0.01
40T41 5,851 9,215 0 3 0.10
5,874 9,250 1 6 0.10 0.00
45 13,617 31,015 74 286 0.26 14,986 34,132 132 492 0.27 0.01
50T55 51,883 80,896 560 5,373 0.10
47,374 73,866 839 7,420 0.11 0.01
60T64 22,048 37,633 3,418 17,704 0.19 18,382 31,375 4,221 19,023 0.22 0.03
65T67 10,440 14,683 14 357 0.04
11,003 15,475 26 595 0.04 0.01
70T74 30,759 41,004 46 966 0.05 32,418 43,217 86 1,584 0.05 0.01
75T95 49,991 71,714 33 329 0.10
52,688 75,583 58 540 0.11 0.01
Total 217,161 374,703 9,498 39,846 0.24 214,638 369,201 14,098 53,589 0.26 0.02
4.2 Test 2: Effect of different national ICIO balancing starting points on VS and bilateral exports
35. The results of our second test are more elaborate to report. We discuss the findings at
various levels of aggregation, moving from the country level, to the country × industry level, and
finally to the country × industry × partner level.
Country level results
36. The results at the country aggregate level are displayed in table 6. Note that in this case,
value-added, output and total exports are exactly the same in full RAS procedure as the partial RAS
procedure that constraints to only a single country’s national perspective. We observe that the core
descriptive statistics of globalisation, i.e. the degree of Vertical Specialization, remains very stable at
the total economy level, regardless of the RAS procedure that is used. For 44 out of 57 countries, the
differences are less than 1 percent. Only Cambodia, Malta, Israel and the Philippines show slightly
higher differences.
Page 16
STD/CSSP/WPTGS(2015)24
16
Table 6. VS share for total economies, 2005, USD millions and percent
COU VA GO TEXP
VS
Share
full
RAS
(1)
VS
Share
country
view
(2)
Diff.
(2)-(1)
COU VA GO TEXP
VS
Share
full
RAS
(1)
VS
Share
countr
y view
(2)
Diff.
(2)-(1)
ARG 169,726 325,816 44,482 0.128 0.127 -0.001
JPN 4,541,977 8,616,279 634,222 0.15 0.149 -0.001
AUS 704,453 1,470,719 144,050 0.124 0.123 -0.001
KHM 5,965 13,116 3,831 0.38 0.413 0.033
AUT 272,865 556,281 141,099 0.318 0.32 0.002
KOR 757,619 1,863,670 320,872 0.365 0.358 -0.007
BEL 335,571 783,109 215,356 0.415 0.418 0.003
LTU 23,491 46,109 14,210 0.38 0.365 -0.015
BGR 24,454 63,497 10,819 0.307 0.296 -0.011
LUX 33,668 94,007 54,593 0.56 0.561 0.001
BRA 756,762 1,557,087 133,091 0.129 0.128 -0.001
LVA 14,114 32,155 7,241 0.287 0.299 0.012
BRN 9,341 14,602 6,606 0.067 0.081 0.014
MEX 823,342 1,451,870 218,745 0.304 0.311 0.007
CAN 1,056,764 2,039,851 413,907 0.257 0.261 0.004
MLT 5,161 11,205 4,262 0.394 0.424 0.03
CHE 350,577 693,817 175,001 0.292 0.29 -0.002
MYS 135,195 327,625 156,397 0.421 0.431 0.01
CHL 111,473 232,817 47,267 0.165 0.168 0.003
NLD 567,306 1,196,376 302,615 0.341 0.333 -0.008
CHN* 2,257,006 6,761,571 802,684 0.285 0.285 0
NOR 268,832 491,376 132,813 0.139 0.143 0.004
CYP 15,193 25,485 6,129 0.151 0.172 0.021
NZL 108,881 234,563 29,972 0.191 0.191 0
CZE 111,667 303,564 80,222 0.409 0.411 0.002
PHL 101,010 207,573 45,151 0.457 0.424 -0.033
DEU 2,516,901 5,054,014 973,072 0.268 0.269 0.001
POL 267,759 603,798 107,576 0.304 0.311 0.007
DNK 218,255 443,296 109,374 0.306 0.308 0.002
PRT 165,251 360,513 49,889 0.264 0.271 0.007
ESP 1,012,008 2,199,966 261,389 0.266 0.271 0.005
ROU 87,599 188,418 30,887 0.276 0.282 0.006
EST 12,304 30,393 10,294 0.472 0.48 0.008
RUS 654,694 1,308,870 262,308 0.083 0.084 0.001
FIN 169,950 377,125 79,521 0.331 0.335 0.004
SAU 309,271 442,592 189,229 0.034 0.035 0.001
FRA 1,914,994 3,851,835 539,316 0.248 0.253 0.005
SGP 119,724 340,359 160,827 0.518 0.518 0
GBR 2,030,279 4,104,732 542,935 0.202 0.211 0.009
SVK 42,511 107,255 35,974 0.476 0.483 0.007
GRC 217,161 374,703 54,614 0.217 0.235 0.018
SVN 31,310 71,367 21,209 0.405 0.42 0.015
HKG 174,770 403,894 79,747 0.266 0.259 -0.007
SWE 324,123 692,253 168,971 0.322 0.327 0.005
HUN 94,323 227,332 64,929 0.483 0.485 0.002
THA 158,717 403,514 124,172 0.385 0.385 0
IDN 280,151 569,699 93,694 0.177 0.187 0.01
TUR 425,515 903,308 99,871 0.209 0.212 0.003
IND 781,681 1,602,346 154,264 0.181 0.179 -0.002
TWN 354,569 857,197 217,972 0.422 0.4 -0.022
IRL 177,717 412,321 158,334 0.446 0.447 0.001
USA 11,695,10
0
21,606,56
6
1,203,11
8 0.132 0.134 0.002
ISL 13,518 28,674 4,912 0.362 0.388 0.026
VNM 51,859 120,659 34,742 0.352 0.35 -0.002
ISR 122,407 249,553 54,933 0.373 0.343 -0.03
ZAF 220,317 509,096 64,696 0.165 0.175 0.01
ITA 1,597,329 3,472,846 457,535 0.272 0.271 -0.001
Note that the VS share is similar to TiVA Foreign valued-added content as a share of exports.
*Chinese VS is 8 percent lower than TiVA foreign value-added content as a share of exports. The main differences come from
how the three Chinese production component are aggregated.
Page 17
STD/CSSP/WPTGS(2015)24
17
Country-Industry level results
37. When we explore the findings at a more detailed level, by focusing on the results by
industry, we continue to find that the VS measure is rather robust to different balancing starting points.
We present here again the results for the UK and Greece, but findings for other countries are available
upon request. Table 7a, for the UK, shows that the VS shares that are derived from a full (global)
balancing (RAS) of the ICIO are not much different from the results we obtain when keeping the UK
data fixed in the final balancing. The overall results of VS are 20.2 percent versus 21.1 percent,
respectively. When looking at the industry level, we see that the highest differences are associated
with the industries with relatively high VS shares, such as Chemicals and minerals (23T26) and
Transport equipment (34T35). For Greece, we find that the VS share when using the Greek data fixed
equals 24 percent, or only 0.2 percent higher compared to the results of the full (global) RAS
procedure (Table 7b). The industry of Transport equipment (34T35) showed the largest difference,
with a 3 percent downwards adjustment.
Table 7a. United Kingdom, Value-added, Gross outputs, exports and VS share by industry, 2005
Industry VA GO TEXP
VS share full
RAS (1)
VS country
perspective (2)
Difference (2)-
(1)
01T05 13,691 36,615 3,967 0.118 0.123 0.005
10T14 49,915 71,866 28,282 0.082 0.089 0.007
15T16 40,033 119,655 18,687 0.146 0.151 0.005
17T19 7,780 19,953 11,656 0.216 0.219 0.003
20T22 40,433 94,147 15,281 0.198 0.207 0.009
23T26 58,112 189,576 91,915 0.332 0.354 0.022
27T28 29,261 76,776 23,512 0.264 0.274 0.010
29 22,264 57,954 33,796 0.236 0.245 0.009
30T33 29,986 78,052 38,832 0.275 0.281 0.006
34T35 29,484 108,307 63,260 0.376 0.389 0.013
36T37 11,944 30,610 6,587 0.218 0.225 0.007
40T41 30,337 108,742 944 0.106 0.118 0.012
45 127,033 327,205 1,239 0.096 0.100 0.004
50T55 296,909 576,312 43,055 0.111 0.114 0.003
60T64 146,000 325,623 30,861 0.123 0.129 0.006
65T67 144,643 294,944 45,023 0.061 0.064 0.003
70T74 472,941 731,617 69,106 0.073 0.076 0.003
75T95 479,514 856,781 16,932 0.077 0.079 0.002
Total 2,030,279 4,104,732 542,935 0.202 0.211 0.009
Page 18
STD/CSSP/WPTGS(2015)24
18
Table 7b. Greece, Value-added, Gross outputs, exports and VS share by industry, 2005
ISIC VA GO TEXP
VS share full
RAS (1)
VS country
perspective (2)
Difference (2)-
(1)
01T05 10,450 16,773 2,192 0.112 0.121 0.009
10T14 998 1,863 - 0.132 0.143 0.011
15T16 5,741 19,410 2,336 0.182 0.191 0.009
17T19 2,186 5,187 1,548 0.217 0.219 0.002
20T22 2,000 4,970 335 0.206 0.214 0.008
23T26 5,117 20,699 5,925 0.439 0.529 0.090
27T28 2,694 11,039 3,329 0.357 0.375 0.018
29 891 2,193 724 0.242 0.254 0.012
30T33 897 2,335 775 0.266 0.276 0.010
34T35 815 2,045 559 0.273 0.300 0.027
36T37 786 2,031 186 0.231 0.236 0.005
40T41 5,852 9,215 43 0.094 0.108 0.014
45 13,617 31,015 563 0.209 0.222 0.013
50T55 51,883 80,896 7,293 0.100 0.103 0.003
60T64 22,048 37,633 24,707 0.214 0.225 0.011
65T67 10,440 14,683 827 0.047 0.047 0.000
70T74 30,759 41,004 2,367 0.055 0.055 0.000
75T95 49,991 71,714 905 0.093 0.095 0.002
TOTAL 217,161 374,703 54,614 0.217 0.235 0.018
38. When summarizing the results for all 57 countries and 18 industries (i.e. for 1026
observations) and comparing the results from the overall RAS procedure as is used in TiVA, and the
procedures that is constraint to each individual national country constant, we see that in three quarters
of the cases, the difference is less than 1% (positively or negatively), as displayed in Table 8 (rows 2,
3, and 4). This table also shows on average, the full RAS procedure has a pushes down the VS shares
(i.e. gives conservative estimates of the importance of imports for exports): in the majority of cases
(rows 4 and 5) the country-industry level results are smaller in the full RAS procedure than when
keeping than original country values fixed as in the partial RAS procedures.
Table 8. VS as a share of exports - the difference between full RAS and national perspectives
VS share from RAS with country fixed perspective, minus VS from full RA’ N Percent
(1) less than 0.01 71 7%
(2) greater or equal to -0.01 but less than 0 168 16%
(3) equal to 0 96 9%
(4) greater than 0 but less or equal to 0.01 499 49%
(5) greater than 0.01 192 19%
Total. 1026 100%
Page 19
STD/CSSP/WPTGS(2015)24
19
Country-Industry-Partner level results
39. As a final step in our analysis, we examine the results at the partner level. It is important to
realize that VS shares do not vary beyond the industry level, since industry production functions are
not assumed to be dependent upon the export destination. For example, Chinese textiles exports to the
US and to South Africa have the same import content, since the same textiles industry produces these
garments. Therefore, in this section, we focus on the effects of the different balancing procedures on
bilateral export relationships.
40. Table 9 provides an example of the results for the exports of the UK to the Netherlands.
(Again, detailed results for other countries are available upon request). In the table, we compare the
‘full RAS” (i.e., the global, TiVA perspective) with the “UK perspective” – which is the solution
when the UK’s data are kept constant when balancing the ICIO, and the “Dutch perspective”, which
gives the solution when the Dutch data are kept fixed during RAS. This shows that the full RAS
solution, when redistributing discrepancies proportionally, can come to a value of exports that is
higher (or lower) than reported by either the UK or the Netherlands. Take for example Chemicals and
minerals (23T26): table 9 shows that from the UK perspective, the exports to the Netherlands are 4.5
billion, while the Netherlands data reports 3.9 billion. The global RAS finds a solution higher than
both countries’ starting points, at 4.9 billion.
Table 9. RAS bilateral solutions for – UK exports to Netherlands (Netherlands imports from UK)
Ind Full RAS
“GBR
Perspective”
“NLD
Perspective”
Stats for exports (excl. GBR and NLD perspectives)*
Mean SD Min Max
01T05 145.0 133.6 109.9 145.8 1.8 145.0 155.1
10T14 2267.0 2105.4 2292.1 2266.8 23.3 2147.0 2374.5
15T16 731.4 692.7 658.4 733.3 3.2 730.7 746.5
17T19 189.2 186.6 127.8 190.4 2.1 189.2 199.5
20T22 584.0 598.5 377.9 587.5 7.5 584.0 621.5
23T26 4855.2 4506.1 3901.2 4877.1 49.0 4855.0 5101.8
27T28 563.7 551.1 612.1 562.9 1.9 553.0 564.2
29 693.9 701.9 640.6 695.0 3.4 692.9 718.1
30T33 1067.1 1055.9 1147.1 1065.8 2.8 1052.4 1068.4
34T35 2533.6 2391.4 1312.7 2557.8 44.6 2533.8 2756.5
36T37 180.0 184.5 230.9 179.0 2.2 169.4 181.3
40T41 46.1 41.0 92.4 45.0 4.2 21.8 47.7
45 88.9 84.5 143.4 87.8 3.6 67.4 90.1
50T55 2770.3 2818.2 1815.9 2783.8 22.2 2764.8 2882.9
60T64 944.9 964.8 1111.3 942.0 5.9 913.9 945.2
65T67 2849.4 2789.6 3455.4 2840.4 26.7 2656.3 2855.6
70T74 3269.1 3383.8 4853.9 3239.6 70.1 2774.9 3268.9
75T95 662.2 639.9 979.0 657.8 6.6 625.8 663.0
Note: RAS bilateral solutions for other country-partner are available upon request.
41. The second (right-hand) half of the table shows the statistics for the value of UK-NL exports
by summarizing all 57 partial RAS solutions (excluding the Dutch and British perspectives). This
Page 20
STD/CSSP/WPTGS(2015)24
20
shows that the mean estimate from those 55 solutions is also around 4.9 billion, with relatively small
standard deviations. The minimum and maximum values shows in the table for Chemicals and
minerals are achieved when respectively fixing Viet Nam, and France. The interpretation of this is as
follows: when fixing the data using a French perspective, the exports to France given by the initial
estimate is lower than expected, meaning that the solution for other countries should be higher.
Similarly, while the original exports data for Viet Nam are is higher than expected, the solution for
other countries should be lower. The degree of the adjustment in the partial RAS in the end depends
on the starting position (i.e. sum of exports is greater or smaller than the constraints) and which
number we choose to hold.
42. When examining the results for all 57 TiVA countries (and partner countries) and all 18
industries, we find that in nearly all instances (98%), the estimated value of bilateral exports relations
subject to RAS with respect to different benchmark countries, the majority of the data (98%) reject the
hypothesis that it is significantly different from the mean estimates when the wide definition is used
(i.e. including the observations of exports from a country perspective and from a partner perspective).
The standard deviations of bilateral trade are also fairly narrow – a small tails for the distributions: 69%
of the observation has the deviation within ±10% of their means; and an additional 17% within ±20%
boundaries
5. Discussion and conclusion
43. The aim of this study was to offer insights into the stability of TiVA indicators that are
derived from the ICIO, with a focus on the measure of Vertical Specialization (as introduced by
Hummels et al, 2001) and bilateral export relationships. We highlighted that given the prevalence of
trade asymmetries and national data asymmetries, the process of constructing an ICIO means that
many decisions have to be taken regarding the correct statistical approach to combining and
harmonizing the conflicting data, nationally and internationally. We in particular examined two steps
in the ICIO construction process in more detail for their consequences for both VS and bilateral
export relationships, which is the benchmarking of national SUTs to national accounts, and the final
RAS procedure to balance the ICIO.
44. The first comparison involved the question of whether or not to benchmark national SUTs
and IOs to national account aggregates. We found that the effects of restricting a country’s SUT to
SNA data has only small effects at either the country or country-industry level, with the majority of
variations being less than 1% of the final VS share in exports. It should be highlighted though that VS
shares continue to be sensitive to how the data is converted from SUT to IO tables, including e.g. the
way in which the aggregation from products to industries is performed, how adjustments are made to
move from purchasers to basic prices, and the treatment of negative values and outliers. Still, the
relative robustness of the main indicators to sometimes quite substantive revisions in national
accounts data should not only reassure users of TiVA indicators, but also gives a positive signal
regarding the feasibility of a related project, which is the nowcasting or projection of the ICIO in
order to derive more timely TiVA indicators (see e.g. STD/CSSP/WPTGS(2015)22).
45. The second test involved comparing the balanced TiVA results for VS and bilateral exports
(‘full RAS’) with the results that are obtained when constraining the RAS procedure in such a way
that the ICIO that is created fully reflects the national statistics of one single country, while permitting
Page 21
STD/CSSP/WPTGS(2015)24
21
other countries’ trade flows to change to respond to the needs of global balancing (i.e. essentially
creating 57 – the number of countries in TiVA – different ICIO tables) (‘partial RAS’). The first and
foremost important conclusion can be drawn from this experiment is that the full RAS procedure as
used in TiVA, and the partial RAS procedures that iteratively holds one country’s data constant,
converge to the same solution. The ratio of Vertical Specialization as a share of exports is stable at
both the country and at the country-industry level. For the majority of the countries (46 out of 57) and
country-industry pairs (74%) in question, the differences of full RAS and partial RAS results is less
than 1%. Also the results for total exports proved to be stable, even if in the context of trade
asymmetries, the estimates could differ from the data reported by countries and partners involved.
Implications for trade statistics
46. Data discrepancies, both nationally and internationally, remain one of the big obstacles for
TiVA to tackle. It has been highlighted many times in this paper: at the aggregate level SUT/SNA
definitions and values differ from those in merchandise trade and trade in service statistics; at bilateral
level, trade asymmetries are the rule rather than the exception. Continued efforts to better understand
and if possible reduce these asymmetries are vital. Also the trade balancing process within TiVA can
be improved. At the moment, the total exports by industry (as given by the national IO table) are used
as constraints in the ICIO. Subsequently, partner breakdowns are produced using import partner
shares. Clearly, in the presence of asymmetries, the summation of imports from all partners will result
in different export figures than observed from the bilateral exports recorded in the merchandise and
services trade statistics. Future work on balancing trade statistics can consider information on both
exports and imports in combination with information from SUTs, which provide insights into the use
of imported products by industry.
Page 22
STD/CSSP/WPTGS(2015)24
22
REFERENCES
Allen, R. (1974) ‘Some Experiments with the RAS Method of updating Input-Output Coefficients’,
Oxford Bulletin of Economics and Statistics, 36, 217-228.
Allen, R. and Lecomber, J (1975) Some Tests on a Generalised Version of RAS.
Bacharach, M (1970) Biproportional Matrices and Input-Output Change. Cambridge, UK, Cambridge
University Press.
Battern, D. and Martellato (1985) Classical Versus Modern Approaches to Interregional Input-Output
Analysis
Eurostat (2008) Eurostat Manual of Supply, Use and Input-Output Tables.
Gilchrist, D. and St Louise, L. (1999) ‘Completing Input-Output Tables Using Partial Information,
with an Application to Canadian Data’, Economics Systems Research, 11, 185-193.
Gilchrist, D. and St Louise, L. (2004) ‘An Algorithm for the consistent Inclusion of Partial
Information in the revision of Input-Output tables’, Economic Systems Research, 16, 149-156.
Hammels, D., Ishii, J. and Yi, Kei-Mu (2001) ‘The Nature and Growth of Vertical Specialization in
World Trade’, Journal of International Economics 54(1): 75-96.
Jackson , R. and Comer J. (1993) ‘An alternative to Aggregated Base Tables in Input-Output Table
Regionalisation’, Growth and Change, 24, 191-205.
Junius, T. and Oosterhaven, J. (2003) ‘The Solution of Updating a Matrix with Both Positives and
Negatives Entries’, Economic Systems Research, 15, 87-96.
Lahr, M (2001) A strategy for Producing Hybrid regional Input-Output Tables.
Lahr, M and de Mesnard L. (2004) ‘Biproportional Techniques in Input-Output Analysis: Table
Updating and Structural Analysis’, Economic Systems Research, 16, 115-134.
Lecomber, J (1975) A critique of Methods of Adjusting, Updating and Projecting Matrices, Together
with Some New Proposals.
Lenzen, M. Gallego, B. and Wood, R. (2006) ‘A Flexible Approach to Matrix Balancing under
Partial Information’, Journal of Applied Input-Output Analysis, 11-12, 1:24.
Lenzen, M. Wood, R. and Gallego, B. (2007) ‘Some comments on the GRAS Method’, Economic
Systems Research, 19, 461-465.
Lenzen, M. Gallego, B. and Wood, R. (2009) ‘Matrix Balancing under Conflicting Information’,
Economic Systems Research, 21:1, 23-44.
Page 23
STD/CSSP/WPTGS(2015)24
23
Leontief, W. (1941) The structure of American Economy, 1919-1929: An Empirical Application of
Equilibrium Analysis, Cambridge, UK, Cambridge University Press.
Miller, R. E. and Blair, P. D. (1985, 2009) Input-Output Analysis – Foundations and Extensions. 1ed
and 2ed. Cambridge University Press.
Oosterhaven, J., Piek G. and Stelder, D (1986) ‘Theory and Practice of Updating Regional versus
Interregional Interindustry Tables’, Papers of the Regional Science Association, 59, 57-72
Stone, R (1961) Input-Output and National Accounts, Organization for European Economic
Cooperation, Paris
Stone, R. and Brown, A. (1962) A Computable Model of Economic Growth, London, Chapman and
Hall.
Page 24
STD/CSSP/WPTGS(2015)24
24
ANNEX 1. PRODUCT AND INDUSTRY CLASSIFICATIONS
CPA codes and products definition for goods
Code Description
01 Products of agriculture, hunting and related services
02 Products of forestry, logging and related services
05 Fish and other fishing products; services incidental of fishing
10 Coal and lignite; peat
11 Crude petroleum and natural gas; services incidental to oil and gas extraction excluding surveying
12 Uranium and thorium ores
13 Metal ores
14 Other mining and quarrying products
15 Food products and beverages
16 Tobacco products
17 Textiles
18 Wearing apparel; furs
19 Leather and leather products
20 Wood and products of wood and cork (except furniture); articles of straw and plaiting materials
21 Pulp, paper and paper products
22 Printed matter and recorded media
23 Coke, refined petroleum products and nuclear fuels
24 Chemicals, chemical products and man-made fibres
25 Rubber and plastic products
26 Other non-metallic mineral products
27 Basic metals
28 Fabricated metal products, except machinery and equipment
29 Machinery and equipment n.e.c.
30 Office machinery and computers
31 Electrical machinery and apparatus n.e.c.
32 Radio, television and communication equipment and apparatus
33 Medical, precision and optical instruments, watches and clocks
34 Motor vehicles, trailers and semi-trailers
35 Other transport equipment
36 Furniture; other manufactured goods n.e.c.
37 Secondary raw materials
40 Electrical energy, gas, steam and hot water
TiVA May 2013 industry classification
Index ISIC Rev.3 Description
1 01T05 Agriculture, hunting, forestry and fishing
2 10T14 Mining and quarrying
3 15T16 Food products, beverages and tobacco
4 17T19 Textiles, textile products, leather and footwear
5 20T22 Wood, paper, paper products, printing and publishing
6 23T26 Chemicals and non-metallic mineral products
7 27T28 Basic metals and fabricated metal products
8 29 Machinery and equipment, nec
9 30T33 Electrical and optical equipment
10 34T35 Transport equipment
11 36T37 Manufacturing nec; recycling
12 40T41 Electricity, gas and water supply
13 45 Construction
14 50T55 Wholesale and retail trade; Hotels and restaurants
15 60T64 Transport and storage, post and telecommunication
16 65T67 Financial intermediation
17 70T74 Real estate, renting and business activities
18 75T95 Community, social and personal services
Page 25
STD/CSSP/WPTGS(2015)24
25
ANNEX 2. NUMERICAL EXAMPLE OF RAS PROCEDURE
d e f Current sum DESIRED sum
a 5.0 8.0 3.0 16.0 19.0
b 6.0 8.0 9.0 23.0 21.0
c 8.0 5.0 5.0 18.0 20.0
Current sum 19.0 21.0 17.0
DESIRED sum 20.0 24.0 16.0
d e f Current sum DESIRED sum
a 5.9 9.5 3.6 19.0 19.0
b 5.5 7.3 8.2 21.0 21.0
c 8.9 5.6 5.6 20.0 20.0
Current sum 20.3 22.4 17.3
DESIRED sum 20.0 24.0 16.0
d e f Current sum DESIRED sum
a 5.8 10.2 3.3 19.3 19.0
b 5.4 7.8 7.6 20.8 21.0
c 8.8 6.0 5.1 19.8 20.0
Current sum 20.3 22.4 17.3
DESIRED sum 20.0 24.0 16.0
d e f Current sum DESIRED sum
a 5.7 10.0 3.2 19.0 19.0
b 5.4 7.9 7.6 21.0 21.0
c 8.8 6.0 5.2 20.0 20.0
Current sum 20.0 23.9 16.0
DESIRED sum 20.0 24.0 16.0
d e f Current sum DESIRED sum
a 5.7 10.0 3.2 19.0 19.0
b 5.4 7.9 7.6 21.0 21.0
c 8.8 6.0 5.2 20.0 20.0
Current sum 20.0 24.0 16.0
DESIRED sum 20.0 24.0 16.0
iteration 4) Align matrix content (green) with the desired totals of rows using the proportional
method (e.g. for cell a/d in the topleft corner, we calculate (5.7 / 20.0)*20)
=> now, the current sums of the rows and columns in the matrix are equal to the desired sums (at
1 digit). No more iterations are needed.
Original matrix (in green) with its current sums of all columns and sums of all rows, with the new
desired (in blue) sums of all columns and sums of all rows
iteration 1) Align matrix content (green) with the desired totals of columns using the proportional
method (e.g. for cell a/d in the topleft corner, we calculate (5 / 16)*19)
iteration 2) Align matrix content (green) with the desired totals of rows using the proportional
method (e.g. for cell a/d in the topleft corner, we calculate (5.9 / 20.3)*20)
iteration 3) Align matrix content (green) with the desired totals of columns using the proportional
method (e.g. for cell a/d in the topleft corner, we calculate (5.8 / 19.3)*19)