USING FIRM-LEVEL DATA TO COMPARE PRODUCTIVITIES ACROSS COUNTRIES AND SECTORS: POSSIBILITIES AND CHALLENGES

Sarah Oliver

Caroline Peters

ECONOMICS WORKING PAPER SERIES

Working Paper 2018-07-A

500 E Street SW

Washington, DC 20436

July 2018

Office of Economics working papers are the result of ongoing professional research of USITC Staff and are solely meant to represent the opinions and professional research of individual authors. These papers are not meant to represent in any way the views of the U.S. International Trade Commission or any of its individual Commissioners. Working papers are circulated to promote the active exchange of ideas between USITC Staff and recognized experts outside the USITC and to promote professional development of Office Staff by encouraging outside professional critique of staff research.

Using Orbis to compare productivities across countries and sectors: possibilities and challenges

Office of Economics Working Paper 2018-07-A

July 2018

Abstract

A five-year panel of cross-country data for 2012-2016 drawn from the Orbis database is used to evaluate the advantages and shortcomings of this data source in calculating firm level productivity. We find that conditional on the productivity measure employed, country and sector coverage can vary widely in the Orbis database due to different national reporting requirements across countries. This paper also compares the average productivity of the same sector across countries and the average productivity of domestic and foreign owned firms in the same sector. In every type of productivity calculation employed in this analysis, foreign firms are significantly more productive than their domestic counterparts.

Office of Economics

Sarah Oliver

Office of Industries

sarah.oliver@usitc.gov

Caroline Peters

Office of Economics

caroline.peters@usitc.gov

Introduction

Theoretical and empirical work have shown that the productivity of a country’s firms is an important factor in determining its place in the global economy, as a country’s most productive firms are more likely to become exporters (Melitz 2003). This stylized fact has led to a renewed focus in the trade literature on measuring productivity, at the sector- or firm-level, in an accurate and consistent manner. One significant hurdle impeding this research endeavor, however, is finding readily accessible databases that cover a large set of countries, industries, and firms, allowing for meaningful analysis of firm productivity dynamics across countries and sectors.

Our purpose in this paper is to demonstrate the usefulness of firm-level data from the Orbis database by constructing and analyzing measures of firm productivity at the country- and sector-level for both manufacturing and services sectors. In order to evaluate the advantages and shortcomings of Orbis for productivity analysis, we document our data coverage by country and two-digit NACE sector, and discuss the variety of productivity measures that can be applied to our dataset. Since different measures of productivity require different financial variables, such as assets or depreciation, country coverage varies by the method choosen to calculate or estimate productivity. Of the methodologies considered, labor productivity provides the best coverage: from 2012 to 2016, 49 countries in our data report total revenue and employment data for at least 30 firms per two-digit sector, while the Levinsohn-Petrin method for estimating total factor productivity (TFP), which requires data on intermediate inputs, had the lowest country coverage at only 27 countries in our sample.

The productivity measures computed from our dataset are also useful in demonstrating firm heterogeneity within a country and sector. Specifically, the Orbis database provides information on firm ownership, which can be used to distinguish between domestic firms and foreign-owned affiliates in a particular market. The second part of this paper finds that on average, foreign firms are significantly more productive than domestic firms in the markets where they operate. This result is consistent across all of our methodologies for calculating productivity.

The remainder of this paper is divided into four sections. The first section provides an overview of the previous literature that relies on firm-level data, including the Orbis database, to analyze productivity. The next section describes the data gathered from Orbis and compares it to other sources of firm-level data. The third section provides an overview of the methodologies used to calculate productivity.

The fourth section, which presents the results of our analysis, is divided into three parts. Part A looks at which countries are most productive and how productivity in particular sectors varies across countries. Part B examines the dispersion of productivity within a country and sector and across different estimation methods. Part C exploits Orbis’ firm ownership information to examine the productivity differences between domestic and foreign firms.

Literature Review

There is a growing consensus in the productivity literature around the advantages of using firm-level data in conducting productivity analysis (Bartelsman and Doms 2000; Syverson 2004; Bartelsman et al. 2009). As discussed in Bartelsman and Doms (2009), firm-level data can be used to establish stylized facts about the dispersion of productivity across firms, the uniformity of changes in productivity, the persistence of productivity differentials, the consequences of entry and exit, and the importance of changes in resource reallocation across firms to aggregate productivity growth. Firm-level data also avoids the issues that accompany productivity analysis when using sector-level data, such as adjustments made for missing establishments by applying productivity assumptions to employment statistics from labor force surveys. Input extrapolation is a frequent occurrence in the services sector where data are thinner overall (OECD 2001). As such, the firm-level nature of the database make Orbis a valuable resource for estimating the productivity of services firms in particular.

There have been several studies that have drawn upon the Orbis database to create firm-level datasets for the purposes of estimating productivity. Some recent works that rely on Orbis as their main data source include Gal (2013), which looks at OECD countries from 2000-2008; Kalemli-Ozcan et al. (2015), which looks at European firms from 1999-2012; and Gopinath et al. (2017), which looks at manufacturing firms in Spain from 1999-2012. Our paper follows in the direction of Gal (2013) who examines Orbis in the context of firm productivity analysis and proposes several imputation strategies to account for coverage issues in Orbis when measuring TFP along with other methods such as re-sampling and PPP-conversion adjustments to make these productivity measures internationally comparable.

While our dataset from Orbis is more current than the datasets used in many of these papers, the total number of years of firm data that we have pulled is less extensive, and we have only queried the database once rather than creating a panel of draws from different database vintages as in Kalemli-Ozcan et al. (2015) and Gopinath et al. (2017). Rather than attempting to build a panel of similar length to conduct more detailed time-series analysis, our primary purpose in this paper is to provide a flavor of the cross-sectional analysis of firm-level productivity that can be performed using the Orbis database.

Data

Bureau van Dijk’s Orbis dataset reports firm-level financial data that varies in coverage based on the reporting requirements of particular countries (Bureau van Dijk 2017). For our sample, we only include firms that had non-missing revenue and employment data for 2013, 2014, and 2015, and our overall coverage of these firms includes the 5-year span from 2012-2016. We use the EU’s Statistical classification of economic activities (NACE) codes to classify firms by industry at the two digit-level, and we include manufacturing sectors corresponding to codes 10-33 and service sectors codes under NACE 41-93, excluding public service (84) and banking and insurance activities (64-66) where revenue is not a good predictor of productivity. After excluding country-sector pairs where there are fewer than 30 firms in a given year, the sample consists of 4.3 million firms, 49 countries and 1,898 country-sector pairs.

To adjust for differences in prices of goods and services across countries, we convert financial variables to purchasing power parity (PPP)-adjusted figures, using the World Bank PPP conversion rate for each year. In order to calculate TFP using the index method, we take the share of labor in total output as reported in either the World KLEMS or OECD Structural Analysis (STAN) databases. A more detailed explanation of these data sources is available in appendix A3.

Among firm-level datasets, Orbis is unique in its coverage of the corporate ownership structure of firms. Collecting data from a variety of reporting sources, Orbis provides fine-grain information on a company’s financials. The coverage of firms within a country depends upon reporting requirements and the difficulty of accessing information. In the subset of countries analyzed in this paper (table A.2), Bureau van Dijk reports Orbis as having at least 75 percent coverage of all firms in each country as of January 2018, with the exception of Iceland, Poland, and Luxembourg where the database only captures 50-74 percent of firms, and Greece where less than 25% percentof all firms are captured. Independent verification of this coverage for past versions of the Orbis database has varied, however; in their comparison of the 2008 Orbis database to the OECD’s Structural and Demographic Business Statistics (SDBS) database, Ribeiro et al. (2010) found much poorer coverage for certain countries in Orbis, while for other countries, like the United States, there were more business records in Orbis than were reported in official figures from the SDBS.

The countries explored in our analysis are almost exclusively developed countries, with a European bias. Still, compared to other firm-level datasets, Orbis’s country coverage is more exhaustive and more current. The World Bank Enterprise Surveys, for example, have data for over 139 countries, but focus on firms in emerging economies, and updated their last full panel wave only as recently as 2002-2006. The European Central Bank’s CompNet collects data directly from central banks and national statistical agencies, but its database is limited to 17 European countries with data from 1995-2012. Like Orbis, CompNet does not provide total firm coverage for each country in each sector. The OECD created and distributed a micro-level dataset of firms in 10 countries, developed by extracting raw country-level data from a combination of business registers, enterprise surveys, social security databases, corporate tax rolls, annual industry surveys and manufacturing censuses. The World Bank supplemented this dataset with information from 14 additional countries (Bartelsman et al. 2009).

Methodology

This paper considers three methodologies for measuring industry-level productivity using firm-level data: labor productivity, TFP computed using the index method, and TFP estimated from firm-level data (using three different specification frameworks). These three methodologies are used to illustrate the tradeoffs between methodological rigor and country-sector coverage when using data from the Orbis database to measure productivity. We briefly discuss these methodologies below, with more technical details available in Gal (2013) and Biesebroeck (2007).

Labor productivity

As shown in equation 1, labor productivity, or output per worker, is simply measured by dividing the operating revenue of a firm by its number of employees in each year . Since the dataset includes the same sample of firms in all five years, labor productivity is calculated for 2014, the midpoint in the data.

${\mathrm{LaborProductivity}}_{\mathrm{it}}=\frac{{\mathrm{OperatingRevenue}}_{\mathrm{it}}}{{\mathrm{NumberofEmployees}}_{\mathrm{it}}}$

(1)

Index method

Although the simplicity of output per worker as a measure of firm productivity is appealing, it does not account for differences in other inputs across firms. As such, we need to consider measures of productivity like TFP, which accounts for the relative contribution of capital and labor to a firm’s output. Following Gal (2013), this paper uses a Cobb-Douglas production function with labor and capital inputs captured by the number of employees and the value of tangible fixed assets, respectively. Such a specification assumes perfect competition and constant returns to scale (Bernard and Jones 1996), and assumes that firms make input choices optimally (Biesebroeck 2007). Equation 2 shows the log linearization of this equation that we used to compute TFP for firm i at time t.

${\mathrm{TFP}}_{\mathrm{it}}=\mathrm{log}⁡\left({\mathrm{OperatingRevenue}}_{\mathrm{it}}\right)-\alpha \mathrm{log}⁡\left({\mathrm{NumberofEmployees}}_{\mathrm{it}}\right)-$

$\left(\mathrm{1- \alpha }\right)\mathrm{log}⁡\left({\mathrm{TangibleFixedAssets}}_{\mathrm{it}}\right)$

(2)

The coefficient shows the share of labor as an input to total operating revenue, and is based on two-digit NACE sector-level estimates of the share of labor input taken from the World KLEMS or OECD Structural Analysis Database (OECD STAN). In these computations, we assume that the rest of the value of total output comes from capital, and also that composition of capital is similar across firms. Following Gal (2013), tangible fixed assets are used to approximate capital goods in the productivity equation for this method as well as the subsequent estimation methods. As with labor productivity, we only present results for the TFP index calculated for 2014.

Index methods of calculating TFP have been found to be among the best measures for estimating productivity levels, particularly in cases when measurement error is small or there is a great deal of variation in the production technology across firms within a sector. In order to calculate a measure of productivity from observables, perfect competition in both input and output markets is generally assumed. If factor shares within a sector vary widely across countries, however, then comparisons of the TFP index are problematic, as these factor shares may be a function of technological constraints across countries. Without knowing or controlling for the level of technology, there is no way of knowing how to attribute differences in productivity to firm performance versus the capital-labor ratio of a sector, and thus there is no way of directly comparing firm-level productivities within the same sector across countries (Bernard and Jones 1996).

Estimation Approaches

While index methods provide a useful snapshot of the relationship between a firm’s current input and the efficiency of its operations, determining TFP from estimation methods allows researchers to incorporate the dynamic nature of firm decisions that are undertaken to maximize profits. Under the estimation approaches, firm profits are a function of the inputs in preceding periods, and productivity is an unobservable, firm-level characteristic, expressed as a component of the error term of a Cobb-Douglas production function.

Individual productivity can be estimated from standard OLS residuals alone as shown in equation 3a, hereafter referred to as the pooled OLS method.

${\mathrm{log \left(Revenue}}_{\mathrm{it}}\mathrm{\right)=}{\beta }_{l}\mathrm{log}⁡\left({\mathrm{NumberofEmployees}}_{\mathrm{it}}\right)+{\beta }_{k}\mathrm{log}⁡\left({\mathrm{TangibleFixedAssets}}_{\mathrm{it}}\right)+{\mu }_{\mathrm{it}}$

(3a)

The downside to using standard OLS is that the coefficients are biased upwards since productivity levels are known by the firm, but unobserved by the researcher (Biesebroeck 2007). As firm-level input choices are likely informed by firm-level productivity, the econometric relationship that results is one where the independent variables are likely correlated with the error term (Del Gatto et al. 2011).

One way to handle this issue of simultaneity is to treat productivity as a fixed, time-invariant firm characteristic θi. TFP estimates can then be obtained from a fixed-effect OLS estimation of equation 3b using either least-square dummies or first-differencing methods:

$\mathrm{log}⁡\left({\mathrm{Revenue}}_{\mathrm{it}}\right)={\beta }_{l}\mathrm{log}⁡\left({\mathrm{NumberOfEmployees}}_{\mathrm{it}}\right)+{\beta }_{k}\mathrm{log}⁡\left({\mathrm{TangibleFixedAssets}}_{\mathrm{it}}\right)+{\theta }_{\mathrm{it}}+{\mu }_{\mathrm{it}}$

where ${\mu }_{\mathrm{it}}={\mathrm{TFP}}_{i}+{\epsilon }_{\mathrm{it}}$

(3b)

However, the assumption that firm productivity does not change over time has been proven false in the extensive literature documenting the effects of technical improvements and technological innovation on firm-level productivity (see Cardarelli and Lusinyan (2015) and Heshmati and Rashidghalam (2016) for recent examples). By holding firm TFP fixed over time, the fixed effect estimation framework is not able to account for productivity shocks, like changes to regulations at the industry-level or the breakdown of machinery at the firm-level. As such, researchers who employ an estimation strategy to model firm-level productivity as time invariant will have to accept that the accuracy of their TFP estimates may not hold over longer time periods, which in turn limits the usefulness of these estimates in empirical applications.

Olley and Pakes (1996) provide a semi-parametric framework to address the simultaneity concerns in TFP estimations arising from the fact that variations in productivity are known by the firm, but unobservable in the data. They account for the unobserved productivity by treating the firm's investment behavior as a state variable in the firm's dynamic optimization problem that depends on the firm’s level of capital and productivity. Thus, the firm’s investment decisions can be used as a proxy for unobserved time-varying shocks to productivity. Following Gal (2013), we calculate firm investment by the perpetual inventory method so that capital in the current period is the sum of investment in the current period and the capital in the previous period less depreciation ${\delta }_{\mathrm{it-1}}$:

${\mathrm{Investment}}_{\mathrm{it}}={\mathrm{TangibleFixedAssets}}_{\mathrm{it}}-{\mathrm{TangibleFixedAssets}}_{\mathrm{it-1}}*\left(\mathrm{1-}{\delta }_{\mathrm{it-1}}\right)$

(4)

The Olley-Pakes method then uses a two-step procedure for estimating TFP. In the first stage, a function of investment and capital is used to control for unobserved productivity:

$\mathrm{log}⁡\left({\mathrm{Revenue}}_{\mathrm{it}}\right)={\beta }_{l}\mathrm{log}⁡\left({\mathrm{NumberofEmployees}}_{\mathrm{it}}\right)+f\left(\mathrm{log}⁡\left({\mathrm{Investment}}_{\mathrm{it}}\right),\mathrm{log}⁡\left({\mathrm{TangibleFixedAssets}}_{\mathrm{it}}\right)\right)+{\epsilon }_{\mathrm{it}}$

(5a)

Here ${\epsilon }_{\mathrm{it}}$ is the idiosyncratic error term for unexpected shocks to firm’s revenue. Since the function f() is unknown, the Olley-Pakes method uses a third or fourth order polynomial in investment and capital as an approximation during the estimation to get consistent estimates of the labor elasticity βl.

In the second stage, the estimated values of βl and residuals eit from the first stage are used to get consistent estimates of the capital elasticity βk. The Olley-Pakes method makes use of the fact that the fitted value $\mathrm{f̂}$ of the function f() is just the actual revenue minus the residuals and βl times the number of employees. We can then estimate βk from equation (5b) using non-linear squares:

$\mathrm{log}⁡\left({\mathrm{Revenue}}_{\mathrm{it}}\right)={\beta }_{l}\mathrm{log}⁡\left({\mathrm{NumberofEmployees}}_{\mathrm{it}}\right)+{\beta }_{k}\mathrm{log}⁡\left({\mathrm{TangibleFixedAssets}}_{\mathrm{it}}\right)+g\left({\mathrm{f̂}}_{\mathrm{it-1}}-{\beta }_{k}\mathrm{log}⁡\left({\mathrm{TangibleFixedAssets}}_{\mathrm{it-1}}\right)\right)+{\zeta }_{\mathrm{it}}+{\epsilon }_{\mathrm{it}}$

(5b)

Here ζit is an unexpected innovation that is uncorrelated with productivity and capital in period t. As in the first stage, the unknown function g() in equation (5b) is treated as a nonparametric term and approximated by a third or fourth order polynomial. The two stages gives us consistent estimates of both βl and βk that can be used to compute the firm’s TFP:

$\mathrm{log}⁡\left({\mathrm{TFP}}_{\mathrm{it}}\right)=\mathrm{log}⁡\left({\mathrm{Revenue}}_{\mathrm{it}}\right)-{\beta }_{l}\mathrm{log}⁡\left({\mathrm{NumberofEmployees}}_{\mathrm{it}}\right)-{\beta }_{k}\mathrm{log}⁡\left({\mathrm{TangibleFixedAssets}}_{\mathrm{it}}\right)$

(6)

Levinsohn and Petrin (2003) extend the Olley-Pakes framework by incorporating intermediate inputs, such as electricity or materials, instead of investment as proxies for unobserved time-varying shocks. Including intermediate inputs in the estimation may be preferable to investment due to the comparative smoothness with which they respond to production shocks. Further, unlike Olley-Pakes, this approach can account for periods with zero reported investment, or if the costs to capital adjustment are non-convex. Accounting for intermediate inputs also becomes important when using gross output (as we do in this paper) rather than value-added measures in the TFP estimations. The coverage of intermediate inputs in Orbis is uneven across countries, however, with firms in some countries like the United States not recording material costs as a separate account. In those instances, the Olley-Pakes method is the better option for estimating TFP using data from the Orbis database.

Table 1 summarizes the data requirements for each of the five productivity methods, and gives the number of countries for which the necessary data are available, as well as the number of country- sectors with at least 30 firms with the necessary data. Not surprisingly, labor productivity has the widest coverage across countries and sectors followed by the TFP estimations using simple OLS using pooled estimates or fixed effects. Coverage drops for estimations that require more information: labor share for the index method, investment for Olley-Pakes, and material costs for Levinson-Petrin. As discussed, the Levinson-Petrin method will not be a feasible alternative to Olley-Pakes for a number of countries in our sample and thus we only present results for Olley-Pakes in the sections that follow.

Table 1: Data requirements and coverage for productivity measures

Methods

Variables Required

(and source)

Number of years of data required Countries with necessary data Country-sectors with necessary data
Labor Productivity

Total firm revenue (Orbis)

Number of employees (Orbis)

1 49 1898
Index Method

Total firms revenue (Orbis)

Number of employees (Orbis)

Tangible fixed assets (Orbis)

Labor Share (KLEMS/OECD STAN)

1 34 1413
Estimation Methods
(a) OLS/FE

Total firms revenue (Orbis)

Number of employees (Orbis)

Tangible fixed assets (Orbis)

3 years 47 1722
(b) Olley-Pakes

Total firms revenue (Orbis)

Number of employees (Orbis)

Tangible fixed assets (Orbis)

Depreciation rate (Orbis)

3 years 43 1445
(c) Levinsohn-Petrin

Total firms revenue (Orbis)

Number of employees (Orbis)

Tangible fixed assets (Orbis)

Materials cost (Orbis)

3 years 27 1156
Source: Authors’ calculations using data from Bureau van Dijk’s Orbis database

Results

Productivity at the country and sector levels

We start our analysis by looking at the differences in the computed productivity measures across countries. Using simple averages, we summarize firm productivity at the NACE 2-digit sector-level for each country in our sample. We focus on labor productivity in this analysis since of the three types of productivity measures discussed in the paper, labor productivity provides the widest country and sector coverage. Further, labor productivity may be more easily compared across countries than more sophisticated measures such as TFP as it requires less information about a firm’s capital stock. Although using PPP-adjusted output data does help correct for cross-country differences in prices of goods and services, we note that differences in sectoral allocations within national economies still make labor productivity an imperfect measure of productivity differences across countries.

We start by identifying the most productive countries in each of the 65 NACE 2-digit sectors in our dataset. Figure 1 shows the number of times a country, based on its average labor productivity, is ranked as a top 3 country for a given sector. For brevity, figure 1 only includes countries that have been ranked as a top 3 country in at least 1 2-digit NACE sector. We find that highly-developed countries such as the United Kingdom (28 times) , Korea (24 times), and Belgium (23 times) dominate these rankings, with nearly half of all possible spots taken by these three countries. It is not surprising that these advanced economies are the most productive for a large number of sectors, although China (17 times) is quickly becoming a strong competitor in a number of sectors. By contrast, we see smaller and less advanced European countries like Hungary and Greece ranking among the most productive countries for only a single sector.

Figure 1: Number of top 3 rankings by country (labor productivity, 2014)

Source: Authors’ calculations using data from Bureau van Dijk’s Orbis database

As discussed above, differences in sectoral composition and factor prices may limit cross-country comparisons of productivity. However, we can also use our constructed productivity measures to examine differences across sectors within a country, and thus identify the top and bottom performing sectors in terms of output per worker.

Table 2 shows the top and bottom 3 manufacturing 2-digit sectors at the country-level. In the table, we only include the countries that had coverage on labor productivity in more than 8 sectors in the dataset. We see a fair degree of heterogeneity in the top 3 manufacturing sectors among countries in the sample, with food and paper products ranking among the most productive sectors in countries like Australia and Serbia that have a strong agricultural base and natural resource endowments; while basic metals, chemicals, and motor vehicles are found to be the most productive sectors in countries with strong manufacturing bases such as Italy, China, and Korea. For the bottom 3 sectors, we see less heterogeneity with wearing apparel, textiles, and furniture ranking among the least productive sectors for a number of countries.

Table 2: Top and Bottom Sectors by Country (Manufacturing)

Top 3 Manufacturing

Bottom 3 Manufacturing

AUS

food products

paper

coke/refined petroleum

repair/install of machinery

furniture

motor vehicles

AUT

repair/install of machinery

rubber/ plastics

machinery/ equipment

furniture

food products

fabricated metal products

BEL

computer, electronic

chemicals

food products

recorded media

machinery/ equipment

other non-metallic

BGR

electrical equipment

chemicals

motor vehicles

other mnf

wearing apparel

furniture

CHN

basic metals

motor vehicles

coke/refined petroleum

furniture

rubber and plastics

wearing apparel

CZE

computer, electronic

paper

motor vehicles

wearing apparel

other non-metallic

beverages

DEU

paper

chemicals

coke/refined petroleum

other mnf

furniture

wood products

ESP

pharmaceutical

chemicals

motor vehicles

furniture

recorded media

fabricated metal

EST

other mnf

fabricated metal

repair/install of machinery

wearing apparel

textiles

wood

FIN

rubber and plastics

electrical equipment

chemicals

wearing apparel

textiles

leather

FRA

pharmaceutical

beverages

other non-metallic

recorded media

other mnf

repair/install of machinery

GBR

wood products

beverages

chemicals

recorded media

furniture

fabricated metal products

HRV

electrical equipment

computer, electronic

paper

wearing apparel

furniture

leather

HUN

pharmaceutical

chemicals

computer, electronic

wearing apparel

other mnf

furniture

ITA

pharmaceutical

basic metals

coke/refined petroleum

recorded media

fabricated metal

furniture

JPN

beverages

wood

coke/refined petroleum

recorded media

repair/install of machinery

other transport equipment

KOR

basic metals

chemicals

coke/refined petroleum

other transport equipment

other mnf

recorded media

LTU

fabricated metal

wood products

rubber and plastics

wearing apparel

other non-metallic

other mnf

LVA

rubber and plastics

food products

paper

wearing apparel

textiles

furniture

MKD

paper

beverages

electrical equipment

leather

furniture

repair/install of machinery

PRT

motor vehicles

chemicals

pharmaceutical

furniture

wearing apparel

other mnf

ROU

motor vehicles

basic metals

coke/refined petroleum

wearing apparel

leather

other mnf

RUS

basic metals

motor vehicles

coke/refined petroleum

leather

recorded media

wood products

SRB

paper

food products

chemicals

wearing apparel

textiles

other mnf

SVK

other non-metallic

computer, electronic

motor vehicles

wearing apparel

food products

furniture

SVN

rubber and plastics

machinery and equipment

computer, electronic

wearing apparel

other non-metallic

furniture

SWE

machinery and equipment

other non-metallic

paper

wearing apparel

recorded media

repair/install of machinery

UKR

basic metals

electrical equipment

coke/refined petroleum

wearing apparel

other mnf

furniture

Source: Authors’ calculations using data from Bureau van Dijk’s Orbis database

Productivity distribution across countries and sectors

We next examine the difference in the distribution of productivities across manufacturing and service sectors within each country. A number of studies have documented large amounts of heterogeneity across firms in terms of their productivity and have explored the key factors behind this heterogeneity within the framework of firm behavior (Bartelsman et al. 2013). Our goal in this section is to establish whether a similar heterogeneity in productivity exists within our dataset.

Figure 2 shows the distribution of labor productivities across manufacturing (NACE codes 10-33) and services sectors (NACE codes 41-93, excluding codes 64-66 and 84). In manufacturing (top panel) countries with the largest interquartile range (the middle 50 percent of firms from that country in the sample) are the Netherlands and Ireland, while in services (bottom panel), Luxembourg and China have the largest interquartile ranges. This suggest that within manufacturing, there is more room for aggregate productivity increases in the Netherlands and Ireland as resources get reallocated from less productive to more productive firms, while Luxembourg and China have the most room for aggregate productivity increases in services. Other countries, such as Russia, have tighter productivity distributions within manufacturing and services, indicating fewer possible productivity gains from reallocation.

We next turn to TFP measures of productivity in order to control for capital inputs across countries and sectors. Figure 3 shows the dispersion in TFP computed by the index method across countries. Here we find a similar pattern as in figure 2. Ireland, the Netherlands, and Belgium have the biggest inter-quartile range for manufacturing, while Lithuania and Romania have the smallest levels of dispersion in computed TFP. In services, China and Luxembourg stand out again with very dispersed productivities.

Using estimation methods for obtaining TFP values, we observe that the levels of dispersion for manufacturing and services firms change, and are no longer consistent across methodologies. One explanation for this result could be the variance in the sample of firms across estimation methods, which may be causing changes in estimated TFP values and the level of dispersion across countries.

While figures 2 and 3 show productivity distribution at the country level, we can also compare productivity distribution across sectors for our sample countries. Figure 4 compares the distribution of labor productivity for manufacturing sectors in Germany (top panel) and Russia (bottom panel), two countries with very different levels of economic development that also are well represented in the data. In manufacturing, the distribution of productivities across two-digit NACE sectors are similar: in both countries, coke and refined petroleum products have high average output per worker and a wide distribution of firm productivities. Other sectors that see high levels of dispersion include chemicals, pharmaceuticals, and food products for Germany, and basic metals and motor vehicles for Russia. Leather and wood products have less dispersion in both countries.

Figure 5 compares the distribution of labor productivity for services sectors in Germany (top panel) and Russia (bottom panel). We see that in services, there is a fairly even distribution of labor productivity in Russia, while in Germany, water transport has by far the highest level of dispersion and average output per worker. Real estate and broadcasting are other German sectors that have relatively high levels of dispersion.

Figure 2: Labor productivity in manufacturing and services by country, 2014.

Manufacturing (NACE 10-33)

Services (NACE 41-64, 66-83, 85-99)

Source: Authors’ estimates using data from Bureau van Dijk’s Orbis database

Figure 3:TFP distribution(index method) in manufacturing and services by country, 2014.

Manufacturing (NACE 10-33)

Services (NACE 41-64, 66-83, 85-99)

Source: Authors’ estimates using data from Bureau van Dijk’s Orbis database. Contact authors for information on labor share estimates used for individual country-sectors.

Figure 4: Labor productivity by manufacturing, 2014.

Germany

Russia

Source: Authors’ estimates using data from Bureau van Dijk’s Orbis database

Figure 5: Labor productivity by services, 2014.

Germany

Russia

Source: Authors’ estimates using data from Bureau van Dijk’s Orbis database

Foreign vs Domestic firms

While previous sections of this paper demonstrated the usefulness of Orbis firm-level data for constructing and comparing sector-level measures of productivity, we now use the firm-specific characteristics to better understand differences in productivity across categories of firms within specific countries and sectors. Modern trade theory predicts that only the most productive firms will be able to make an investment to set up operations in foreign countries (Helpman et al., 2004) and so we can use our computed productivity measures to test if this holds true for the firms in our dataset. The additional benefit of examining foreign and domestic firms within a specific country, is that the foreign firms in a particular market should face the same prices for factor inputs as domestic firms, so even non-PPP adjusted financial variables are comparable. We use both a two-sample t-test of means and a Kolmogorov-Smirnov (K-S) test of distribution to compare the two types of firms across all of our estimated productivity variables. For each of the estimation methods used in this analysis, foreign firms are significantly more productive than domestic firms.

To illustrate the differences in domestic and foreign firm productivity, we compare labor productivity of foreign and domestic firms in two sectors: computer programming, consultancy, and related activities (NACE 62) and manufacture of fabricated metal products, except machinery and equipment (NACE 25). We choose to focus on these two particular sectors because of the relatively high number of countries that meet the data requirements to calculate labor productivity in these sectors (26 and 23 countries, respectively), and because these are two sectors in which Orbis provides relatively good coverage in terms of the overall employment. Overall, there are approximately 29,000 firm observations in the manufacture of fabricated metal products, and 20,000 firm observations in computer programming.

In Orbis, the name and country of origin of the global ultimate owner (GUO) of a firm is provided if that GUO controls at least 50 percent of a company observation. If the GUO is located in a different country than the firm, that firm is considered foreign. Domestic firms are those with a GUO located in the same country. Firm observations for the GUO itself, or firm observations that do not have any ownership information are excluded from this analysis. For more information on the classification of foreign and domestic firms, see appendix A2.

Table 3 provides summary statistics for the sample of firms in NACE code 25 and 62, separated by foreign and domestic firms. In both cases, the sample includes more domestic firms than foreign firms, but there are more than 1,500 foreign firm observations in each sector. In both sectors, foreign firms have higher revenues, more employees and higher-value tangible fixed assets than domestic firms.

Table 3: Summary statistics for foreign and domestic firms in NACE codes 25 and 62

Manufacturing Sector: NACE 25 - Manufacture of fabricated metal products except machinery and equipment

Number of employees

Firm type

Number of firms

Mean

SD

Mean

SD

Mean

SD

Foreign

1,791

129.8

(569.4)

$41,237 ($151,568)

$8,364 ($31,751)

Domestic

27,222

31.5

(215.9)

$8,582 ($115,443)

$2,026 ($22,974)

Services Sector: NACE 62 - Computer programming, consultancy, and related activities

Number of employees

Firm type

Number of firms

Mean

SD

Mean

SD

Mean

SD

Foreign

2,982

166.6

(665.9)

$63,109 ($503,011)

$3,723 ($31,930)

Domestic

16,884

35.8

(441.9)

$8,310 ($104,679)

$1,261 ($39,886)

Source: Authors’ calculations using data from Bureau van Dijk’s Orbis database

The difference in labor productivity between foreign and domestic firms is also apparent looking at the data at the firm and country level. Figure 6 compares the distribution of domestic and foreign firms’ labor productivities, pooled across countries in the fabricated metals sector. In that sector, domestic firms tend to be more tightly concentrated at lower productivity levels than foreign firms. In computer programming, the difference in the distribution of labor productivities between domestic and foreign firms is even more pronounced, as shown in figure 7. Again, domestic firms tend to be more tightly concentrated at lower productivity levels than foreign firms.

Figure 6: Distribution of productivity by firm ownership (Fabricated metals – NACE 25)

Note: For clarity, excludes firms in the top 95th percentile of observations in this sector.

Source: Authors’ estimates using data from Bureau van Dijk’s Orbis database

Figure 7: Distribution of labor productivity by firm ownership (Computer Programming – NACE 62)

Note: For clarity, excludes firms in the top 95th percentile of observations in this sector

Source: Authors’ estimates using data from Bureau van Dijk’s Orbis database

At the country level, average labor productivity of foreign and domestic firms also lend support to the idea that foreign firms tend to be more productive than domestic firms. In the fabricated metals sector, labor productivity is only higher for domestic firms than for foreign firms in 2 of the countries covered in the sample: Australia and Hungary. Figure 9 compares the average labor productivity in computer programming for foreign-owned and domestic companies across countries. Productivity outliers (Ireland and the Netherlands) left-skew the labor productivity distribution of output per worker. In every country but Hungary, Japan, and the United States, however, the average productivity of foreign firms exceeds the average productivity of domestic firms. This result is not surprising given the prominence of U.S. and Japanese companies in computer services.

Figure 8: Average labor productivity by country (Fabricated Metal – NACE 25)

Note: There are no foreign-owned firms in U.S, Icelandic, or Japanese fabricated metal sector in our dataset.

Source: Authors’ estimates using data from Bureau van Dijk’s Orbis database

Figure 9: Average labor productivity by country (Computer Programming – NACE 62)

Source: Authors’ estimates using data from Bureau van Dijk’s Orbis database

To test the relationship between foreign ownership and firm productivity, we use the two-sample t-test and the K-S test, which are sensitive to differences in both the mean and the shape of the distribution of the two samples. The two-sample t-test indicates whether there is a significant difference in the average labor productivity across domestic and foreign firms, while the K-S test indicates whether the distribution of the two samples is significantly different. Table 4 presents results for all of the methods for calculating and estimating firm level productivity used in this paper.

Table 4: Two-sample tests for all reporting countries in 2014

Services Sector: NACE 62 - Computer programming, consultancy, and related activities

Foreign

Domestic

Methodology

Number of countries

Mean TFP Estimate

Number of firms

Number of countries

Mean TFP Estimate

Number of firms

Difference in Means

K-S Stat

Labor Productivity

35

714.69

4,367

33

166.98

39,168

547.71***

0.25***

Index Method

26

4.8

2,982

27

4.0

16,884

0.79***

0.31***

OLS Method

32

29.53

3,344

33

15.09

18,554

14.44**

0.14**

OLS FE Method

32

5.02

3,344

33

1.82

18,554

3.2***

0.36***

Olley-Pakes Method

26

496.06

2,134

28

257.09

7,135

238.96***

0.24***

Manufacturing Sector: NACE 25 - Manufacture of fabricated metal products except machinery and equipment

Foreign Domestic

Methodology

Number of countries

Mean TFP Estimate

Number of firms

Number of countries

Mean TFP Estimate

Number of firms

Difference in Means

K-S Stat

Labor Productivity

29

339.08

2,318

31

196.15

52,893

142.92***

0.19***

Index Method

22

4.1

1,791

24

3.84

27,222

0.25***

0.18***

OLS Method

27

10.07

1,959

29

6.94

29,388

3.13***

0.22***

OLS FE Method

27

4.31

1,959

29

1.59

29,388

2.72***

0.39***

Olley-Pakes Method

22

456.35

1,297

24

251.29

13,867

205.06***

0.14***

Note: Both GUOs and firms with no ownership are excluded from calculations in the table above. For estimation-based productivity methods, we dropped results where the estimated labor or capital coefficient in the regression equation (i.e. the labor and capital elasticities within a sector) were less than zero, as that would imply that additional units of capital and labor result in diminished revenue. While in practice one can imagine a variety of circumstances in which a firm’s revenue might decrease over time despite additional units of input, this possibility is not allowed for under our assumption of perfect competition and constant returns to scale under the Cobb-Douglas production function.
Source: Authors’ estimates using data from Bureau van Dijk’s Orbis database

The higher productivity of foreign-owned firms persists, regardless of the method of estimation. The trend is common to both the services and manufacturing sectors we are exploring. Further, under all productivity estimation methods, the difference in the mean productivity between foreign and domestic firms is larger within the computer programming and consultancy sector than within the fabricated metal products sector. Overall, the finding that foreign-owned firms are, on average, more productive than domestic firms, is consistent with modern trade theory and indicates that foreign-owned firms are present in a country because they are able to compete, at least in terms of productivity, with domestic firms.

Conclusion

This paper considers the utility of firm-level data for constructing cross-country and sector measures of firm productivity. Variation in productivity estimates across the three methods considered in this analysis show that country and sector coverage in the Orbis database is contingent on choice of estimation strategy. While the Olley-Pakes method may be a more methodologically rigorous way to calculate TFP than labor productivity, TFP index, or simple OLS-based estimation methods, the need for multiple years of data and more detailed financial information decreases the firm sample size, preventing the estimation of productivity for some country-sectors in our dataset entirely.

One of the advantages of using Orbis as a source of firm-level data is the ability to distinguish between domestic and foreign-owned firms operating in particular country markets. We use this distinction to test whether foreign-owned firms have significantly different average productivities and productivity distributions. Using two-sample tests of means we find that foreign firms tend to be more productive than domestic firms on average. We also find that foreign and domestic firms have significantly different productivity distributions.

Because we are calculating productivity from cross-sections rather than focusing on growth rates, the issue of true comparability between country-sectors looms large. Factors precluding balanced comparisons of productivity measures between country-sectors stem from both the nature of the data collected and the limitations of the estimation strategies used to overcome it. Our estimates here should be viewed with these caveats in mind. Our hope for future work in this area is to build on these findings by considering specific sectors and testing the empirical relationship between productivity and international trade.

References

Bartelsman, Eric J., and Mark Doms. “Understanding Productivity: Lessons from Longitudinal Microdata.” . 38, no. 3 (Sept. 2000): 569-594. http://www.jstor.org/stable/2565420

Bartelsman, Eric J., John Haltiwanger, and Stefano Scarpetta. “Measuring and Analyzing Cross-country Differences in Firm Dynamics.” In 15-76. Edited by Timothy Dunne, J. Bradford Jensen, and Mark J. Roberts. University of Chicago Press: 2009. http://www.nber.org/chapters/c0480

Bartelsman, Eric J., John Haltiwanger, and Stefano Scarpetta. “Cross-Country Differences in Productivity: The Role of Allocation and Selection.” The American Economic Review. 103, no. 1 (Feb. 2013): 305-334. http://www.jstor.org/stable/23469644

Bernard, Andrew and Charles I. Jones. “Comparing Apples to Oranges: Productivity Convergence and Measurement Across Industries and Countries.” 86, no. 5 (Dec. 1996): 1216-1238. http://www.jstor.org/stable/2118287

Biesebroeck, Johannes Van. “Robustness of Productivity Estimates.” 55, no. 3 (Sept. 2007): 529-569. http://www.jstor.org/stable/4622397

Bureau van Dijk. “Untangling the world of private company information.” White Paper. (2017). https://www.bvdinfo.com/en-gb/contact-us/view-white-paper?White%20Paper%20Title={BF133221-3609-4A43-A375-C5F02E0DDFBE}

Cardarelli, Roberto and Lusine Lusinyan. “U.S. Total Factor Productivity Slowdown: Evidence for the U.S. States.” WP/15/116. May 2015.

Caselli, Francesco and Daniel J. Wilson. “Importing technology.” . 51, (2004): 1-32.

Del Gatto, Massimo, Adriana Di Liberto, and Carmelo Petraglia. “Measuring Productivity.” . 25, no. 5 (2011): 952-1008.

Doms, Mark and Timothy Dunne. “Capital Adjustment Patterns in Manufacturing Plants.” . 1, (1998): 409-429.

Gal, Peter N. “Measuring Total Factor Productivity at the Firm Level using OECD-Orbis.” , No. 1049, OECD Publishing,Paris. (2013).http://dx.doi.org/10.1787/5k46dsb25ls6-en

Gopinath, Gita, Sebnem Kalemli-Ozcan, Loukas Karabarbounis, and Carolina Villegas-Sanchez. “Captial Allocation and Productivity in South Europe.” 132, no. 4 (June 2017): 1915-1967.https://doi.org/10.1093/qje/qjx024

Hall, Robert E., and Charles I. Jones. “Why Do Some Countries Produce So Much More Output Per Worker Than Others?” . 114, no. 1 (Feb. 1999): 83-116. http://www.jstor.org/stable/2586948

Helpman, Elhanan, Marc Melitz, and Stephen Yeaple. “Export Versus FDI with Hetrogenous Firms.”. 94, no. 1 (Mar. 2004): 300-316.https://www.aeaweb.org/articles?id=10.1257/000282804322970814

Heshmati, Almas, and Masoomeh Rashidghlalam. “Estimation of Technical Change and TFP Growth based on Observable Technology Shifters.” GLO Discussion Paper, no. 8. Global Labor Organization, Maastricht. (2017). http://hdl.handle.net/10419/152325

Jorgenson, Dale W. “The World KLEMS Initiative: Measuring Productivity at the Industry Level.” In , 1: Ch. 20. Oxford: The Oxford University Press, Forthcoming.

Kalemli-Ozcan, Sebnem, Bent Sorenson, Carolina Villegas-Sanchez, Vadym Volosovych, and Sevcan Yesiltas. “How to construct nationally Representative Firm Level Data from the Orbis Global Database.” NBER Working Paper No. 21558. (Sept. 2015)

OECD. . https://stats.oecd.org/Index.aspx?DataSetCode=STANI4 (Accessed April 2018)

OECD. “Measurement of Aggregate and Industry-Level Productivity Growth.” In . (2001)

Olley, G. Steven, and Ariel Pakes. “The Dynamics of Productivity in the Telecommunications Equipment Industry.” . 64, no. 6 (Nov. 1996): 1263-1297.https://www.jstor.org/stable/2171831

Melitz, Marc J. “The Impact of Trade on Intra-Industry Reallocations and Aggregate Industry Productivity.” Econometrica. 71, no. 6 (Nov. 2003): 1695-1725. https://www.jstor.org/stable/1555536

Levinsohn, James, and Amil Petrin. “Estimating Production Functions Using Inputs to Control for Unobservables. . 70, no. 2 (Apr. 2003): 317-341. http://www.jstor.org/stable/3648636

Ribeiro, Samuel Pinto, Stefano Menghinello, and Koen De Backer. “The OECD Orbis Database: Responding to the Need for Firm-Level Micro-Data in the OECD.” OECD Statistics Working Papers, 2010/01. OECD Publishing, Paris. (2010) http://dx.doi.org/10.1787/5kmhds8mzj8w-en

Syverson, Chad. “Market Structure and Productivity: A Concrete Example.” . 112, no. 6 (Dec. 2004): 1181-1222. http://www.jstor.org/stable/10.1086/424743

Appendix A: Data Construction

Creating a productivity dataset from Orbis.

The dataset used in this paper covers financial information from 2012-2016 for all companies in the Orbis database with non-missing operational revenue and employment data for 2013, 2014 and 2015. Data was downloaded from Orbis’ online portal (https://Orbis4.bvdinfo.com) in September and October 2017. This portal provides 10 years of firm-level financial data, depending on the last available year for a firm’s financial data. For example, a firm that has ten years of observations in Orbis where the last available year is 2016 has data from 2006 to 2016, while a firm that has ten years of observations in Orbis where the last available year is 2017, has data from 2007-2017. The Orbis online portal also updates company data when new data becomes available rather than annually. This makes it more difficult to construct a representative time series from the Orbis online portal than through annual versions of the Orbis database released on CDs, as was the case the compilation of datasets in Kalemli-Ozcan et al (2015) and Gopinath et al (2017). Table A.1 lists the number of firms available by country in our dataset.

Table A.1 Number of firm observations in our dataset pulled from the Orbis database in November 2017

Italy

1410750

United States

8896

Sri Lanka

194

Zambia

21

Russia

616588

China

7450

Liechtenstein

192

Zimbabwe

21

Romania

474281

Poland

6949

South Africa

132

Brazil

18

Spain

376121

Bosnia and Herzegovina

6919

Philippines

105

Ghana

18

Australia

211548

Ireland

6308

Malaysia

104

Kenya

18

Finland

207900

Netherlands

4296

Jordan

100

Chile

17

Bulgaria

201768

Greece

4284

Monaco

95

Colombia

15

Ukraine

201069

Denmark

3907

Taiwan

91

Thailand

15

Portugal

194362

Hong Kong

3408

88

Angola

12

Germany

176969

Kazakhstan

2231

Nigeria

88

Albania

12

Hungary

168401

Iceland

2126

Malta

78

Kuwait

12

Sweden

138237

Belarus

1877

Mexico

74

Marshall Islands

12

Japan

127222

Montenegro

1361

Oman

49

Peru

12

Czech Republic

87660

Vietnam

746

Cuba

44

Uganda

11

Latvia

81332

Luxembourg

733

Moldova

39

Nepal

10

United Kingdom

61778

Cayman Islands

610

35

Lebanon

9

Slovakia

56043

Israel

35

United Arab Emirates

8

Serbia

51525

Switzerland

35

Ethiopia

8

Croatia

51180

Indonesia

370

Egypt

34

Gibraltar

8

Korea

49991

Turkey

365

British Virgin Islands

30

Panama

8

Macedonia, FYR

39760

Bermuda

334

Palestine

26

Azerbaijan

7

Slovenia

36768

Cyprus

281

Saudi Arabia

26

Botswana

7

Lithuania

35384

New Zealand

262

Tanzania

24

Kosovo

7

France

26055

Iran

258

Bahrain

23

Curacao

6

Estonia

22513

India

235

Costa Rica

22

Mongolia

6

Belgium

16954

Norway

232

Dominican Republic

21

Austria

12214

Pakistan

224

Singapore

21

Note: the dataset contains 14 additional observations with the ISO codes of II, which indicates an international organization such as the World Bank.
Source: Authors’ calculations using data from Bureau van Dijk’s Orbis database

The BvDID number is used as a unique identifier for firms, and firms are classified into sectors by two-digit NACE codes. The subsample of this dataset used for our analysis only includes country-sector pairs with at least 30 firm-level observations. Additionally, we limit the NACE codes covered in our analysis to those falling under NACE two-digit codes 10-33 (manufacturing) and 41-93 (services), excluding codes 64-66, which include banking and insurance activities, and 84, which covers public administration and defense. Banking and insurance are excluded not because of poor coverage, but because return on assets (rather than output per worker) tends to reflect productivity of these sectors.

These restrictions produce a sample of 49 countries (listed below) that can be used for our calculations of county-sector productivity.

Table A.2: Country observations for productivity analysis

Country

Number of sectors

Country

Number of sectors

Italy

543

Lithuania

176

Russia

462

Macedonia, FYR

165

Spain

427

Estonia

133

Germany

425

Belgium

127

Romania

362

Austria

92

Ukraine

361

China

72

Hungary

346

Poland

44

Portugal

324

Bosnia and Herzegovina

38

Bulgaria

301

Greece

34

Finland

289

Ireland

26

Japan

260

Hong Kong

22

United Kingdom

256

Netherlands

20

Latvia

254

Denmark

16

Czech Republic

249

Island

16

Sweden

242

United States

12

Korea

216

Belarus

7

Australia

214

Kazakhstan

7

Croatia

210

Montenegro

6

Slovakia

204

Israel

1

Serbia

200

Luxembourg

1

Slovenia

182

Vietnam

1

France

178

Source: Authors’ calculations using data from Bureau van Dijk’s Orbis database

Foreign Ownership

We use Orbis’ Global Ultimate Owner (GUO) variables to determine foreign ownership. A GUO owns at least 51 percent of a company, either directly or through at least 51 percent ownership of a subsidiary that owns the company. In addition to identifying the GUO, the dataset includes information on the GUO country of origin, which allows us to classify subsidiaries as either domestic or foreign-owned. Firms for which the firm country and the GUO country match are considered domestic firms, while firms for which the firm country and the GUO country do not match are considered foreign firms. One limitation of the Orbis database’s prioritization of up-to-date information over historical information is that the GUO variable only reflects the latest ownership information, so we do not know whether firms have changed ownership during our sample timeframe. Additionally, we are unable to distinguish between foreign acquisitions of companies and greenfield investment.

This methodology can be misleading in cases where large multinational companies have GUOs that are holding companies in a separate country for tax purposes. For example, because Baidu’s GUO is a holding company in the Cayman Islands, Baidu’s main operating arm would be considered foreign in China. To correct for this problem, we considered firms to be domestic if the GUO was a holding company (classified under primary NACE code 6420) located in either the Cayman Islands or Bermuda.

There are additional cases of firms where assigning a classification of foreign or domestic is less straightforward. Some firms in the sample have BvDID numbers that match the GUO ID number, indicating that these firm observations are Global Ultimate Owners. Since many of these GUOs have consolidated accounts that include their global operations, it is difficult to classify them as foreign or domestic firms, since their financial variables may reflect conditions in a market other than their headquarters market. Additionally, there are company observations with no information on the GUO. This could indicate that there are no shareholders and therefore that the firm is a domestic entity. However, this could also indicate that although the firm is majority foreign-owned, it is owned by multiple foreign entities, none of which have a total share above 51 percent. As a result, both GUOs and firms with no ownership are included in our calculations of country and sector level productivity, but excluded when comparing the productivity of foreign and domestic firms.

Finally, for about 270,000 firms, the GUO country variable is marked n.a. This indicates that the GUO is an individual, trust, or investment firm (such as a private equity firm). In these cases, the following rules were used to assign country codes to the GUO:

1. When there was only one company associated with the GUO, the company was considered a domestic firm (212,032 observations).

2. When there were multiple companies associated with the GUO but all were located in the same country, all companies associated with that GUO were considered domestic firms (57,451 observations).

3. When there were multiple companies in different countries associated with the GUO, we conducted internet searches to assign country codes to GUOs based on the headquarter location of the individual’s primary company, based on sources such as company websites, Bloomberg’s Executive profiles, Forbes Billionaires lists, and news articles. For example, while the Walton family is listed as the GUO of Walmart’s subsidiaries, since we can connect the Walton family to Walmart, we assigned the Walton family the United States as their country code. This technique was applied for 911 GUOs in our sample, and we successfully assigned country codes to 42 percent of these firms. The remaining unassigned firms were not included in our analysis.

Purchasing Power Parity (PPP) country-level deflators come from the World Bank World Development Indicators. These estimates are based on the 2011 International Comparison Program benchmark estimates in most cases, but are supplemented by annual conversion factors for 47 countries through Eurostat and OECD data.

The share of labor as an input into the gross output of a sector is obtained for each country from the World KLEMS or OECD STAN. Industry labor shares are defined as specifically as possible: at the two-digit ISIC level at its finest level of detail, or at the ISIC section or range of sections where specification at the two-digit was not provided. Out of 3,471 country-sectors, 197 are missing labor share data. Data for the estimation of labor share is from 2012 or from the next most recent complete year of data available.

Table A.3: Data sources for labor share by country and sector

Country

Year

Number of ISIC divisions (2-digit NACE) available (89 total)

Data Source (release partner)

Australia

2012

86

World KLEMS (Australian Bureau of Statistics)

Austria

2012

86

World KLEMS (EU KLEMS)

Belgium

2012

88

OECD STAN

Bulgaria

2012

78

World KLEMS (EU KLEMS)

2008

88

China

2010

79

World KLEMS (RIETI)

Cyprus

2012

76

World KLEMS (EU KLEMS)

Czech

2012

89

World KLEMS (EU KLEMS)

Germany

2012

81

World KLEMS (EU KLEMS)

Denmark

2012

89

World KLEMS (EU KLEMS)

Spain

2012

89

World KLEMS (EU KLEMS)

Estonia

2012

89

World KLEMS (EU KLEMS)

Finland

2012

89

World KLEMS (EU KLEMS)

France

2012

81

World KLEMS (EU KLEMS)

United Kingdom

2012

86

World KLEMS (EU KLEMS)

Greece

2012

89

World KLEMS (EU KLEMS)

Croatia

2012

83

World KLEMS (EU KLEMS)

Hungary

2012

89

World KLEMS (EU KLEMS)

India

2012

88

World KLEMS (Reserve Bank of India)

Ireland

2011

88

OECD STAN

Iceland

2012

78

OECD STAN

Israel

2012

62

OECD STAN

Italy

2012

89

World KLEMS (EU KLEMS)

Japan

2009

86

World KLEMS (RIETI)

Korea (the Republic of)

2012

88

World KLEMS (Korea Productivity Center)

Lithuania

2011

88

OECD STAN

Luxembourg

2012

84

World KLEMS (EU KLEMS)

Latvia

2012

51

World KLEMS (EU KLEMS)

Netherlands (the)

2012

83

World KLEMS (EU KLEMS)

Norway

2012

88

OECD STAN

Poland

2012

79

World KLEMS (EU KLEMS)

Portugal

2012

86

World KLEMS (EU KLEMS)

Romania

2012

86

World KLEMS (EU KLEMS)

Russian Federation (the)

2012

86

World KLEMS (GGDC and HSE)

Slovakia

2012

89

World KLEMS (EU KLEMS)

Slovenia

2012

83

World KLEMS (EU KLEMS)

Sweden

2012

84

World KLEMS (EU KLEMS)

Taiwan

2010

84

World KLEMS (Asia KLEMS)

United States

2012

89

World KLEMS (EU KLEMS)

Descriptives

Table A.4. Descriptives for all firms in NACE sectors 25 and 62 in 2014

Manufacturing Sector: NACE 25 - Manufacture of fabricated metal products except machinery and equipment

Number of employees

Firm type

Number of firms

Mean

Median

Max

Min

Mean

Median

Max

Min

Mean

Median

Max

Min

Foreign

1,791

129.8

47

19,850

1

$41,237$9,259

$3,749,334$2

$8,364$1,585

$2,546,534$1

Domestic

27,222

31.5

10

14,187

1

$8,582$1,224

$8,901,959$2

$2,026$193

$943,201$1

Services Sector: NACE 62 - Computer programming, consultancy, and related activities

Number of employees

Firm type

Number of firms

Mean

Median

Max

Min

Mean

Median

Max

Min

Mean

Median

Max

Min

Foreign

2,982

166.6

33

15,516

1

$63,109$6,829

$22,304,144$2

$3,723$116

$1,227,034$1

Domestic

16,884

35.8

4

43,726

1

$8,310$441

$7,753,017$1

$1,261$25

$3,452,168$1

*sample includes all firms with adequate data to perform the TFP Index Method calculation.

Appendix B: Supplemental tables and figures

Figure B.1: Labor shares for Select Industries

Manufacture of fabricated metal products, except machinery and equipment (NACE 25)

Computer Programming, consultancy and related activities (NACE 62)

Source: Authors’ calculations from KLEMS and OECD STAN databases.

Table B.1: Count of Sectors in Each Estimation Method, by country

Country Code The Labor Productivity Sample (2014) The Index Method Sample (2014) The OLS Sample (at least 3 years) The Olley-Pakes Sample (at least 3 years) The Levinsohn-Petrin Sample (at least 3 years)
AUS

214

3

3

3

AUT

92

75

76

22

3

BEL

127

124

126

125

104

BGR

301

166

292

230

265

BIH

38

38

32

36

BLR

7

CHN

72

68

68

20

CYM

1

1

CZE

249

130

161

117

156

DEU

425

235

333

126

109

DNK

16

14

18

14

ESP

427

422

425

415

419

EST

133

98

110

93

95

FIN

289

177

186

173

163

FRA

178

106

178

169

141

GBR

256

244

247

245

GRC

34

33

34

22

HKG

22

HRV

210

194

210

191

210

HUN

346

328

334

330

85

IRL

26

22

22

20

ISL

16

10

11

10

ISR

1

1

1

1

ITA

543

510

512

510

508

JPN

260

258

261

46

1

KAZ

7

7

KOR

216

214

214

201

197

LTU

176

42

34

LUX

1

1

1

1

LVA

254

77

265

8

8

MKD

165

165

129

150

MNE

6

7

3

4

NLD

20

16

18

1

POL

44

28

29

23

23

PRT

324

317

318

309

279

ROU

362

339

388

352

368

RUS

462

385

462

SRB

200

219

185

197

SVK

204

189

204

188

203

SVN

182

170

177

175

175

SWE

242

7

10

8

4

UKR

361

360

91

87

USA

12

12

12

VNM

1

1

Total

7521

5003

6538

4601

3990

Source: Authors’ estimates using data from Bureau van Dijk’s Orbis database

Table B.2: Four digit NACE codes contained in two-digit code 25

25 Manufacture of fabricated metal products, except machinery and equipment
25.11 Manufacture of metal structures and parts of structures
25.12 Manufacture of doors and windows of metal
25.21 Manufacture of central heating radiators and boilers
25.29 Manufacture of other tanks, reservoirs and containers of metal
25.30 Manufacture of steam generators, except central heating hot water boilers
25.40 Manufacture of weapons and ammunition
25.50 Forging, pressing, stamping and roll-forming of metal; powder metallurgy
25.61 Treatment and coating of metals
25.62 Machining
25.71 Manufacture of cutlery
25.72 Manufacture of locks and hinges
25.73 Manufacture of tools
25.91 Manufacture of steel drums and similar containers
25.92 Manufacture of light metal packaging
25.93 Manufacture of wire products, chain and springs
25.94 Manufacture of fasteners and screw machine products
25.99 Manufacture of other fabricated metal products n.e.c.

Table B.3: Four digit NACE codes contained in two-digit code and 62

62 Computer programming, consultancy and related activities

62.01 Computer programming activities
62.02 Computer consultancy activities

Figure B.2: Dispersion of pooled OLS productivity estimates in Manufacturing and Services by county, 2012-2016.

Manufacturing (NACE 10-33)

Services (NACE 41-64, 66-83, 85-99)

Source: Authors’ estimates using data from Bureau van Dijk’s Orbis database

Figure B.3: Dispersion of OLS fixed effects productivity estimates in Manufacturing and Services by county, 2012-2016.

Manufacturing (NACE 10-33)

Services (NACE 41-64, 66-83, 85-99)

Source: Authors’ estimates using data from Bureau van Dijk’s Orbis database

Figure B.4: Dispersion of Olley-Pakes productivity estimates in Manufacturing and Services by county, 2012-2016.

Manufacturing (NACE 10-33)

Services (NACE 41-64, 66-83, 85-99)