Development and the Labor Share

Abstract

There is a vast literature, both theoretical and empirical, linking income inequality to economic development, starting with Kuznets (1955) and examining a potential inverted U-shaped pattern of inequality along the development path. In this paper, an original perspective concerning this debate is adopted by focusing on the labor share of income; that is, the ratio of wage bill to value added. The example of China is illustrative. Since the beginning of the economic boom, the increase of wages has not been as high as the spectacular increase of productivity. This has reinforced competitiveness in the industry. Conversely, during the past few years, wages have increased substantially more than productivity. For example, according to the Chinese National Bureau of Statistics, Chinese private sector wages rose by 14 percent in 2012 as the result of the labor market. As shown in this paper, this pattern may be inherent to the development process and is not specific to the China case. Panel data with total wage bill and value added are used at the manufacturing sector level (hereafter “aggregate level”) and also in the 28 manufacturing subsectors. The labor share follows a U-shaped relationship at the aggregate manufacturing level and also within the 28 manufacturing subsectors. This suggests that the wedge between productivity and wages observed at the aggregate manufacturing level is only partially the result of composition effects that are likely to occur during the development process.

These empirical findings are then rationalized with the help of a simple model that illustrates our stylized fact. It is now established that dualism in the organization of production activities is pervasive in developing countries (DCs) where traditional, low-productivity methods of production coexist with modern, higher-productivity methods. This is at the heart of many models in development economics (see Banerjee and Newman [1993], for instance), where agents making occupational choices face constraints to enter the modern, high-productivity sector. We show that it is quite natural for the labor share to exhibit such a U-shaped pattern in those classes of model characterized by firm heterogeneity.

Traditionally, economists have devoted much attention to the pattern of personal income inequality despite the recent interest in the issue of factor income distribution. We see at least two reasons why focusing on the labor share of income along the development path is important. First, if wage increases do not compensate for productivity gains, this could affect competitiveness and net exports. Second, development is often seen as a necessary condition for people's well-being. But we also have known since Kuznets (1955) that the first stage of development may be associated with an increase in inequality. In this paper we focus on the factor distribution of income, which may be a key component of personal income inequality since capital distribution is more concentrated than wage distribution.¹ Most of the papers in the literature on inequality have focused on wage or income inequality, but little attention has been paid to factor income distribution. The main reason for the neglect of the labor share relies on Kaldor’s stylized fact (1956) in favor of constant labor shares across time and space, in spite of Solow’s skepticism (1958). This fact—mainly inspired by the US experience—is not supported by recent empirical studies. The labor share has sharply decreased in many European countries since the 1980s but also in the United States more recently (see Karabarbounis and Neiman 2014) as well as in developing countries (Decreuse and Maarek [2015] or Harrison [2002]). In addition, the labor share remains substantially higher in developed countries than in developing ones (see Rodríguez and Ortega [2006], for instance).

This article offers several contributions. First, it exploits the panel dimension of the labor share for developing countries at the aggregate manufacturing level and at the subsector manufacturing level. To do so, a specific dataset from The United Nations Industrial Development Organization (UNIDO)—a subdivision of the UN—is used, which allows us to have reliable data on the labor share for developing countries at the disaggregated level. This dataset allows us to study the evolution within countries at the aggregate manufacturing level but also within the 28 manufacturing subsectors. To our knowledge, it is the first time that such a dataset has been exploited to study the impact of development on the labor share, which is not possible with other datasets on the labor share because of very poor coverage for developing countries. A clear U-shaped pattern of the labor share is identified at the aggregate manufacturing level and also at a more disaggregated level. Data also allow us to discriminate between several potential causes affecting the labor share that are generally identified in the literature. Indeed, the fact that the U-shaped pattern is observed at the aggregate manufacturing level but also within the 28 manufacturing subsectors suggests that this pattern is not simply related to the structural changes that occur during the development process, and which could affect the aggregate labor share through a simple composition effect across manufacturing subsectors. Of course, composition effects which impact the labor share are very likely to occur at a more aggregated level; that is, for the whole economy, but this dataset does not capture them.

This paper also provides some evidence on the channel through which development is likely to impact the labor share. Indeed, changes in the labor share can reflect various mechanisms. In a competitive world where technology is not Cobb-Douglas, changes in the labor share are exclusively related to capital accumulation and hence reflect changes in the competitive price of capital (Karabarbounis and Neiman 2014). In an imperfectly competitive world, the capital share is partially made of super-profits. Generally, the empirical literature on the labor share makes no difference in the nature of the capital share, considering that it is simply the complement of the labor share (see Karabarbounis and Neiman [2014] or Blanchard [1997], to cite the most famous examples).

Since estimations are controlled for factor accumulation that is likely to impact the labor share through the competitive channel and since it does not significantly affect the U-shaped relationship between development and the labor share, the observed U-shaped pattern of the labor share during the development process is likely to originate in an alternative channel. Any change in the wage determination process that modifies the wedge between wages and labor productivity impacts the labor share through the super-profit component. However, one cannot claim that factor accumulation never plays a role. First, by focusing only on the manufacturing sector, one cannot observe reallocations across sectors of the whole economy that happen when capital accumulates (Acemoglu and Guerrieri 2008). Second, and more importantly, estimations show that capital accumulation indeed impacts the manufacturing labor share, in line with Bentolila and Saint-Paul (2003) or Karabarbounis and Neiman (2014) for OECD countries. Nevertheless, despite a significant and robust effect, the capital stock only marginally affects the U-shape. In other words, the capital accumulation channel is valid but not sufficient to explain the observed pattern of the labor share.

It is not a controversial idea in the development economics literature that labor market is dual. For instance, Lewis (1954) argues that labor supply is unlimited in the modern sector of the economy where firms have higher productivity than in the traditional one, generating relatively low wages with respect to productivity in this sector. La Porta and Shleifer (2008) provide strong evidence in favor of this dual view of economic development.

Such a duality generates a U-shaped relationship between development and the labor share under plausible assumptions. The intuition is the following: When the proportion of modern firms is insufficient to absorb labor supply, wages remain stuck at a low level, reflecting the low productivity in the traditional sector. As the proportion of modern firms to old firms increases, so does value added in the economy, but since wages remain low the labor share decreases. This “Lewis effect” implies that the labor share decreases at the first stages of development. Once the proportion of traditional firms is sufficiently low and the proportion of modern firms is sufficiently high to absorb the labor supply, wages jump to a higher level. The labor share increases as a result, generating our U-shaped relationship. Actually, we argue that any workhorse model of development economics exhibiting duality in the labor market can quite naturally generate this U-shaped relationship. Here, it is worth mentioning that the contribution of this paper is not theoretical. The model is purely illustrative and is used to derive factor shares in a dual economic environment. The main contribution of the paper is empirical and documents a clear U-shaped pattern of the labor share at the aggregate manufacturing level. The commonly accepted dual view in development economics generates such a pattern quite naturally, as illustrated theoretically in this paper. Unfortunately, given the absence of panel data at the firm level for developing countries, it is not possible to directly test the mechanisms of the model; that is, the ability of the modern sector to extract rents along the development path. Nevertheless, some cross-sectional statistics at the firm level from the World Bank Entreprise Survey (2017) highlight that the ratio of employment in big and high-productivity firms over total employment increases sharply with GDP per capita.

This paper relates to different strands of the literature. First, it belongs to the growing literature on the labor share, as emphasized by the contributions of Bentolila and Saint-Paul (2003), Blanchard and Giavazzi (2003), Acemoglu and Guerrieri (2008), Maarek and Orgiazzi (2013), Bazillier and Najman (2017), Decreuse and Maarek (2015), Rodrik (1997), Karabarbounis and Neiman (2014), Bridgman (2014), and Piketty (2013). None of these authors focuses on the role of the economic development in developing countries. However, a notable exception is Gollin (2002), who argues that after correcting the labor share data for self-employment income, the link between the labor share and development disappears. The corrected data of Gollin do not allows to study the evolution of the labor share over time in developing countries. When using data that allows for time comparisons as it is done in this paper, a strong correlation between development and the labor share is observed. Indeed, by correcting the labor share for self-employment income, Gollin is able to compare the labor shares for a dozen developing economies at one point in time only. This dataset on the manufacturing sector does not require such a correction since the self-employed are excluded from the sample by UNIDO, and this permits us to run regressions on more than 1500 observations (see below). The debate on the evolution of the factor shares of income has become more intense during the past couple of years with the two important publications of Piketty (2013) and Karabarbounis and Neiman (2014) that stress the recent decrease of the labor share in most OECD countries. This paper does not directly contribute to this debate since it deals with developing countries. In the illustrative model presented in this paper, once the whole economy has adopted modern production techniques, the labor share remains constant. Thus, the recent decreasing trend in the labor share observed in mature economies is not likely to be related to the process of economic development described in the paper. Second, the paper is related to the literature on inequality along the development path. This literature started with Kuznets (1955), who argues that inequality increases during the first stages of development as the population shifts from the agricultural sector to the urban sector. During later stages of development, this force is more than offset by the fact that within the urban sector income becomes more equally distributed. This idea was formalized later on by Robinson (1976), Knight (1976), and Fields (1979), who argued that the rural-urban income differential is constant and equally distributed, but that the share of the population in the agricultural sector changes with development, producing the familiar inverted U-shape for evolution of income inequality over time. Nevertheless, Bourguignon and Morrisson (1998) review the empirical evidence of such an inverted U-shaped relationship and show that it is not robust. This paper adopts a quite different perspective by focusing on the labor share, which is one possible component of personal income inequality but not the only one. Actually, the link between the factor shares and income inequality is not so simple and relies on other factors, as shown by Karabarbounis et al. (2014) in a recent working paper, or by García-Peñalosa and Orgiazzi (2013). More precisely, a decomposition of income inequality index by factor components shows that the inequality index is impacted by changes in the factor shares but also by changes in factor income inequality and by the correlation between factor income and total income.² All these components can evolve with economic development, possibly in opposite directions, and this may explain why Bourguignon and Morrison (1998) review that the link between development and income inequality is not clear in the empirical literature. One should look at this relationship in more detail, looking at some specific channels in order to understand more precisely the mechanisms at work. This is the main goal of this paper in looking at the labor share of income. Actually, the very basic question of this paper is just “how factors’ rewards change when developing economies grow.” By answering this question, we highlight an important but specific channel of changes in income inequality.

1. Empirical Evidence

The empirical analysis consists of estimating a reduced form equation on sectoral panel data. The dependent variable is the labor share, and the regressor of interest is the logarithm of gross domestic product per capita (variable GDP hereafter), which captures the level of development.

Empirical Strategy

Two terms appear. The first one represents the within-subsector effect and equals the sum of the variations of the labor share within each subsector, weighted by the initial subsector share. The second term corresponds to what we call the “composition effect” and equals the sum of the variations of the shares of each subsector in the economy, weighted by the final value of the labor share. This term captures the extent to which the variation in the aggregate manufacturing labor share is due to changes in the structure of the manufacturing sector. Descriptive statistics show that the standard deviation within subsectors of the manufacturing sectors is 42 and that the between-subsector standard deviation is 48. Therefore, the aggregate variance of the labor share is due to both within- and across-subsectors variability.

In order to disentangle between the within and the composition effect, estimations on subsectoral data are performed by controlling for crossed fixed effects: country-subsector fixed effects a_is. These regressions are weighted by the average share of the subsectors over the period. Thus, the estimated coefficients correspond to the within-subsector effect of development on the aggregate manufacturing labor share. Doing that, variations of the labor share in small subsectors of manufacturing do not contribute much to the variation of the aggregate manufacturing labor share and to the estimated coefficients of GDP per capita. This allows us to capture the within-subsector variations of the labor share due to development. Moreover, subsector-year fixed effects a_ts capture symmetric shocks that occur in some specific subsectors.

Let’s now discuss in detail the variables that are included in the regressions to test for specific channels through which development could affect the labor share. First, it is important to have in mind that changes in the capital (labor) share can reflect various mechanisms at work in an economy that develops. In a competitive world, if the technology is not Cobb-Douglas, the elasticity between labor and capital can be different from one, and in this case the level of the capital (labor) share is exclusively related to capital accumulation. In this case, movements in the labor share are related to changes in the competitive price of capital. Conversely, in a world in which markets are not competitive, the capital share can also be made of super-profits. In order to disentangle between those two mechanisms, it is necessary to control for capital accumulation. If the GDP per capita affects the labor share only through the factor accumulation channel, the inclusion of such a variable should capture the entire effect of development on factor shares. In the same spirit, the accumulation of human capital is also controlled by including the average year of schooling of the population (School).³ Second, it is necessary to control for globalization. Indeed, various studies have shown that those variables are negatively correlated to the labor share (see Rodrik [1997]; Harrison [2002]; Jayadev [2007]; Francisco y Daniel Ortega [2001]; or Guscina [2006]) and are very likely to evolve with development.⁴ Therefore, omitting openness variables would imply that the GDP variable is potentially capturing the effects of globalization on the labor share of income. Globalization is controlled by including in the set of regressors a trade openness OPENT variable (the ratio of imports and exports over GDP) and a capital openness OPENK index (a de jure index). Finally, there are some other channels related to development that could affect the labor share. First of all, migration from rural to urban areas may affect the labor share, as it affects the labor supply in cities and thus the bargaining power of workers (if there is an excess supply, for instance). Hence, estimations are controlled for the share of urban population (Urban) in some regressions. Moreover, economic development can impact the labor share through demographic factors. The ratio of population between the ages 15 to 64 (Work.age) is included in some regressions since this ratio is related to the proportion of youths, and the latter evolves sharply with development. This may affect the labor share since youths may have a lower bargaining power, for instance. Moreover, estimations are controlled for the government size (proxied by the government consumption share of GDP, G/Y, available for many developing countries), which increases with development (development of state capacity). Indeed, government provides many services to the population, such as unemployment insurance or social security, which may reinforce the bargaining power of workers. Unfortunately it is not possible to control separately for all those channels, so we assume that for social protection programs to citizens grow as the size of government increases. Finally, estimations are controlled for the ability of workers to organize into unions. Those rights to organize and join a union appear with development and affect the bargaining power of workers. Since the unionization rate is not available for developing countries, we use an index for freedom of assembly and association that should be related to the ability of workers to organize into unions (Asso).

Data

The labor share is computed from UNIDO data. These data cover 180 countries over the period 1963–2003.⁵ However, the wage bills are often missing in several developing countries, so there is a total of 106 countries (28 developed and 78 developing) where the labor shares are available. The UNIDO data provide various variables at the aggregate manufacturing level, as well as at the 3-digit level for 28 subsectors.⁶

The labor share is defined as the ratio of wages and salaries over value added.⁷ As argued by Gollin (2002), this definition implies that all the income of the self-employed is treated as capital income, which underestimates the labor share. This is particularly problematic in this study because it could bias the impact of development. Indeed, during development the share of population being self-employed declines, which mechanically implies that the overall wage bill increases. Therefore, using data where the value added contains self-employment income and defining the labor share naively as the ratio of the wage bill over value added is likely to lead to mistakenly interpreting an increase in the labor share as a change in the “real” sharing of value added. On the contrary, data from UNIDO allow us to avoid this problem. Indeed, the surveys sent by UNIDO are designed to collect data only in the corporate manufacturing sector, specifying a cutoff point below which economic activity is not measured.⁸ Consequently, this selection removes to a large extent the problem of self-employment.

One could have chosen to use as the main dataset one that includes the self-employed like the UN data, and would have adjusted the labor share for self-employment income, as suggested by Gollin (2002) in seminal contribution. However, there would have been major problems. First, self-employment income is available for very few developing countries, as detailed by Trapp (2015). Second, the availability is restricted to very few years, which does not allow for time comparisons (see Trapp 2015). For instance, Gollin is able to correct the naive definition of the labor share by taking into account self-employment income for only 12 developing countries at one time point only. Third, there are several competing methods to correct for self-employment income, which are not totally satisfying, which lead to different measures (sometimes aberrant), and which accentuate the variance of the labor share.

Finally, UNIDO data are available at a disaggregated level for a larger panel of developing countries, and for a longer period than any other data on developing countries. Indeed, the labor share can be observed for 83 developing countries, which corresponds to 19 observations per country on average at the aggregate manufacturing level (1578 observations). Using the UN data would have allowed us to work on 29 developing countries only, with a total number of 195 observations, meaning a time coverage of one-eighth of the UNIDO data.

Of course this approach has major drawbacks. First, it leads to examining the effect of development on the manufacturing sector only, and not on the whole economy. Hence neither reallocations across the sectors at a more aggregated level, nor the evolution of the labor shares in primary and tertiary sectors, are observable. Second, self-employment represents a large share of the population in the developing world that we do not observe. Nevertheless, as stressed above, using another database including self-employment income would have required correcting the naive definition of the labor share, which is, once again, at best prone to measurement inconsistencies, at worst not possible for many countries, and not possible in order to have a time dimension. Third, UNIDO data are based on surveys and the sample of firms may change over time. As a result, the characteristics of surveyed firms such as their size may change, which could affect the aggregate labor share. However, this does not systematically bias the results. Indeed, the number of surveyed firms represents a substantial fraction of total activity. Therefore, due to the law of large numbers, firms’ characteristics should remain stable and representative of the whole manufacturing sector if the firms’ selection process is random. Moreover, even if firms’ characteristics do not remain equal from one year to the other because of sample selection, there is no reason to think that a change in firms’ representativity in the sample would be related to the development process.

Another problem with the UN National Accounts Data is that it includes the nonprofit sector, whose value added is often computed by considering that it corresponds to labor costs, inducing a labor share of 100 percent for such activities. Once again, as these activities develop during the development process, they imply an increase in the labor share, which does not reflect the forces behind the determination of the labor share of the value added. Of course one can subtract this sector from the overall value added, but this would imply a very sharp decrease in the already low availability of these UN data.

In conclusion, this is in our mind the only dataset that has a sufficient coverage for developing countries to allow us to observe variations of the labor share along the development path. For example, in the UN dataset it is not possible to perform estimations of the model described in equation (1) on the two lowest-income groups of countries because of lack of data, even at the aggregate level.

Note that one of the problems faced with UNIDO data is that the way in which the manufacturing sector is disaggregated in subsectors can change over time and countries, which is problematic when focusing on subsectors. For instance, in Bolivia in 1981 the manufacturing sector is disaggregated in 28 subsectors, whereas it is only disaggregated in 27 subsectors in 1979 and in 26 in 1976. Hence, when focusing on subsectors in disaggregated data, only years with the same sectoral structure have been retained, selecting those that cover the maximum number of observations. Doing this prevents us from mistakenly interpreting a change in the labor share within sectors as the reflection of a change in the way the manufacturing sector is disaggregated in the data.

The variable used for computing the logarithm of the GDP per capita is the GDP per capita measured at constant 2011 prices from the Penn World Table (hereafter PWT), see Zeileis (2017) and Feenstra et al. (2015). It has impressive coverage. Since the regressions include country fixed effects, it is not a problem that GDP series are not measured at PPP. Indeed, the coefficients of interest are identified using variations within countries. However, as robustness checks, some regressions using the GDP per capita measured at PPP 2005 from the World Development Indicators from the World Bank (2013) and some others using the GDP per capita at PPP from the Maddison project are presented in appendix A1. Note that these later series of GDP per capita have a much lower coverage than ones from the Penn World Table.

The variable used in order to control for capital accumulation is the capital stock of the whole economy from the Penn World Table. The main advantage of the PWT is impressive coverage. Combined with the UNIDO dataset on the labor share, this allows us to perform regressions with many more observations than the existing literature on developing countries. The main drawback is that it does not correspond exactly to capital accumulation in the manufacturing sector. Nevertheless, it is a good proxy since the manufacturing sector is the more capital-intensive sector in the economy, especially in developing countries. Therefore, most of the capital accumulation should be located in the manufacturing sector along the development path. On the contrary, the service sector, which expands during the development process, has a low capital intensity. An alternative option would have been to use the investment output ratio in the manufacturing sector from the UNIDO dataset. The main advantage of this control variable is that it is available at the subsector level, allowing us to control for capital accumulation at the subsectoral level. Moreover, the investment output ratio can also capture some factor bias in technological change: For instance, in Karabarbounis and Neiman (2014), technological changes result in a decrease in the price of the investment good, which enhances investment. Most of the technological changes that are not country specific should be captured by time dummies or sector-time dummies when performing estimations at the subsector level. However, this reduces the sample size, and moreover this variable does not exactly correspond to the capital stock.⁹ Note that in other versions of the estimated model, the investment output ratio in the manufacturing sector from UNIDO data is used in order to control for capital accumulation. In order to control for human capital accumulation, the average years of schooling in the total population aged 25 and more is added in some regressions (see Barro and Lee 2001). The proxy used for trade openness is the ratio of import plus export to GDP from the World Development Indicators. The measure for financial openness is the de jure index of Chinn and Ito (2006), which captures how policies are restrictive toward capital flows. It is a continuous composite index available from 1960 to 2006 for more than 200 countries. The share of urban population and the share of population between the ages of 15 to 64 come from the World Development Indicators of the World Bank. The ratio of government consumption to GDP is from the Penn World Table, and the freedom of assembly and association index is from the Cingranelli-Richards (CIRI) Human Rights Dataset and goes from 0 (not free) to 2 (totally free).

Econometric Analysis

The first specification, equation (1), regresses the labor share on the level of the logarithm of GDP per capita (GDP variable) and on its square for 78 developing countries.

It is first estimated at the aggregate manufacturing level, and the results are reported in table 1 when using (K/Y)_it, the capital-output ratio of the whole economy, as a proxy for capital accumulation. Table 2 displays the results when (I/Y)_it, the investment-output ratio in the manufacturing sector, is used as a proxy for capital accumulation. The first four columns correspond respectively to specifications without any controls and without fixed effects (1), with country fixed effects only (2), with time dummies only (3), and with both (4). All the other regressions include both fixed effects. In table 2, columns labeled with a B refer to the alternative regressions that are performed at the subsector level in order to capture the within-subsector effect. This allows us to compare the results with the estimations at the aggregate manufacturing level. Since the only control for capital accumulation available at the subsector level is the investment-output ratio, results for regressions at the subsector level are reported in table 2 only. As many observations are lost by controlling for this variable, results with and without controlling for it are provided. Estimations are performed controlling or not controlling for it, on the same subsample (see columns 5 and 6, respectively). Results displayed in column 5 allow us to understand if the variation of the observed coefficient for GDP reflects a reduction of the size of the sample or the inclusion of the control for human capital. The other control variables are included progressively.

Whichever the specification of the model presented in equation (1), both the coefficient of GDP and of its square are significant, the former being negative and the latter being positive. This suggests a U-shaped relationship between development and the labor share.

Figures 1, 2, and 3 plot this relationship as partial correlations between the labor share and GDP obtained from the estimates. In figure 1(a) the vertical axis is |$LS_{it}-\widehat{a_{i}}$|⁠, where |$\widehat{a_{i}}$| are the estimated coefficients of country fixed effects from the regression presented in column 2 of table 1. In figure 1(b) the vertical axis is |$LS_{it}-\widehat{a_{i}}-\widehat{a_{t}}$|⁠, where |$\widehat{a_{i}}$| are the estimated coefficients of country fixed effects and |$\widehat{a_{t}}$| are the estimated coefficients of time fixed effects from the regression presented in column 4 of table 1. In figure 2(a) the vertical axis is |$LS_{it}-\widehat{a_{i}}$|⁠, where |$\widehat{a_{i}}$| are the estimated coefficients of country fixed effects from the regression presented in column 6 of table 1. In figure 2(b) the vertical axis is |$LS_{it}-\widehat{a_{i}}-\widehat{a_{t}}-\widehat{\beta }(K/Y)_{it}$|⁠, where |$\widehat{a_{i}}$| are the estimated coefficients of country fixed effects, |$\widehat{a_{t}}$| are the estimated coefficients of time fixed effects, and |$\widehat{\beta }$| is the estimated coefficient of (K/Y)_it from the regression presented in column 6 of table 1.

Figure 1.

Partial correlations between the labor share and GDP per capita (in logarithm), estimations from columns 2 and 4, table 1 or 2

Source: Authors' estimates from UNIDO data and Penn World Tables (2015).

Note: In figure 1(a) the vertical axis displays |$LS_{it}-\widehat{a_{i}}$| where |$\widehat{a_{i}}$| are the estimated coefficients of country fixed effects from the regression presented in column 2 of table 1 or 2. In figure 1(b) the vertical axis is |$LS_{it}-\widehat{a_{i}}-\widehat{a_{t}}$|⁠, where |$\widehat{a_{i}}$| are the estimated coefficients of country fixed effects and |$\widehat{a_{t}}$| are the estimated coefficients of time fixed effects from the regression presented in column 4 of table 1.

Open in new tabDownload slide

Figure 2.

Partial correlations between the labor share and GDP per capita (in logarithm), estimations from columns 6, table 1

Source: Authors' estimates from UNIDO data and Penn World Tables (2015).

Note: Figure 2 displays results from the regression presented in column 6 of table 1. In figure 2(a) the vertical axis is |$LS_{it}-\widehat{a_{i}}$|⁠, where |$\widehat{a_{i}}$| are the estimated coefficients of country fixed effects from this regression. In figure 2(b) the vertical axis is |$LS_{it}-\widehat{a_{i}}-\widehat{a_{t}} - \hat{\beta}(K/Y)_{it}$| where |$\hat{\beta}$| is the estimated coefficient of |$(K/Y)_{it}$| from this regression.

Open in new tabDownload slide

Figure 3.

Partial correlations between the labor share and GDP per capita (in logarithm), estimations from columns 6, table 2

Source: Authors' estimates from UNIDO data and Penn World Tables (2015).

Note: Figure 3 displays results from the regression presented in column 6 of table 2. In figure 3(a) the vertical axis is |$LS_{it}-\widehat{a_{i}}$|⁠, where |$\widehat{a_{i}}$| are the estimated coefficients of country fixed effects from this regression. In figure 3(b) the vertical axis is |$LS_{it}-\widehat{a_{i}}-\widehat{a_{t}} - \hat{\beta}(I/Y)_{it}$| where |$\hat{\beta}$| is the estimated coefficient of |$(I/Y)_{it}$| from this regression.

Open in new tabDownload slide

In figure 3(a) the vertical axis is |$LS_{it}-\widehat{a_{i}}$|⁠, where |$\widehat{a_{i}}$| are the estimated coefficients of country fixed effects from the regression presented in column 6 of table 2. In figure 3(b) the vertical axis is |$LS_{it}-\widehat{a_{i}}-\widehat{a_{t}}-\widehat{\beta }(I/Y)_{it}$| from the regression presented in column 6 of table 2.

In appendix A1, alternative versions of these relationships are presented where the estimated coefficients of country and time fixed effects are subtracted from the labor share (⁠|$LS_{it}-\widehat{a_{i}}-\widehat{a_{t}}$|⁠), or alternatively only the estimated time fixed effects (⁠|$LS_{it}-\widehat{a_{t}}$|⁠).

Not only is the relationship non-monotonous, but it is also U-shaped. Moreover, there are a lot of observations in the decreasing part of the curve as well as in the increasing part. The turning point above which development has a positive impact on the labor share is around |${\$}$|2200 per capita (see column 4 of table 1).

Necessary and sufficient conditions for this U-shaped relationship have been checked by using the test of Lind and Mehlum (2010).

The model described by equation (1) is also estimated on sectoral data; that is, at the 28 manufacturing subsectors level. Once again the labor share is regressed on GDP and its square to appraise the impact of development on the sharing of the value added. As explained earlier in this section, controlling for country-sector fixed effects, and using the average of sector shares as weights, allows us to control for individual heterogeneity and so to clean the effect that structural changes may have on the aggregate manufacturing labor share when the economy develops. Doing so allows us to estimate the coefficients of interest on the within-subsectors variations of the labor share. Results are reported in columns 2B, 4B, and 6B of table 2 (to be compared with columns 2, 4, and 6).

Several lessons can be derived from those regressions. First, the coefficients of GDP and of the square of GDP per capita (GDP²) are significantly negative and significantly positive, respectively, confirming the previous results. Second, since the estimations are weighted and controlled for country-sector fixed effects, it means that it is a within effect that is observed. Since the coefficients of GDP are very similar at the aggregate level (columns 2, 4, and 6 of table 2) and at the subsector level (columns 2B, 4B, and 6B of table 2), the aggregate manufacturing variations of the labor share are mostly due to variations within subsectors (within effect) rather than to changes in the weight of the subsectors (composition effect). That means that the reason why development impacts the labor share is marginally related to composition effects at the manufacturing level. We do not claim that there are no reallocations across the manufacturing subsectors, but we argue that they do not explain the main part of the pattern of the manufacturing labor share.¹⁰ Third, the inclusion of the various controls does not affect in a significant manner the observed U-shaped pattern: This suggests that development impacts the labor share through some other channels than those included in the regressions as controls. Note that the capital-output ratio has a negative and significant coefficient in table 1 and that the investment-output ratio has a positive but insignificant sign in table 2. This is consistent with the findings of Karabarbounis and Neiman (2014) that the elasticity of substitution between capital and labor is higher than one and confirms that capital accumulation plays a role in explaining the evolution of factor shares, as the neoclassical theory suggests.

However, by comparing columns 5 and 6 of tables 1 and 2 (regressions run on the same sample without and with capital accumulation), it can be seen that the U-shaped relationship is only marginally affected by the inclusion of the proxy for capital accumulation. This means that despite the validity of the capital accumulation channel, the latter marginally explains the U-shaped pattern that is observed: Other forces playing during the development process dominate.¹¹ At this stage it is worth mentioning that the empirical evidence provided in this paper concerns only the manufacturing sector. Capital accumulation may have a stronger effect on the aggregate labor share when focusing on the whole economy.

Concerning other controls, none of them are significant in tables 1 and 2, except the freedom of assembly and association index and the ratio of government spending, which have a positive and significant sign. This is consistent with the view that the welfare state and institutions provide some bargaining power to workers and tend to emerge with development.

To conclude, there is a strong empirical evidence of a U-shaped relationship between development and the labor share within sectors and it is related neither to factor accumulation nor to other controls.

2. A Simple Model

This section presents a very simple model of a dual economy in the spirit of Lewis (1954). Actually, any model of this kind easily generates a U-shaped pattern of labor share within the subsectors of the manufacturing. The model features a modern sector with high-productivity firms and a traditional sector with low-productivity firms within manufacturing subsectors. In other words, within each subsector of the economy there are modern production techniques that compete with traditional ones (what we call “sectors” in the model).

The model shows that a progressive decrease in the number of traditional firms along the development path generates first a decrease in the labor share and then, at some point, an increase. The mechanism behind such a relationship is quite intuitive. When the number of high-productivity firms is low, firms are unable to absorb the entire workforce, and the outside options for workers are located in the traditional low-productivity firms. As a result, in such an environment, labor demand is low, and so modern firms only need to pay the workers the wage they would have received in the traditional firms. As the modern sector grows with the process of development, labor demand from modern firms increases and firm competition to attract workers makes wages and the labor share increase.

Here the simplest and minimal model that highlights a very plausible rationalization of our findings is presented. It exhibits a competitive labor market and an exogenous share of modern firms that increases along the development path. In the supplementary online appendixes S1 and S2, two richer models are presented.¹²

The first subsection displays some stylized facts in order to better justify Lewis's view of development used in the model. More specifically, this subsection shows that there exists a huge heterogeneity of productivities in developing countries and that the proportion of modern firms tends to increase along the development path.

Stylized Facts

In this subsection several stylized facts about firms in developing economies are presented. They support Lewis's dual view of development and the assumptions made in the model. Most of the empirical evidence provided in this subsection comes from the World Bank Enterprise Survey, the World Bank Micro Survey, and the World Bank Informal Survey. Most of the statistics and variables used here have been constructed by La Porta and Shleifer (2008, 2014), but some of them are also from our own computations.

The World Bank Enterprise Survey corresponds to registered firms only, which are classified as small (lower than 20 employees), medium, or large (more than 99 employees).¹³ The informal survey and the micro survey cover both formal and informal firms, and target establishments with fewer than five employees.^14,15 Micro surveys contain mostly registered firms, however. Several striking facts, highlighted by La Porta and Shleifer, emerge from these surveys. First, the labor market is dual in developing countries. There are high-productivity differences between the modern sector and the traditional sector of the economy. For La Porta and Shleifer, the size of firms is a key characteristic to distinguish the traditional and modern sectors of the economy.¹⁶ The productivity gap (the value added per employee) between large and small firms of the enterprise survey is about 95 percent. The productivity gap between small firms (enterprise survey) and very small firms (fewer than 5 employees, micro survey) is 103 percent. However, these productivity gaps do not entirely translate into wage gaps. Indeed, the wage gap between small and very small firms is not as high and corresponds to 65 percent. Moreover, and even more surprisingly, the wage gap between large firms and small firms in the enterprise survey is even negative (−13 percent). This means that in low-income countries, productivity differences are far from being compensated by wages.

For La Porta and Shleifer (2008), productivity differences between large and small firms reflect differences in the access to key resources. First, only 3 percent of the financing for firms in the micro surveys comes from banks whereas the proportion amounts to 9.6 percent for small firms and to 20.9 percent for large firms in the enterprise survey. Second, and more important, is the difference in the human capital of managers.¹⁷ On a scale from 0 to 4 for human capital of managers, firms in the micro surveys have a score of 2.1 against 2.7 and 3.8 for small and large firms in the enterprise survey.¹⁸ On the other hand, the characteristics in terms of human capital of the employees other than managers are very similar across small and large firms in the enterprise survey (2.3 vs. 2.5) and close to the ones of firms in the micro survey (2.3). For La Porta and Shleifer, given this evidence, most self-employed and employees would accept any job offer in the modern sector of the economy (large firms in the economy).

It is difficult to document the evolution of the size distribution of firms along the development path given the fact that the distribution of firms surveyed may not be a good representation of the real distribution of the total number of firms. However, the informal sector that concentrates an important share of very small and unproductive firms tends to disappear with economic development, as documented by La Porta and Shleifer, among others. Also, by using various surveys from the enterprise survey (2007–2008), one can compute the share of total employment from big firms (more than 100 employees) where productivity is much higher than in the rest of the economy. This share is graphed on the log of GDP per capita (measured in PPP) in fig. 4.

Figure 4.

Share of Big Firms’ Employment in Total Employment

Source: The share of big firms' employment in total employment comes from the World Bank Enterprise Survey (2007–2008) and the GDP per capita (in PPP) comes from the world development indicators.

Note: The share of big firms' employment in total employment corresponds to employment in firms with more than 100 employees over total employment in the World Bank Enterprise Survey.

Open in new tabDownload slide

Figure 4 shows that employment in high-productivity/big firms as a share of total employment increases with economic development. This evidence is compatible with Lewis's dual view of development. The traditional sector is mainly a subsistence sector that exists because of the low labor demand from high-productivity firms in the modern sector of the economy. The fact that the modern sector is too small to absorb the entire workforce is due to the lack of some key resources in developing countries, such as educated managers (La Porta and Shleifer 2008, 2014; Gennaioli et al. 2013) or finance (Banerjee and Newman 1993). Traditional sectors of the economy tend to disappear with the development of modern ones that are able to provide high-productivity jobs using modern production techniques (they can pay more than low-productivity jobs). In such an environment, people endowed with those scarce resources may enjoy some rents, which could explain the pattern of the labor along the development path that is observed in the manufacturing sector.

A Simple and Illustrative Model

There are two production technologies, and population is normalized to one. The traditional one does not need any particular inputs and uses two units of labor to produce y_T. Everyone can enter and operate in the traditional sector. The modern technology also uses two units of labor to produce y_M with y_M > y_T. However, running a firm in the modern sector requires some specific inputs and only a part of the workforce is endowed with such inputs. These inputs can be, for instance, capital if the economy is characterized by credit constraints (Banerjee and Newman 1993) or educated managers are necessary to run large and complex firms in the modern sector (La Porta and Shleifer 2008, 2014). More precisely, only a fraction 1 − G(a*) is endowed with the threshold level of input a* (capital or educated managers) and can run a firm in the modern sector. One unit of labor is provided by the entrepreneur, and the other by a worker. In order to simplify the model, it is assumed that such an input is only valuable to run a firm in the modern sector of the economy (this does not have any implications for the results). An entrepreneur in the modern sector gets a payoff of y_M − w, with w corresponding to the equilibrium wage in the competitive labor market. Similarly, an entrepreneur in the traditional sector gets a payoff of y_T − w and a worker gets a payoff of w. Labor market equilibrium wage takes two possible values in such an environment. First, |$\underline{w}$| is the wage rate at which an individual in the traditional sector is indifferent between working as a wage-earning worker and becoming an entrepreneur. Therefore, |$\underline{w}$| is such that |$\underline{w}=y_{T}-\underline{w}$|⁠, which implies |$\underline{w}=y_{T}/2$|⁠. Second, |$\overline{w}$| is the wage rate at which an individual in the modern sector is indifferent between working as a wage-earning worker and becoming an entrepreneur. Therefore, |$\overline{w}$| is such that |$\overline{w}=y_{M}- \overline{w}$|⁠, which implies |$\overline{w}=(y_{M})/2.$| Of course |$\overline{w }>\underline{w}$| given that y_M > y_T.

It is obvious to see that under the low wage regime (⁠|$w=\underline{w}$|⁠, when the modern sector is not able to absorb the entire workforce), any increase in the proportion of modern firms in the economy (1 − G(a)) or any increase in the productivity of modern firms y_M leads to an increase in GDP per capita but has no impact on the total wage bill. This is due to the fact that outside options for workers remain in the low-productivity traditional sector, which keeps wages stuck at a low level despite economic development. This is the case in China in the 1990s and the 2000s. At one point, when the economy enters the “high wage regime” in which the modern sector is sufficiently developed to absorb the entire workforce, wages jump to |$\overline{w}$| at a given output and the labor share jumps to its long-run constant value. This is the process that China is actually experiencing with wages that grow much faster than productivity. With the end of massive migration from rural to urban areas, this is also the end of the unlimited labor supply environment Lewis describes.¹⁹

In the supplementary online appendix, two richer models with entrepreneurial choices and matching frictions are provided and both generate a similar pattern of the labor share.

3. Conclusion

This paper highlights a strong correlation between development and the labor share of income at the aggregate manufacturing level but also within manufacturing subsectors, which suggests that this correlation is not due to reallocation forces across subsectors. Moreover, controlling for factor accumulation allows us to show that changes in the labor share during the development process are not related solely to changes in factor intensities.

Actually, any standard models in development economics that feature duality can easily account for such a pattern. It is shown in this paper that, for example, a very simplified version of the Banerjee and Newman (1993) or Ghatak and Nien-Huei Jiang (2002) frameworks generates a U-shaped pattern for the labor share under plausible assumptions. The intuition is: At early stages of development, many agents are credit constrained and only a few of them can enter the modern sector to manage high-productivity firms. At this stage, the labor demand from high-productivity firms is low and so wages reflect the low productivity of traditional firms. Therefore, the number of high-productivity firms increases but wages remain stuck at low levels due to insufficient labor demand. The aggregate labor share decreases as a result. At one point, enough entrepreneurs enter the modern sector and the labor demand becomes too high. The traditional sector disappears, wages increase, and the labor share increases as a result.

This pattern is relevant when thinking of the Chinese economy. Recently, strong tensions have started to emerge in the labor market due to an insufficient number of job-seekers compared to vacant jobs. As a result, wages have started to increase quickly, deteriorating the competitiveness of many firms in the economy and leading to the slowdown that the Chinese economy currently experiences and leading to the slowdown recently observed in China.

The link between development and inequality à la Kuznets has always been controversial and not very robust empirically. This lack of consensus can be seen as going against the validity of the here-observed U-shaped relationship, but actually, the link between factor and personal income distribution is not a one-to-one relationship. This paper stresses that the Kuznets curve is valid when examining the link between development and the labor share, a particular measure of income inequality.

Several research issues linked with this paper are left to future work. For example, it would be interesting to explore the impact of development using micro-level data on firms to study the sharing of the value added directly at the firm level. Moreover, it would be worth modelizing many implications that such a pattern of the labor share may have: competitiveness of firms in the exports market, global imbalances, or within-countries migration led by economic development. These questions still remain open.

Appendix A1: Data, descriptive statistics and robustness checks

Data

The labor shares and the investment-output ratios are computed from the industrial statistics of UNIDO (United Nations Industrial Development Organization). All data are originally stored in national currency at current prices. Since these variables are a ratio, it is pointless to convert them from national currency into a common currency.

Wages and salaries: All payment in cash or in kind paid to “employees,” including direct wages and salaries, remuneration for time not worked, bonuses and gratuities, housing and family allowances paid directly by the employer, and payment in kind. Despite UNIDO recommendations, there can remain employers’ social security contributions, pensions, and insurance schemes, as well as the benefits received by employees under these schemes, and severance and termination pay.

Value added: Value of the output less value of the inputs, which covers the value of materials and supplies for production and cost of industrial services received. Can be at factor cost (i.e., excluding indirect taxes minus the subsidies) or at market cost (including indirect taxes minus the subsidies), depending on the treatment.

Gross fixed capital formation: Refers to the value of purchases and own-account construction of fixed assets during the reference year less the value of corresponding sales. The fixed assets covered are those (whether new or used) with a productive life of one year or more.

Descriptive Statistics

Table A.1.1. provides descriptive statistics of the key variables. The ones from the UNIDO database (the labor share and the investment output ratio) are at the aggregate manufacturing level. Tables A1.2. and A1.3 provides the GDP per capita and the logarithm of the GDP per capita for each country of our sample at the beginning and at the end of our sample period.