Kuznets Revisited: What Do We Know about the Relationship between Structural Transformation and Inequality?

In recent decades, most developing and emerging economies have seen large shifts of workers from agriculture to the manufacturing and service sectors (Dabla-Norris et al. 2013; Felipe, Mehta, and Rhee 2015). At the same time, in several economies, inequality has increased (Berg and Ostry 2011, United Nations Development Programme 2013, Milanović 2016). In a famous 1955 paper, Kuznets argued that as low-income economies industrialize, inequality will increase over time as workers move from low-productivity agriculture to high-productivity manufacturing, which may lead to an increase in overall inequality, though the process of industrialization will also accelerate economic growth (Lewis 1954).

Two complications arise when considering Kuznets’ thesis from the viewpoint of today. Firstly, very few economies have followed successful industrialization strategies since Kuznets published his article, and some economies may well be undergoing “premature deindustrialization” (Rodrik 2016). It is not clear what would be the inequality implications of the mixed record on industrialization in developing economies. Secondly, as we will show in this paper, much of the shift of workers from agriculture has been to services and not to manufacturing. Services, in general, tend to have lower levels of productivity than manufacturing, so it is not obvious that structural change that is biased toward services is necessarily as inequality enhancing as the agriculture-to-manufacturing shift in employment.

In this paper, we revisit the stylized facts of structural transformation and inequality, using comparable data on these measures for a range of low-, middle-, and (now) high-income economies in Asia, Africa, and Latin America for the period 1950–2010.¹ We ask whether there is a positive relationship between structural transformation and inequality, as hypothesized by Kuznets, and whether this relationship differs across economies that have followed different paths of structural transformation, and across regions.

In section II, we first provide a summary of the main theoretical underpinnings of the Kuznets process. In section III, we describe the data used in the paper. In section IV, we document the patterns of structural transformation across economies and classify them according to their stage of structural transformation. In section V, we present the stylized facts on the relationship between structural transformation and inequality. In section VI, we look at the regional differences in the structural transformation–inequality relationship. In section VII, we present our conclusions.

In his classic 1955 paper, Kuznets suggested that in the early phase of economic development, inequality will increase. At a later phase of economic development, as governments follow redistributive policies combining progressive taxation with welfare spending, inequality may decrease. The core of Kuznet's argument on the relationship between inequality and development is captured in the following paragraph extracted from his 1955 paper:

The Kuznets process of widening inequality with structural transformation—that is, the movement of workers away from agriculture—can be described as comprising two subprocesses: (i) the movement of the population from a sector characterized by lower mean income to a sector characterized by higher mean income, and (ii) the movement of the population from a sector with low within-sector inequality to a sector with higher within-sector inequality. If both subprocesses work in the same direction—that is, if the movement of workers is from a sector with both a low mean and low variance in incomes to a sector with a higher mean and high variance in incomes—then structural transformation will unambiguously increase inequality. However, if the movement of workers is from a sector with a low mean but higher variance in income to a sector with a higher mean but lower variance in income, then it is less obvious that inequality will necessarily increase.

In Kuznets’ view, the sector from which workers were moving out from is clearly agriculture. However, the sector that is absorbing the labor movement is left ambiguous in the 1955 paper; while it is most likely industry, it could be services as well. The two defining features of this sector are that it should have both higher mean income and within-sector inequality than the agricultural sector for the Kuznets process to hold. Both these features may not hold for any particular economy in the process of structural transformation. For example, if the movement of workers away from agriculture is mostly to the informal service sector such as retail trade and restaurants, it is not clear that such a transfer is necessarily a move to a sector with higher mean income. It is also possible that the agricultural sector in any particular economy has high inequality if the land distribution is concentrated among a few elites. In this case, if the movement of workers is from agriculture to a sector with relatively low inequality such as a labor-intensive manufacturing, inequality may not increase with structural transformation and may even decline.

What the above discussion shows is that whether the Kuznets process holds for any particular economy depends on the specific characteristics of the path of structural transformation that the economy follows. For example, are workers moving from an agricultural sector that has high land inequality or is the agricultural sector in this economy characterized by more equal land distribution? And, is the movement of workers to a sector with relatively low mean incomes such as low-productivity services or to a sector with high within-sector inequality such as mining or capital-intensive manufacturing? Previous empirical research on the Kuznets process does not have an unambiguous finding of inequality first increasing and then decreasing with economic development (Anand and Kanbur 1993a, 1993b; Milanović 2000; Lindert and Williamson 2003; Roine and Waldenström 2015). However, much of this literature has focused on the growth–inequality relationship, and there is a large gap in the literature on understanding the structural transformation–inequality relationship. This is a crucial omission, given the relevance of the debates around structural transformation and inequality in contemporary development policy.

In this section, we describe the data used in the analysis of structural transformation, inequality, and poverty.

Data on structural transformation in economies are taken from the Groningen Growth and Development Centre's (GGDC) 10-Sector Database. The GGDC database includes data from 42 economies covering the 1950–2012 period. We have excluded advanced market economies from Europe, along with Japan and the United States, which left us with 32 economies from four geographic regions. Table 1 provides a list of economies in our sample with the time period that the data cover for each economy. The GGDC database consists of annual series for the gross value-added output and the number of people employed in agriculture, mining, manufacturing, utilities, construction, trade services, transport services, business services, government services, and personal services. We have grouped these 10 sectors into four main categories:

• Agricultural sector = agriculture

• Manufacturing industry = manufacturing

• Nonmanufacturing industry = mining + utilities + construction

• Service sector = trade services + transport services + business services + government services + personal services

Gross value-added data are taken from national income accounts of the various economies and compiled according to the United Nations System of National Accounts. The 10 sectors have been classified using the International Standard Industrial Classification, Revision 3.1. Using this classification of manufacturing instead of the narrower Standard International Trade Classification implies that primary processed products are also included in the definition of manufacturing. Employment is defined as “all persons engaged,” thus including all paid employees as well as self-employed and family workers. This implies that the GGDC employment data include both the formal and informal sectors. The primary source of the employment data is the population census, supplemented by labor force and business surveys (Timmer and de Vries 2009; Timmer, de Vries, and de Vries 2016).

The share of employment for the four main categories is calculated by dividing the number of people employed in each category by the total number of people employed in the economy in a given year. Productivity in each category is calculated by dividing the value-added output in constant 2005 local currency by the number of people employed.

As noted by Diao, McMillan, and Rodrik (2017), GGDC provides the highest quality data available on sectoral output and employment for developing economies. However, it is also subject to certain limitations, which can raise concerns when the data are used to calculate productivity. The first set of limitations relates to the quality of the source data and the extent to which they include the informal sector. The quality of data on the sectoral value-added output published by national statistical agencies of underdeveloped economies can be unsatisfactory, and whether the data successfully account for the informal sector depends on the quality of the national sources. On the other hand, as the annual series on the number of people employed in each sector are obtained from census data and household surveys by the GGDC researchers, they are more likely to capture informal employment. (Appendix 1 discusses other sources of sectoral employment data and their limitations.)

Income inequality data are taken from the standardized income inequality dataset computed by Baymul and Shorrocks (forthcoming). The Gini coefficient, calculated from household surveys, is the most commonly used measure of inequality. However, due to conceptual and methodological differences between household surveys, the comparability of inequality data is an issue that troubles empirical researchers. The standardized dataset used in this research tries to enhance comparability by adjusting all available data that exceeds a quality threshold from various sources through a regression adjustment method that includes an extensive list of independent variables. Despite generating the highest number of individual annual observations per economy compared with any other available dataset, the number of observations still varies between economies.

In this paper, we use Gini coefficients that indicate the net income per capita inequality.² However, standardized income inequality data are prone to measurement errors made in source data. Measurement errors could be especially problematic in least developed economies where the quality of the data collection methods is questionable.

A second striking feature of structural transformation is that the shift of employment from agriculture to services has been accompanied by falling productivity in the service sector compared with agriculture (Figure 1).³ The large shift of employment from agriculture to services accompanied by the falling relative productivity of services suggests that structural transformation in most developing economies (barring a few economies in Asia) has not been growth enhancing. This has implications for the possible effects that structural transformation may have on inequality, which we explore in the next section. We also observe a similar falling ratio of manufacturing to agricultural productivity, though the relative productivity of manufacturing is far higher than that of services.

Economies in our sample show three different paths or stages of structural transformation. There are economies where agriculture is still the largest sector in terms of the share of employment in the most recent time period available. In our sample, these economies are Ethiopia, India, Kenya, Malawi, Nigeria, Senegal, Tanzania, and Zambia. These economies are all in Sub-Saharan Africa except for India. We call these economies structurally underdeveloped. The next set of economies are where more people are employed in the service sector than in agriculture, with agriculture being the second-largest sector. These economies are Bolivia, Botswana, Brazil, Colombia, Costa Rica, Egypt, Ghana, Indonesia, Morocco, the People's Republic of China, Peru, the Philippines, Thailand, and South Africa. We call them structurally developing economies. These economies span all three continents included in our study: Africa, Asia, and Latin America. The final set of economies has more people employed in the manufacturing sector than in agriculture. These economies are Argentina; Chile; Hong Kong, China; Malaysia; Mauritius; Mexico; the Republic of Korea; Singapore; Taipei,China; and Venezuela. These economies are either in East Asia or Latin America, with the exception of Mauritius, which is in Africa. We call these economies structurally developed.

As we have noted in the previous section, the movement of labor away from agriculture in the process of economic development can either be toward manufacturing or services. We first look at the relationship between structural transformation and inequality when the share of employment in agriculture is falling. We then look at the manufacturing employment–inequality relationship, followed by the services employment–inequality relationship. In each case, we first look at the pooled relationship between structural transformation and inequality, where we measure inequality by the net income per capita Gini. We then focus on our three economy groups, which we have categorized by their stage of structural transformation: (i) structurally developed, (ii) structurally developing, and (iii) structurally underdeveloped.⁴

However, one problem in assessing the relationship between manufacturing employment share and inequality is that the share of manufacturing in total employment does not show a clear monotonic relationship with time. This is in contrast with the behavior of the shares of agriculture and services in total employment, both of which show a clear monotonic relationship with time. (In the case of agriculture, its share in total employment falls over time for our sample economies; in the case of services, its share increases more or less continuously over time for our sample economies.)

Economies exhibit the following patterns in the share of manufacturing in total employment over time: (i) a “hump” (increasing, then decreasing); (ii) continuously increasing; (iii) continuously decreasing; and (iv) no discernible movement. This suggests that a scatter plot of inequality against the manufacturing employment share may simply be capturing cross-sectional differences in the relationship of inequality with the manufacturing employment share across the sample economies, in contrast to the scatters of inequality against the agricultural and services employment shares, which capture both time series and cross-sectional variation in the relationship. (In the case of the inequality–agriculture scatter, a movement in the graph from right to left in the horizontal axis is a movement in time; in the case of the inequality–services scatter, a movement in the graph from left to right in the horizontal axis is a movement in time.)

Thus, we observe a very different behavior of inequality to increases in the services employment share compared with what we observed with the increases in the manufacturing employment share.

We conduct three further tests to check the robustness of our results. Firstly, we use the gross Gini instead of the net Gini to see the direct effect of structural transformation on market inequality, prior to taxes and transfers. We present the regression results in Table A2.2. Next, we confine our analysis to the post-1970 period as there was not a significant structural transformation in most developing economies during the 1960s. We present the regression results in Table A2.3. Finally, we use the sectoral employment data from the International Labour Organization (ILO). These data are seen as being poor quality as they are directly obtained from the statistical agencies of the economies concerned and are not subject to consistency checks in the same way as the GGDC data. (For a discussion of the weaknesses of these data, see Diao, McMillan, and Rodrik [2017].) By using these data, we more than double the number of observations to 1,148. We present the regression results in Table A2.4 and the plots of the marginal effects of manufacturing and services on inequality in Figures A2.1 and A2.2, respectively.

When we use gross Gini instead of net Gini, we do not find any difference in our results in terms of the manufacturing and services employment shares on inequality, either by structural transformation group or region. The sign and significance of the coefficients of the manufacturing employment share and its square, and the interaction of these two variables with structural transformation groups and with regions, generally remain the same compared with the results in Table A2.1 and columns (I) and (II) in Table A2.2. Similarly, we do not find any discernible difference in the sign and significance of the coefficients of the services employment share and its square and the interaction of these two variables with structural transformation groups and with regions compared with Table A2.1 and columns (III) and (IV) in Table A2.2. We also get identical results with the post-1970 employment data (Table A2.3).

When we use the ILO data, we find that the marginal effect of the manufacturing employment share on inequality changes from negative to positive for structurally developed economies, but that there is no change in the effect of the manufacturing employment share on inequality for structurally underdeveloped and developing economies for relevant ranges of the manufacturing employment shares compared with Figure 7 (Table A2.4 and Figures A2.1, A2.2).

A long-held view in the literature on economic development is that inequality increases with structural transformation as workers move from a low-inequality sector such as agriculture to high-inequality sectors such as manufacturing and services. This is commonly known as the Kuznets process. We revisit the relationship between structural transformation and inequality using comparable data for 32 developing and recently developed economies for the period 1950–2010.

Firstly, we find that structural transformation in the majority of our 32 economies has entailed a move of workers from agriculture to services, and not to manufacturing. Further, the move of workers from agriculture to services (and, wherever it has occurred, to manufacturing) has been accompanied by a fall in the relative productivity of services and manufacturing compared with agriculture (barring a few economies in East and Southeast Asia). The economies in our sample have shown different paths of structural transformation that cut across geographical regions. A set of economies can be categorized as structurally developed if the number of workers employed in manufacturing exceeds the number of workers employed in agriculture. Five Asian economies figure in this list—Hong Kong, China; Malaysia; the Republic of Korea; Singapore; and Taipei,China—along with Argentina, Chile, Mauritius, Mexico, and Venezuela. Structurally underdeveloped economies have agriculture as the largest sector in terms of the number of people employed in the most recent time period available. In our sample, only one Asian economy figures in this list, India, along with Ethiopia, Kenya, Malawi, Nigeria, Senegal, Tanzania, and Zambia. Structurally developing economies are those where more people are employed in the service sector than in agriculture, with agriculture being the second-largest sector. Four Asian economies figure in this list—Indonesia, the People's Republic of China, the Philippines, and Thailand—along with Bolivia, Botswana, Brazil, Colombia, Costa Rica, Egypt, Ghana, Morocco, Peru, and South Africa.

If we look at the relationship of the employment share of agriculture in total employment and inequality, we see a Kuznets-type, inverted-U relationship for structurally developed economies. For structurally developing and underdeveloped economies, a lower employment share in agriculture is accompanied by higher inequality. However, we do not observe a Kuznets-type relationship between the share of manufacturing in total employment and inequality. This is particularly evident when we take into account the different paths of industrialization that developing economies have followed. In fact, in contrast to what was postulated by Kuznets, there is a fall in inequality with an increase in the manufacturing employment share for all economies. We also see clear regional differences in the structural transformation–inequality relationship in the case of manufacturing, with an increase in the employment share of the latter associated with falling inequality in Africa and Asia, but with increasing inequality in Latin America.

In the case of services, we see that the effect of an increase in the services employment share on inequality is unambiguously positive, irrespective of an economy's stage of structural transformation. However, we also find that there is a decline in the marginal effect of the increase in the services employment share on inequality across all economy groups, so that the rate of increase in inequality as the services employment share increases declines over time. We also find that an increase in inequality with services-driven structural transformation is evident for Africa, Asia, and Latin America. However, the rate of increase in inequality falls over time in Asia and Latin America, in contrast with Africa where the rate of change of inequality increases over time, suggesting that the evolution of inequality will be very different in Asia and Latin America compared with Africa as all three regions see significant shift of workers from agriculture to services.

Our paper did not attempt to explain why we see such a heterogeneous response of inequality to structural transformation across economies, and why the effect of manufacturing-driven structural transformation on inequality is different from that of services-driven structural transformation. For Asia, the high rates of manufacturing-driven structural transformation and the relatively benign effect of such a pattern provides more of a win–win scenario of structural transformation into manufacturing, leading to both economic growth and falling inequality. This is a scenario that is very different from that envisaged by Kuznets and many others in the development community in which structural transformation was inevitably associated with rising inequality. Why Asia has had such a favorable scenario is an avenue for further research.

There are two additional sources of data, apart from the GGDC database, on sectoral employment at the economy level. The first is the World Bank's World Development Indicators (WDI), which covers more economies than the GGDC. However, the WDI only reports total shares of labor in agriculture, industry, and services. The industry sector consists of mining, construction, public utilities, and manufacturing. The service sector consists of wholesale and retail trade and restaurants and hotels; transport, storage, and communications; financing, insurance, real estate, and business services; and community, social, and personal services. The WDI dataset does not break down industry employment data by manufacturing and nonmanufacturing (e.g., mining, construction, utilities) and services employment by subsectors. The aim of our analysis is to examine the impact of manufacturing and service subsectors on inequality. Since the WDI does not offer information on subsectoral allocations of employment, we are unable to use the data it provides.

A second source of employment data is the ILO's database, ILOSTAT, which provides detailed information on the number of people working in each sector for a majority of the economies in our sample since the 1950s. The data are based mostly on labor force surveys and supplemented by censuses and other minor sources. However, even though ILOSTAT offers the largest sample size and time scale, the comparability of this dataset is limited as concept definitions and population coverage differ between economies and over time. The frequency of the data collected also varies between economies and disregards all impacts of seasonality on the labor force. For these reasons, the GGDC 10-Sector Database is our preferred data source.