1. Globalization and inequality

Globalization—or the increased interaction and interdependence between different geographical areas around the globe—has created large secular changes in many economies (Stiglitz, Reference Stiglitz2002). Comparative advantage has led to gains from trade across countries (Feenstra and Hanson, Reference Feenstra and Hanson1997; Feenstra, Reference Feenstra2015), although recent studies have found that such gains have not been reallocated homogeneously across the income distribution (Hays, Reference Hays2009; Rodrik, Reference Rodrik2011). Globalization has also been associated with a number of other outcomes, such as a reduction in global poverty (Ravallion, Reference Ravallion2003), decreased cost of goods and services, and increased democratization (Schwartzman, Reference Schwartzman1998).

There is a considerable body of literature on the relationship between globalization and inequality. Though globalization may lower absolute poverty through rising incomes (Ravallion, Reference Ravallion2003), there is a growing consensus that some individuals are likely to experience relatively greater gains than others (Alderson and Nielsen, Reference Alderson and Nielsen2002; Stiglitz, Reference Stiglitz2002). While scholars present a variety of causes for economic inequality—ranging from land ownership (Miller, Reference Miller2011), inherited wealth (Boserup et al., Reference Boserup, Kopczuk and Kreiner2016), intellectual ability (Huggett et al., Reference Huggett, Ventura and Yaron2011), and individual risk aversion (Azzimonti et al., Reference Azzimonti, De Francisco and Quadrini2014)—many works point to political and economic globalization as a major determinant of inequality (c.f., Stiglitz, Reference Stiglitz2002). Garicano and Rossi-Hansberg (Reference Garicano and Rossi-Hansberg2005), for example, show that lowering the costs of communication and technological change affect wage inequality. Beckfield (Reference Beckfield2006) finds that economic and political “regional integration explains nearly half of the increase in income inequality” in Western Europe (p. 964). In a cross-country analysis, Lundberg and Squire (Reference Lundberg and Squire2003) find a positive relationship between economic openness and inequality.

Scholars suggest several reasons why globalization may exacerbate inequality. Piketty (Reference Piketty2014) stresses that, overall, inequality is intrinsic to market economies. He also claims that the gap between the rate of return on capital and economic growth explains extreme income inequality until World War I, while political and institutional shocks as well as the supply of and demand for skills account for it thereafter (Piketty, Reference Piketty2015). Other scholars have proposed that increased international trade has exacerbated inequality—most noticeably in the developed world—due to decreasing returns to unskilled labor and diverging levels of human capital across different nations (Wood, Reference Wood1995; Corak, Reference Corak2013).

Globalization may also have non-uniform effects across income distributions based on the level of economic development. Advanced industrialized economies, in particular, have seen a resurgence in inequality after a period of consistent decrease in their disparity of wealth (Alderson and Nielsen, Reference Alderson and Nielsen2002).Footnote ¹ In these countries, comparative trade advantages in capital-intensive industries have caused relative declines in the return to unskilled labor. As a result, highly-skilled, better-educated workers typically have benefited from increased economic integration. Such “college earnings premiums” have been associated with decreased income mobility for unskilled workers (Corak, Reference Corak2013). Developed countries may also exacerbate inequality by failing to adequately redistribute the gains from trade. For instance, low levels of redistribution, inadequate trade adjustment assistance, or policies that favor skilled workers exposed to trade at the cost of others, might also lead to increased economic inequality (Hays, Reference Hays2009; Rodrik, Reference Rodrik2011; Acemoglu et al., Reference Acemoglu, Naidu, Restrepo and Robinson2013).Footnote ²

Empirical evidence suggests that in OECD countries, globalization has increased inequality (Dreher and Gaston, Reference Dreher and Gaston2008). Examining 27 EU countries, Asteriou et al. (Reference Asteriou, Dimelis and Moudatsou2014) find an increase in inequality since 1995, mostly attributable to financial globalization stemming from foreign direct investment. Schneider and Soskice (Reference Schneider and Soskice2009) emphasize the influence of distinct political institutions on within-country inequality across the developed world. Smeeding (Reference Smeeding2002) observes that, although rich countries have experienced a rise in inequality, these trends have been driven by increasing income shares to the top income brackets, not necessarily a decrease to lower quantiles of the distribution. However, Luttmer (Reference Luttmer2004) suggests that this is exactly the type of relative difference that individuals care about when making well-being assessments. Thus, even though an entire society may be better off in terms of real inter-temporal comparisons, rising inequality may have serious political and social consequences.

In developing countries, in contrast, globalization might have an entirely different effect on inequality. The poorest 80 percent of individuals on earth have experienced large increases in real incomes over the past 30 years, while those in the next 15 percent have seen very little (The Economist).Footnote ³ Thus, at first glance globalization appears to be helping out the very bottom of the income distribution in developing countries. At its most simple, comparative advantage suggests that low-skilled, labor-intensive industries will flourish.Footnote ⁴ Therefore, we might expect that lower incomes will experience relative gains, while the highest incomes—typically taking advantage of capital rather than labor—will experience relative decreases. However, few studies have examined this relationship in the developing context. Those that have have generally found mixed results (e.g., Wood, Reference Wood1997). For instance, Dreher and Gaston (Reference Dreher and Gaston2008) do not find evidence of a globalization-inequality effect in developing countries.

2. Theory

As outlined in the previous section, there is abundant evidence that inequality has increased in developed nations during the recent periods of rapid globalization, with the highest earning segments of such societies experiencing substantial relative gains. However, it is not clear which income segments are losing out as the top makes relative gains. The same trends are not apparent in less developed nations where the empirical evidence on the impact of globalization on inequality has been somewhat mixed.

We argue that to better focus our thinking about the impact of globalization on relative income shares in a nation, we first need to think about the income composition as a whole, and the likely consequences of globalization on different segments of society within that composition. To do this, we turn to figures developed by Boix (Reference Boix1998) for depicting the relative skill and income distribution in a hypothetical society. As per Boix, we begin with Figure 1, in which we depict the levels of skills and wages on the vertical axis, and the rank distribution thereof on the horizontal axis. The s-shaped curve in Figure 1 reflects the initial distribution of wages and skills before an increase in globalization.

Figure 1. Wages and skills by percentile rank.

Figure 1 can be connected to theoretical models of trade by recognizing that individuals who gain most of their livelihood from labor will be clustered in the lower-left quadrant of this figure, while individuals who gain most of their livelihood from capital will be clustered in the upper-right. A key insight from economic models of trade (e.g., Heckscher and Ohlin, Reference Heckscher and Ohlin1991) is that a country's imports and exports will be determined by their factor endowments relative to the rest of the world. Essentially, exports will come from sources of relative abundance and advantage while imports will be driven by sources of relative scarcity and disadvantage.

We recognize that the connection between international trade and a society's relative economic development is more complicated than this. For one, low skilled labor in one country might not be perfectly substitutable for low skilled labor in a different country. In other words, there are legitimate differences in factor endowment levels across (and within) countries. Reinert (Reference Reinert2007) points out that, for these very reasons, factor price equalization has failed to materialize with the advent of modern globalized trade. Reinert argues that countries that have advanced their economies have done so through emulation rather than comparative advantage. While these arguments are certainly reasonable, addressing the effects of globalization on economic development and global poverty is beyond the scope of this paper. Instead, we focus on different aspects of the income distribution, assuming that in a country that employs more capital relative to labor, capital-rich products are more likely to be exported (to the benefit of individuals who employ capital). Similarly, in a country that employs more labor relative to capital, labor-rich products are more likely to be exported (benefiting those who employ greater amounts of labor relative to capital). We contend that, depending on the level of factor endowments in a nation, globalization is likely to affect relative gains and losses differently across the income distribution.

To convey our logic, in Figure 2 we present a hypothetical scenario in which two countries, one capital-rich (dark lines) and one labor-rich (grey lines), are both exposed to an increase in globalization. In both countries, we focus on the top and bottom income quantiles, with the solid lines representing the pre-globalization absolute real incomes, and the dashed lines representing the post-globalization absolute incomes of the four groups. As we show in Figure 2, we expect the absolute real income in all income groups in this scenario to increase from gains from trade—more specifically, all groups can be expected to benefit from lower prices for consumer goods. But, as this figure displays, we do not expect all groups in this scenario to gain equally.

Figure 2. Pre- and post-globalization absolute income shifts among top and bottom groups in relatively labor-rich and relatively capital-rich countries.

In particular, we expect the top income group in the capital-rich country and the bottom income group in the labor-rich country to gain the most from this increase in globalization. The top income group in the capital-rich country will benefit the most because they now have a wider range of investment opportunities and they have the capital needed to take advantage of these opportunities. The wealthiest income quantile in the labor-rich country will also benefit, but their gains will be considerably more modest. We expect this to be the case because, although they have greater investment opportunities, they now face competition from the top income group in the capital-rich country as they pursue these opportunities. As we mentioned above, all groups will benefit from lower prices for consumer goods. This drives the absolute real income of all groups up but has a particularly strong effect on the two bottom income groups in this scenario. The bottom income group in the capital-rich country will also experience an increase in competition from the lower paid labor in the labor-rich country. Thus their absolute gains from this increase in globalization will be somewhat modest. In contrast, the bottom income group in the labor-rich country will benefit a lot in absolute terms because they will simultaneously experience increased demand for their labor and lower prices for consumer goods.

As can be seen in Figure 2, these gains in absolute income across all four groups translate into relative gains and losses across the two groups within each country. We can see this from the brackets on the outside of this figure depicting the gap between the top and bottom groups in each country before and after the increase in globalization. The brackets on the left, for the capital-rich country, show that the top quantile gains relative to the bottom quantile. The brackets on the right, for the labor-rich country, show that the bottom quantile gains relative to the top quantile. Although absolute real incomes have increased across all strata within each nation, the relative measure of within-country inequality has moved in opposite directions. In the labor-rich country, globalization has led to a shrinking of the income gap between the top and bottom income groups. In the capital-rich country, globalization has lead to a larger income gap between the top and bottom income groups.

If we return to the Boix-style figures like Figure 1, we expect that the impact of an increase in international trade on the relative income distribution in a society depends on its relative factor endowments. These expectations are reflected in Figures 3(a) and 3(b). In Figure 3(a), we depict the expected changes from an increase in trade openness on a nation with a relative abundance of capital. Given this abundance, we expect gains from international trade to disproportionately benefit upper income groups relative to the lower income groups. In a nation with a relative abundance of labor, our expectations are the opposite; as depicted in Figure 3(b), we expect relative increases on the lower end of the income distribution and relative declines at the upper end. Therefore, we expect that inequality will decrease in countries with a relatively abundant supply of labor.

Figure 3. Expected impact of an increase in globalization on the relative income distribution. (a) Nation with large capital endowment. (b) Nation with large labor endowment.

The combined effects of relative factor endowments and trade patterns are likely to shape national income distributions in several ways, as shown in Figures 3(a) and 3(b). In Table 1, we show theoretical expectations for the effect of globalization on the composition of income. For simplicity of exposition, we split income categories into top, middle, and bottom.Footnote ⁵ As shown in Table 1, we expect that in countries with high capital endowments, holding other relevant factors constant, increased economic openness increases the proportion of income for the highest earners and decreases the proportion of income for the lowest earners. While we expect the middle group to experience relative decreases in income shares as well, we expect that this decline is not as large as the decrease for the bottom income group. In contrast, in countries with high labor endowments, our expectation is that, holding all else constant, increased economic openness increases the proportion of income for the lowest and middle earners, while decreasing the proportion of income for the highest earners. We expect this increase to be largest for the bottom group.

Table 1. Expected effects on income composition

Note: Each income category column shows the relative gains and losses in income shares in response to an increase in the independent variables shown on the left. ++: strong relative increase, +: weak relative increase, −: weak relative decrease, −−: strong relative decrease.

In Table 1, we also display our theoretical expectations for the effects of increases in the other relevant variables. We have mixed expectations about the relative impact of economic growth. On the one hand, similar to our expectations for an increase in globalization, it seems reasonable to expect that increased economic growth will benefit whichever factor a country has relatively more of. For instance, in Brazil, increased economic growth would benefit the poor relative to the rich. On the other hand, we might expect that economic growth disproportionately benefits the lower and middle strata of the income distribution relative to the top. In line with the argument of Bonica et al. (Reference Bonica, McCarty, Poole and Rosenthal2013), we expect that increased political polarization will lead to partisan gridlock, which may slow passage of regulatory laws in the legislature. This is expected to increase the share of the income distribution for the top income groups at the expense of the middle and, especially, the bottom. Similarly, we expect the effects of increased political constraints to benefit the top income groups at the expense of the middle and bottom groups. Although our theoretical ideas are about relative income shares before taxes and redistribution, we still expect other policies of governments on the ideological left to increase the relative share of bottom income groups at the expense of the top. We do not have specific expectations about the impact of the age dependency ratio; we include it in our models as a control for the possible effects of demographic change on relative income shares.

In this paper, we are interested in understanding the short- and long-term trade-offs between different layers of the income distribution—including the poor, the lower middle class, the upper middle class, the rich, and the super rich. We believe that globalization affects each of these groups differently over time. Yet, by virtue of being compositions of the same income distribution—the remaining error terms are correlated across each of these groups. Furthermore, we suspect that the relationship in question varies based on relative capital intensity. We are unaware of other academic articles that measure these complex relationships dynamically treating income as a compositional dependent variable.

In order to test the theoretical expectations laid out in Table 1, we need data on the relative income shares of all quantiles within each country-year as well as a measure of each control variable. Because all of the relative variables are likely to be fluctuating over time, we need annual, or more finely-grained temporal data. Given the high persistence in income composition data over time, a good model for testing our theoretical propositions would be

(1)$$ \eqalign{ \hbox{Income Composition}_{it} & =f\,(\hbox{Income Composition}_{it-1} \cr &\quad +\hbox{Globalization}_{it}\times \hbox{High Relative Capital}_{i} +\hbox{Controls}_{it}) } $$

where High Relative Capital_i is a dummy variable identifying countries with relatively high levels of capital endowments.

3. Data and research design

While summary measures of inequality such as the Gini coefficient are available for a wide number of countries over a long period of time (Solt, Reference Solt2016), data that disaggregate the income distribution into quantiles are more rare and often incomplete in terms of annual coverage. As such, we are severely limited in the cases we can analyze. Since we are interested in understanding how all quantiles of the income distribution are affected by changes in economic globalization over time, we cannot utilize common measures such as Piketty's (Reference Piketty2014) top income shares, or the Luxembourg Income Study's (LIS) reported household income information. The former only includes information on top income quantiles, while the latter has substantial breaks in the series that eliminates the possibility of using a time series model.Footnote ⁶ Given our data requirements, we were able to obtain income data for four country-cases: the United States, Canada, Sweden, and Brazil.Footnote ⁷ Although a larger multi-country analysis would be ideal, as Figure 4 shows, these four countries provide us with three cases that have high relative capital endowments (Canada, Sweden, United States) and one case (Brazil) that has a high relative endowment of labor over the period for which we have appropriate data on income quantiles.Footnote ⁸ As we can see from Figure 4, most of the variation in relative factor endowments for our cases is between the high labor case and the three high capital cases rather than over time for the individual country time series.

Figure 4. Capital to labor ratio, 1976–2012.

Despite the similarity of Canada, Sweden, and the United States in terms of relative factor endowments, there are major differences across these nations in terms of other relevant characteristics. For instance, it is well known that, by most measures, Sweden has more generous social welfare provisions than Canada and that Canada is more generous than the United States. To the extent possible, we have tried to remove these types of issues by using pre-tax measures of income as the basis for our dependent variable (more on this below). Despite this, there are doubtless other differences between these nations that persist. Given these, rather than estimate a pooled model such as that specified in Equation 1, we instead estimate a series of single-country equations, each specified as

(2)$$ \hbox{Income Composition}_{t}=f\,(\hbox{Income Composition}_{t-1}+\hbox{Globalization}_{t}+\hbox{Controls}_{t}). $$

We then compare the results from the three countries with relatively high capital endowments with the results from our one country with a relatively high labor endowment. Although this approach does not allow for formal hypothesis tests of the interaction between globalization and high relative capital, it does provide estimates of the effects of globalization at the two different values of the conditioning variable (relative capital), and it avoids the pooling assumptions imposed in typical time series cross sectional models.Footnote ⁹

Before discussing the technical details of our model estimation, we provide a brief discussion of the measure of our dependent variable for each country as well as the operationalization of our independent variables.

3.1. Dependent variable

United States: Income distribution data for the United States were obtained from the US Census Bureau (2017) Current Population survey, and are available between 1947 and 2013. These data represent the share of all pre-tax taxable household income held by each of the five income quintiles as well as the top 5 percent.Footnote ¹⁰ Due to additional available data, we were able to disaggregate the top quintile into two categories, the top 80-94th percentile as well as the top 5 percent, for a total of six income categories.

Canada: Income distribution data for Canada were obtained from the Income Statistics Division (N.d.) of Statistics Canada. These data include pre-tax household income shares of the bottom four income quintiles as well as the top 80-90th percentiles and the top 10 percent between 1976 and 2015. These estimates are based on numerous household surveys including the Survey of Consumer Finances, Survey of Labor and Income Dynamics, and the Canadian Income Survey. The measure of income in question encompasses all market income such as wages, net investments, and rental earnings in addition to government transfers, such as social assistance, less income taxes.

Sweden: Swedish income distribution data were obtained from Hicks et al. (Reference Hicks, Jacobs and Matthews2016) who acquired the original data from Statistics Sweden. These data include total disposable pre-tax income shares of the bottom four income quintiles, as well as the top 80–90th percentiles and the top 10 percent between 1975 and 2011. Documentation from Statistics Sweden defines disposable income as earnings from factor income (wages and capital) as well as government transfers.Footnote ¹¹

Brazil: Our inequality data for Brazil comes from the Institute of Applied Economic Research (Instituto de Pesquisa Econômica Aplicada – IPEA, N.d.), a government agency associated with the Brazilian Ministry for Planning, Development and Management. Income shares were calculated from annual national household survey responses (Pesquisa Nacional por Amostra de Domicílios Contínua – PNAD, N.d) conducted by the country's federal statistical agency, IBGE, from 1976 through 2014 (excepting 1980, 1991, 1994, 2000, and 2010).Footnote ¹² The PNAD questionnaire requests gross income information for retirement, pensions, social assistance, rental properties, donations, investments, and wages from each household member for the reference month. The data we utilize include the bottom four income quintiles, the top 80–99th percentiles, as well as the top 1 percent of Brazilian households.

Figure 5 illustrates the advantages of a compositional approach compared to examining only the trajectory of the top income group. In the left panel in Figure 5, we show the relative share of the top 1 percent over the time period for which we have data. This is the type of dependent variable typically analyzed in studies of inequality. The conclusions from this approach are that income inequality declined slowly across the depicted time period. The general assumption from such a figure would be that this decline, in turn, benefited the rest of the distribution. In the right panel, however, we can see that this was not the case; there was also a decline in the relative income share of the top 80–99th percentiles (the top line in the right panel of Figure 5 depicted with patterning of four dots and a dash). Moreover, we can see that while the relative shares of the remaining quantiles all went up across this period, they did so on fairly different trajectories.

Figure 5. Brazil (1976–2014) comparison between standard measures and a compositional approach.

4. Modeling strategy

Using the recommendations of Philips et al. (Reference Philips, Rutherford and Whitten2015, Reference Philips, Rutherford and Whitten2016a,Reference Philips, Rutherford and Whittenb), we model the relative effects of globalization across income quantiles using a dynamic pie modeling strategy. Such an approach closely mirrors our theoretical expectations about how different income categories respond to increasing globalization. Our strategy for modeling the J income categories measured over t time points, y _tj, is as follows. Since we cannot estimate all J income categories simultaneously (if we did, one of the categories would fall out of the equation, since it would be perfectly predicted by the other income categories), we instead select one category of the compositional dependent variable (say, y _t1), and divide all of the other categories by it.Footnote ¹⁴ We then take the natural log of each resulting ratio, thus creating J−1 logged compositions:

(3)$$ s_{tj} = ln\left( {{y_{tj}} \over {y_{t1}}} \right) \forall j\neq 1 $$

this has the advantage of “unbounding” the series from a compositional space; now, instead of ranging from zero to one (as y _tj does by definition), s _tj can take on any range of possible values. Thus, standard multivariate approaches may be used to estimate models of s _tj (Katz and King, Reference Katz and King1999; Tomz et al., Reference Tomz, Tucker and Wittenberg2002).

Next, we estimate J−1 error correction models (ECMs) for each of the s _tj income compositions. These are estimated simultaneously using a seemingly unrelated regression approach which increases the efficiency of estimates by allowing for correlated errors across equations. These equations are specified as

(4)$$ \Delta s_{tj}=\beta_{0j} - \alpha_j s_{\,jt-1} + {\bi \beta}_{Lj} {\bi x}_{t-1} + {\bi \beta}_{Sj} \Delta {\bi x}_{t} + \epsilon_{\,jt} $$

where:

• • Δs _tj is the change in the logged ratio of dependent variable category “j” for j>1 to baseline category j = 1 from time “t-1” to time “t”,

• • β_0j is a constant term,

• • x_t is a vector of independent variable values at time “t”,

• • − α_j is a vector of adjustment parameters that measure the long-run error correction processes,

• • β_Sj is a vector of short-run effects,

• • β_Lj, is a vector of parameters that can be combined with − α_j to calculate estimated long-run effects of changes in each independent variableFootnote ¹⁵,

• • and ε_jt is a stochastic disturbance term that may be correlated across the “j” equations.

When working with time series data, the potential for non-stationary series prompts one to question whether the relationships we observe may be a product of spurious inferences. Recent work has advocated for checking first for non-stationary series, then performing multiple cointegration testing procedures (Philips, Reference Philips2018). Several unit root tests produced strong evidence that all series in our analyses contain unit roots. We also found evidence of cointegration using a variety of approaches.Footnote ¹⁶ We tested for autocorrelation using the Harvey test for global autocorrelation and were unable to reject the null hypothesis that no autocorrelation remains in our system for each of the countries analyzed.Footnote ¹⁷

Instead of presenting a sizable table of regression results of categories of a dependent variable in logged ratios, we use the Stata program dynsimpie to produce expected values of the proportion of income across all categories in response to a counterfactual increase in globalization (Philips et al., Reference Philips, Rutherford and Whitten2016b). A stylized version of this approach is shown in Figure 6. The program first creates 1000 parameter estimates using Clarify (Tomz et al., Reference Tomz, Wittenberg and King2003), which come from the coefficients and covariance matrix from the estimated model. At t = 1, it then sets all lagged dependent variables to their sample means, the first difference and lag of globalization, $x^*_{t-1}$, to zero and its sample means, respectively; the same is done for the control variables. Then the program generates expected values, which become the new value of the lagged dependent variable for t = 2 and t = 3. At time t = 4, the program introduces a counterfactual one-period shock into the system through variable $x^*_t$ (globalization in our example). The lag of all other control variables remain at their sample means, and their first-differences at zero, effectively rendering them “all else equal”. The counterfactual shock first appears in the first difference of $x^*_t$, then the lag of $x^*_t$ when t = 5. The resulting predictions are then “untransformed” and plotted as short- and long-run changes from starting values (i.e., analogous to marginal effects in standard regression models). Tables of the parameter estimates from these models are included in Section 7 of the Supplemental Appendix.

Figure 6. Dynamically simulating the effect of globalization on income compositions; $x^*_t$ represents globalization at time t.