1. Theoretical and empirical context

At the end of World War II, advanced economies in Western Europe re-embedded market forces into the domestic social and institutional order to achieve an “embedded liberalism,” a compromise between embracing free trade and pursuing interventionist welfare policies domestically (Ruggie, Reference Ruggie1982). A related literature illustrates how globalization and structural economic changes correspond to higher levels of welfare compensation in OECD countries (Cameron, Reference Cameron1978; Katzenstein, Reference Katzenstein1985; Garrett, Reference Garrett1998; Rodrik, Reference Rodrik1998; Iversen and Cusack Reference Iversen and Cusack2000; Iversen Reference Iversen and Soskice2001). In contrast, developing economies liberalized much later during the 1980s and 1990s, and the effects of liberalization have been more mixed. Some studies do find a positive correlation between trade openness and compensation (Rodrik, Reference Rodrik1998; Adserá and Carles, Reference Adserá and Carles2002; Baker, Reference Baker2003, Reference Baker2005; Brooks, Reference Brooks2008; Nooruddin and Rudra, Reference Nooruddin and Rudra2014). However, the majority of studies conclude that liberalization failed to expand or even contracted the welfare state in the developing world (Rudra, Reference Rudra2002; Wibbels and Arce, Reference Wibbels and Arce2003; Nooruddin and Simmons, Reference Nooruddin and Simmons2009; Wibbels and Ahlquist, Reference Wibbels and Ahlquist2011).

What helps to situate this debate is an understanding of the identities of the “losers” of trade. In Latin America’s “truncated” welfare states, the dualism of the labor market segments the population into formal and informal workers. The trade-affected constituencies were members of the working and middle class engaged in largely formal sector labor in manufacturing and agriculture. Social security benefits, therefore, represent a natural form of compensation for this group, while broader social assistance also comprises noncontributory benefits for informal workers (Rudra, Reference Rudra2002; Huber and Stephens, Reference Huber and Stephens2012). However, recent research shows even formal workers, especially those with lower skills or less job mobility, may increasingly view social assistance programs as relevant to their economic security (Menéndez, Reference Menéndez2021). Anticipating difficulty with remaining in the formal labor market, trade-affected workers may also turn to demands for broader social assistance.Footnote ⁶

I assume individual voters form preferences based on material interests (Scheve and Slaughter, Reference Scheve and Slaughter2001). Such demands may be particularly strong among workers with industry-specific skills, for whom displacement entails substantial losses of income and human capital (Reference Iversen and Soskice2001)(Reference Iversen2005). Facing heightened risks of wage declines or job loss, trade-exposed workers demand welfare compensation to hedge against these sources of uncertainty. Developing countries were also especially sensitive to the effects of trade, because this neoliberal era coincided with the earlier years after the democratic transitions of “the third wave.” Failure to respond to demands for compensation could have irrevocable consequences on social stability for newly elected democratic governments (Avelino et al., Reference Avelino, Brown and Hunter2005; Dean, Reference Dean2022). Because trade-affected groups were large and economically vulnerable, and liberalization coincided with the first decade of democracy, I expect governments even in developing regions to have expanded welfare programs alongside liberalization.

Given these contextual factors, it is, at face value, puzzling that much of the extant literature finds limited evidence for the “embedded liberalism” or “compensation” thesis in developing nations. At the same time, an important macro-level constraint, however, qualifies this. Many developing countries pursued trade liberalization under the auspices of International Monetary Fund Structural Adjustment Programs, which simultaneously mandated fiscal austerity and public spending cuts (Nooruddin and Simmons, Reference Nooruddin and Simmons2006). Several studies note this constraint and provide other explanations for why evidence of “compensation” is limited in developing contexts (see Rodrik, Reference Rodrik1997; Baker, Reference Baker2003, Reference Baker2005; Wibbels, Reference Wibbels2006; Baker, Reference Baker2009; Richardson, Reference Richardson2009; Hart, Reference Hart2010; Campello, Reference Campello2013, Reference Campello2015 for theories for the Latin American region and Appendix A7 for theories for the South Asia region). Building on this literature, this paper first makes an empirical advancement that could partly explain the contradictory findings: addressing the endogeneity of existing measures of trade openness. Measures commonly used in the literature not only imprecisely capture trade integration (Avelino et al., Reference Avelino, Brown and Hunter2005) but also fail to account for the endogenous nature of trade policy. Thus, different measures of trade openness used produce considerably different results (Avelino et al., Reference Avelino, Brown and Hunter2005). For example, findings based on exchange rate conversions switch signs when purchasing power parities trade measures are used instead (Avelino et al., Reference Avelino, Brown and Hunter2005). In addition, measures used in the literature also tend to conflate trade openness with financial market liberalization. Since trade liberalization in the developing world coincided with other major political and economic transitions, it is difficult to precisely isolate its effects.

Trade protection is also theorized to be endogenous. Endogenous tariff theory posits that tariff levels are shaped by domestic political bargaining, with more influential or organized industries securing greater trade protection through lobbying (Grossman and Helpman, Reference Grossman and Helpman1994). By the same logic, the extent of trade liberalization may, therefore, also be politically determined, not by exogenous policy shocks. This creates an identification problem: tariff changes may be driven by the same political conditions that influence social spending outcomes. Yet, studies in the vast literature estimate correlational relationships without addressing these concerns.

In this paper, I first address these measurement and methodological challenges by leveraging an important exception to endogenous tariff dynamics: Brazil’s experience with trade liberalization in the early 1990s. Liberalization in Brazil was part of a sweeping, unilateral reform package implemented by technocratic policymakers, largely bypassing domestic interest group negotiations (Kume et al., Reference Kume, Piani and Bráz de Souza2003; Kingstone, Reference Kingstone2010) (see Figure 1 for empirical evidence). By March of 1990, Brazil had eliminated most nontariff barriers, making tariffs the primary instrument of trade policy (Kume et al., Reference Kume, Piani and Bráz de Souza2003). The subsequent reductions in tariffs during the first half of the 1990s were both unprecedented and sustained (Dix-Carneiro et al., Reference Dix-Carneiro, Soares and Ulyssea2018), providing a quasi-natural experiment for identify trade-induced microregional economic shocks (Dix-Carneiro et al., Reference Dix-Carneiro, Soares and Ulyssea2018).

Figure 1.

The industries most protected under pre-liberalization tariffs in 1990 became the ones that were most exposed to post-liberalization trade shocks (i.e., correlation of −0.899).

1 Long description

The scatter plot illustrates the relationship between 1990 pre-liberalization log(1 plus t) on the horizontal axis and the 1990-95 change in log(1 plus t) on the vertical axis. The horizontal axis ranges from 0 to 0.4, while the vertical axis ranges from 0 to negative 0.25. A dashed diagonal line indicates a negative trend, suggesting that as the 1990 pre-liberalization log(1 plus t) increases, the 1990-95 change becomes more negative. Data points are clustered around x values of approximately 0.2 to 0.33 and y values of negative 0.12 to negative 0.20, indicating a dense area of industries. Notable outliers include Agriculture near y equals 0 and Rubber and Apparel with the most negative changes at higher x values. The plot does not show a legend, implying all points share the same symbol. Key industries labeled include Agriculture, Petroleum, Gas, Coal, Mineral Mining, Paper, Publishing, Printing and Apparel, illustrating the trend and outliers effectively.

The structure of pre-liberalization protection in Latin America more broadly also supports this identification strategy. Under the prior model of import substitution industrialization (ISI), high tariffs coupled with overvalued exchange rates disproportionately benefited certain industries. With the onset of liberalization, the most protected industries under ISI faced the steepest tariff cuts and, thus, the largest exposure to international trade. In Brazil, tariff policy changes were almost entirely determined by the pre-liberalization tariff levels under ISI (see Figure 1) (Kume et al., Reference Kume, Piani and Bráz de Souza2003; Kovak, Reference Kovak2013). In making clear that this relationship is almost a one-to-one correlation, Figure 1 provides direct support for the identification assumption of the exogeneity of the size and distribution of tariff cuts across sectors (and geographies) (Kovak, Reference Kovak2013). Benefiting from decades of trade protection, the “trade losers” were politically powerful constituencies. Yet, the decision to liberalize was orthogonal to their political leverage, but rather reflected structural economic pressures (Stallings, Reference Stallings, Haggard and Kaufman1992; Kingstone, Reference Kingstone2010). As a result, tariff reductions can be treated as exogenous to counterfactual industry performance and manipulation by political influence (Kovak, Reference Kovak2013).

Existing theories in the literature assume that a country uniformly embraces “compensation” over “efficiency” or vice versa. However, the economic effects of integration with international markets are not uniform, but rather exhibit considerable variation across subnational regions. Figure 2 provides evidence for this variation across microregional economies within Brazil. In addition, Figure 3 illustrates that geographic distribution of economically vulnerable trade-affected industries, likewise, exhibits considerable subnational variation. I argue that this subnational heterogeneity in voter exposure to trade integration is an overlooked dimension that explains the contradictory findings in the literature.

Figure 2.

Incidence of microregional tariff shocks in Brazil during the 1990–1995 period of trade integration.

Figure 2 Long description

Thematic choropleth map titled, Micro Regional Tariff Shocks, with a subtitle, Weighted Average Tariff Change. Brazil is shown with internal boundary lines dividing many subnational areas, but no area names are labeled. North is at the top. A legend labeled, Tariff Cuts, shows a continuous scale from 0 percent to 100 percent. The map uses a single light to dark shading ramp to encode the tariff cut value. Higher tariff cut shading appears in many scattered patches across the country, with frequent clusters along the eastern side and in parts of the southern portion. Lower tariff cut shading appears in multiple interior areas, including several larger contiguous zones in the central and northern interior. No north arrow or scale bar is shown.

Note: Mapped using R and calculated using data and methodology from Kovak (2013) and data from Kume et al. (2003).

Figure 3.

Geographic concentration of “trade-affected” populations.

Figure 3 Long description

Thematic choropleth map of Brazil with internal boundary lines and a patchwork of shaded areas representing HH Index. A legend labeled, HH Index, shows a continuous scale from 0 percent at the left end to 100 percent at the right end. No state names are shown on the map. North is at the top. Higher legend-end shading appears in scattered clusters across the map, including parts of the far north, sections of the central interior and multiple pockets in the eastern half. Lower legend-end shading appears in many small areas spread throughout the country, including parts of the central interior and several areas along the eastern and southern portions. The map contains no point symbols, arrows, or scale bar.

In addition, political response to these differentiated effects across each subnational jurisdiction is, likewise, not uniform. This raises an important question: how and to what extent do subnational politicians in Brazil influence social spending? Although redistribution is often characterized as a federal responsibility, a growing body of literature documents how Brazilian states exercise substantial discretion in directing social spending (Arretche, Reference Arretche1999; Niedzwiecki, Reference Niedzwiecki2016; Rogers, Reference Rogers2021; Dunn, Reference Dunn2022). Using their own coded measure of subnational authority developed for countries across the world, Hooghe et al. (Reference Hooghe, Marks, Schakel, Osterkatz, Niedzwiecki and Shair-rosenfield2016) classify Brazil as a highly decentralized federation in terms of both fiscal and administrative autonomy. State governments in Brazil finance their own revenue streams, most notably using the ICMS (i.e., a value-added sales tax), which accounts for well over half of their income (Ter-Minassian, Reference Ter-Minassian2012). When considered in combination with other sources of state-generated revenue (e.g., the motor vehicle tax), over 70 percent of state government revenue comes from their own sources, not from federal transfers (Ter-Minassian, Reference Ter-Minassian2012).Footnote ⁷

Importantly, these own-source revenues give state governments leverage in initiating or expanding social programs beyond federally mandated transfers. Although certain earmarks constrain expenditures, state governments can and do implement their own social assistance programs (Niedzwiecki, Reference Niedzwiecki2018; Dunn, Reference Dunn2022). Existing work by Rogers (Reference Rogers2021) demonstrates that Brazilian state governments exhibit considerably different redistributive social policies across stat. For instance, while the state of Maranhão reduced market inequality by nearly 22 percent, Goiás achieves only a 12 percent reduction (Rogers, Reference Rogers2021). These differences underscore subnational agency.

But what does subnational agency look like in practice? First, governors coordinate with federal ministries to shape how intergovernmental transfers are deployed.Footnote ⁸ In states with weaker capacity or limited fiscal effort, such as Pará or Alagoas, decentralized implementation results in less robust coverage and administrative quality, even for nationally funded programs (Niedzwiecki, Reference Niedzwiecki2018). Thus, even when programs are federally funded, their reach and efficacy depend on local administrative capacity and political will (Arretche and Rodden, Reference Arretche and Rodden2004; Gibson et al., Reference Gibson, Calvo and Falleti2004). A second form of subnational agency operates through political veto power. Federalism introduces subnational political veto points (Beramendi, Reference Beramendi2012), allowing governors, in particular, to veto or repurpose federal discretionary transfers. Finally, subnational agency also takes the form of policy supplementation. Governors frequently supplement federal initiatives with matching funds, such as supplementary income supports, public job training schemes, and emergency aid programs (Arretche, Reference Arretche1999; Dunn, Reference Dunn2022). The legal authority for states to design and implement their own social policies is clarified in Brazil’s 1993 LOAS Law No. 8742 and 12435.Footnote ⁹

To give an example, the state of São Paulo operates its own targeted social assistance program, Renda Cidadã, designed to supplement federal transfers by providing monthly cash transfers to low-income families not fully covered by Bolsa Família.Footnote ¹⁰ Introduced in 2001 by the former governor of São Paulo, Geraldo Alckmin (PSDB) and targeted using state-specific criteria, the program illustrates how governors can expand redistributive coverage to politically and economically strategic populations.Footnote ¹¹ Similarly, the state of Pernambuco develops and sustains social protection programs, such as Chapéeu de Palha, which provides income support and vocational training for rural workers during the sugarcane off-season (Bezerra, Reference Bezerra2014). Introduced in the 1980s under former Governor Miguel Arrães and later once again under Governor Eduardo Campos in 2007, this state-level initiative addresses seasonal unemployment and complements federal transfers (Bezerra, Reference Bezerra2014). Thus, especially during times of economic dislocation from shifts in local labor conditions, I expect governors to have considerable agency in offering welfare “compensation.”

This paper recognizes the autonomy of subnational governments, particularly in federal systems, as an overlooked dimension in the trade and compensation debate. It responds to calls for advancing the literature by analyzing the “relationship between [subnational] geographical socioeconomic and political trends and their expression in … the behavior of individuals” (Frieden, Reference Frieden2022). Voters’ demand for compensation is not purely instrumental. Rather, their material interests are also shaped by their local environments (Schaffer and Spilker, Reference Schaffer and Spilker2019; Frieden, Reference Frieden2022). Even if they are not personally affected, voters embedded in regional contexts of economic decline—e.g., rising unemployment, business closures—can still develop support for compensation policies through sociotropic (i.e., community-level) concerns (Schaffer and Spilker, Reference Schaffer and Spilker2019). These sociotropic perceptions often substitute for incomplete information about personal impacts, a dynamic Schaffer and Spilker’s (Reference Schaffer and Spilker2019) term “obfuscated egocentric preferences.” Thus, the individual voter calculus includes both personal risk and communal vulnerability, and perceptions of the latter also inform the former.

To specify the argument more precisely, I argue that variation in the subnational distribution of political and economic identities are key contextual dimensions that mediate the welfare compensation effect. First, when geographically concentrated groups are united under one economic identity (e.g., “trade-affected”), the location of these concentrated industries also reflects increased labor market risk. Especially in regions with a concentration of specialized sectoral composition, a trade shock affecting one industry may also affect the local economy of the entire region (Menendez, Reference Menendez2016; Iversen, Reference Iversen2005). For instance, with the experience of European cases with trade integration, (Menendez, Reference Menendez2016) argues that the concentration of the manufacturing industry makes these trade-affected industries more susceptible to unemployment risk. In contrast, in regions with diversified (dispersed) composition, the multiplicity of industrial sectors offsets the effects of a shock to any one of them (Longui et al., Reference Longui, Nijkamp and Traistaru2005). Thus, I expect that geographically concentrated economic interests (e.g., “trade losers”) receive more compensation.

While economic geography marks territories of high risk and, thus, demand for compensation, political geography reflects the capacity to make demands heard. I expect political concentration to play an even more critical role than economic concentration in predicting compensation for several reasons. First, the geographic concentration of like-minded political interests enables the communication and coordination of shared political preferences. The proximity of like-minded voters in regionally concentrated homogenous enclaves enables them to form stronger vested interests groups and to more easily engage in political mobilization. Community-level decline or prosperity colors individual perceptions of risk. For example, two individuals with similar job profiles may respond differently to trade if one resides in a region experiencing concentrated plant closures and the other in a diversified labor market. In addition, the concentration of political identities provides the networks to collectively mobilize and articulate cohesive demands in response to these risks. Morever, concentrated constituencies can more easily use retrospective voting to send a clearer signal of their preferences to politicians (Stokes, Reference Stokes2016).

Similarly, the literature on social policy in Latin America provides accounts of how the density of like-minded voters in local contexts enables the adoption of local participatory institutions. For instance, Wampler (Reference Wampler2010) and Goldfrank (Reference Goldfrank2015) demonstrate that in Brazil, municipalities that have a greater density of civil society organizations (CSOs) were more likely to implement participatory budgeting. Patterns of adoption of participatory governance, in turn, correspond to higher levels of local public and social expenditure (Abers, Reference Abers2000; Baiocchi, Reference Baiocchi2005; Goldfrank, Reference Goldfrank2007, Reference Goldfrank2015; Wampler, Reference Wampler2010; Touchton et al., Reference Touchton, Sugiyama and Wampler2017). This literature observes that, overall, collective mobilization increased social policy expansion in the region (Zarate Tenorio, Reference Zarate Tenorio2014; Nooruddin and Simmons, Reference Nooruddin and Simmons2015; Garay, Reference Garay2016; Cini et al., Reference Cini, Della Porta and Guzmán-concha2021). Although not directly theorized, such forms of agglomerated “bottom-up” grassroots mobilization are only feasible because of the proximity and geographic concentration of voter bases.

In Brazil, labor unions, such as the Central Única dos Trabalhadores and Força Sindical, are an example of CSOs in the 1990s that facilitated collective demand-making for unemployment benefits and early retirement schemes (Keck and Smith, Reference Keck and Smith1992). Importantly, union membership and activity exhibited geographic variation (Keck and Smith, Reference Keck and Smith1992). I argue that where there is a concentration of politically aligned voters, the robustness of these organizations in concentrated enclaves, in turn, amplifies collective action by lowering the costs of coordination and demand-making with local politicians. Where political constituencies are geographically fragmented, union infrastructure is likely to be absent and collective action for welfare compensation is harder to sustain. More broadly, I expect that geographically homogeneous political blocs boost higher levels of voter turnout among this constituency compared to among a group of equal size that is geographically dispersed. In contrast, geographically dispersed constituencies face greater difficulty with overcoming the collective action problem when making demands for compensation.

2. Data and methodology

2.4. Measuring localized trade liberalization shocks

As discussed, Brazil’s experience with trade liberalization was a once-and-for-all event in that tariff reductions during the early 1990s were unprecedented and have remained relatively constant since liberalization (Dix-Carneiro et al., Reference Dix-Carneiro, Soares and Ulyssea2018). I borrow from the empirical literature focused on estimating the subnational labor market effects of increased foreign competition (see Kovak, Reference Kovak2013; Autor et al., Reference Autor, Dorn and Hanson2013; Dix-Carneiro, Reference Dix-Carneiro2014). Because there is usually subnational variation in patterns of specialization in production, trade shocks, in turn, have heterogeneous effects across industries and regions. This literature uses the interaction between sectoral composition at the regional level and sector-specific trade shocks to estimate trade-induced shocks to local labor demand (Dix-Carneiro et al., Reference Dix-Carneiro, Soares and Ulyssea2018).

In particular, I follow Kovak’s (Reference Kovak2013) methodology of estimating the local shock of liberalization as the weighted average microregional-level tariff change in microregion r. This regional-level tariff change (RTC_r) reflects the average tariff change faced by microregion r during the 1990–1995 period, weighted by the industrial allocation of labor specific to each microregion:

(1)\begin{equation}RT{C_r}\, = \,\sum\limits_{i \in T} {{\gamma _{ri}}} \,{{\Delta }}ln\,(1 + {\tau _i})\end{equation}

where

(2)\begin{equation}{\gamma _{ri}}\, = \,\frac{{\frac{{\lambda ri}}{{{\theta _i}}}}}{{{\sum _{j \in T}}\frac{{\lambda rj}}{{{\theta _j}}}}}\end{equation}

where τ_i is the tariff on industry i, λ_ri is the share of microregion r’s workforce employed in industry i, θ_i equals 1 minus the wage share of industry i, and T denotes the set of all industries affected by trade. Using data from the 1991 census provided by the IBGE, the method calculates the importance of each sector for the regional employment of each microregion. These weights by microregion and industry are denoted as γ_ri above. To be clear, the measure captures a reduction in microregional tariffs. Hence, estimated coefficients that are negative indicate tariff cuts have a positive effect on the outcome variable. Again, Figure 2 maps the microregional tariff shocks to localized economies across Brazil during the 1990–1995 period.

The outcome measure is calculated as the change in the level of microregional social expenditure between 1990 and 2000. I then estimate the effects of localized trade shocks on compensation as follows (Kovak, Reference Kovak2013):

(3)\begin{equation}log\left( {S{S_{r,t}}} \right) - log\left( {S{S_{r,}}_{1990}} \right) = {\xi _t} + {\varphi _t}RT{C_r} + {\varepsilon _{r,t}}\end{equation}

where SS_r,t denotes social assistance and social security expenditure in microregion r at time t > 1990, RTC_r is the microregional tariff change in microregion r, and φ_t captures the estimated causal effect of trade shocks on welfare compensation. All standard errors are clustered at the state level, and each specification is reported both with and without state fixed effects. This approach is essentially a reduced-form analysis in which microregional tariff changes serve as an exogenous instrument for the price changes faced by local producers (Kovak, Reference Kovak2013).

To consistently estimate φ_t, the identification assumption is that the error term, ε_r,t, must be uncorrelated with tariff changes, RTC_r (see Appendix A6 for discussion). This identification assumption is violated if there is an omitted variable that (1) induces social expenditure, (2) is also correlated with RTC_r, and (3) is not captured by pre-liberalization outcome trends. To demonstrate the validity of the identification assumption, I run a series of estimations to confirm that the results are robust to a variety of potential confounders and alternative specifications. First, I show that the results are robust to the inclusion of pre-liberalization social expenditure trends. Second, I follow Kovak (Reference Kovak2013) by confirming that the results are robust even when using a series of alternative RTC_r measures: (1) RTC_r omitting the labor share adjustment (i.e., assuming labor demand elasticities are identical across industries), and (2) RTC_r including a zero price change for the non-tradable sector.

Next, I modify Kovak’s (Reference Kovak2013) estimation strategy with an interaction to observe how the effects of trade integration on welfare compensation are mediated by geography:

(4)\begin{equation}log\left( {S{S_{r,t}}} \right) - log\left( {S{S_{r,}}_{1990}} \right) = {\xi _t} + {\varphi _t}RT{C_r} + {\gamma _t}hhi + {\phi _t}RT{C_r} \cdot hhi + {\varepsilon _{r,t}}\end{equation}

where hhi is the HHI for either the geographic concentration of the governors’ constituency or of the trade-affected sectors. Last, I code four indicator variables that capture whether each index is high or low (e.g., high political concentration–high sectoral concentration, high political–low sectoral, low political–high sectoral, etc.). I replace the hhi in the interaction estimation above with the indicator variables, running the estimation for each indicator. To be clear, unlike in Equation 3, the interactions of the endogenous geographic concentration measures with RTC_r do not have a causal interpretation. Yet, it still sheds light on how geography mediates the association between trade liberalization and compensation.

3. Empirical results

3.1. Main results

I first consider the results disregarding the role of geography to discern whether trade integration has a causal effect on social welfare expenditure (Hypothesis 1). I then move on to examine the extent to which the effects of local trade shocks are mediated by the geographic distribution of political and economic identities (Hypotheses 2 and 3). Table 1 presents the estimates for the effects of microregional tariff shocks on expenditure on social security and social assistance (in million Brazilian reals) measured as absolute spending, not in per capita terms. Standard errors are clustered at the level of the state, and Models (4)–(6) also includes state fixed-effects. Model 1 shows the results for the main specification (Equation 3), Model 2 omits the labor share adjustment, and Model 3 includes the non-trade-affected sectors (e.g., services) in the calculation of regional tariff change (RTC), with non-trade sector price change set to zero.Footnote ¹³

Table 1.

Effects of trade liberalization on social assistance and social security expenditure

Table 1 Long description

Dependent variable is Welfare Compensation, measured in R$ 1 mil, reported for three models labeled 1, 2, and 3. Regional Tariff Change is minus 611.363 with standard error 205.159 in model 1, minus 690.765 with standard error 226.305 in model 2, and minus 932.446 with standard error 431.910 in model 3, each with star markers. Constant is minus 13.944 with standard error 9.466 in model 1, minus 11.734 with standard error 8.517 in model 2, and minus 0.046 with standard error 6.628 in model 3. State Fixed Effects are marked with a check in all three models. Observations are 456 for each model. R Squared is 0.107 in model 1, 0.121 in model 2, and 0.064 in model 3.

Notes: Robust standard errors in parentheses, clustered at the state-level. Fixed effects included at the state-level.

*p < 0.1; **p < 0.05; ***p < 0.01.

Robust across all models, the results indicate that the effect of RTCs on social expenditure is negative. As discussed, because the independent variable is coded as a reduction in microregional tariffs, this implies that Brazilian microregions facing the largest reductions in tariffs experienced more welfare compensation (i.e., negative coefficients indicate a positive effect). The provide direct evidence for the “compensation hypothesis”: localized trade liberalization shocks have a large positive effect on welfare compensation. Microregional economies that experienced more exposure to competition with global markets, in turn, experienced an increase in welfare compensation. The “compensation effect” is large as welfare expenditure is in units of million Brazilian reals. In Appendix A3, I test the robustness of this observed “compensation effect” using an alternative measure of subnational social spending.

In addition, I argue that geographically concentrated political identities have greater capacity for collective mobilization. When like-minded political interests are geographically clustered, it facilitates their communication and coordination. Spatial proximity allows voters within these politically homogeneous enclaves to organize and demand “compensation” more effectively. Existing work demonstrates that collective mobilization is, in turn, linked with social policy expansion in the Latin American region (Zarate Tenorio, Reference Zarate Tenorio2014; Garay, Reference Garay2016; Cini et al., Reference Cini, Della Porta and Guzmán-concha2021). To test this claim, I estimate Equation 4 to examine the mediation effects of the HHI measuring the concentration of political identities. If there is no subnational variation in political response to trade, then the index would not have a mediation effect on the results. Table 2 presents the results that evaluate the mediation effect of this dimension in comparison to that of the spatial concentration of industries affected by trade. Consistent with the argument that political mobilization is critical for translating risk into policy response, I expect that political concentration exerts a stronger influence on compensation than economic concentration.

Table 2.

The role of spatial concentration

Table 2 Long description

Dependent variable is Welfare Compensation, measured in R$ 1 mil, with four models labeled (1) to (4). Regional Tariff Change has coefficients of minus 389.513 with standard error 113.059 in model (1), minus 436.423 with 137.481 in model (2), minus 322.767 with 135.763 in model (3), and minus 397.921 with 189.076 in model (4). HHI Political Concentration appears in models (1) and (3) with coefficients of minus 141.763 with standard error 72.401 and minus 176.237 with 95.822. The interaction of Regional Tariff Change with HHI Political appears in models (1) and (3) with coefficients of minus 3259.586 with standard error 1461.848 and minus 3575.492 with 1710.438. HHI Industry Concentration appears in models (2) and (4) with coefficients of minus 104.949 with standard error 58.898 and minus 131.799 with 76.653. The interaction of Regional Tariff Change with HHI Industry appears in models (2) and (4) with coefficients of minus 2712.880 with standard error 1458.318 and minus 2841.097 with 1859.444. The constant is minus 12.020 with standard error 6.030 in model (1), minus 9.832 with 5.129 in model (2), 50.932 with 27.826 in model (3), and 45.679 with 26.573 in model (4). State Fixed Effects are marked No for models (1) and (2) and marked with a check for models (3) and (4). Observations are 456 for all models, and R squared values are 0.165, 0.130, 0.185, and 0.145 for models (1) through (4).

Notes: Robust standard errors in parentheses, clustered at the state-level. Results shown with and without state-level fixed effects.

*p < 0.1; **p < 0.05; ***p < 0.01.

The results show that the concentration of trade-affected industries increases the magnitude of the positive effect of trade openness on compensation. This could be observed by taking the sum of the coefficients of RTC and of the interaction term, RTC × HHI (Industry), in Model 2. Again, because the independent variable reflects a reduction in tariffs, negative coefficients indicate a positive effect on compensation. As discussed, the coefficients estimated here do not have a causal interpretation, given the inclusion of the interaction terms. The results show how political and economic geography mediates the relationship between trade and welfare compensation. The latter marks pockets of labor market risk, while the former reflects capacity for making collective demands for compensation.

As Model 1 shows, the concentration of political identities also augments the positive association of trade openness with social spending, providing evidence for the theory that such concentration reflects the capacity for demand-making. More importantly, Model 2 provides a useful grade by which to evaluate the effects of political geography in Model 1. Comparing Model 1 with Model 2, it is clear that the mediation effect of increasing political concentration is of comparable magnitude to that of the concentration of labor market risk. The results also hold in Models 3 and 4 with the inclusion of state fixed-effects.

To examine this further, I code four indicator variables that capture the overlay of the geographic concentration or dispersion of political identities with that of industries affected by trade, and I estimate Equation (5). The indicator denotes whether each microregional economy has high or low concentration of political support and of trade-relevant industries, resulting in four classifications: (1) high HHI for political concentration, high HHI for industry concentration; (2) high political HHI, low industry HHI; (3) low political HHI, high industry HHI; and (4) low political HHI, low industry HHI. Instead of using the full range of each index, I dichotomize concentration to more clearly observe the cross-cutting and overlapping effects of the two dimensions. Figure 4 presents the results.

Figure 4.

The effect of trade liberalization on compensation is mediated by the overlay of political and economic geographies.

Note: Model 1 plots the sum of the coefficients of trade liberalization, RTC, and of the interaction term, RTC × High Pol, High Ind. See also Table A1 in the Appendix.

When high levels of political concentration overlay high industry concentration, the positive association of localized trade shocks and compensation is of the greatest magnitude. Alternatively, when political concentration is high while industry concentration is low, the positive association of trade and welfare is reduced by around 75 percent. In Model 3, when instead political concentration is low and industry concentration is high, the positive association is reduced even more than it is in Model 2. More specifically, the sum of the coefficient of RTC and of RTC × Low, High,—144.081, constitutes a 84 percent reduction—in comparison to Model 1—in the association of trade and welfare state expansion. This result lends direct evidence in support of the hypothesis that the geographic concentration of like-minded voters mediates welfare compensation. Even where the concentration of trade-relevant industries is high, the dispersion of the political constituency substantially undercuts the compensation effect.

Last, Model 4 presents the results for when both the geographic concentration of political identities and of trade-affected industries is low. As Model 4 shows, the geographic dispersion of both dimensions almost completely negates the positive impact of trade shocks on local welfare states. It is clear from the results that the state’s response to integration with global markets is largely contingent on geography. Although Brazil’s experience with trade liberalization, overall, supports the compensation hypothesis, the experience of subnational political economies within the country varies extensively depending on the geographic dispersion or concentration of political identities and of labor market risk.

3.2. Do concentrated constituencies have greater capacity for collective action?

I now explore more direct evidence of the extent to which the political concentration of voters enhances their capacity for collective action in regard to turning out at the polls. The geographic proximity of members from the same constituency facilitates their mobilization and political engagement. In contrast, the dispersion of a group renders it more difficult for them to coordinate their political demands. The observable implication is that geographically concentrated constituencies have a higher rate of turnout at the polls than does a dispersed constituency that is of equal group size.Footnote ¹⁴ All else equal (i.e., controlling for population size, etc.), it is the concentration of a group that drives more of their members to the polls on election day. The results, estimated with standard errors clustered at the state level, are presented in Table 3.

Table 3.

Political concentration and voter turnout

Table 3 Long description

Dependent variable is voter turnout, measured as thousand votes, reported for four models labeled 1 through 4. HHI political concentration is 204.395 with standard error 89.765 in model 1, 140.486 with 66.182 in model 2, 194.429 with 83.508 in model 3, and 212.079 with 102.129 in model 4. Population density log appears in models 2 to 4 with coefficients 106.266 and standard error 28.341, 111.478 and 28.665, and 117.890 and 35.615. Mean income appears in models 3 and 4 with coefficients shown as minus 0.000, with standard errors 0.000 in both models. Percentage urban appears only in model 4 with coefficient minus 71.266 and standard error 88.453. Constants are minus 112.183 with standard error 53.688 in model 1, minus 1218.504 with 332.229 in model 2, minus 801.400 with 181.597 in model 3, and minus 907.884 with 244.765 in model 4. State fixed effects are included in all models, observations are 456 in each model, and R squared values are 0.070, 0.346, 0.375, and 0.377.

Note: Standard errors in parentheses.

* p < 0.1, ** p < 0.05, *** p < 0.01.

The table presents results from the analysis estimating the association between political concentration and voter turnout. The results show that among constituencies of equal size, geographically concentrated constituencies have a higher rate of turnout at the polls for gubernatorial elections than does the dispersed one.

As the results show, the concentration of political support for the governor corresponds to a much higher level of voter turnout for this candidate. The dependent variable, voter turnout, is in units of thousand votes; hence, the effect of concentration is quite large. Geographically concentrated enclaves of political support could also reflect a higher density of voters, for example, in metropolitan areas. Additionally, income or educational background could shape voter preferences for living closer together in large numbers or for political engagement. To rule out these alternative effects that could confound the results, I control for these demographic attributes at the microregional level in Models 2–4. The results remain robust across the models. In Appendix A4, I present additional results that show the concentration of political support predicts a higher incidence of local participatory councils, as an alternative measure of collective action.

In a country where voting is compulsory, there is concern that measures of turnout are high. However, while compulsory voting likely boosts turnout for the country overall compared to a country without such a rule, it applies uniformly across all subnational jurisdictions. The analysis pursued in Table 3 examines, instead, subnational variation in voter turnout. Given there is no variation in the compulsory voting rule subnationally, compulsory voting does not confound the analysis; it cannot explain the subnational variation observed.