The Conditional Direct and Unconditional Indirect Effect of Income on Preferences for Centralization of Authority

Transcription

1 The Conditional Direct and Unconditional Indirect Effect of Income on Preferences for Centralization of Authority Diogo Ferrari September 1, Introduction There are two trends of research that have developed more or less in parallel. One such track of research details the association between income distribution and preferences for redistributive policies. Although some authors have demonstrated that variation in the preferences of members of the middle and affluent classes on redistribution seems to be affected by factors such as class identity, inequality, and social identity, the basic element determining preferences for redistributive policies is the relative position of the individual on income distribution: poor individuals are more likely to support redistribution than rich individuals or the middle class (Romer, 1975; Meltzer and Richard, 1981; Lupu and Pontusson, 2011; Shayo, 2009; Rueda and Stegmueller, 2015). In parallel, researchers have focused on the effect of income as well as preference for interpersonal and interregional redistribution on preferences for allocation of political authority in multilevel polities. While some authors have argued that trust in the central government or strong sentiment of regional identity drive preferences for decentralization of political authority (Weissert and Jones, 2015; Henderson et al., 2013; Jeffery, 2009; Moreno, 1997; Wlezien and Soroka, 2010; Wolak and Palus, 2010; Kincaid and Cole, 2011), the basic argument of the political economy of fiscal federalism states that the preference for centralization depends on the expected effect of different institutional designs on the individual s income (Bolton and Roland, 1997; Beramendi, 2007, 2012). Recent literature on the political economy of fiscal federalism has develop the theoretical connection between income distribution, preferences for allocation of political authority, and preferences for redistribution (Balcells et al., 2015; Beramendi, 2012, 2007). The central argument is that the individual s anticipation of the redistributive consequences of the allocation of political authority affects the formation of preferences for centralized or decentralized institutional design. Furthermore, the preferred redistributive effect is driven by the position of the individuals and their region in the income spectrum. This article builds upon this literature and contributes to the debate in two ways. First, we disentangle the two core ideas present in the argument outlined in the previous paragraph. In line with that literature, we argue that income has an indirect effect on preferences for centralization, mediated by preferences for interregional redistribution. Additionally, we argue that income has a conditional direct effect on preferences for centralization, maybe mediated by factors not considered here. That is, the effect of income depends on the expected distributional consequences of different institutional designs. Our explanation for the conditional direct effect is in line with arguments developed by the political 1

2 economy of fiscal federalism literature which highlight the effect of short-term income maximization concerns on the formation of redistributive preferences. Individuals evaluate the different electoral pressures for redistribution that emerge in different scenarios and tend to prefer the design that is closer to their preferred levels of redistribution. Second, we employ the potential outcome and the mediation analysis framework to gain empirical access to the hypotheses. To the best of our knowledge, such arguments have not been empirically tested on the basis of data on individual s preferences. That is, the mediation role of preferences for redistribution have not been tested by means of information provided by voters themselves. We conducted that much-needed analysis, based on a representative sample of 2,285 adults interviewees in Brazil. As our first finding, we show that income is negatively associated with preference for redistribution. We also show that preference for redistribution, particularly interregional redistribution, is positively associated with preference for centralization of political authority. Our findings indicate that preferences for redistribution mediates the effect of income on preference for centralization in the direction expected. Second, we show that there is a direct but conditional effect of income on preferences for centralization, perhaps mediated by other mechanisms not treated here. We propose a simple measure to be used as proxy for capturing the different political economic environment and the differences in electoral pressure for redistribution under different designs. The proxy is associated with preferences as expected. Income is negatively associated with preferences for centralization in the extent that such institutional arrangement means more electoral pressure for redistribution. The article is divided as follows. The first section presents theories about determinants of preference for redistribution, preferences for centralization, and our alternative explanation that connects the arguments presented for each case and state the hypothesis about the mediation effect of preferences for interregional redistribution. We also develop the rationale behind the conditional direct effect of income. Next, we present the research design, formalize the hypotheses presented in the preceding section, and describe the data and variables used in the analysis. The following section presents the empirical analyses. We conduct a mediation analysis and show evidence supporting the hypotheses of conditional direct effect of income on preferences for centralization as well as the mediation effect of preferences for redistribution. The final section presents our conclusions. 2 Theory Preference for Redistribution By definition redistributive policies, broadly understood, transfer resources from upper to lower income groups. Considering its immediate effect on income, taxation creates clear winners and losers. There is no dissent in the literature about redistributive preferences of those who are the beneficiaries of redistributive policies. As beneficiaries, they tend to support redistribution. There is some disagreement both among hypotheses and among empirical findings in the literature, however, about the preferences of those whose taxation potentially imposes short-term income lost. Before reviewing the arguments, it is important to emphasize a distinction regarding redistributive policies. The majority of the literature that focus on determinants of redistributive preferences has focused on a particular type of redistribution, namely interpersonal or inter-goup redistribution. In 2

3 multi-tier polities, the possibility of regional-based redistributive policies, that is, those that redistribute from relatively wealthy to relatively poor regions, becomes salient. The causes and effects of preferences for interregional redistribution have been comparatively understudied. This section discusses first some hypotheses about formation of preferences for interpersonal redistribution, or redistribution in general, in order to derive implications about preferences for interregional redistribution. Then, it focuses in some hypothesis in the literature that deals directly with preferences for interregional redistribution. The simplest argument regarding causes of support for redistributive policies states that those whose income would be negatively affected by the redistributive policies are more likely oppose those policies. The individuals that have a short-term income maximization preference underlying their preferences for redistribution evaluate the effect of taxation on their pre-tax income and support it proportionally to their net gain or loss (Romer, 1975; Meltzer and Richard, 1981; Corneo and Grüner, 2002; Rueda and Stegmueller, 2015). The research on the topic, then, suggests a negative association between income level and support for redistribution. A great deal of empirical evidence points in that direction, such as Bean and Papadakis (1998); Finseraas (2008); Shayo (2009); Gilens (2005); Rueda and Stegmueller (2015). Authors have given different explanations for why the middle or the affluent class would support redistribution despite immediate resultant income costs. One of the explanations emphasize the effect of affinity based on race, religion, or ethnicity. The argument states that affluent classes (that is, the middle class or the rich) tend to be more supportive of redistribution if they and the beneficiaries of the policies (the poor) belong to the same race, religion, or ethnicity (Luttmer, 2001; Gilens, 2009; Alesina and Glaeser, 2004). Luttmer (2001) has shown that in U.S. not only self-interest, i.e., immediate income-related concerns, but also racial group loyalty affects support for redistribution. The more racially heterogeneous the affluent and beneficiaries of redistribution are, the less inclined the affluent members are to support redistribution. We refer to this class of argument as the social affinity hypothesis. Another argument posits the effect of affinity that results from class proximity. A version of this argument is based on the idea of class coalition. It states that members of the middle class tend to support redistribution when the beneficiaries of the redistributive policies have income closer to those of the middle class than of the affluent class (Lupu and Pontusson, 2011). The reason is that the lifestyle of the poor and the middle class tends not to be much different, which increases the class solidarity between them., leading the former to increase their support redistribution 1. Class solidarity can emerge either as an effect of the perceived distance in income between the individual and the prototypical member of the group, because of the perceived relative status of the group in question, of because the income proximity make it more likely for the classes to create network ties since they tend to share same social spaces (Lupu and Pontusson, 2011; Shayo, 2009; Wilensky, 1974; Kristov et al., 1992; Goette et al., 2012; Fowler and Kam, 2007). According to the class affinity literature, the rich oppose redistribution, but the members of the middle class preferences are conditional on its relative distance to the other two classes, the rich and the poor, therefore determining the class coalition that emerges (Lupu and Pontusson, 2011). We refer to this set of arguments as the class affinity hypothesis. Class proximity, however, can have the opposite effect on preferences regarding redistribution through the mechanism of class competition. The middle class perhaps support less redistribution if the change in the standard of living of the poor threaten the hierarchical position of the former and 1 Although the literature has not fully explored this point, a distinction between pre- and post-taxation distance must be made. We explore this point somewhere else. 3

4 its access to services, goods, and networks. Redistribution can increase the consumption power of the lower class and it can affect the social environment of the middle class, increasing its competition for services and social spaces (Corneo and Grüner, 2002; Cole et al., 1992). Finally, the middle and/or the affluent class may be more supportive of redistributive policies because of the negative externalities of inequality or because they are driven by altruistic values (Fowler and Kam, 2007). Rueda and Stegmueller (2015), for instance, have found a more constant support for redistribution among the poor, but preferences of rich individuals depend also on the levels of inequality due to its negative externalities, such as high crime rate. So, rich in more unequal regions are more supportive of redistribution than rich in more equal regions. In sum, the relative preference of the income beneficiaries hasn t been contested, but the negative association between income and preference for redistribution can be moderated by - or even conditional on - class affinity, social affinity, generalized feeling of altruism, or negative externalities of inequality. Such hypotheses focus primarily in the interpersonal or inter-group redistributive preferences of individuals. In order to use them to enlighten hypothesis about how individuals form their preferences for interregional redistribution, some adaptations must be made. Lets consider a multi-tier polity and possibility of redistributing income between jurisdictions, for instance, through intergovernamental transfers of fiscal resources. In such context, those hypotheses can help to understand preferences for interregional redistribution if there is an overlap between the groups and the jurisdictions. If individuals expect that the majority of the beneficiaries of the interregional rdistribution in the region are out-group individuals, then the in-group should oppose redistribution (Balcells et al., 2015). In the case of the social affinity argument, for example, individuals of one region may oppose interregional redistribution if the majority of the people in the beneficiary region belongs to a different race or ethnicity (Balcells et al., 2015). The most straightforward hypothesis about interregional redistributive preferences of individuals with short-term income maximization concerns is that those living in rich regions tend to oppose interregional redistribution. The opposite is expected among those living in the beneficiary regions (Beramendi, 2012, 2007; Balcells et al., 2015). Balcells et al. (2015) present some evidences supporting this hypothesis in Spain. They also test the hypothesis that income effect is conditional on the regional income. That is, assuming a progressive tax system 2, poor individuals in poor regions tend to support interregional redistribution while rich in rich regions tend to oppose it. The reason, according the them, is that poor individuals in poor regions always benefit from interregional redistribution. Rich individuals in rich regions always lose from it. The preference of the poor in rich regions and the rich in poor regions somewhere between those two groups, and it depends on the structure of the inequality in the region. They have found no evidence supporting the conditional effect of income (Balcells et al., 2015, p. 1335). Finally, they also find that out-group concerns matter only in the context of other relevant information about relative regional income (idem, pg.1337) A different version of the argument about the interactive effect of income and the structure of inequality relies on the median voter hypothesis (Beramendi, 2007, 2012; Balcells et al., 2015), which we explore in the next section because it involves its relation with distribution of authority. Our main goal in this article is to show how preferences, particularly regarding interregional redistribution, are one of the mechanisms toward which income affect preference for allocation of authority among levels of government in federal systems. We expect a negative association between income 2 One should add, assuming also that individuals are short-term income maximizers 4

5 and preference for redistribution, particularly after controlling for class and social affinity, inequality, or more directly its effect on the emergence of negative externalities such as crime rates. Preference for Allocation of Authority Some have argued that preferences for or against centralization of authority in multilevel polities are driven by how strong certain groups identify with their local communities. Strong intergroup identity among regionally concentrated groups leads to higher support for decentralization of political authority There is a long tradition in the political economy of (fiscal) federalism that emphasizes the distributional consequences of the decentralization of authority in multilevel polities (Wibbels, 2005; Weingast, 1995; Oates, 2005; Besley and Coate, 2003; Gibson et al., 2003; González, 2012; Dragu and Rodden, 2011; Rodden and Wilkinson, 2004; Rodden, 2002; Prud homme, 1995; Peterson, 1995). In federal systems, the problem of redistribution intersects with the problem of allocation of fiscal authority because centralization enables interjurisdictional-based redistributive programs and intergovernmental transfers, whereas in decentralized federations such redistributive channels are relatively reduced (Persson and Tabellini, 1996; Alesina and Perotti, 1998; Bosch et al., 2010; Ansolabehere and Snyder, 2006; Beramendi, 2007). Additionally, rescaling authority changes allocative pressure for redistribution from the electorate since local and national elections, in general, have different constituencies. These factors explain why some have argued that the preferences of individuals toward centralization are mainly driven by their anticipation of the expected distributional consequences that emerge under the different allocation of fiscal authority (Bolton and Roland, 1997; Beramendi, 2007). Basically, the causal story connecting preferences for redistributive policies and preferences for centralization begins with the concept that vertical distribution of fiscal authority in multilevel polities has redistributive consequences. Interregional redistribution, for instance, is possible only if the decision on fiscal allocation is the prerogative of the federal government (Bosch et al., 2010). Voters recognize the importance of the distribution of authority among the levels of government and support the configuration that most likely will produce policies that are close to their preferences (Beramendi, 2012). Beramendi (2012) have emphasized the effect of the interaction between within- and between-region income distribution on preference for types of redistribution (interpersonal or interregional), and then derived theoreically the effect of such preferences on support for centralized allocation of authority. Poor individuals in poor and highly unequal regions tend to prefer centralization if, for instance, it enlarges the available income for taxation and therefore increase levels of redistribution to their benefit thought interregional transfers. Finally, another factor that affect preference for centralization is the risk of income loss. Higher specialization and the hability of individuals in the region to move accross regions and industry affect demand for protection against income loss. Conditional on the economic activity of the region, if region is highly specialized workers are more exposed to income loss, and therefore support centralization to diminish the economic risk (Beramendi, 2012). In summary, some argue that individual preferences for centralization can be driven by class identity or by income-related concerns especially regarding redistribution. When the argument emphasizes the effect an individual s income has on preferences for allocation of fiscal authority, it involves an indirect association: anticipation of the redistributive consequences of the distribution of authority 5

6 affect the formation of preferences for centralization, and the preferred redistributive effect is driven by the position of the individuals and their region in the income spectrum. The next section explores our hypothesis regarding this indirect influence of income. Direct and Indirect Effect of Income on Preference for Centralization As reflected by the above literature, an individual s income appears to be an important factor affecting preferences for redistribution as well as preferences for allocation of authority in fiscal systems due to the anticipation of the distributional consequences under different designs. As the literature on the topic notes, preferences for redistribution may mediate the effect of an individual s income level on preferences for centralization. If, for instance, the affluent class tend to oppose high levels of redistribution, it may also influence the degree to which they support institutional designs that lead to higher pressure for redistribution. In simple terms, our first hypothesis is that one of the mechanisms by which income affects preferences for centralization is through its effect on preference for redistribution, particular interregional redistribution. To the best of our knowledge, no previous work has investigated this causal mechanism empirically. Income can also have a direct effect on preferences, although this effect may be mediated by other elements not considered in this exploration. Our argument regarding the direct effect of income on preferences for centralization relies on the idea that in majoritarian elections, as it is the case for the chief of the Executive Branch, in the absence of large polarization of preferences among the electorate the position of candidates and incumbents tends to converge to the preferences of the median voter (Downs, 1957). When the preferences in question refer to attitudes toward redistributive policies, we follow the argument put forward by Romer (1975), Meltzer and Richard (1981), and more recently by Beramendi (2007, 2012). Briefly, the highest electoral pressure for redistribution comes from the preferences of the median voter, and her/his preference for levels of redistribution depends on how poor s/he is compared to the average income in the polity s/he belongs. Given identical average income in both centralized and decentralized scenarios, there is more electoral pressure for redistribution under the design in which the poor population is more numerous that is, where the median income is lower. However, given the same median under decentralized and centralized scenarions, there is more pressure for redistribution in the design in which the average income is higher. We argue that the direct effect of income on preference for centralization is conditional on the relation of the median income and the mean income, in a given location, under different distributions of authority. More precisely, our argument is that underlying the voter s preference for institutional design there is his/her perception of the relation between the electoral pressure from the median income preferences and the average income available for taxation. If, for example, the national median is richer than the local one, and the average national and local income are the same, poor people prefer a decentralized scenario because the pressure for redistribution is higher as a result of that particular income distribution. Rich people in the same scenario, on the other hand, prefer a centralized distribution of authority. As another example, suppose that the national average income is higher than the local average and the national median is also richer. In case of centralization, the change in the electoral demand for more redistribution due to the increase in the average income can be compensated for by the increase of the median income in such a way that the net change in the electoral pressure is 6

7 zero. In sum, our hypothesis is that the direct effect of income on preferences for centralization (perhaps mediated by other factors omitted here) is conditional on the perception of the individual about the level of electoral pressure for redistribution under different scenarios. Preference for centralization is a positive function of income only if individuals associate centralization with less pressure for redistribution. Our hypothesis about the mediated and direct effect of income on preferences does not imply that the actual level of redistribution or the actual distribution of authority will be determined simply by electoral pressure. Researchers have used levels of redistribution as a dependent variable and income distribution as independent variable to evaluate the median voter hypothesis and they have found evidence for relationships in both directions (Alesina and Glaeser, 2004; Gouveia and Masia, 1998; Iversen and Soskice, 2006, 2009; Moene and Wallerstein, 2003; Rodrigiuez, 1999; Rueda and Stegmueller, 2015; Beramendi, 2012). The level of redistribution actually adopted depends on other features of the system, as the political institutions, party system, and legislative bargaining, and such features must be taken into account (Baron and Ferejohn, 1989; Dragu and Rodden, 2011; Rodden and Arretche, 2004; Rodden, 2010; Gibson et al., 2003; Ansolabehere et al., 2003; Persson and Tabellini, 2005; Kristov et al., 1992; Lindert et al., 2007). In terms of the theorized mechanism underlying the observed relationship, we only claim that the preferences are affected. The perception about the electoral pressure under different scenarios drives the support for different institutional designs. Our outcomes of interest are preferences for redistribution, and the actual levels of redistribution adopted by the government in question are not within the scope of our theory or hypotheses. In what follows we present the research design, formalize the hypothesis, and present the data used to access the hypothesis stated in this section. 3 Research Design Our goal is to estimate the direct conditional and (potentially unconditional) indirect effect of income on preferences for centralization. To evaluate the hypotheses we use a national survey conducted in 2013 in Brazil. The data and the variables are described in the next section. First we formalize the strategy used to gain empirical access to the hypothesis. We can formalize the indirect (mediation) effect using the potential outcome framework and mediation analysis in the following way (see Imai et al. (2010a, 2011) for details). Let R i denote the mediator, that is, the preference for redistribution of individual i. Let C i denote the support of i for centralization. Let y i denote the income of i, and y ik1, y ik2, k 1, k 2 R, k 1 k 2, two different levels of income. Denote R i (y k ) the potential outcome for the mediator when y i = y k, and C i (y a, R(y b )) the potential outcome, i.e., preference for centralization, under certain level of income and mediator values. Although we cannot recover the individual effects, we can estimate the average mediation effect. Denote the average (causal) mediation effect (ACME) of R by δ(y). Then, the quantity of interest can be expressed as δ(y) = E[C i (y, R(y 1 ))] E[C i (y, R(y 0 ))] According to the hypothesis outlined about, we expect that income has a negative effect on preference for redistribution (y 1 > y 0 implies R i (y 1 ) < R i (y 0 )), and that preference for (interregional) redistribution have a positive effect on preference for centralization (R i (y 1 ) < R i (y 0 ) implies C i (y, R i (y 1 )) < C i (y, R i (y 0 ))). The working hypothesis (H M ) about the mediation effect can be 7

8 expressed as H M : δ(y) < 0 We can formalize the conditional direct effect of income in the following way. We posit that the individuals perceive how the income is distributed among the voters under centralized and decentralized scenarios. They perceive, for instance, that their region is richer than other regions or the country on average, and they perceive where the bulk of the voters are located in terms of income in the two different scenarios, local and national elections. So, they understand the different distributive preferences of the two different constituencies, national and local 3. If the highest electoral pressure comes from the median income voters, and the redistributive level demanded by the median depends on its distance from the mean, we can capture approximately the different electoral pressures under different institutional designs empirically by quantifying the relation between the change in the mean and median income in the different scenarios. Let Y n and Ỹn denote, respectively, the average and median national income, and let Y s and Ỹs denote the same variable measured at the local level. Let ρ = Y s Y n Ỹs Ỹ n We refer to ρ as the difference in ratio factor. It captures the relation between local and national medians and means. If the hypothesis stated here is correct, the direct conditional effect of income on preference for centralization is positive when ρ > 0, that is, rich people (more income) prefer centralization when, for instance, an increase in the average income under centralization (Y s /Y n ) is not compensated by an increase in the median income (Ỹs/Ỹn). So ρ > 0 means that centralization reduces the pressure for redistribution on the rich in the region, and ρ < 0 means that the electoral pressure is expected to be higher under centralized design. regarding the direct effect of income on preference for centralization. Table 1 summarizes our theoretical hypothesis Table 1: Conditional effect of income on preference for centralization of authority Possible Cases ρ < 0 ρ = 0 ρ > 0 Meaning More pressure on the Same pressure on the Less pressure on the local affluent class local affluent class local affluent class under centralization under any design under centralization Direct Effect of Income Negative near zero Positive Using the potential outcome framework as before, denote the average direct effect by ζ. Then we can write ζ(y) = E[C i (y 1, M(y))] E[C i (y 0, M(y))]. The hypothesis can be written as 3 For an empirical investigation of the effect of providing individuals with information about the actual distribution of regional income on their preferences for redistribution see Balcells et al. (2015). Here we assume that individuals have a relatively accurate perception of the regional redistribution of income. In our understanding, such assumption is not very strong for the empirical case under investigation. Further studies can explore that assumption empirically 8

9 H D : ζ(y, ρ > 0) = E[C i (y 1, M(y)) ρ > 0] E[C i (y 0, M(y)) ρ > 0] > 0 ζ(y, ρ < 0) = E[C i (y 1, M(y)) ρ < 0] E[C i (y 0, M(y)) ρ < 0] < 0 We can, alternatively, state the hypothesis about the conditional direct effect in the following way. As before, let y i denote the income of individual i, C i indicates her/his preference for centralization, and X i other systematic factors of i that affects his/her preference. Let ɛ i denote idiosyncratic factors of i affecting his/her preference and/or measurement errors not associated with y i or X i. Let C i = f(y i, X i, ɛ i ρ). The hypothesis regarding the conditional direct effect of income, denoted ζ i, can be stated as H D :ζ i = f(y i, X i, ɛ i, ρ < 0) = ζ i < 0 y i ; ζ i = f(y i, X i, ɛ i, ρ > 0) = ζ i > 0 y i ; f(y i, X i, ɛ i, ρ) > 0 ρ y i Traditional approaches use linear structural equation modeling (LSEM) to investigate mediation effect (Shadish et al., 2002; Baron and Kenny, 1986; Hayes, 2009). The outcome variable C i in our data is a binary measure indicating support for centralization (see section 4). Hence, the LSEM for our problem can be formulated by approximating f(y i, X i, ɛ i, ρ) linearly using a logistic transformation of the probability of support for centralization, and by using ρ as the interactive term. Let the index j denote the variable measured at the regional level. The structural equation model that captures the association we are interested in is ( ) p ln = α 0 + (α 1 + α 2 ρ ij )y i + α 3 ρ ij + αx i1 (1) 1 p R ij = β 0 + (β 1 + β 2 ρ ij )y i + β 3 ρ ij + βx i2 + ε ij2 (2) ( ) p ln = γ 0 + (γ 1 + γ 2 ρ ij )y i + γ R R i + γ 3 ρ ij + γx i3 (3) 1 p where ρ is = Y s Y n Ỹs Ỹ n ; p = P (C ij = 1 y i, R i, X i ) The conditional direct effect of income is given by ζ(y, ρ j ) = γ 1 +γ 2 ρ j. According to our hypothesis, we must have ζ(y, ρ j < 0) = γ 1 + γ 2 ρ j < 0 and ζ(y, ρ j > 0) = γ 1 + γ 2 ρ j > 0. The mediation effect in this structural equations is generally estimated by the product ( ˆβ 1 + ˆβ 2 ρ ij )ˆγ R. However, for cases with nonlinear models among the structural equations, as here with the logistic model, the product estimator of interest is inconsistent (Imai et al., 2010a; VanderWeele, 2009). Hence, in order to evaluate the mediation effect of preferences for redistribution, we performe a mediation analysis as advanced by Imai et al. (2010a, 2011). 9

10 In order to interpret the results of the mediation analysis and attach causal meaning to the results, we need to assume sequential ignorability (Imai et al., 2010a). Sequential ignorability means that, first, given the observed covariates, the main independent variable (the treatment ) is statistically independent of the potential outcomes representing the dependent variable. This first assumption is referred to as ignorability assumption in the literature (Rubin, 1977; Imbens and Rubin, 2015). The second part of the assumption states that the potential values of the mediator is statistically independent of the potential values of the outcome given the treatment and the (pre-treatment) observed covariates. In the context of this research, ignorability means that given age, education, and all the other covariates used in the analysis, the earnings of the respondents are not affected by their potential preference for redistribution and centralization. That is, whatever determines their income, it does not depend on which levels of redistribution or centralization they would prefer under different states of the world. We believe this assumption is reasonable to adopt in this study. The interpretation of the second part of the assumption in the context of this work is that, for a given level of income, ρ, age, and all the other covariates, individuals do not tend to prefer more (less) interregional redistribution because of their preference for centralization. In other words, they form they preference for interregional redistribution regardless of their preference for allocation of political authority. This second assumption can be more problematic. For instance, if some exogenous factor increase preference of individuals for centralization, say because they dislike the ideological position of the local representative, and they know that centralization will increase chances of interregional redistribution, it can cause the person to be less supportive for redistribution. If that is the case, sequential ignorability is violated. Such assumption is untestable and we need to rely of theoretical justification, particularly in the case of observational studies, as it is the case here. We take two standard measures to deal with the problem. First, we include all covariates regarded as important by the literature and that are available in our data. Additionally, we conduct a sensitivity analysis in the empirical section to test how a set of possible violations of the sequential ignorability assumption changes the results. 4 Data To evaluate the hypothesis we use an original national survey conducted in 2013 in Brazil. This survey includes questions about attitudes toward centralization of political authority, support for redistributive policies, and demographic characteristics of the respondent and his or her household. The variables used are the following. The outcome variable is Preference for Centralization of Authority, It is a binary variable combining different measures of the survey regarding preferences for distribution of political authority among the levels of government as used in Arretche et al. (2016), which they call preferences toward the architecture of the state. The variable denotes those that express support for centralization of the authority and those that oppose it. We use per capita household income (y i ) to study the direct and indirect effect of income on our outcome variable. The survey contains two questions to measure income: the approximate total income in the household and the number of people living in the house. The measure used is the inverse hyperbolic transformation of the per capita household income. We used the national 2010 Census to validate the income information in the survey. In order to construct the Difference in Ratio factor (ρ), we used the 2013 Household National 10

11 Survey (PNAD) conducted by the Brazilian Bureau of Statistics (IBGE 4 ). We use state level variables for Y s and Ỹs. We use two different mediators. The first is Preference for Interregional Redistribution (mediator 1). It is constructed using dimension reduction techniques (Factor Analysis) applied to three questions about interregional redistribution, all of them measured using a 5 points scale. The questions are 5 : (Q1) Do you X that the government should redistribute from wealthy to poor states? (Q2) Do you X that the government should redistribute public resources so that the poor states have public servies with the same quality of the wealthy states? (Q5) Do you X tha the government must approve legislation so that people have public services similarly in all the states?. The distribution of the mediator scores on the original variables can be seen in the figure 7. For the second mediator, Preference for Interregional and Interpersonal Redistribution (mediator 2) we performed the same procedure used in the mediator 1, but we include in the index a question about interpersonal redistribution. The questions is (Q6) Do you X that the government must redistribute to reduce the gap between rich and poor people? The distribution of the mediator scores on the original variables can be seen in the figure 8 We include the GDP per capita of the regions to control for the effect of regional development 6. Trust in the federal government is constructed using a binary indicator combining four questions that asked How much do you trust in the Y, where Y was substituted by municipal, state, and federal government. Finally, we include age measuered as discrete scale, a binary indicator for gender, and also a binary variable for education indicating individual has completed high school or more. 5 Empirical Analysis Figure 1 shows the empirical distribution of household per capita income at the national level for three different states, Santa Catarina (SC), Maranhao (MA), Tocantins (TO), and for the Federal District (DF). The dotted line represents the median income and the solid line the mean. The concentration of voters according to their income varies considerably among these states. In SC and DF, for instance, both the median and the mean income are higher than the national values. In MA and TO, those values are comparatively lower. Also, for DF there is a long tail in the right side of the distribution, and in MA most of the population is concentrated in the left side of the range considered. ***** Figure 1 HERE ***** The left panel of figure 2 shows the distribution of the states according to their mean and median household per capita income. Some states are shown in black dots, others using their abbreviation to facilitate visualization. If there was a direct negative association between wealthy of the region and preference for centralization, we would expect to see that people in DF, SC, SP, RJ and in all the states in the upper-right quadrant would be more likely to oppose centralization. Our results shows that this is not the case, as discussed below. The right panel of the figure 2 shows a quite different picture. The y-axis represents the difference in ratio, that is ρ. There we see that SC and DF are in opposite sides. For DF we have ρ DF > 0. 4 Available at 5 Identification number are arbitrary. X is a placeholder for strongly disagree, disagree, neither disagree or agree, agree, strongly agree 6 Data available at 11

12 It means that the reduction in the average income when we go from the local to the national values (Y DF > Y n ) more than compensate the reduction in the median income (ỸDF > Ỹn). Therefore, the redistributive pressure would be closer to the preference of a poorer median, who nevertheless is close to the average income. According to our conditional direct effect hypothesis (H D ), rich people would prefer a centralized scenario to mitigate the local pressure for redistribution on the local government, while poor people would prefer a decentralized design. Hence, in Df we would see a positive association between income and preference for centralization. The opposite relationship would be expected in SC. ***** Figure 2 HERE ***** We first estimate the equation (1) and equation (3) separately. The controls used are age, education, gender, trust in the federal government, and the level of development of the region. Table 2 presents the models that use preference for centralization as a dependent variable. It shows both the significant negative association between income and preference for centralization (as shown in our models denoted with odd numbers) as well as the change in the association when the preferences for interregional redistribution 7 (mediator 1) is included. The inclusion of the mediator in the models reduces the effect of income, as expected for variables that mediates the dependence between the other variables. We see that preference for redistribution increases the likelihood that the individual prefers a centralized design. The association is significant for all the models that includes this mediating variable. Models 3,4, and 5 as well as 8, 9, and 10 in table 2 present the association of the preference for centralization and the difference in ratio variable (ρ). The association is significant and has the expected sign according to our hypothesis. ***** Table 2 HERE ***** The conditional effect of income given ρ is presented in figure 3, which presents the predicted probability of supporting centralization as a function of income given different levels of the difference in ratio (ρ). The model used in the figure is number 10 of table 2. The left panel of the figure shows the positive slope of the probability of favor centralization when ρ is bigger than zero, particularly set at the DF level. The confidence interval is large in the figure, showing the uncertainty introduce by small number of cases in that range when compared to other values (see figure 2). It seems that ρ must be large so that the negative effect of income is reversed. The central panel of the figure shows the effect of income when ρ is near zero, which encompass the bulk of the cases (again, see right panel of figure 2). The right panel show the same income effect when ρ takes the most extreme negative case of our sample, that is, the one corresponding to SC. The figure supports the hipothesis about the conditonal direct effect of income. We have ˆγ 2 = ˆf(yi, X i, ɛ i, ρ j ) > 0, that is, the decreasing ρ j y i effect of income diminishes as ρ j increases. ***** Figure 3 HERE ***** Table 3 shows the regression of the preferences for redistribution on income, ρ, and the interaction between the two. Income is significant and negatively associated with preference for interregional redistribution as predicted. Also ρ and its association with income are not significant in any of the models, again as predicted. 7 Results for the mediator 2 are very similar and can be found in the appendix 12

13 ***** Table 3 HERE ***** As discussed, the product estimator ( ˆβ 1 + ˆβ 2 ρ ij )ˆγ R for the mediated effect is inconsistent when there is nonlinear equations in the structural model. In order to estimate the effect consistent, we adopt the procedure proposed by Imai et al. (2010a, 2011) and Tingley et al. (2014). Figure 4 shows the direct effect of income ( ζ) in the left column for different values of ρ. In the right panel, we see the mediation effect of income ( δ). We select a base value of the income variable y 0 = 0.23 to use as the control condition in all the panels. Different treatment conditions y 1 are shown in the x axis. The top two panels shows the direct (at left) and mediation (at right) effect of income. It shows that on average there is a negative effect when ρ is negative, as expected. The 95% confidence interval include zero, probably due to few cases in that range of ρ. The central panels clearly show the significant negative direct and mediated effect, also as expected by the hypothesis. In the bottom panels we have ρ > 0 and the negative trend vanishes. In other words, when centralization increase the pressure for redistribution (ρ < 0) more income is associated with less support for centralization. When centralization means less pressure for redistribution relative to the pressure in the region (ρ > 0), the evidence of negative association of income an preference for centralization vanishes. Using the central panels to compute the share of the mediation effect, on average.8% of the effect of the income on preference for centralization is mediated by the preference for redistribution. The biggest share of the variation on preferences for redistribution is due to the direct effect. ***** Figure 4 HERE ***** We go further and do the same exercise using levels of ρ as the control and treatment variable, and income as the conditional variable. Figure 5 shows the results. When the income is positive, increasing ρ produces a positive and direct effect on preference for centralization, as we see in the bottom and center left panels. When the income is negative (top-left panel), the trend disappears and seem to reverse. In other words, the bottom and center left panels show that, among the local rich, the lower the relative pressure for redistribution under centralization (ρ increasing), the more those individuals favor centralization. The right panel shows that the mediation effect of redistribution for ρ is near zero for any level of income, as predicted by the hypothesis. ***** Figure 5 HERE ***** We also repeat the analysis using three groups of income instead of a continuous measure. The first group is composed by those individuals with income below the first decile of the distribution. The second group includes those with income aboe the first and below the ninth deciles. Finally, the third group includes those in above the ninth decile. This coding scheme captures the largest conditional direct and unconditional indirect effect of income shown in the central panel of figure 4. We hold ρ at zero. We use that recoded income variable to performe the sensitivity analysis. The left panel of figure 6 shows the average direct (ADE) and average mediation (ACME) effect of the income when we use the first decile as the control and the ninth as the treatment group. The regression tables using the categorical income variable are shown in the appendix. The results are the same as before. As discussed, sequential ignorability is a fundamental assumption for the interpretation of the results of the mediation analysis. In order to verify how the unobserved covariates that affect both the mediator 13

14 and the outcome would affect the estimated quantities for ζ (ADE) and δ (ACME), we use a sensitivity parameter ν. That parameter represents the correlation between the residuals of equations (3) and (2). The parameters is not zero if the sequential ignorability is violated (see Tingley et al. (2014); Imai et al. (2010a, 2011, 2010b) for details). The figure 6 shows how sensitive our results are for breaking the sequential ignorability assumption as captured by the parameter ν. The two panels at the bottom-right of figure 6 shows that for ν.2, the 95% conficence interval starts crossing the zero line for the ADE. The average value is almost always negative, except for values of ν near the maximum possible negative correlation. The two top-right panels of the same figure show the sensitivity of the mediation effect ζ (ACME). As we can see, in this case the result is very sensitive to violations of the sequential ignorability assumption. Together, the tables 2 and 3 and the mediation analysis presented in the figures 4 and 5 show trends and tell a story that support our hypothesis. Briefly, income has an unconditional (on ρ) negative indirect effect and it has a conditional (on ρ) direct effect on preferences for centralization. In both cases, the evidences support the hypotheses. In the case of the direct effect, as the electoral pressure for redistribution tend to be higher under centralization, the rich prefer decentralized design and poor people tend to prefer centralization. The opposite relationship occurs when centralization diminishes the pressure for redistribution. Finally, the effect of income on preference for centralization is partially mediated by preferences for redistribution. Income is negatively associated with preferences for interregional redistribution, and the latter in positively associated with preferences for centralization. 6 Conclusion Our results show that the mediation effect of preferences for redistribution on preferences for the (de)centralization is small, but present as predicted by the theory. We have also shown that the direct effect of income is conditional on ρ. That variable intends to capture the perception of individuals about alternative contexts of electoral pressure for redistribution that would most likely emerge under different allocations of authority. Income has a positive direct effect on preferences for centralization when pressure for redistribution under a centralized design tends to be smaller than under the alternative design. More precisely, the larger the potential electoral pressure for redistribution under decentralized design (the larger the ρ), the more income tends to be positively (less negatively) associated with centralization. The results are consistent with the argument stating that underlying the voters preference for institutional design there is their perception of the relation between (de)centralization and the electoral pressure for redistribution. As final words, it is important to mention that we capture the voters perception of consequences of centralized design rather indirect using the variable ρ. Further research can be conducted in order to evaluate how ρ actually matches the perception of individuals about the regional distribution of income. We believe the way we have framed the problem as well as our findings can contribute to open the debate in that direction and suggest interesting hypotheses for future investigation. 14