Trust, Welfare States and Income Equality: What Causes What?

Transcription

1 Trust, Welfare States and Income Equality: What Causes What? Andreas Bergh Lund University and the Research Institute for Industrial Economics Christian Bjørnskov * Aarhus University July 26, 2011 Abstract: The cross-country correlation between social trust and income equality is well documented, but few studies examine the direction of causality. Reviewing the literature and using a prisoner s dilemma that allows payoff inequality, we show that under plausible circumstances, trust aids the creation of welfare states, which reduces inequality. Trust may also directly affect the income distribution. Using a structural equations model estimated with data for 104 countries, we confirm three hypotheses: 1) Trust has a direct positive effect on gross and net income equality; 2) Larger welfare states lead to higher net equality; and 3) Neither net income equality nor welfare state size causally affect trust, but gross inequality negatively affects trust. We conclude that while welfare states reduce net inequality, this will not increase trust. Keywords: Social trust; Inequality, Welfare State * Bergh: The Research Institute for Industrial Economics, Grevgatan 64, SE Stockholm, Sweden, and Lund University, andreas.bergh@ratio.se. Bjørnskov: Aarhus University, Department of Economics and Business, Hermodsvej 22, DK-8230 Åbyhøj, Denmark, chbj@asb.dk. Bergh thanks Vetenskapsrådet for their financial support. 1

2 1. Introduction The positive cross-country correlations between social trust, welfare state size and the level of income equality are well documented in the literature. Few studies, however, examine the direction of causality in these relationships. While most studies of trust and income distribution assume that growing inequality leads to a decrease in trust (Uslaner, 2002; Delhey and Newton, 2005; Bjørnskov, 2007, 2008), the theoretical mechanisms involved are still not sufficiently explored, as shown in overviews by Jordahl (2008) and Nannestad (2008). A similar situation applies to research regarding the correlation between welfare state size and trust, where for example Barr (2004) and Kumlin and Rothstein (2005) argue that universal welfare state policies create trust. Nannestad (2008), however, notes that the correlation might as well reflect the opposite causal direction, as high levels of trust and trustworthiness would insulate welfare states from immediate moral hazard, free riding and incentive problems. Indeed, recent studies provide evidence for a causal link from historical trust levels to current welfare state size (Bergh and Bjørnskov, 2011). This finding does not exclude the possibility of bi-directional causality in which equality also increases trust, but this link has yet to be empirically explored. Doing so will give information on the extent to which redistributive policies, central to the modern welfare state and intended to increase income equality, thereby also increase trust. In this paper, we therefore investigate the causality between inequality and trust, and the potential role of welfare state institutions and policies as moderators of the causal associations. Our aim is illustrated in Figure 1, where full arrows denote associations argued for in the literature while dotted arrows denote the additional possibilities explored in the following. Insert Figure 1 about here 2

3 Theoretically, this paper contributes by using a simple prisoner s dilemma game to infer plausible directions of causality. We use this theoretical setup to structure our discussion of the different possible mechanisms that may explain the associations one observes. We then estimate structural equations models that account for potentially substantial feedback effects between income equality and social trust. We start the paper by outlining a number of potential theoretical links between social trust and inequality, based on recent studies. In section 3, we describe our empirical strategy to sort out the causal directions. Section 4 describes the data used in section 5 to estimate the relation. Section 6 discusses our findings and concludes. 2. Theoretical links between social trust and inequality Theoretical and experimental analyses of trust typically use a cooperation game such as the prisoners dilemma (PD). Well-known examples include Axelrod (1984) and Bendor and Swistak (1997). In disciplines including political science, economics and theoretical biology, a large literature studies the circumstances under which cooperation can evolve and be stable. For overviews, see Gintis, et al. (2005), Hammerstein (2003), Ostrom and Walker (2003) and Cox et al. (2009). 2.1 Trust and cooperation The literature shows that countries with high levels of social trust achieve better outcomes in several dimensions such as economic growth and quality of governance. The standard explanation, stated either explicitly or implicitly, is that trust aids societies in reaching cooperative outcomes in large-n cooperation problems (Sønderskov, 2011). Interestingly, while cross-country correlations are well established, the individual level evidence is less clear, and experimental results have been mixed. For example, Glaeser, et al. (2000) show that survey-measured trust predicts trustworthy rather than 3

4 trusting behavior in economic experiments, and Holm and Danielson (2005) find that survey-measured trust predicts experimental behavior in Sweden but not in Tanzania. 1 Outside of economic laboratories, Sønderskov (2011) nevertheless finds a link between the standard survey measure of social trust and actual cooperation, measured as recycling behavior. Importantly, the effect of trust is significant only when people perceive recycling to be a true collective action dilemma. Likewise, Sapienza et al. (2007) find that trust is a significant determinant of behavior in experiments, but only when the stakes in the games are economically significant. Theoretically, using the prisoners dilemma to analyze trust and cooperation means that distributional issues are ignored. In the prisoner s dilemma it is implicitly assumed that the gains from cooperation are shared equally, and the same holds for the outcome when agents fail to cooperate. To see this, consider the standard two-player prisoner s dilemma shown in Table 1. Here, α is the reward for cooperation, β is the temptation pay-off and γ is the payoff when agents fail to cooperate. The pay-offs are normalized so that the outcome for cooperating when the other agent defects (the sucker payoff ) is 0. When both players cooperate they each get α, and when both fail to cooperate they get γ. This is clearly enough to analyze highly complex issues regarding efficiency and stability (when the game is repeated), but it says little about whether societies where agents cooperate more often will be more or less equal compared to less cooperating societies. Insert Table 1 about here 1 Although the experimental results in Glaeser et al. (2000) have received substantial attention, they ought to be treated with care. As stressed by Bjørnskov (2010), Glaeser et al. failed to anonymize their experiments, instead allowing people who knew each other to play head to head. Behavior in the experiments is therefore not likely to reflect social trust, but more likely captures the degree of particularized trust that subjects have in each other. A similar problem may cause Holm and Danielson s (2005) findings in Tanzania. 4

5 2.2 A prisoner s dilemma with payoff inequalities To analyze issues of distribution using the PD-game, assume that the payoffs when both agents cooperate are (α 1, α 2 ) rather than (α, α). Similarly, the outcomes when players do not cooperate will be (γ 1, γ 2 ). Relative inequality can then be analyzed by comparing the ratio α 1 /γ 1 to α 2 /γ 2. Figure 2 illustrates the possible outcomes. Without cooperation, agents 1 and 2 end up in point a, which is the inefficient allocation (α 1, α 2 ). In point b they cooperate and receive payoffs (β 1, β 2 ). If societies with higher social trust reach the cooperative outcome b more often, the consequences of trust for inequality depend on the distribution in point b relative to point a. Formally, we call this the assumption of equalizing cooperation: Assumption 1 (Equalizing cooperation):, implying that inequality will be decreasing in the degree of cooperation. In the case shown, point c is the cooperative outcome of the PD-game when payoffs are restricted to the 45-degree line of equality. In the situation depicted, allocation b is less equal than c, but more equal than a. Insert Figure 2 about here 2.3 Is cooperation equalizing? To assess the relevance of Assumption 1, consider first the relative difference between the two agents if they fail to cooperate the Hobbesian state of nature. Unless agents are identical, γ 1 = γ 2 is unlikely. At least two equally simple assumptions seem more appropriate. One is that in the absence of cooperation resources are distributed proportionally to some individual ability that differs between agent 1 and agent 2. In a primitive society, this ability may be physical strength, as in the jungle equilibrium 5

6 analyzed by Piccione and Rubinstein (2007). In a developed society it may be some cognitive, communicative or political skill. A second possibility is that, in the absence of cooperation, a conflict arises where the stronger agent wins all resources, leaving the other one with nothing. In other words, a society without cooperation will most likely not only be poor but also relatively non-egalitarian. Consider now the gains from cooperation, i.e. α 1 - γ 1 for agent 1 and α 2 γ 2 for agent 2. In political philosophy, there are numerous accounts of how the gains from cooperation should be shared. Normative accounts have been provided for the maximin solution (Rawls, 1971), a utilitarian solution (Harsanyi, 1955) and a distribution based on equal relative concession corresponding to the Kalaia-Smorodinsky bargaining solution (Kalai and Smorodinsky, 1975) advocated by Gauthier (1987). There is also some theoretical work on how evolution will lead agents to share the gains from cooperation. Binmore (1994) argues that evolution will lead agents to use the Nash bargaining solution. Similar attempts to naturalize various conceptions of normative political theory have also been made by for example Skyrms (1996) and Sugden (2005). As discussed at length in Roemer (1996), when the utility possibility set is symmetric and convex for example a straight line all of the approaches mentioned lead to similar outcomes, and in many cases they coincide with the egalitarian outcome. Only when the utility possibility set is highly asymmetric and/or concave will different normative theories and bargaining solutions lead to clearly different outcomes. There is also some experimental evidence on how agents actually share the gains from cooperation in various experimental situations. Interesting examples include Hennig- Schmidt (2002) and Ciriolo (2007), both supporting the notion that sharing equally is a focal point. Hennig-Schmidt (2002) shows that people are eager to argue that the gains from cooperation should be divided in a way that is fair and equal, but also detect a selfserving bias in exactly how fairness is interpreted by different agents. Ciriolo (2007) re- 6

7 analyzes data from a trust-experiment and argue that the behavior that was being interpreted as trust and reciprocity was in fact much easier to explain: participants simply acted in order to share equally the gains from participation in the experiment, including the show-up fee paid by the experiments. Finally, an important reason for expecting the allocation b to be more equal than a is if the gains from cooperation derive from equalizing insurance arrangements that trust and trustworthiness help sustain and protect against free-riding. This holds for both informal sharing arrangements in primitive societies as well as for the formal social insurance systems of modern welfare states. As rightly noted by Rothstein (2001), the universal welfare state is a social dilemma, and as shown by Bergh and Bjørnskov (2011), countries where the population was historically more trusting have been more successful in creating and maintaining such welfare states. The notion that trust aids the implementation of insurance schemes is also demonstrated by the findings in Zhang et al. (2006) regarding community based health insurance in rural China. In this case, the gains from cooperation come from a more equal income distribution across states, as described in any text book on the economics of the welfare state, such as Barr (2004). An argument along the same lines is outlined by Rothstein and Uslaner (2005) who suggest that social trust leads to lower levels of corruption, and that corrupt public bureaucracies and polities reduce the efficiency of redistributive policies. As such, this is a leaky bucket type argument well-known from public choice theory, although with a leakiness depending on the level of trust in society (Okun, 1975). 2.4 The link from the welfare state to trust Another argument proposed by Kumlin and Rothstein (2005) is that the impartiality of universal welfare states will lead people to infer that most people can be trusted. If this holds, there is a positive, reinforcing trust-welfare state dynamic. On the other hand, 7

8 universal welfare states provide several opportunities for free-riding on publicly financed goods and benefits (Buchanan, 1975; Coate, 1995; Lindbeck, 1995), and if people believe that such temptations make us less trustworthy, trust will decrease. In this case, there will be a negative trust-welfare state dynamic, where trust aids the construction of the welfare state, but the existence of the welfare state erodes trust. As we do not have time varying data, we cannot discriminate between these two cases, and consequently we have no theoretical prior regarding the effect of welfare state size on trust. However, these are not the only ways in which welfare states and trust may be associated. Taking the recent literature seriously first of all means that one has to question the basis of Kumlin and Rothstein s hypothesis. Central to it is the assessment that people generalize from the observable behavior of bureaucrats to most other people. As the welfare state rests on bureaucrats alleviating social wrongs, the existence of generalized redistributive systems means that one observes honest behavior from the system and infers that most people are trustworthy. However, there is very little evidence that people actually generalize in this way (e.g., Uslaner, 2002; Naef and Schupp, 2009). Instead, even without feedback mechanisms creating free-rider problems such as those in Buchanan (1975) and Lindbeck (1995), the existence of a welfare state could in principle cause ambiguous effects on social trust. As sketched in Bjørnskov (2008), having a strongly redistributive state could from a specific ideological point of view mean that the wrong that is inequality is corrected, which means that fairness norms are enforced by society as a whole. Conversely, in opposite ideological weltanschauungen, the existence of a welfare state could be seen as a signal that society is unfair, by punishing those who through effort, skill and talent succeed economically. With a leftwing conviction, people might see its existence as evidence that most people, not least those better off, are unfair and should not be trusted. With a rightwing conviction, redistribution of income that was created in a fair way would appear unfair, and 8

9 perhaps a signal that groups in society are parasitical and therefore not to be trusted. What one can conclude is merely that the welfare state trust association is far from as unambiguous as most of the political science literature would have it. 2.5 The link from equality to trust Even in our simplified framework, there are several reasons to expect a negative feedback from inequality to trust. First, it may appear likely that people infer the trustworthiness of others by observing past actions and outcomes. If assumption 1 holds, then inequality will be a signal that others have behaved less trustworthy in the past, in which case trust may decrease. We nevertheless note a complication with this type of argument: if inequality provides a signal that people were actually less trustworthy and cooperative in the past, or if they were perceived to be so, the signal will be transmitted through levels of gross inequality. Conversely, net inequality includes the original signal, but also whatever quasiredistribution occurs due to individual and welfare state responses to the original signal. In case assumption 1 holds, net inequality will only affect trust if quasi-redistribution somehow rights the wrong that observed levels of gross inequality signal. If not, we would expect to see only gross inequality affecting trust while redistributive policies would be inconsequential. A second possible mechanism is the following: regardless of whether assumption 1 holds or not, there are reasons why β, the temptation payoff, may be increasing in inequality. For example, when the gains from cooperation come from trade and property rights, the poor will have more to gain from defection when inequality is high. This is the argument outlined by Gustavsson and Jordahl (2008) who note that when income differences are relatively large, poorer people have more to gain and richer people more to 9

10 lose from theft and robbery as well as from forceful, redistributive policies. 2 In this type of scenario, redistributive policy will therefore cause lower crime rates, which will subsequently increase trust levels of those potentially affected by property crime. Third, Bjørnskov (2008) outlines a number of normative-ideological reasons for expecting inequality to affect social trust. Uslaner (2002, 181) assesses that the rich and poor have little reason to believe that they share common values, and thus might well be wary of each other s motives if the income distance is sufficiently great. However, individuals assessments of other s motives and their association with inequality are likely to rest on a normative-ideological basis (Bjørnskov, 2008). While chances for misinterpretation and suspicion towards motives and actions are much more pronounced when people are different, the salience of that difference depends on how it comes about. In a Marxian mental model, there is no reason to trust the rich as they will have become rich at the expense of the poor both motives and actions are thus dishonest. However, in a libertarian model, the income distance holds no information of the motives or honesty of people belonging to different income groups. In a broader optic, individuals may arguably find it difficult or impossible to assess the motives and evaluate the basis of social phenomena when a substantial share of fellow citizens does not share their general weltanschauung (cf. Rokeach et al., 1960). As such, income inequality may affect social trust to the extent that it is related to inequality / intra-country variability in values, mental models and beliefs about society. 2.6 Hypothesis regarding trust, inequality and welfare state size 2 While this idea is consistent with rational behavior in a standard economic framework, we note one major problem. The argument implies that richer people should rationally be less prone to trust their fellow citizens, as they will be the object of dishonest behavior. However, in the vast majority of countries, the opposite is the case rich people tend to be more trusting. 10

11 Based on the reasoning above, we now formulate the following hypotheses for empirical testing: H1: Countries with higher trust levels will have bigger welfare states where larger social insurance systems result in lower inequality of disposable income. H2: If Assumption 1 holds, countries with higher trust levels will have lower gross as well as net inequality also when controlling for welfare state size. Importantly, these two hypotheses do not exclude the possibility of a feedback from inequality to trust. On the contrary, the links reviewed in section 2.5 suggest that H3: Inequality decreases trust Regarding the link from the welfare state to trust, we have identified two theoretical mechanisms with incompatible implications. We conclude that the link from welfare state size to trust is theoretically ambiguous, and proceed to empirical testing. 3. Identifying causality using a structural equations model To test our hypotheses as formulated above, we employ a structural equations model, where the long-run equilibrium determinants of trust and inequality are estimated simultaneously, allowing for a varying degree of endogeneity. Figure 3 outlines the overall model that we estimate. Insert Figure 3 about here The validity of our approach rests on the assumption that there are variables specific to either part of the system that are properly exogenous, i.e. factors that affect trust but not inequality and vice versa. It also necessarily rests on sufficiently strong identification of both parts. As such, the structural approach is in principle similar to two linked instrumental variables estimates with the specific variables as instruments. It is nevertheless more flexible than standard instrumental estimators by allowing explicitly for two-way causality and covariance of the error terms. 11

12 To infer exogenous variation in inequality, we use a standard Kuznets curve (GPP per capita and its squared term), a political Kuznets curve (the degree of democracy and its squared term), a set of dummies capturing variations in religious affiliation, and a control for common law countries (i.e., countries with some form of British heritage). Finally, we include IQ scores as proxies for educational quality to control for inequality arising from skills-based technological progress (Johnson, 1997; Lemieux, 2006). Similarly, exogenous variation in trust is inferred using a dummy variable for monarchies and a measure of religiosity. We include religiosity following Berggren and Bjørnskov (in press) while monarchy is included based on previous research in Bjørnskov (2007, 2010) and Robbins (in press). Variables common to both inequality and trust are two dummies for postcommunist countries that are both less trusting and more equal, and for Nordic countries, that have more equal income distributions, large welfare states, and are substantially more trusting, all other things being equal. The system is then estimated simultaneously with the predicted values for trust as an independent variable when explaining inequality, and predicted inequality values to explain trust. This allows us to identify either a two-way causality, or a one-way causality in either direction. 3 However, while this type of modeling may be more flexible than standard twostage least squares, it obviously rests to an equal extent on having variables specific to one equation that are exogenous to the residual of the other equation. We have therefore used separate two-stage least squares regressions to test the exclusion restrictions of the structural system. We in general find no evidence of overidentification problems with the excluded variables in either part of the system. As we note that the goodness-of-fit statistics 3 Note that the causal relation from trust to welfare state size was demonstrated in Bergh and Bjørnskov (2011). We therefore do not model welfare state size explicitly, but only treat it as a potentially moderating variable in the following. 12

13 also reveal that both parts are properly predicted and that specific variables hold sufficient predictive power, we are unlikely to suffer from underidentification problems. A potential problem might be that a substantial part of the identification occurs as a difference between rich, middle-income and very poor countries, for which redistributive policies are not a real option. This problem would make measures of net and gross inequality almost identical in part of our sample, which would prevent proper identification of the likely effects of redistributive policy that is captured through the difference between net and gross measures. We therefore report all main results in the full 104-country sample as well as a reduced sample in which we only keep countries with a GDP per capita above 2000 USD. 4. Data Measuring income inequality and social trust both represent potential problems. Inequality in particular is measured in a variety of ways, and estimated from surveys and other information of very varying quality and coverage (cf., Deininger and Squire, 1996). To be properly comparable, the large variety of measures therefore needs to be standardized. We use data from Solt (2008), who categorizes income inequality calculated from different types of income measures into 21 categories. He thereafter estimates all differences between these inequality measures, thereby obtaining a large set of ratios. These ratios are then used to generate predictions in other categories of inequality, generated through multilevel modeling. The final measures of gross and net income inequality therefore not only take advantage of all available information, but also of the quality of that information. The result is a large database of Ginis that are as comparable as practically possible, given the information available from each country and each year. This provides us with a large dataset from which we take the average of all available measures of net and gross income inequality between 2002 and This gives us a 13

14 dataset with full information on all main variables of 104 countries, spanning the globe and all levels of economic and political development. In order to test the associations between inequality and social trust, we combine data from a number of sources with the Solt (2008) measures of inequality. First, we follow the literature by employing what has become the standard measure of social trust over the last decade. Trust is measured as the percent of a population that answers yes to the question In general, do you think most people can be trusted? In our full sample of 104 countries, this measure varies between a low of 3.4% in Cape Verde and a high of 68.1% in Denmark; the other Scandinavian welfare states, Norway and Sweden, have only marginally lower trust scores. The full sample average is 25.3%, putting into perspective how high the Nordic trust scores are. We draw these data from the five waves of the World Values Survey, combined with data from the AfroBarometer, Asian Barometer, East Asian Barometer, LatinoBarometro, and the Arab Barometer. We get the full dataset from Berggren and Bjørnskov (in press). Despite theoretical criticism, the trust measure in general does well in various validity tests and is remarkably stable over time. 4 Although the question a priori seems unclear and imprecise, it correlates strongly with return rates in wallet-drop experiments, which capture the honesty of people in a situation that is unobservable to formal institutions (cf. Knack and Keefer, 1997). In-depth interviews and extended surveys show that this simple question measures respondents trust in strangers, while their trust in individuals they know or have specific information on is a quite different concept (Bjørnskov, 2007; Naef and 4 The examples of obviously non-stationary trust scores since the early 1980s include the much-discussed declines in the USA and the UK (though this drop only occurs in the World Values Survey but not in the contemporaneous British Social Attitudes Survey), but also positive trends in Denmark and Uruguay. Yet, trust is trendless in almost all other countries (Bjørnskov, 2007; Uslaner, 2008). We therefore use the average of all available observations. 14

15 Schupp, 2009; Uslaner, 2002, in press). Nannestad (2008, 419) also emphasizes that respondents do not in general seem to find the generalized trust question difficult, as only roughly five percent of respondents in developed countries do not or cannot respond to the question. Recent evidence furthermore indicates that trust questions do well in explaining observed behavior in experiments. Actual trusting behavior in ultimatum games, trust games and public goods games correlates with respondents answers to the trust question, although perhaps primarily in situations in which the stakes are of economic significance (Sapienza et al., 2007; Cox et al., 2009). Yet, one exception is Holm and Danielsson s (2005) of trust in Sweden and Tanzania, which suggests that the trust question may be a rather noisy measure in the least developed countries. As for control variables, we keep the trust specification to a relative minimum. As Nannestad (2008) stresses, only a few variables have proved to be robustly associated with trust. We predict social trust by dummies for postcommunist countries, constitutional monarchies and the Nordic countries, and the Gallup (2009) measure of religiosity, following recent work in Berggren and Bjørnskov (in press). To this simple specification, we add one of the two measures of inequality. As determinants of inequality, we first of all include variables controlling for two types of Kuznets curves. We include the logarithm to GDP per capita, measured in PPP adjusted US dollars from Heston et al. (2006), as well as it squared term. This effectively controls for the traditional Kuznets curve (Kuznets, 1955), which states that inequality is increasing in economic development at low levels of development, due to sectorally unbalanced growth, while it will be decreasing in development from some level, as other sectors and political redistribution outweighs the initial effects. Second, we also add the more recent explanation of a political Kuznets curve (cf., Chong, 2004). The rationale is that with increasing democratic influence, larger groups in society will become politically influential and thus be able to tweak redistributive policies in their favor. At some point, 15

16 the electorate and thus redistributive politics becomes sufficiently representative of the entire population that further democratic development will cause a more even distribution of incomes. As such, the subsequent inclusion of poorer groups of voters in politics gives rise to a non-linear version of Director s Law (Stigler, 1970). A number of other potential explanations need inclusion. We add further variables capturing common law systems (i.e., Anglo-Saxon cultural backgrounds), religious affiliations, intelligence quotients that proxy for the quality of education and hence skillbias explanations to income inequality, and dummies for postcommunist and Nordic countries that for obvious reasons may be different. We also include a set of regional dummies capturing differences between Asia, Latin America and the Caribbean, North Africa and the Middle East, and Sub-Saharan Africa; the comparison group is thus countries belonging to the Western Hemisphere. Finally, we consider welfare state characteristics as potential moderators of the trustinequality association. To capture the relevant factors, we employ three variables in subsequent analysis. Government final expenditures (percent of GDP) from Heston et al. (2009) proxy for the provision of public goods; government full expenditures from CIA (2010) capture the full activities of the government sector including both public goods and tax-transfer schemes, i.e. all redistributive effort; and transfers and subsidies (percent of GDP) from Gwartney et al. (2010) specifically capture directly redistributive policy. In a set of robustness tests, we include further control variables informed by the literature: ethnic diversity, measured as one minus the Herfindahl-Hirschman index of ethnicity, deriving from Alesina et al. (2003), the absence of corruption from Transparency International (2010), urban population (percent of total) and population density from World Bank (2010), coup intensity (coups per decade) from Marshall and Marshall (2009), and trade volumes (percent of GDP) from Heston et al. (2009). 16

17 5. Results As a first impression, Figures 4 and 5 provide scatter plots of the inequality-trust association, measured by gross and net inequality, respectively. Both figures depict the clear negative association identified in the literature, although they also exhibit one particular feature: while the Nordic welfare states have some of the lowest levels of net inequality, gross inequality, i.e. pre-transfer, pre-tax inequality, is not particularly low in these countries. It is this regularity that has caused some of the special attention to these countries and the welfare state in particular. Yet, the record Nordic trust levels could be caused by the obvious redistribution or they might reflect that trust allows a Nordic level of income redistribution (cf., Nannestad, 2008). Insert Figure 4 about here Insert Figure 5 about here Table 2 presents the formal results of estimating our two-way system between income inequality and social trust. Throughout the tables, we employ net income inequality in columns 1-2 and gross inequality in columns 3-4. Odd-numbered columns report the results of the trust specification while even-numbered columns report the inequality results. Insert Table 2 about here First, the estimates of trust in general provide support for the simple specification. In columns 1 and 3, the specification explains about 67% of the variation, and all variables are highly significant. Monarchies and Nordic countries are more trusting while postcommunist and more religious countries are less so. The results thus replicate the main findings in the most recent studies in the field (Berggren and Bjørnskov, in press; Robbins, in press). With respect to the main variable, we find that both net and gross income inequality are strongly significant, yet the point estimate of gross inequality is substantially larger than 17

18 that of net inequality. At first sight, gross inequality may therefore seem more important than net inequality by simply comparing coefficients, but this conclusion is not clearly valid. As the former is less variable than the latter across the full sample, the raw point estimates cannot be compared. When comparing the marginal effects of a similar one standard deviation shock, the effects of net and gross income inequality are almost identical. The results across the even-numbered columns provide support for a Kuznets curve in the full sample. With net inequality, the top point of the Kuznets curve top occurs at a GDP level of 4587 USD per capita; with gross income, the top point is at 8443 USD. We also find ample evidence of a political Kuznets curve as in Chong (2004) with a top point at a Polity score of approximately across columns 2 and 4. Furthermore, the results show consistently lower net and gross inequality in postcommunist countries but no significant difference in terms of redistribution, as the point estimates do not differ significantly. As such, lower inequality seems a structural remain of communism more than a current policy outcome. We also find evidence of more redistribution and lower net inequality in the Nordic welfare states and less redistribution in common law countries as well as substantial religious differences. Finally, the evidence points to a skills premium (captured by IQ) as suggested in recent studies, which is curiously only significant for net inequality. With respect to our main variables, the estimates suggest that social trust exerts a significant influence on net and gross inequality. In terms of quasi-redistribution, i.e. the difference between net and gross inequality, the results reflect that the effect of trust on gross income inequality is significantly stronger than that on net inequality. Indeed, comparing marginal effects (betas), that on gross inequality is almost double of that on net inequality. The results in the full sample thereby point to a clear two-way causality between income inequality and social trust, which is nonetheless statistically and quantitatively 18

19 stronger for gross inequality than for inequality after redistribution has taken place. For two reasons, we now repeat this analysis for a sample of countries that excludes the absolutely poorest countries (defined as those with a GDP per capita below 2000 USD): 1) these countries have very little political redistribution and the difference between net and gross measures is thus likely to be due primarily to noise; and 2) although the trust measure works very well in most cases, there are reasons to believe that the measure itself as well as the surveys it derives from are questionable for the very poorest countries (cf. Holm and Danielson, 2005). We report the results in Table 3. Insert Table 3 about here Although the restricted sample covers only 89 countries we exclude the 15 least developed countries of the full sample the results are remarkably similar. The traditional Kuznets curve now has a consistently negative slope, reflecting that the bulk of the sample is middle-income and rich countries. The political Kuznets curve continues to receive substantial support and intelligence remains weakly significant in column 2. The results in columns 1 and 3 also remain significant with one important difference: in the reduced sample where the two types of inequality are more readily separable, net inequality is no longer a significant determinant of social trust. Conversely, gross inequality remains significant and trust remains a significant determinant of both measures of inequality. Using the restricted sample, if we test the long-run effects of some shock that provokes a one-standard deviation increase in social trust (13.5 percentage points), it eventually turns into a 14.5 point increase in trust due to the feedback from gross income inequality. In the long run, gross inequality drops by 3.7 points (50% of a standard deviation). Conversely, a one standard deviation shock to gross income inequality (7.6 points) in turn becomes an increase of 8.2 points through feedback effects from trust and induces a drop in trust of 2.3 percentage points (17% of a standard deviation). Hence, it 19

20 seems rather clear that the mechanism from trust to inequality is substantially stronger than the other way. As such, the evidence mainly speaks against the standard interpretation of the trustinequality association. In particular, we find no evidence of effects of redistribution, as net inequality does not affect trust. We therefore devote the next section to testing the robustness of these findings and probing into whether we can identify likely transmission channels suggested in the existing literature. 5.1 Robustness tests In the following, we explore the robustness of the findings to a number of potential other variables and factors. In Table 4, we first test the findings when including three measures of government size: government full expenditures, including all transfers and subsidies, government final consumption expenditures, which exclude transfers and subsidies, and transfers and subsidies as percent of GDP. The table therefore tests for direct effects of welfare states and redistributive efforts that all other things being equal will be associated with larger government sectors. While we find an unsurprising effect of full expenditures on net inequality, as these expenditures include all redistributive transfers, there is no effect on gross inequality, and hence no likely effect on social trust (cf. Bergh and Bjørnskov, 2011). Government final expenditures that will include all public goods provision are never significant while we find that transfers and subsidies reduce both measures of inequality, though the effect is substantially larger for net inequality. We thus find no evidence for direct effects of welfare state spending, nor do our results indicate an indirect effect through redistribution. Insert Table 4 about here A set of further results are reported in the appendix. Table A2 first tests for the robustness to two characteristics often associated with both trust and inequality: ethnic 20

21 diversity, which is likely to weaken the support for redistribution, and corruption (measured as absence of corruption) that may arguably reduce the effectiveness of redistribution and could also benefit elite rent-grabbing, which would tend to increase gross inequality (Glaeser, 2006; Rothstein and Uslaner, 2005). While we find only insignificant support for the effects of ethnic diversity and some indication of effects on corruption although insignificant, the difference between the point estimates is significant we note that their inclusion does not change the main results. Table A3 instead provides evidence for the robustness to including measures of political instability and openness. We find that politically less stable countries tend to have significantly larger net inequality, yet the main difference is that the estimate is insignificant for gross inequality. On the other hand, trade openness significantly increases gross inequality, but consistent with Dreher and Gaston (2007), does not affect net inequality. Again, our main results remain unchanged. Table A4 tests for two other factors potentially associated with inequality and trust: the relative size of the urban population and population density. We find that countries with larger urban populations have relatively less equal income distribution as the association is only weakly significant for gross inequality. However, consistent with cost explanations of redistributive efforts, we also find that more densely populated countries have more gross inequality but not more net inequality. In other words, quasi-redistribution appears to be substantially more effective in countries with high population density. Yet, these inclusions do not affect our main results. Finally, Table A5 provides a particularly strong test of the relevance of our results by excluding all poor and lower middle-income countries. In the table, we only retain countries with a PPP-adjusted level of GDP per capita above 7000 USD. In other words, all 55 countries included are clearly economically able to pursue substantive redistributive policies. We note that while the association between trust and net inequality becomes 21

22 insignificant in both directions in larger samples, the causal direction from trust to net inequality remains significant and the effect of gross inequality on trust likewise losses significant, the association from trust to gross inequality continues to be significant. While we must stress that a sample of only 55 relatively similar countries may not offer enough variation to identify effects with relatively imprecise data, one result remains significant throughout all robustness tests: that social trust affects gross inequality. As such, we must conclude that a set of careful estimates in general provides support for a nonstandard interpretation of the trust-inequality association. We therefore proceed to discuss these results at some length. 6. Discussion and conclusions In a survey of the trust literature, Nannestad (2008, 430) hypothesized that high levels of trust in some countries enable these countries to solve the collective action dilemma created by their welfare systems and that as a consequence, generalized trust is what makes the universal welfare system sustainable and allows equality to coexist with wealth. In this paper, we have focused specifically on the main aim of the modern welfare state: redistributing income to achieve a more equal distribution. We have shown that the hypothesis that trust affects the distribution is both theoretically plausible and empirically supported. In a set of structural equations estimates, we find that social trust has a direct impact on inequality, regardless of whether or not one controls for the size of the welfare state. This impact occurs for pre-tax, pre-transfer (gross) inequality as well as for postredistribution (net) inequality. Exploring the opposite causal direction, we find evidence that gross inequality affects social trust while we see no support for the claim that net inequality has any causal effect on trust. 22

23 The findings that social trust seems to have a direct impact on inequality even when controlling for the size of the welfare state while there is no evidence for feedback from net inequality have important implications. For example, they imply that income equality may be correlated with a number of desirable outcomes, but that the correlation is not necessarily causal but can rather be caused by the omitted variable trust. Among the documented positive effects of trust we find better governance, more education, higher economic growth and higher levels of subjective wellbeing (Bjørnskov, 2010; Helliwell and Huang, 2008; Knack and Keefer, 1997; Papagapitos and Riley, 2009; Putnam, 1993; Zak and Knack, 2001). In addition, Rostila (2007) shows that contexts with low trust, such as most post-socialist societies, are more detrimental for the health of distrustful individuals. In overall terms, these findings suggest that contextual social trust could explain health differences between European welfare regimes. The findings also have implications for the likelihood of policy affecting trust levels. One of the most robust results in the trust literature is the association between social trust and measures of inequality (Knack and Keefer, 1997; Uslaner, 2002; Delhey and Newton, 2005; Bjørnskov, 2007, 2008; Nannestad, 2008). Most studies have explicitly or implicitly interpreted this association as evidence of an effect of large income differences on social trust. When including policy implications, studies have gone one step further and interpreted this finding as evidence of the social effectiveness of redistributive welfare state policies (e.g., Rothstein and Uslaner, 2005). Nevertheless, the causality of the association has gone more or less untested. Yet, taking care in outlining the causal directions between inequality and social trust, we note that if redistributive welfare policy is likely to affect trust, we ought to observe that it is mainly net inequality that is associated with trust, and that causality would run from inequality to trust. As our evidence points to the opposite result, it would imply that direct redistribution is not likely to be effective in influencing social trust in any particular direction. In addition, our results show that while the causality 23

24 is likely to be bi-directional, the effect from trust to inequality is substantially stronger than that from inequality to trust. As such, our results do not fit what has emerged as the dominant interpretation of the inequality-trust association in the trust literature. Instead, we find strong support for an effect that allows high-trust countries to have more equal pre-transfer, pre-tax income distributions than found in other comparable societies. We therefore find no evidence for any simplistic policy advice and welfare state implications as those championed by, e.g., Barr (2004) or Kumlin and Rothstein (2005). Our results again point to social trust as a deeper determinant of formal-institutional and societal characteristics or, put differently, as an institutional feature belonging to Williamson s (2000) first level of social analysis, the social embeddedness level, where a culturally stable set of norms, traditions and basic beliefs are located, instead of being a policy variable as argued by Jackman and Miller (1998). We can only speculate on what type of effect is responsible for the influence of social trust on gross inequality, although formal theoretical work such as that in section 2 may provide input. For example, most of the trust literature stresses that social trust reduces the incidence of various forms of rent-seeking, from different public morals and more inclusion of voter interests to better bureaucratic quality (Putnam, 1993; Boix and Posner, 1998; Uslaner, 2002; Bjørnskov, 2010). The prime suspect must therefore be some form of rent-seeking such as elite rent-grabbing or other similar activities that increase elite shares of the income distribution, but are historically limited in high-trust countries. While we leave the specific mechanisms for future research, we stress that such mechanisms are more relevant than the simple policy advice focused on in the literature so far. Appendix Insert Table A1 about here 24

25 Insert Table A2 about here Insert Table A3 about here Insert Table A4 about here References Alesina, Alberto, Arnaud Devleeschauwer, William Easterly, Sergio Kurlat and Romain Wacziarg. (2003). Fractionalization. Journal of Economic Growth, 8, Axelrod, Robert. (1984). The Evolution of Cooperation. New York: Basic Books. Barr, N. (2004). The Economics of the Welfare State. Stanford, California: Stanford University Press. Bendor, Jonathan and Piotr Swistak. (1997). The Evolutionary Stability of Cooperation. American Political Science Review, 91, Berggren, Niclas and Christian Bjørnskov. (in press). Is the Importance of Religion in Daily Life Related to Social Trust? Cross-Country and Cross-State Comparisons. Forthcoming in Journal of Economic Behavior & Organization. Bergh, Andreas and Christian Bjørnskov. (2011). Historical trust levels explain current welfare state size. Kyklos, 64, Binmore, Ken. (1994). Game Theory and the Social Contract Volume I: Playing Fair. Cambridge, Massachusetts: MIT Press. Bjørnskov, Christian. (2007). Determinants of Generalized Trust. A Cross-Country Comparison. Public Choice 130, Bjørnskov, Christian. (2008). Social Trust and Fractionalization: A Possible Reinterpretation. European Sociological Review, 24, Bjørnskov, Christian. (2010). How does Social Trust lead to Better Governance? An Attempt to Separate Electoral and Bureaucratic Mechanisms. Public Choice 144,