1. Introduction

Why democracy fails to confront rising inequality remains a significant puzzle in the relevant literature. Neither within nor across countries have voters consistently responded to rapidly increasing inequality in the way mainstream models, notably 1, suggest. The most recent attempts to address this puzzle point to a widespread failure to comprehend the actual extent of inequality, but also show that voters support redistribution when they perceive high inequality. I argue that voters stay rationally ignorant of aggregate income structures, and focus instead on inequality in local observables—particularly housing—to make inferences about inequality.

2. Background and Argument

Democracy is suppose to cure inequality, but it has not 2. Leading scholars in this long literature recently conclude that the “current consensus….[is] that inequality does not matter for the politics of redistribution, at least not in any direct and particularly significant way” 3. Another claims that “results have consistently called into question…that pressures for redistribution increase with inequality” 4. This weak empirical relationship between inequality and the demand for redistribution particularly flies in the face of workhorse academic models that predict growing support for redistribution as inequality rises 5. This paper offers a new political economy theory that provides a direct mechanism that links rising inequality to the demand for redistribution at the local level.

In short, voters’ demands for redistribution increase only with the inequality that is (1) observable and (2) relevant to voter utility. On the former, different forms of inequality vary in how costly they are to observe. The first observability dimension is distance; inequality in one’s neighborhood is more observable than inequality in the city, state, country and so on. The second important dimension comes from the type of inequality; relative income is less observable then relative housing wealth, especially at very local levels where discrepancies in housing values become apparent through sight and vibrant on-line markets. In contrast, average employees lack basic information about earnings within their own firms, let alone across firms, sector, and distance. These observability conditions represent the costs of information on inequality.

On the benefits side, information on inequality varies in its relevance to voter utility. Over the spatial dimension, changes in local earnings and housing price inequality are more consequential to investments and consumption decisions than are changes in distant markets. This is a kind of NIMBYism. Furthermore, for most voters, housing eats up the highest share of income, and represents the most significant investment 6. Unlike income, property values can swing rapidly in response to changes in the local market structure. Combined, the costs and benefits outlined above result in the following propositions.

  1. Voters’ median demand for redistribution is more sensitive to increases in housing inequality than to income inequality.
  2. Voters’ median demand for redistribution is more sensitive to local inequality than aggregate (e.g. national) inequality.

Figure A1 in the Appendix stylizes these propositions graphically. It shows the declining effect of inequality on the demand for redistribution as the cost of acquiring information about inequality increases. This is a function of distance, or aggregation \(\delta\) and skill specificity \(s\), which is particular to earning. 7

3. Data and Empirical Strategy

3.1 Data:

In an attempt to explain the weak empirical evidence for the relationship between inequality and redistribution, scholars point to the poor data quality that pervade this literature and prevent rigorous testing of these arguments 8. National and cross-national survey instruments used to measure demands for redistribution lack clear construct validity, and suffer from potentially severe social desirability bias. Other so called ``pre-post’’ approaches to measuring redistribution inherently incorporate institutional and political features that cannot be easily teased out of the estimations.

To help resolve these concerns, I take advantage of two California referenda, Propositions 30 and 55, commonly referred to as the ``millionaires tax.’’ Prop 30 was put before California voters in 2012 with a sunset clause. After its passage, voters returned to the polls in 2016 to vote for an extension. In total, I collected data for 7963 Census tracts nested within 58 counties across two years for a total sample size of 15,926. It is at this level of analysis that I construct the dependent variable: Census tract vote shares for redistribution. These data come from the UC Berkeley Statewide Database. Figure 1 shows the geographic distribution of support for redistribution at the precinct level 9 in only Los Angeles. 10

Figure 1: Los Angeles Support for Redistribution, 2016

This measure has several advantages over the current literature. First, it is a behavioral measure of the demand for redistribution, with real-world policy implications. Second, the use of referenda on a particular policy isolates preferences over the issue area rather than for politicians that represent a basket of policies. Third, these referenda occur in general election years, which maximizes turnout of the politically relevant population of interest. 11 Fourth, time series and cross-sectional variation at the level of Census tracts accommodates more rigorous panel data designs, and tests of theories at highly granular levels.

3.2 Design:

I use a longitudinal design that features selection on observables and time and unit invariant unobservables. The empirical model is a generalized difference in differences setup shown in the following equation.

\(Y_{igt} = D_{igt} \tau + X_{igt}\beta + \alpha_i + \lambda_t + \epsilon_{igt}\) (1)

Where

  • \(Y_{igt}\): (NG \(\times\) 1) vector of vote shares for redistribution in tract \(i\), county \(g\), and time \(t\);
  • \(D_{igt}\): (NG \(\times\) 1) vector of mean-median income measure of inequality;
  • \(\tau\): (1 \(\times\) 1) scalar causal parameter of interest;
  • \(X_{igt}\): (NG \(\times\) K) matrix of time-varying observables;
  • \(\alpha_i\): (NG \(\times\) 1) vector of time-invariant unobservables;
  • \(\lambda_t\): (NG \(\times\) 1) vector of unit-invariant unobservables;
  • \(\epsilon_{igt}\): (NG \(\times\) 1) vector of time-varying unobservables.

4. Preliminary Findings

Due to current data limitation, I only test one aspect of the theory here. I show that the demand for redistribution is largely insensitive to local changes in the structure of income. However, heterogeneity exists in this effect with average income areas responding positively to inequality and very wealthy areas strongly opposing redistribution.

Unlike a classic DD framework, this more general two-way fixed effects model imposes a linearity assumption that we must check. Figure 2 below shows that this assumption fails. In all specifications below, I correct for this non-linearity with a 2nd degree polynomial on inequality.

Figure 2: Redistribution and Inequality

Figure 3 reports the regression coefficients from the above model. Coefficients are standardized to mean zero and one standard deviation units to make coefficients comparable. The error bands reflect 95% confidence intervals generated from standard errors that are robust to arbitrary forms of heteroscedasticity and dependence at the county level. I use these H.A.C. standard errors at the county level in order to account for likely cross-sectional dependencies between Census tracts.

Figure 3: Generalized DD Coefficients

The control variables shown were chosen in careful consideration of selection processes that would confound estimation of the causal effect of inequality on redistribution. Through various processes related to migration, gentrification, and segregation, voters sort into Census tracts by observable covariates. They sort for reasons related to income, race, ideology and age. Because these factors partially determine who receives what treatment in terms of the level of inequality, not controlling for these would bias the results. Table 3 suggests that votes for democratic presidential nominees and average log mean income are stronger predictors of the demand for redistribution than inequality. However, inequality is shown here as a quadratic term, which is difficult to interpret. Therefore, Figure 4 shows the marginal effects plot, which largely confirms the result that income inequality on average does not related to voter demand for redistribution at the Census tract level. This supports the hypothesis that changes in income inequality have little effect on average support for redistribution. However, this model assumes effect homogeneity across the sample. There may exist subsets of the data where the relationship still holds.

Figure 4: Marginal Effect Plot of Inequality on Redistribution

To explore possible effect heterogeneities, I first investigate the relationship between income inequality and redistribution across the 58 counties in the sample. See Figure 5. Upon visual inspection, much heterogeneity exists in the levels of inequality across counties, but not so much in their effects—across counties, the null relationship seems to hold.

Figure 5: Heterogeneous Effects over Counties

Another likely source of heterogeneity would come from the level of average income. Very wealthy tracts may respond differently to changes in inequality than very poor tracts. I plot the marginal effects of this in Figure 6. Inspection of the common support assumption (not shown here) suggests that the negative effects at the low levels of log income (between 9 and 10.5) suffer from severe extrapolation, and are therefore unreliable. However, the other levels of income show significant results: close the the middle earnings range, increases in income inequality results in slight increases in the demand for inequality. For very rich tracts, a strong reversal occurs where increases in inequality are met with strong resistance to redistribution. This simply reflects the fact that the average voter in these wealthy areas vote against potentially harmful redistribution in the way the median voter theory predicts.

Figure 6: Heterogeneous Effects over Levels of Income

5.Summary

This short blog shows that the demand for redistribution at the Census tract level is largely insensitive to changes in the earnings structure. As anticipated by the theory, voters in average income areas only marginally react to inequality with more redistribution. Also in line with median voter predictions, voters in very wealthy areas respond to the greater concentration of earnings with strong opposition to redistribution. This suggests that average wealthy residents fear redistributive taxes in the way we would expect. Future work will test voter sensitivity to changes in earning inequality at greater levels of aggregation. I will also add measures of wealth inequality through housing prices.

Appendix

Figure A1: Theorized Effect of Inequality on Redistribution over Distance

  1. Meltzer, Allan H and Scott F Richard. 1981. “A rational theory of the size of government.” Journal of political Economy 89(5):914-927.↩

  2. Bonica, Adam, Nolan McCarty, Keith T Poole and Howard Rosenthal. 2013. “Why hasn’t democracy slowed rising inequality?” The Journal of Economic Perspectives 27(3):103123.↩

  3. Lupu, Noam and Jonas Pontusson. 2011. “The structure of inequality and the politics of redistribution.”American Political Science Review 105(2):316336.↩

  4. Ansell, Ben W and David J Samuels. 2011. “Inequality and democratization: individual-level evidence of preferences for redistribution under autocracy.”.↩

  5. See Romer, Thomas. 1975. “Individual welfare, majority voting, and the properties of a linear income tax.” Journal of Public Economics 4(2):163185 and Meltzer, Allan H and Scott F Richard. 1981. “A rational theory of the size of government.” Journal of political Economy 89(5):914927↩

  6. Ansell, Ben. 2014. “The political economy of ownership: Housing markets and the welfare state.” American Political Science Review 108(2):383402.↩

  7. Skills that are specific to an occupation, firm, or sector, form a barrier to learning about earnings differentials in occupations that require a separate set of skills. This is particularly the case because there are few individual benefits to learning about such differtials.}↩

  8. Ahlquist, John S and Erik Wibbels. 2012. “Riding the Wave: World Trade and Factor-Based Models of Democratization.” American Journal of Political Science 56(2):447464.↩

  9. Note that for the analysis I aggregate upward from the precinct to the slightly larger Census tract (with populations of between 1,200 and 8,000 people) in order to merge the data with variables from the American Community Survey.↩

  10. Credit for this map is due to the Los Angeles Times. http://www.latimes.com/projects/la-pol-ca-california-neighborhood-election-results/↩

  11. This is opposed to nationally representative or experimental samples of voters who may select out of political participation and who are therefore not relevant to the inequality and redistribution that exists in the observable world.↩