Expressive voting and government redistribution: Testing Tullock s charity of the uncharitable

Transcription

1 Public Choice 119: , Kluwer Academic Publishers. Printed in the Netherlands. 143 Expressive voting and government redistribution: Testing Tullock s charity of the uncharitable RUSSELL S. SOBEL 1 & GARY A. WAGNER 2 1 Department of Economics, West Virginia University, Morgantown, WV , U.S.A.; rsobel2@wvu.edu; 2 A.J. Palumbo School of Business, Duquesne University, Pittsburgh, PA 15282, U.S.A.; wagner@duq.edu Accepted 13 February 2003 Abstract. Models of expressive voting postulate that voters will consume ideological stances on issues by voting for them, even when they are against the voter s own narrow self interest, if the probability of being a decisive voter is low. When a voter is unlikely to sway the outcome, the odds that a voter will incur any real personal cost (a higher tax burden, for example) from her own expressive vote is small. We test and find support for Tullock s straightforward empirical implication of this model, that government welfare (transfer) payments are inversely related to the probability of being the decisive voter. 1. Introduction Traditional rational voter models, following the works of Downs (1957), Tullock (1967), and Riker and Ordeshook (1968, 1973), attempt to show how the voting calculus of an individual is affected by the costs and benefits of voting. In these models, a central factor influencing voter turnout is the probability that the voter s single vote will change the outcome of the election, (i.e., the probability that he or she is the decisive voter). 1 And while the average citizen might consider voter turnout rates in the U.S. quite low, the typical economist generally finds it puzzling as to why so many people still vote given the remote odds that they will influence the outcome of the election. The probability of being the decisive voter is so low for major U.S. national elections, say less than , that many writers have commented on how an individual is more likely to be in an auto accident on the way to vote than being the decisive voter in the election. The apparent inability of the rational voter model to explain why individuals vote has become known more generally as the paradox of voting. The authors would like to thank William N. Trumbull, participants at the 2001 Public Choice Society meetings, and an anonymous referee of this journal, for helpful comments and suggestions. Any remaining errors are the sole responsibility of the authors.

2 144 One of the most prominent lines along which the rational voter model has been extended to rectify the paradox of voting is to allow for expressive voting. First proposed by Buchanan (1954), and further developed by Tullock (1971) and Brennan and Lomasky (1993), the theory of expressive voting holds that when there is a relatively low probability of casting the decisive vote, individuals will chose to vote as an act of expressive behavior, often voting along ideological or moral lines for what might be considered public minded policies that are apparently against the voter s own narrow personal self interest. In essence, the notion of expressive voting extended the traditional rational voter model by allowing voters to receive direct utility from the act of voting itself. At first glance, the motives of a typical voter seem to run counter to the theory of expressive voting, After all, why should an individual vote in favor of a public-minded policy if it is against their own personal self interest? An explanation for this apparent clash in motives was noted by Tullock (1971: 387) to be the individual s attempt to reduce internal dissonance. Tullock elaborates using an example in which an individual has the choice to give $100 to charity directly or vote on whether to be taxed the $100. According to Tullock, an individual that derives little satisfaction from charity will not voluntarily contribute, but would be willing to vote in favor of the tax to reduce internal dissonance because the probability that they will influence the outcome is so small. Eichenberger and Oberholzer-Gee (1998) also address the more psychological factors behind expressive voting that account for why voters apparently behave more in line with social norms for fairness in their voting behavior than in market behavior. In this paper we make no attempt to directly address the theoretical merit of the expressive voting hypothesis itself, but rather to test a straightforward implication of the model first postulated by Tullock (1971: ) in his famous article The Charity of the Uncharitable : Some further implications can be drawn from this phenomenon. As the size of the constituency in which I am voting increases, the likelihood that my vote will have any effect on the outcome decreases. Looked at from the standpoint of the voter, he can obtain the satisfaction of behaving charitably... muchcheaper. Tullock goes on to explain that because of this effect, welfare spending should be higher at the state than the local level because of the larger size of the constituency (and also correspondingly explains that national level welfare spending should be even higher). Despite the rather large literature on expressive voting that has evolved since Tullock s article, there have been no direct tests of his hypothesis that

3 145 reductions in the probability of casting the decisive vote will result in an increase in welfare spending (and vice versa). While there appears to be a straightforward line of reasoning suggesting this inverse relationship between the probability of casting the decisive vote and government welfare spending, several theoretical complications arise. First, is it the case that both Republicans and Democrats (and members of other political parties as well) all vote ideologically in the same direction? Second, as the size of the constituency changes, altering the probability of casting the decisive vote, does this alter the likelihood that a voter will participate in the voting process, and if so does a selection bias come into play in terms of those who participate? Third, as the size of the constituency expands, how do welfare recipients themselves alter their votes? Finally, does a change in the constituency size alter the position of the median voter in a manner that might also affect welfare spending? Hopefully we have raised enough issues here to make the reader aware of the value of empirically testing Tullock s hypothesis and providing some evidence on its merit. We test Tullock s hypothesis using U.S. state level data from Our results provide very strong evidence that the probability of casting the decisive vote in a state is an important determinant of welfare spending. Our fixed-effects, panel estimation technique allows us to confirm that not only do differences across states in this probability explain cross-state differences in welfare spending, but also that changes in the probability of casting the decisive vote in a given state through time is an important determinant of changes in that state s welfare spending through time. The paper continues by describing the data and estimation techniques used in our analysis. We then present the results of our analysis and conclude with a discussion of the policy implications of our results. We begin by first reviewing the previous studies of expressive voting and the probability of casting the decisive vote. 2. Expressive voting and the probability of being decisive 2.1. Previous research The number of empirical studies testing for the presence of expressive voting has been rather small, but they have all generally provided support for the hypothesis. A recent study by Copeland and Laband (2002), for example, explores the relationship between actual voter turnout and expressive behavior. Using data from the National Election Survey, they find strong evidence correlating the act of voting with political expressiveness (i.e., wearing campaign buttons, posting political signs, and contributing to the

4 146 Federal Election Commission via income tax returns). 2 Feigenbaum, Karoly, and Levy (1988) use county-level voting data from California on the 1982 Nuclear Freeze Referendum to test whether a model of expressive voting better explains the data than a model containing only standard economicconsequences variables. Their results strongly support the idea that votes on this issue can be better explained by a model of expressive voting. 3 More recently, Kan and Yang (2001) find evidence in favor of expressive voting using data from the 1988 U.S. presidential election. In addition to linking expressive behavior and voter turnout, several studies have investigated how being the decisive voter affects individuals choices. Sobel (1992), using data on state legislatures, finds that the probability that an individual legislator will be decisive in the legislature is a significant determinant of whether a legislator will shirk (i.e., vote against their constituents interests). Sobel finds that legislators are more likely to shirk when the probability of being decisive in the legislature is higher (i.e., thus shirking should be more apt to shirk in smaller legislative bodies). Moreover, several experimental studies, conducted by Carter and Guerette (1992) and Fischer (1996), claim to provide evidence that individuals vote expressively. In these studies, individuals are given the choice to vote on earmarking funds for charity or for themselves. The authors find that individuals are indeed more likely to earmark the funds for chanty as the probability of influencing the outcome declines, providing support for expressive voting. Despite this literature, our paper represents the first empirical test using real-world data of the hypothesis postulated by Tullock regarding the relationship between the probability of an individual being the decisive voter and welfare spending that works through this mechanism of expressive voting The probability of being the decisive voter The probability that an individual voter s vote will actually change the final outcome has been the subject of several technical papers. 4 Essentially, in say a two-candidate election, this probability may be thought of as the probability that exactly half of the other voters will vote for Candidate 1, while the other half vote for Candidate 2. Mueller (1989: 350) summarizes this mathematical work and presents a unified formula for calculating the probability that any one voter will be decisive (P): 3 P = 2 2π(N 1), (1) where N is the number of voters participating in the vote. 5 For our empirical work, we calculate this probability for each state, for each year from 1972 to Using data from the Statistical Abstract of the United States, we

5 147 employ the actual ex post voter turnout for N. Following the methodology of Husted and Kenny (1997), we used turnout in presidential elections during presidential election years, followed by turnout for the Governor s race, Senate race, or House race in non-presidential years, in descending order of preference. 6 Our sample period begins in 1972 in order to avoid including periods where substantial legal changes were made in the way of poll taxes, literacy tests, and legal voting age restrictions. While these institutional changes significantly altered the size of the voting franchise, and thus dramatically altered the probability of being a decisive voter, such institutional changes also had the effect of moving the pivotal median voter down the income distribution, which should also increase the demand for state welfare payments. This is precisely the hypothesis presented and confirmed in Husted and Kenny (1997). They strongly conclude that these expansions in the voting franchise, resulting from institutional change, led to a sharp rise in state welfare spending because of the systematic effect of moving the median voter down the income distribution. 7 The periods of institutional change exploited by Husted and Kenny (1997) is thus problematic here because the expressive voting model also predicts that a larger voting franchise would result in more welfare spending (but here because it lowers the probability that any one voter is decisive). It is likely that both effects were present during the pre-1972 period, but they would be very difficult to separate. We believe that by restricting our sample to the post-1972 period, so that the variation in the voting franchise is due more to population changes, immigration, and mobility than it is to institutional changes, and by including variables to capture changes in the preferences and characteristics of the median voter, we increase the certainty that the correlation we find between welfare spending and the probability of being the decisive voter is not alternatively explained by institutionally-driven changes in the median voter. Thus, our empirical model closely follows the Husted and Kenny model, but we use a more recent sample period and incorporate an additional independent variable into the model to capture the effect of changes in the probability of being a decisive voter in the state. 3. Empirical model and results The dependent variable in our model is state welfare spending and we estimate the model measuring welfare spending both as a share of general expenditures and as a real per capita dollar amount. Following Husted and Kenny (1997) we present results for both the linear and double log specifications. The double log specifications are advantageous because the coefficients

6 148 can be directly interpreted as elasticities. We present only the results of the double log models in the text to conserve space; however, the results of the linear specifications are included as Appendix Tables B1 and B2. As independent variables in the model we include the probability of being a decisive voter in the state in addition to the general set of independent variables used by Husted and Kenny (1997). These independent variables for each state include the percent black, the percent over age 65, the poverty rate, real median income, Gini coefficient, total per capita real federal aid, and indicator variables for whether the states governor and both houses of the legislature were controlled by Republicans or Democrats. 8 In addition, we also include the state s population as a regressor to control for the possibility that our measure of the probability of being decisive is not simply capturing the impact of a state s population on welfare spending. As Faith and Tollison (1990) note, increases in the population reduce the per capita price of charity, which should increase the quantity of welfare spending demanded. Since voter turnout, used to construct the probability of being the decisive voter, may be correlated with a state s population, including population separately as a regressor allows the empirical model to distinguish between the impact of voter turnout and population. 9 The complete list of dependent and independent variables, as well as descriptive statistics for each variable, can be found in Appendix Table A1. Our sample is biennial data for 49 states over the period, resulting in 637 observations. Nebraska could not be included in the sample because the state s legislators are elected without party designation, making it impossible to determine the controlling political party. We first estimate each model by pooled OLS, then we extend each model to include fixed time effects, fixed state effects, and finally fixed time and state effects. The results of our estimations are provided in Tables 1 and 2. Tables 1 and 2 differ only in the way the dependent variable is measured. In Table 1 the dependent variable is state welfare spending as a share of the state budget, while in Table 2 the dependent variable is real per capita state welfare spending. Each table presents the coefficient estimates obtained from the different specifications of the model. The first column shows the results obtained on the pooled data set with no variables included for either time or state fixed effects, The second and third columns of coefficient estimates show how these results change as either time fixed effects, or state fixed effects, are included individually. The final column shows the results from the two-way fixed effects model. This is the most stringent model, and provides the most possible correction for any factors specific to any state or to any given year.

7 149 Table 1. Regression results using log of real public welfare spending as a share of state general expenditures as a dependent variable (double-log model) Pooled OLS Fixed time Fixed state Fixed time & effects effects state effects Constant (1.6596) (1.7911) (2.1226) (2.4750) log (%Black) (0.0135) (0.0123) (0.0426) (0.0356) log (%Over65) (0.0709) (0.0655) (0.0295) (0.0271) log (Poverty rate) (0.0715) (0.0756) (0.0791) (0.0837) log (Real median income) (0.1244) (0.1193) (0.2199) (0.1891) log (Gini coefficient) (0.2661) (0.3358) (0.2465) (0.2520) log (Real per capita federal aid) (0.0624) (0.1191) (0.0499) (0.1098) log (Probability of being decisive voter) (0.1369) (0.1761) (0.1120) (0.1174) log (Population) ( ) (0.1022) (0.1048) (0.1442) Democratic control (0.0312) (0.0361) (0.0256) (0.0247) Republican control (0.0412) (0.0368) (0.0360) (0.0341) F-test of model R Notes. Standard errors corrected for heteroskedasticity. Significance levels are as follows: denotes the 1% level, denotes the 5% level, and denotes the 10% level. All models exclude Nebraska. Estimated coefficients of the state and time effects are not reported. Fixed effects models excluded one year (1996) and/or one state (Wyoming) to avoid perfect collinearity with the constant term.

8 150 Table 2. Regression results using the log of real per capita public welfare spending as a dependent variable (double-log model) Pooled OLS Fixed time Fixed state Fixed time & effects effects state effects Constant (1.5481) (1.7063) (2.0058) (2.4199) log (%Black) (0.0147) (0.0134) (0.0394) (0.0322) log (%Over65) (0.0541) (0.0515) (0.0360) (0.0317) log (Poverty rate) (0.0687) (0.0745) (0.0779) (0.0823) log (Real median income) (0.1183) (0.1118) (0.2116) (0.1876) log (Gini coefficient) (0.2743) (0.3439) (0.2502) (0.2571) log (Real per capita federal aid) (0.0628) (0.1199) (0.0511) (0.1119) log (Probability of being decisive voter) (0.1328) (0.1773) (0.1117) (0.1141) log (Population) (0.0803) (0.1037) (0.1067) (0.1435) Democratic control (0.0312) (0.0346) (0.0248) (0.0239) Republican control (0.0409) (0.0363) (0.0377) (0.0353) F-test of model R Notes. Standard errors corrected for heteroskedasticity. Significance levels are as follows: denotes the 1% level, denotes the 5% level, and denotes the 10% level. All models exclude Nebraska. Estimated coefficients of the state and time effects are not reported. Fixed effects models excluded one year (1996) and/or one state (Wyoming) to avoid perfect collinearity with the constant term. The results are similar to those obtained in Husted and Kenny for the other control variables. Most importantly, our new variable of interest, the probability of casting the decisive vote, is negative and significant at the 5% level or better in every specification except the model with state fixed effects that does not correct for fixed time effects. These same models estimated in

9 151 linear, rather than log form (presented in Appendix Tables B1 and B2) find the probability of being decisive to be significant in all eight specifications, even the one that does not control for fixed time effects. Thus, regardless of whether welfare spending is measured as a share of the state budget (Table 1) or in real per capita terms (Table 2), we find strong evidence that the probability of casting the decisive is an important determinant of welfare spending. Considering the 8 double log and 8 linear models jointly, we find the probability of being decisive to be significantly correlated with welfare spending (at least at the 5% level) in 14 of the 16 specifications. Moreover, in an attempt to ensure other factors were not driving our results, we also restricted our sample to include only presidential and again to include only non-presidential years (the results of which may be found in Appendix Tables B3 and B4 respectively) and find the probability of being decisive to be statistically significant in all but one of those regressions as well. The magnitude of the coefficient on the probability of being the decisive voter (which is an elasticity in the double log specifications) is fairly consistent across different specifications and different measures of welfare spending, falling somewhere between 0.27 and It appears generally that a 10% reduction in the probability of casting the decisive vote would lead to a 3 to 5% increase in state welfare spending, ceteris paribus, regardless of whether it is measured as a percent of the state budget or as a real per capita amount. 4. Policy implications The results presented here confirm a strong and statistically significant relationship between the probability of being a decisive voter and the level of state welfare spending, a linkage originally hypothesized by Tullock some 30 years ago. Thus, despite all of the potential complicating factors in the relationship between constituency size and welfare spending, his charity of the uncharitable appears to hold very strongly for the U.S. states. Because one determinant of this probability is simply the number of registered voters, it is clearly the case that expansions in the voting franchise can be a driving force in causing growth in public sector welfare spending. Given the rather rapid growth of U.S. Federal transfer expenditures during the past century, a growth that did not occur in the early and mid 1800s, one must wonder if the reductions in this probability that resulted from expansions in the U.S. population through immigration and birth, and also the institutional changes in voting restrictions during this time, were not responsible for some of this growth in transfer spending.

10 152 The lower the probability that any single voter will be decisive, the less incentive there is for voters to turnout, and the less of an incentive there is for voters to become informed. There are some potential ways to restructure the political system to intentionally increase this probability. One such method that has been suggested by authors such as Page (1992) is specialized elections in which only a relatively small randomly drawn subset of voters would be eligible to vote on each election or item on the ballot. There would be many of these randomly drawn small subsets and each would, in a sense, specialize on a single issue to decide it. These groups would be small enough that voters would have a higher incentive to become informed, and their burden would only be to become informed on that single issue for which they are eligible to vote. While we are not suggesting this reform, and there are many other reasons why it may or may not be good, the point here is that election reform proposals that reduce the voting group (and thus increase the probability of being decisive) would also result in smaller government transfer spending. 5. Conclusion Models of expressive voting postulate that voters will consume ideological or moral stances on issues by voting for them, even when they are against the voter s own narrow self-interest, if the probability of being a decisive voter is low. When a voter is unlikely to sway the outcome of the vote, the odds that a voter will incur any real personal cost (a higher tax burden, for example) from his or her own expressive vote is small. In 1971, Tullock originally used this line of reasoning to postulate that welfare spending should be related to the size of the constituency because of the effect is has on the probability of a voter being decisive. We test and find support for Tullock s hypothesis, that public sector welfare (transfer) payments are directly related to the probability of being a decisive voter. Using a biennial panel data of states over the period from 1972 to 1996, we find that a ten percent reduction in the probability of casting the decisive vote generally leads to approximately a 6 percent increase in public welfare spending produced by the political process. This finding begs further inquiry into the role that changes in this probability played in the transition of the federal government to a large welfare state during the 20th century. Notes 1. See Mueller (1989: 350) for a discussion of this probability and the related literature. Also, a recent summary of the factors influencing voter turnout may be found in Aldrich (1993).

11 The expressive voting framework is used as an assumption underlying other broader public choice models. For example, Clark and Lee (1999) employ the notion of expressive voting in a model of the optimal trust in government. Their model derives a Laffercurve type relationship between the trust citizens have in government and government performance. Thus, increased trust in government reduces government performance over some ranges and increases it over others. 3. For additional evidence on the expressive voting model see Faith and Tollison (1990). 4. See, for example, Riker and Ordeshook (1968, 1973), Tollison, Crain, and Pautler (1975), Barzel and Silberberg (1973), and Owen and Grofman (1984). 5. Here we have simplified the equation by assuming a diffuse prior with respect to the voter s ex ante knowledge of other voters preferences across the two outcomes. Even though in any one specific election this may not hold, our intent is to obtain the average probability across all relevant races on that election ballot. In essence, state-level welfare spending is not directly up for vote, but rather many candidates for many offices are on the ballot and the combined average probability of being decisive across all of these relevant elections is what is of interest. 6. Since turnout during presidential years is typically much larger than non-presidential years, we also estimated our empirical model using only presidential year data and only non-presidential year data to ensure that variation in voter turnout between presidential and non-presidential years was not driving our results. The results of the presidential and non-presidential year estimates, which revealed no significant qualitative differences, may be found in Appendix Tables B3 and 4 respectively. 7. The impact of moving the median voter down the income distribution on government expenditures on services is less clear than the impact it has on welfare spending. Kenny (1978) explores this issue, and it is again tested in Husted and Kenny (1997). Essentially they find no increase in other areas of spending. This is consistent with studies that find the estimated income elasticity for government services generally exceeding the elasticity of substitution between government services and private goods (the opposite condition would have to be true for expenditures on government services to rise as the median voter moves down the income distribution). 8. Following Husted and Kenny (1997), intercensus year data was interpolated. In addition, the Gini coefficient for each state was estimated for each year following the methodology of Langer (1999). Detailed variable descriptions, including particular years that were interpolated, are available in Appendix Table A1. 9. We computed the correlation between voter turnout and population for each state over our sample period ( ) and find that the correlations range between 0.47 in Vermont to 0.94 in Alaska. Averaging across states we find the correlation to be 0.33, which suggests that voter turnout (and thus the probability of being decisive) is not simply a proxy for the state s population. References Aldrich, J.H. (1993). Rational choice and turn-out. American Journal of Political Science 37: Barzel, Y. and Silberberg, E.L. (1973). Is the act of voting rational? Public Choice 16:

12 154 Brennan, H.G. and Lomasky, L.E. (1993). Democracy and decision. Cambridge: Cambridge University Press. Buchanan, J.M. (1954). Individual choice in voting and the market. Journal of Political Economy 62: Carter, J.R. and Guerette, S.D. (1992). An experimental study of expressive voting. Public Choice 73: Clark, J.R. and Lee, D.R. (1999). Trust in government as a constitutional consequence. Unpublished manuscript. University of Tennessee at Chattanooga. Copeland, C. and Laband, D.N. (2002). Expressiveness and voting. Public Choice 110: Downs, A. (1957). An economic theory of democracy. NewYork:Harper& Row. Eichenberger, R. and Oberholzer-Gee, F. (1998). Rational moralists: The role of fairness in democratic economic politics. Public Choice 94: Faith, R.L. and Tollison, R.D. (1990). Expressive versus economic voting. In W.M. Crain and R.D.Tollison (Eds.),Predicting politics: Essays in empirical public choice, Ann Arbor: University of Michigan Press. Feigenbaum, S., Karoly, L. and Levy, D. (1988). When votes are words not deeds: Some evidence from the nuclear freeze referendum. Public Choice 58: Fischer, A.J. (1996). A further experimental study of expressive voting. Public Choice 88: Husted, T.A. and Kenny, L.W. (1997). The effect of the expansion of the voting franchise on the size of government. Journal of Political Economy 105: Kan, K. and Yang, C.C. (2001). On expressive voting: Evidence from the 1988 U.S. presidential election. Public Choice 108: Kenny, L.W. (1978). The collective allocation of commodities in a democratic society: A generalization. Public Choice 33: Langer, L. (1999). Measuring income distribution across space and time in the American states. Social Science Quarterly 80: Mueller, D.C. (1989). Public choice II: A revised edition of public choice. Cambridge: Cambridge University Press. Owen, G. and Grofman, B. (1984). To vote or not to vote: The paradox of nonvoting. Public Choice 42: Page, S.E. (1992). Specialized elections. Northwestern University Center for Mathematical Studies in Economics and Management Science Discussion Paper Riker, W. and Ordeshook, P.C. (1968). A theory of the calculus of voting. American Political Science Review 62: Riker W. and Ordeshook, P.C. (1973). Introduction to positive political theory. Englewood Cliffs, NJ: Prentice-Hall. Sobel, R.S. (1992). Political incentives and legislative voting. Journal of Public Finance and Public Choice 10: Tollison, R., Crain, M., and Pautler, P. (1975). Information and voting: An empirical note. Public Choice 24: Tullock, G. (1967). Toward a mathematics of politics. Ann Arbor: University of Michigan Press. Tullock, G. (1971). The charity of the uncharitable. Western Economic Journal 9:

13 155 Appendix Table A1. Variable descriptions, summary statistics, and sources ( ) Notes. Sample excludes Nebraska. Real public welfare spending and real federal aid were deflated using the CPI (1992 = 100). All of the data and programs used by the authors are freely available upon request.

14 156 Table B1. Regression results using real public welfare spending as a share of general expenditures as a dependent variable (linear model) Pooled OLS Fixed time Fixed state Fixed time & effects effects state effects Constant (4.0032) (5.4211) (4.5887) (5.6411) %Black (0.0201) (0.0186) (0.1300) (0.1113) %Over (0.1068) (0.1241) (0.0978) (0.1030) Poverty rate (0.0765) (0.0803) (0.0696) (0.0764) Real median income ( ) ( ) (0.0008) ( ) Gini coefficient (9.5778) (9.9336) (8.2794) (7.7758) Real per capita federal aid (0.0012) (0.0014) (0.0014) (0.0019) Probability of being decisive voter (1130.3) (1119.8) (1223.7) (1358.1) Population ( ) ( ) (0.0002) (0.0001) Democratic control (0.4299) (0.4248) (0.3486) (0.3130) Republican control (0.7033) (0.6577) (0.4693) (0.4446) F-test of model R Notes. Standard errors corrected for heteroskedasticity. Significance levels are as follows: denotes the 1% level, denotes the 5% level, and denotes the 10% level. All models exclude Nebraska. Estimated coefficients of the state and time effects are not reported. Fixed effects models excluded one year (1996) and/or one state (Wyoming) to avoid perfect collinearity with the constant term.

15 157 Table B2. Regression results using real public welfare spending as a dependent variable (linear model) Pooled OLS Fixed time Fixed state Fixed time & effects effects state effects Constant (101.54) (143.85) (128.50) (172.43) %Black (0.4700) (0.4347) (4.5729) (3.7389) %Over (2.4881) (2.7524) (2.9621) (2.2109) Poverty rate (1.8068) (1.9298) (1.6619) (1.8554) Real median income (0.0013) (0.0013) (0.0021) (0.0020) Gini coefficient (234.57) (243.01) (229.67) (205.20) Real per capita federal aid (0.0442) (0.0525) (0.0488) (0.0654) Probability of being decisive voter (27693) (26098) (47027) (40963) Population (0.0016) (0.0014) (0.0039) (0.0027) Democratic control (9.4265) (9.3883) (7.7840) (6.7589) Republican control (16.271) (14.832) (13.029) (11.087) F-test of model R Notes. Standard errors corrected for heteroskedasticity. Significance levels are as follows: denotes the 1% level, denotes the 5% level, and denotes the 10% level. All models exclude Nebraska. Estimated coefficients of the state and time effects are not reported. Fixed effects models excluded one year (1996) and/or one state (Wyoming) to avoid perfect collinearity with the constant term.

16 158 Table B3. Regression results using real public welfare spending as a share of general expenditures as a dependent variable (double-log model) with sample restricted to only years in which there was a presidential election Pooled OLS Fixed time Fixed state Fixed time & effects effects state effects Constant (2.2690) (2.5178) (2.9133) (3.4698) log (%Black) (0.0198) (0.0170) (0.0521) (0.0464) log (%Over65) (0.0697) (0.0628) (0.0341) (0.0029) log (Poverty rate) (0.116) (0.1100) (0.0127) (0.1215) log (Real median income) (0.1667) (0.1551) (0.3139) (0.2725) log (Gini coefficient) (0.3696) (0.4619) (0.3719) (0.3968) log (Real per capita federal aid) (0.0810) (0.1610) (0.0715) (0.1617) log (Probability of being decisive voter) (0.2513) (0.2662) (0.1382) (0.1272) log (Population) (0.1438) (0.1526) (0.1442) (0.2072) Democratic control (0.0444) (0.0493) (0.0366) (0.0360) Republican control (0.0565) (0.0531) (0.0585) (0.0548) Sample size 343 F-test of model R Notes. Standard errors corrected for heteroskedasticity. Significance levels are as follows: denotes the 1% level, denotes the 5% level, and denotes the 10% level. All models exclude Nebraska. Estimated coefficients of the state and time effects are not reported. Fixed effects models excluded one year (1996) and/or one state (Wyoming) to avoid perfect collinearity with the constant term.

17 159 Table B4. Regression results using real public welfare spending as a share of general expenditures as a dependent variable (double-log model) with sample restricted to only years in which there was not a presidential election Pooled OLS Fixed time Fixed state Fixed time & effects effects state effects Constant (2.1892) (2.7719) (2.3826) (3.3054) log (%Black) (0.0178) (0.1595) (0.0655) (0.0494) log (%Over65) (0.0914) (0.1406) (0.1966) (0.2431) log (Poverty rate) (0.0795) (0.0937) (0.0773) (0.0958) log (Real median income) (0.1935) (0.2083) (0.2723) (0.2202) log (Gini coefficient) (0.3680) (0.4782) (0.3832) (0.3629) log (Real per capita federal aid) (0.1001) (0.1873) (0.0736) (0.1502) log (Probability of being decisive voter) (0.2682) (0.2515) (0.2443) (0.1959) log (Population) (0.1528) (0.1455) (0.1601) (0.1948) Democratic control (0.0441) (0.0491) (0.0367) (0.0349) Republican control (0.0605) (0.0471) (0.0555) (0.0524) Sample size 294 F-test of model R Notes. Standard errors corrected for heteroskedasticity. Significance levels are as follows: denotes the 1% level, denotes the 5% level, and denotes the 10% level. All models exclude Nebraska. Estimated coefficients of the state and time effects are not reported. Fixed effects models excluded one year (1996) and/or one state (Wyoming) to avoid perfect collinearity with the constant term.

18