Community Heterogeneity and Revealed Preferences for Environmental Goods

Transcription

1 SSQ MS REVISED Community Heterogeneity and Revealed Preferences for Environmental Goods Julio Videras Hamilton College Abstract Objectives: This paper studies the links between community heterogeneity and the public provision of environmental goods. Methods: This paper examines 2000 referenda data from California. The models include socio-demographic characteristics, ideology, voter turnout, and measures of racial and educational homogeneity. Results: More heterogeneity is correlated with lower support for environmental initiatives. Whites and Hispanics are more likely to support environmental initiatives as the shares of their groups increase. As homogeneity among African- Americans increases, support increases at a decreasing rate. Conclusions: Within-group homogeneity might lead to greater coordination and overcoming of structural barriers that inhibit participation of minorities in the political process.

2 1. Introduction Social scientists have examined election results to understand the support for public policies. Deacon and Shapiro (1975), Fischel (1979), Matsusaka (1992), and Dubin et al. (1992) are among the seminal papers in this literature. The analysis of referendum data, usually aggregated at the county, city, or census-tract level, is particularly helpful to understand the demand for environmental quality (Kahn and Matsusaka, 1997; Kahn, 2002). The natural environment is a public good and willingness to pay must be inferred through stated preferences or indirect methods such as hedonic price and averting expenditures models. Examining voting on referenda is another approach. The election process allows people to learn about the income and price effects of their decision and support for a ballot proposal is interpreted as an indication that the median voter s willingness to pay is equal to or greater than the cost. Typically, this literature has adopted the view that socio-demographic aggregates change the profile of the median voter across communities. Thus, researchers control for variables that measure the tradeoffs the median voter faces, variables such as community s income, shares of individuals in different occupations, and educational attainment. These models also include the share of the population of a given race if the public good might create differences in well-being, for example, if minorities are disproportionally exposed to environmental hazards. However, there is consistent evidence in the social sciences that there are factors beyond population shares that relate to the social structure of a community and that influence economic outcomes. One of these factors is diversity. Research shows that the degree of racial and class homogeneity in a community can matter for the provision of public goods. Using data from Indonesia, Otken and Okonkwo-Osili (2004) find a negative relationship between ethnic diversity and participation in community groups. Alesina and La Ferrara (2000, 2002) find that 1

3 racial diversity in American cities is negatively correlated with participation in social activities and with the propensity to trust other people. Costa and Kahn (2003) and, in an experimental setting, Glaeser, Laibson, Scheinkman, and Soutter (2000) also find effects of diversity on trust. Other papers examining the links between community composition and the provision of public goods include Sampson, Raudenbush, and Earls (1997), Alesina, Baqir, and Easterly (1999), Vigdor (2004), and Alesina and La Ferrara (2005). There are several non-excusive explanations for these results. Two mechanisms in particular seem relevant for this paper. First, diversity can increase the costs of coordinating and defining common policy goals. Second, diversity can reduce the provision of public goods if interaction with members from other groups reduces utility, that is, if individuals are averse to mixing. Thus, even when different groups have the same preference for a given public good, community heterogeneity can lower the provision of the good if its use requires contact across groups and sharing. These two overlapping explanations imply that diversity might reduce the incentives for collective action and the likelihood that a community can foment solidarity and group action. Such mechanisms might influence voting for costly environmental propositions, in addition to the effect that diversity might have on voter turnout, as the level of support for initiatives depends on social interaction and civic involvement before Election Day. This paper connects these two areas of research: the examination of revealed preferences for environmental quality through referendum data and the analysis of the effects of community composition. Using 2000 census-tract data from California, the results show that racial and social-class homogeneity is positively correlated with support for the public provision of environmental goods, even after controlling for income and price effects, ideology, and voter turnout. Exploring the role of community heterogeneity is important to understand how 2

4 institutional and social factors that affect the socio-demographic mix of a community might facilitate or hinder the public provision of environmental goods. The results indicate that accounting for community homogeneity improves the model s fit and provides substantive insights into the links between community composition and support for the public provision of environmental quality. The paper is organized as follows. Section 2 briefly summarizes relevant papers in each literature. Section 3 presents the environmental initiatives that are the object of study. Section 4 discusses the variables and specifications. Section 5 presents results and Section 6 concludes. 2. Related Literature Deacon and Shapiro (1975) use city-level data to examine voting results for two initiatives in the early 70s to conserve the California coastline and to provide public financing for rapid transit. The authors find that education, political preference, and income were consistent determinants of support for the initiatives. Fischel (1979) is an early study of individual voter responses in New Hampshire. Fischel finds that income, occupation, and education are significant predictors of support for a wood pulp processing mill locating in a community. Matsusaka and Kahn (1997) analyze county-level data from environmental ballot propositions in California between 1970 and The authors use employment in different industries and occupations, educational attainment, and urban population and find that income and price proxies are sufficient to explain the demand for environmental protection while ideology and political preferences add relatively little explanatory power. Kahn (2002) expands this research by analyzing six additional California ballot proposals between 1994 and 1998 using data at the census-tract level. Kahn finds that the more educated, the young, and those not working in polluting industries are more 3

5 likely to demand environmental quality. Communities with higher proportion of Hispanics and African-Americans are more likely to vote pro-environment. These studies provide insights about average preferences for public environmental goods. However, the social composition of a community can also influence individuals support for environmental propositions. Among the relevant papers in the extensive literature on diversity and economic outcomes, Alesina, Baqir, and Easterly (1999) show that the local provision of public goods such as education and roads is lower in more heterogeneous communities. The authors argue that this finding is not necessarily due to minorities having a lower demand for these goods. Rather, they hypothesize that individuals from any given group are unlikely to support using tax revenues to provide goods they would share with members of other groups. Consistent with this insight, Alesina and La Ferrara (2000) develop a model in which individuals prefer to interact with people from their same racial group. In their model, individuals utility increases with the amount of the public good, however, as the number of different groups increases, the level of utility and the marginal utility from the public good decrease. 1 Their model provides the theoretical motivation for the empirical analysis in this paper. Everything else equal, greater community homogeneity can increase support for the public provision of environmental amenities if homogeneity facilitates coordination, defining common policy goals, civic involvement, and the sharing of public goods. To examine this hypothesis, this paper estimates whether racial and class homogeneity influences for environmental ballots after controlling for income and price effects, ideology, and voter turnout. The next section discusses the ballot proposals and Section 4 presents the empirical model, including the measures of community homogeneity. 4

6 3. California Initiatives Research has focused on ballot proposals in California because of the frequency of initiatives in the state. Although the results might not be easily generalized to the U.S. at large, substantial variability within the state facilitates estimating the demand for environmental regulation. This paper examines two environmental initiatives in the March 2000 primary elections: ballot proposals 12 and Proposition 12 provides for a bond issue of $2,100 million dollars to buy, develop, and protect natural areas. The acquisition of land and the development and protection of natural areas could affect the construction industry and forestry and farming workers. Some benefits would accrue to city dwellers since the part of the bond funds is used for recreational areas in urban areas. The proposition was controversial. The California Chamber of Commerce supported it citing its potential befits for tourism. Opponents argued that most of the funding would accrue for pork-barrel spending projects and land conservation rather than for improvements in air quality and recreational opportunities (Skeen, 2000). Among others, the Chairman of the Black Chamber of Commerce of Los Angeles County opposed the proposition. To the extent that different groups might have perceived this proposition to favor special interest projects, it is important to explore whether there are differential homogeneity effects across groups. Section 4 discusses how to account for this possibility. Proposition 13 provides for a bond issue of $1,970 million dollars to drinking water facilities, flood and watershed protection, water pollution treatment and control, water conservation, and water supply reliability. The acquisition and restoration of land for flood protection and watershed protection could affect the construction industry. A diverse group of organizations supported the measure, including the Agricultural Council of California, and the 5

7 California Manufacturers Association. Ballot proposal 12 passed with 63.2 percent of the total votes in favor and ballot proposal 13 with 64.8 percent of the votes. Support for proposal 12 is highly correlated with the votes in favor of proposal 13 (the correlation coefficient is.97). Following Kahn (2002), this paper explores two 2000 non-environmental initiatives for comparison purposes. Ballot proposal 22 adds a provision to the California Family Code to the extent that the state only recognizes marriage between a man and a woman. Ballot proposition 25 proposes campaign reforms. 3 This paper does not attempt to test specific hypotheses about initiatives 22 and 25. Rather, the models that are used for the environmental proposals are estimated for these social and political issues. The goal of this exercise is to explore whether the independent variables have different effects across initiatives in expected ways. Data on election results come from the Institute for Governmental Studies at the University of California Berkeley. Information about the initiatives comes from the California Ballot Measures Database (University of California, Hastings College of the Law). 4. The Empirical Model This paper presents two sets of models. First, models that follow the linear probability specification presented by Kahn (2002) who also uses California census-tract data. These base models are expanded to control for community homogeneity in terms of race and educational attainment. The tables present log-likelihoods and the Bayesian information criterion (BIC) to compare the expanded models to the base models. All the models use population-tract weights and include county fixed effects (the tables do not report these coefficients). Robust regression, median regression, and grouped logit regression models were also used. The results are very similar qualitatively. 6

8 The unit of analysis is census tracts. 4 The dependent variable is the share of tract voters who voted in favor of a ballot proposition (Table 1 presents summary statistics). The base models include median household income and income squared to control for a possible nonlinear relationship between the demand for environmental amenities and wealth along the lines of the Environmental Kuznets Curve hypothesis. The proportions of workers in manufacturing (PCTMANUF), farming or mining (PCTAGRI), and in finance, insurance, and real state (PCTFIRE) measure how the potential economic benefit or loss for workers in these industries might influence voting. Thus, these variables are proxies for price effects. Population density (DENSITY) controls for differential benefits between rural and urban areas. Population density squared controls for non-linear benefits of public goods due to congestion. The variable PCTPOP65 measures the percentage of the population at least 65 years old. Although younger people are usually believed to be more pro-environment, the elderly are likely to suffer more from environmental hazards than younger individuals do. The shares of Hispanics, African- Americans, Asians, and Native Americans in the population of each tract (PCTHISPANIC, PCTBLACK, PCTASIAN, and PCTNATIVE) might matter if minorities are more exposed to environmental problems than whites are. The proportion of individuals over 25 years of age who have at least a college degree (PCTCOLLEGE) or at most a high-school degree (PCTHS) control for effects due to variability in skills within specific occupations and for the potential influence of greater knowledge of the ballot propositions. The proportion of high-school dropouts is the default category in terms of educational attainment. Data for these variables come from the Census 2000 Long Form files. The models include the percent of votes for Republican candidates in the 2000 primary election, PCTREP. Hispanics and African-Americans in California are less likely to vote 7

9 Republican than whites are. If the models do not include political preference then the average effect of being Hispanic might be confounded with the average effect of voting Democrat. The models include voter turnout to account for a potential correlation between racial and class homogeneity and attitudes toward civic participation. 5 In this way, the models that include measures of diversity estimate the direct impact of homogeneity on support for the environmental propositions after controlling for the effect of diversity on turnout. TURNOUT is calculated as total votes divided by voting-age population. 6 Data for these variables come from the Institute for Governmental Studies at the University of California Berkeley. The expanded models add controls for community homogeneity. This paper considers five racial groups (non-hispanic white, non-hispanic black, non-hispanic Asian, non-hispanic Native American, and Hispanic) and three education groups (high-school dropouts, high-school graduates, and college graduates).the indices of homogeneity are calculated as the sum of group shares squared. The homogeneity index equals one in a perfectly homogenous community, for example, all residents are Hispanic, and approaches zero if the community is composed of many small groups. Using this variable, or the alternative fragmentation index defined as one minus the homogeneity index (Easterly and Levine, 1997), is the standard approach to measuring diversity. Despite its limitations (Alesina and La Ferrara, 2005), this variable provides a simple parsimonious measurement of racial and class composition in a community. An alternative to the index is to include group shares squared. If all groups respond to homogeneity in the same way, then the index and the squared terms generate similar insights, but using the index have the advantages of increasing degrees of freedom and reducing potential collinearity. However, if there are differential homogeneity effects across groups, the index 8

10 might obscure these effects. Thus, a second specification decomposes the homogeneity index into each of its components by estimating the coefficients on group shares squared. By decomposing the index one allows for the effects of homogeneity to be group-specific (Vigdor, 2002). The controversy surrounding proposition 12 justifies taking this approach: opponents, including the Chairman of the Black Chamber of Commerce of Los Angeles County, claimed most of the money would go to special interests and land preservation rather than improvements in pollution and recreational opportunities. In this case, greater coordination and information within a group opposing the perceived allocation of money might have created different homogeneity effects. 5. Results Table 2 reports the results of the base models. Table 3 and Table 4 show the results when the models include homogeneity indices and group shares squared, respectively. The tables present parameter estimates of linear probability models. 7 For initiatives 12, 13, and 22, less than 1 percent of the predicted probabilities fall outside the range of 0 to 100 percent while all predicted probabilities fall within the 0 to 100 range for initiative 25. Because the results for weighted least-square grouped logit estimations (that restrict predicted probabilities to the unit circle) are qualitatively very similar and the estimates of the linear probability model are easier to interpret, the latter are discussed here. 8 To account for heterocedasticity, the models are weighted by the population of each tract and robust standard errors are computed. Likelihood-ratio tests indicate that there is overwhelming statistical evidence that the specifications that include either group shares squared or indices of homogeneity fit the data 9

11 better than the base models. For example, comparing the log-likelihood of column 1 in Table 3 to the log-likelihood of column 1 in Table 2, there is overwhelming statistical evidence that a model that includes the racial homogeneity index fits the data better than the base model (the chi-square statistic with 1 degree of freedom is with a p-value below.1 percent). Thus, controlling for racial homogeneity improves the model. This conclusion also holds when accounting for educational attainment homogeneity. The values of the BIC also indicate that even after penalizing the model for the additional explanatory variables, models that include the indices or group shares squared are statistically better than the base models. Overall, models that account for the degree of homogeneity in a community produce results that are statistically better and, as the following sections discuss, substantively richer than the base models. 5.1 The Base Models The negative coefficient on income is consistent with Kahn (2002) who argues that the public provision of environmental goods, rather than the goods themselves, might be considered an inferior good as income increases. The coefficient on income squared reveals a convex relationship. These estimates are statistically significant at the 1 percent level or better. The coefficients in column 1 and column 2 indicate that the turning point occurs when median household income is approximately $100, Consistent with Kahn (2002) and Kahn and Matsusaka (1997), opposition to the environmental initiatives is stronger in tracts with more employment in farming and mining. Opposition comes from those sectors that are more likely to be negatively influenced by the propositions. Employment in manufacturing has a positive and significant effect for initiative 13 (the California Manufacturers Association supported this water bond) while employment in 10

12 industries where workers are likely to be more educated is also positively correlated with support for the initiatives. The proportion of the population 65 years of age and older is positively correlated with votes for initiative 13. The coefficient on density is positive and statistically significant in the models for initiatives 12 and 13. Communities that vote Republican are less likely to support the environmental initiatives than communities that vote Democrat do. The importance of political preference stands in clear contrast to the relatively minor impact that such factors have in the analysis by Kahn and Matsusaka (1997). Besides the fact that the unit of analysis is different, the 2000 elections were a period of political polarization. It is also interesting to note that voter turnout is negatively correlated with votes in favor of ballot propositions 12, 13, and 25 but positively correlated with support for limits on marriage. This pattern of effects aligns with the estimated coefficients on the percent of votes for a Republican candidate. Finally, the independent variables have different effects across ballots proposals in expected ways. For example, PCTPOP65 is positively related to support for the initiative that imposes limits on marriage. Higher educational attainment correlates with more votes in favor of the environmental initiatives but with fewer votes for the limit-on-marriage proposal. Those who vote for a Republican candidate and those employed in farming or mining are more likely to vote for the limit-on-marriage initiative and against the environmental initiatives. 5.2 Racial Homogeneity The results in Table 2 show that communities where Hispanics are a greater fraction of the population are more likely to support the public provision of environmental goods than communities that are predominantly white. There is a similar effect for African-Americans and 11

13 support for ballot proposition 13. On the other hand, the coefficient on PCTBLACK is negative for initiative 12. However, there is evidence that this coefficient is biased when the models do not account for racial homogeneity. Table 3 reports the results of models that include either the homogeneity index or group shares squared in addition to group shares. Consistent with the literature on fragmentation, racial homogeneity is positively correlated with support for the provision of public goods. The coefficients on the homogeneity index are positive and significant at the 1 percent level or better. Interpreting the estimate can be misleading because it is impossible to increase the index without changing the values of the group shares. However, excluding the group shares may bias the estimate on the index (Vigdor, 2002). With this caveat in mind, the coefficient estimate on racial homogeneity is such that one standard deviation increase in the index increases the proportion of votes for initiative 12 by.044 standard deviations, everything else equal. 10 For comparison purposes, one standard deviation increase in PCTFIRE and PCTCOLLEGE increases support for initiative 12 by.016 and.37 standard deviations respectively. Decomposing the homogeneity index into group shares squared offers interesting insights. Whites and Hispanics are more likely to support the environmental initiatives as their share in the population increases. Support for the propositions increases with the share of African-Americans, but at a decreasing rate. Graph 1 and Graph 2 show the relationships between predicted votes for proposals 12 and 13 and race group shares using the estimates from the models in Table The graphs show a convex relationship between expected votes and proportion of whites and Hispanics. However, the relationship between expected votes and proportion of African-Americans is concave. The turning point of the relationship between 12

14 predicted votes and the African-American group share occurs for communities that are approximately 70 percent African-American (41 census tracts above this value in the sample). The turning point of the relationship between Asians and expected votes does not occur in the sample 12. These results clearly illustrate that the strong convex relationship for whites and Hispanics is not present among African-Americans. The results for African-Americans might appear counterintuitive. However, there is documented evidence of opposition to ballot proposal 12 among the African-American business community. In particular, the Black Chamber of Commerce of Los Angeles County was one of the signatories of the rebuttal to the argument in favor of ballot proposition 12. Greater coordination and within-group homogeneity can lower the demand for environmental regulation if one group understands the provision of the public good to be a bad. These results show the benefits of decomposing the homogeneity index and suggest that researchers and policy-makers can gain interesting insights from this modeling approach. 5.3 Educational Attainment Homogeneity Table 4 reports the parameter estimates of the models that control for diversity in educational attainment. Regarding the coefficients on group shares, the results are consistent with previous literature: communities where a larger share of individuals holds college degrees support the public provision of public goods more than communities where educational attainment is lower. The more homogenous the community is in terms of educational attainment the greater the support for the environmental initiatives. As the index increases by one standard deviation, the proportion of votes for the environmental initiatives increases by approximately.03 standard deviations, everything else equal. The models with group shares squared suggest that positive 13

15 homogeneity effects in the high-school dropout group drive these results. The coefficients on group shares squared for the high-school dropout group share squared are positive and significant at the 1 percent level. Thus, after controlling for household income, occupation, and race, members of the most economically disadvantaged group are more likely to support the public provision of environmental goods as their share of the tract population increases. If more homogeneity is necessary to encourage participation in referenda, homogeneity might facilitate and encourage the demand for environmental quality among low-income households that are more likely to be exposed to environmental hazards. (The paper by Grineski, Bolin, and Boone (2007) is a recent examination of environmental equality. The authors find that low-income communities suffer an unequal distribution of air pollution in metropolitan Phoenix.) 5.4 Check for Endogenous Sorting Bias and Multicollinearity A common limitation of studies that explore heterogeneity effects is the possibility that the demographic composition of the unit of analysis is not random as individuals endogenously sort into communities. 13 It is possible to sign the direction of the bias if the following argument holds: individuals with a stronger preference for mixing with people from other groups are more likely to live in less homogeneous communities. At the same time, individuals with a stronger preference for mixing are less likely to suffer a loss of utility from interacting and sharing goods with people from other groups. In this case, sorting bias would decrease point estimates of the homogeneity index and group shares squared relative to the true effects (Vigdor, 2004). The finer the unit of aggregation, the more likely sorting is and the more severe sorting bias should be. Thus, estimates using block-level data should be smaller than point estimates using tract-level data. The results in Table 5 confirm this argument when estimating the effects 14

16 of racial homogeneity at the census-block level. The estimates of the index and group shares squared are smaller. The exception is the estimate for Hispanic share squared that takes on similar values using either tract data or block data. Table 6 presents the results when the models control for educational homogeneity. The estimates of high-school dropout share squared are marginally lower using block data. The estimates on the index barely change and the marginal effects of college share squared are now positive. Overall, the results from models using census block data suggest the point estimates of the controls for homogeneity might be lower bounds of the true effects, in particular for racial homogeneity. The models were also estimated using county-level data (in this case the estimates should be larger). However, given the complexity of the models and the fact that there are only 58 counties in California, the coefficients were estimated imprecisely and 95 percent confidence intervals for the homogeneity indices were notably wide. Multicollinearity is another potential problem in models using census data. The consequences of multicollinearity would be wide changes in standard errors and possibly in point estimates. To explore whether multicollinearity is affecting the results, the models were estimated dropping those variables that are more likely to cause the problem. First, models omitting income squared and density squared produce point estimates and standard errors that are very similar to the original results. Other high pair-wise correlation coefficients occur between PCTCOLLEGE and PCTHISPANIC (-.70) and HINCOME and PCTCOLLEGE (.74). Variance inflation factors can identify more complex linear relationships. The variables with relatively large variance inflation factors are PCTCOLLEGE and HINCOME. Omitting PCTCOLLEGE does not affect the signs and statistical significance of the racial and education indices but biases downwards the estimates of racial group shares. Omitting HINCOME does not 15

17 affect the signs and statistical significance of the racial and education indices and there are minor changes in the estimates of racial group shares. Thus, dropping relevant variables does not affect standard errors, biases down the estimates of racial group shares, and does not affect much other estimates. Finally, a more drastic approach is to estimate models only with county dummies, group shares, and the index or group shares squared. The indices are still positive and statistically significant in these reduced specifications and the concave relationship between support and share of African-Americans still hold. Overall, there is no evidence that multicollinearity is influencing the results. 6. Summary and Conclusions This paper examines the effects of racial and class homogeneity on the public provision of environmental goods. Analysis of 2000 census-tract data of election results in California shows that more heterogeneity is consistently correlated with lower support for environmental initiatives. There are differences across groups in how homogeneity affects election results. Whites and Hispanics are more likely to support environmental initiatives in more homogenous communities. Support is greater in communities with more African Americans than in communities with more non-minorities, but the support increases at a decreasing rate as homogeneity among African-Americans increases. This might be explained by the fact that there was opposition to the initiatives from the African-American business community. In particular, the Black Chamber of Commerce of Los Angeles County opposed ballot proposition 12. Opponents argued that most of the money would go to special interests and land preservation rather than improvements in pollution and recreational opportunities. There is also some evidence in the literature that African-Americans might be more concerned about local pollution 16

18 problems than about nature conservation issues (Mohai and Bryant, 1998). Within-group homogeneity can lead to greater coordination and overcoming of structural barriers that inhibit participation of minorities in the political process (Mohai, 1990), thus reducing the demand for a public bad. Regarding education, the more homogenous the community is the greater the support for the environmental initiatives. The models with group shares squared indicate that these results are mostly due to strong affinity effects in the high-school dropout group. If more homogeneity is necessary to encourage participation in referenda, homogeneity might facilitate and encourage the demand for environmental quality among low-income households that are more likely to be exposed to environmental hazards. Overall, this paper finds that the degree of racial and educational attainment homogeneity in a community influences the demand of environmental goods through voting on referenda and that decomposing homogeneity indices can produce a richer picture of the links between sociodemographic characteristics and the demand for environmental goods. These results can be useful for researchers, advocates, and policymakers studying and implementing policies to address issues of environmental justice and, more generally, for social scientists interested in the determinants of demand for environmental goods. 17

19 Footnotes [1] There are also positive effects of diversity. Alesina and La Ferrara (2005) discuss how diversity can increase innovation and productivity. [2] Regarding the primary system in California, party members can vote only for their party s candidates but people can affiliate to any party during Election Day. In all other partisan contests, registered voters can vote for any candidate. This modified open system does not appear, a priori, to restrict severely who can vote. In addition, the models include voter turnout, controlling for potential biases created by the voting system. [3] Proposal 22 has no fiscal effects while the estimated cost of proposal 25 was $55 million annually as the state government would have provided larger public funds for campaigning. Proposal 22 passed with 61.4 percent of the vote and initiative 25 failed with 34.7 percent of the vote. [4] According to the Census, tracts are defined by permanent visible boundaries (streets, roads, rivers) and contain between 2,500 and 8,000 residents. [5] Civic involvement might also depend on the degree of transiency of the population. However, including in the models percent owner-occupied housing units or percent of population (5+) in same house in 1995 barely changes point estimates or standard errors. The controls are statistically significant at the 5 percent level for proposition 12 but not for proposition 13. [6] This calculation underestimates actual voter turnout because voting-age population includes individuals who are illegible to vote. However, the results are unchanged if the specifications include the percent of non-naturalized foreign-born population in the tract. In addition, county fixed effects will capture some of the possible effects of differences in foreign-born population 18

20 across counties. The coefficient on voter turnout should not be interpreted causally as it is possible that it is the decision to vote on a given ballot proposal that influences turnout. [7] For the linear probability models, the log-likelihood reaches a maximum when the density is greater than 1, thus the log-likelihood is positive and the BIC is negative. Because the loglikelihood is positive, a larger negative BIC indicates a better model. [8] The results from weighted least-squares logistic regression models for grouped data are available from the author upon request. [9] The 99-percentile of the distribution of median household income is approximately $135,000. [10] The specifications in Table 3 were also estimated excluding group shares. The estimates on the racial homogeneity index in the models predicting support for propositions 12 and 13 were very similar,.039 and.018 respectively. Thus, including group shares does not drastically influence the interpretation of these effects. [11] Predicted support is estimated using sample means and setting the dummy for Los Angeles county (the county with most tracts) equal to one. The shape of the relationship is unaffected by the selection a county. [12] The estimate on the Asian group share squared is insignificant in the logistic regression specification. [13] As Alesina, Baqir, and Easterly (1999) argue, some degree of community heterogeneity is likely to persist because of legal constraints, economies of scale in the provision of public goods, limited mobility, and the multidimensional features of most public goods. 19

21 References Alesina, Alberto, Batir, Reza, and William Easterly (1999): Public Goods and Ethnic Divisions, Quarterly Journal of Economics, 114: Alesina, Alberto and Eliana La Ferrara (2000): Participation in Heterogeneous Communities, Quarterly Journal of Economics 3: Alesina, A. and E. La Ferrara (2005): Ethnic diversity and economic performance, the Journal of Economic Literature, XLIII: Costa, Dora L., and Matthew E. Kahn (2003): Understanding the decline in American social capital, , Kyklos, 56(1): Deacon, Robert and Perry Shapiro (1975): Private Preference for Collective Goods Revealed Through Voting on Referenda, American Economic Review, 65: Dubin, Jeffrey A., D. Roderick Kiewiet, and Charles Noussair (1992): Voting on growth control measures: Preferences and strategies, Economics and Politics, 4(2): Easterly, W. and, Levine, R. "Africa's Growth Tragedy: Policies and Ethnic Divisions." Quarterly Journal of Economics, 112 (1997): Fischel, William A. (1979): Determinants of Voting on Environmental Quality: A Study of a New Hampshire Pulp Mill Referendum, Journal of Environmental Economics and Management, 6: Gates, Gary J., and Jason Ost (2004): The Gay and Lesbian Atlas, The Urban Institute Press, Washington, D.C. Glaeser, Edward L., Laibson, David, Scheinkman, Jose A., and Christine L. Soutter (2000): Measuring trust, Quarterly Journal of Economics, 115(3): Grineski, Sara, Bolin, Bob, and Christopher Boone (2007): Criteria Air Pollution and Marginalized Populations: Environmental Inequality in Metropolitan Phoenix, Arizona, Social Science Quarterly, 88(2): Institute for Governmental Studies at the University of California Berkeley, Kahn, Matthew (2002): Demographic Change and the Demand for Environmental Regulation, Journal of Policy Analysis and Management, 21(1): Kahn, Matthew E., and John G. Matsusaka (1997): Demand for environmental goods: Evidence from voting patterns on California initiatives, Journal of Law and Economics, 40(1):

22 Matsusaka, John G. (1992): Economics of Direct Legislation, Quarterly Journal of Economics, 107: Mohai, Paul (1990): Black Environmentalism, Social Science Quarterly, 71(4): Mohai, Paul, and Bunyan Bryant (1998): Is There a Race Effect on Concern for Environmental Quality? Public Opinion Quarterly, 62: Okten, Cagla, and Una Osili-Okonkwo (2004): Contributions in Heterogeneous Communities: Evidence from Indonesia, Journal of Population Economics, 17(4): Sampson, Robert J., Raudenbush, Setephen W., and Felton Earls (1997): Neighborhoods and Violent Crime: A Multilevel Study of Collective Efficacy, Science, vol. 277, no. 5328: Skeen, Jim (2000): Measure has AV provisions opportunities or pork? Daily News (Los Angeles, CA). accessed July 29, University of California, Hastings College of the Law. California Ballot Measures Database, res/ca_ballot_measures_main.htm Vigdor, Jacob L. "Community Composition and Collective Action: Analyzing Initial Mail Response to the 2000 Census." The Review of Economics and Statistics, February 2004, 86(1): Vigdor, Jacob L. (2002): Interpreting Ethnic Fragmentation Effects, Economics Letters, 75:

23 Table 1: Summary Statistics, Census Tracts Variable Description N Mean (Standard Dev.) yes12 Percent votes in favor of proposition 12, (.133) parks, clean water, clean air, and coastal protection yes13 Percent votes in favor of proposition 13, (.125) drinking water, clean water, watershed and flood protection yes22 Percent votes in favor of proposition 22, (.150) limits on marriage yes25 Percent votes in favor of proposition 25, (.059) campaign reform POPULATION Population (2136.5) DENSITY People per square meter (.0036) HINCOME Median Household Income ($10,000) (24.75) PCTWHITE White Population (percent of total ) (.226) PCTBLACK African-American Population (percent of (.116) total ) PCTNATIVE Native American Population (percent of (.018) total ) PCTASIAN Asian and Pacific Population (percent of (.132) total ) PCTHISPANIC Hispanic Population (percent of total ) (.257) RACE Race Homogeneity Index (five racial (.171) groups) PCTPOP65 Population age 65 or above (percent of (.081) total) PCTCOLLEGE Population with college degree (percent of (.189) population 25 years old or older) PCTHS Population with high-school degree or (.193) equivalent (percent of population 25 years old or older) PCTHSDROP High-school dropouts (percent of population (.193) 25 years old or older) EDUCATION Education Homogeneity Index (three (.059) education groups) PCTAGRI Population employed in agriculture and (.060) mining (percent of population 16 years old or older) PCTMANUF Population employed in manufacturing (.079) (percent of population 16 years old or older) PCTFIRE Population employed in finance, insurance, (.039) and real estate (percent of population 16 years old or older) PCTREP Votes for Republican candidates in (.199) primary elections (percent of total) TURNOUT Voting-age turnout rate: total votes divided by voting-age population (.132) 22

24 Table 2: Base Models (1) (2) (3) (4) yes12 yes13 yes22 yes25 PCTMANUF *** *** (.0119) (.0118) (.0285) (.0162) PCTAGRI *** ***.0869*** *** (.0173) (.0170) (.0232) (.0149) PCTFIRE.0644***.0660***.1259*** (.0229) (.0248) (.0419) (.0270) PCTPOP ***.1150***.1789*** (.0092) (.0088) (.0110) (.0087) PCTBLACK ***.0148**.3246*** *** (.0067) (.0067) (.0109) (.0080) PCTNATIVE *.0048 (.0345) (.0545) (.1704) (.0885) PCTASIAN **.2672***.0404*** (.0061) (.0058) (.0098) (.0059) PCTHISPANIC.0444***.0595***.1814*** *** (.0070) (.0070) (.0135) (.0090) PCTCOLLEGE.2499***.2286*** ***.1374*** (.0131) (.0135) (.0186) (.0133) PCTHS *** (.0137) (.0137) (.0155) (.0121) PCTREP *** ***.5880*** *** (.0081) (.0083) (.0141) (.0095) HINCOME *** ***.0251*** *** (.0015) (.0016) (.0025) (.0018) HINCOME* HINCOME.0008***.0007*** ***.0006*** (.0001) (.0001) (.0001) (.0001) DENSITY *** *** * *** (.4777) (.6116) (.6236) (.3944) DENSITY* DENSITY *** *** ( ) ( ) (2.4517) ( ) TURNOUT *** ***.0662*** *** (.0142) (.0139) (.0203) (.0129) Constant.9303***.9154***.1694***.4785*** (.0110) (.0136) (.0195) (.0137) Observations LL BIC Robust Standard errors in parentheses. All models are weighted by the population of each tract and include county dummies. * significant at 10%; ** significant at 5%; *** significant at 1%. 23

25 Table 3: Effects of Race Homogeneity (1) (2) (3) (4) yes12 yes12 yes13 yes13 PCTBLACK ***.0281***.1455*** (.0080) (.0195) (.0081) (.0202) PCTNATIVE (.0350) (.0538) (.0541) (.0677) PCTASIAN.0126*.1161***.0301***.1253*** (.0073) (.0211) (.0072) (.0211) PCTHISPANIC.0448***.0471***.0598***.0506*** (.0071) (.0156) (.0071) (.0155) Race Homogeneity Index.0327***.0232*** (.0060) (.0061) White share squared.0741***.0605*** (.0101) (.0104) Black share squared *** *** (.0185) (.0188) Native share squared (.0788) (.1124) Asian share squared *** *** (.0246) (.0238) Hispanic share squared.0657***.0677*** (.0161) (.0163) Constant.9015***.8610***.7458***.7007*** (.0127) (.0154) (.0128) (.0155) Observations LL BIC Robust Standard errors in parentheses. All models are weighted by the population of each tract and include county dummies. * significant at 10%; ** significant at 5%; *** significant at 1%. 24

26 Table 4: Effects of Education Homogeneity (1) (2) (3) (4) yes12 Yes12 yes13 yes13 PCTCOLLEGE.2547***.3762***.2331***.3623*** (.0139) (.0339) (.0142) (.0346) PCTHS * *** (.0145) (.0560) (.0144) (.0603) Education Homogeneity Index.0644***.0599*** (.0195) (.0188) College share squared (.0235) (.0256) High-school share squared High-school dropout share squared (.0552) (.0550).1632***.1915*** (.0298) (.0306) Constant.8909***.8147***.8788***.7660*** (.0176) (.0262) (.0189) (.0293) Observations LL BIC Robust Standard errors in parentheses. All models are weighted by the population of each tract. All models include county dummies and additional controls in Table 2. * significant at 10%; ** significant at 5%; *** significant at 1%. 25

27 Table 5: Effects of Race Homogeneity using Census Blocks Data (1) (2) (3) (4) yes12 yes12 yes13 yes13 PCTBLACK ***.0392***.0151***.0828*** (.0055) (.0117) (.0052) (.0116) PCTNATIVE (.0203) (.0249) (.0235) (.0268) PCTASIAN.0082*.0536***.0294***.0748*** (.0045) (.0121) (.0044) (.0115) PCTHISPANIC.0318*** ***.0051 (.0043) (.0093) (.0042) (.0091) Race Homogeneity Index.0222***.0190*** White share squared Black share squared Native share squared Asian share squared Hispanic share squared (.0034) (.0034).0285***.0249*** (.0049) (.0049) *** *** (.0123) (.0119) ** (.0846) (.0971) *** *** (.0154) (.0146).0674***.0726*** (.0104) (.0103) Constant.9738***.9624***.7525***.7397*** (.0091) (.0094) (.0132) (.0135) Observations Robust Standard errors in parentheses. All models are weighted by the population of each tract. All models include county dummies and additional controls in Table 2. * significant at 10%; ** significant at 5%; *** significant at 1%. 26

28 Table 6: Effects of Education Homogeneity using Census Blocks Data (1) (2) (3) (4) yes12 yes12 yes13 yes13 PCTCOLLEGE.2150***.2917***.1949***.2772*** (.0079) (.0219) (.0080) (.0212) PCTHS *** *** (.0079) (.0356) (.0076) (.0362) Education Homogeneity Index.0630***.0604*** College share squared High-school share squared High-school dropout share squared (.0119) (.0114).0312**.0414** (.0159) (.0162) (.0359) (.0349).1458***.1645*** (.0183) (.0178) Constant.9576***.8859***.7335***.6427*** (.0120) (.0169) (.0153) (.0195) Observations Robust Standard errors in parentheses. All models are weighted by the population of each tract. All models include county dummies and additional controls in Table 2. * significant at 10%; ** significant at 5%; *** significant at 1%. 27

29 Graph1: Predicted Votes for 12 vs Race Group Shares Group Shares White Hispanic Black Asian Graph 2: Predicted Votes for 13 vs Race Group Shares Group Shares White Hispanic Black Asian 28