DEPARTMENT OF ECONOMICS AND FINANCE COLLEGE OF BUSINESS AND ECONOMICS UNIVERSITY OF CANTERBURY CHRISTCHURCH, NEW ZEALAND

Transcription

1 DEPARTMENT OF ECONOMICS AND FINANCE COLLEGE OF BUSINESS AND ECONOMICS UNIVERSITY OF CANTERBURY CHRISTCHURCH, NEW ZEALAND The Effect of Neighborhood Diversity On Volunteering: Evidence From New Zealand Jeremy Clark and Bonggeun Kim WORKING PAPER No. 09/2009 Department of Economics and Finance College of Business and Economics University of Canterbury Private Bag 4800, Christchurch New Zealand

2 The Effect of Neighborhood Diversity On Volunteering: Evidence From New Zealand * Jeremy Clark a Bonggeun Kim b A growing empirical literature has found that neighborhood heterogeneity lowers people s likelihood of contributing to public goods, but has struggled to address the issues raised by definition of neighborhood size or endogenous neighborhood location. We show that the estimated effect of concave neighborhood characteristics like heterogeneity on outcomes of interest may be biased if boundaries are defined too broadly, or underestimated if they are defined too narrowly. We also argue that fixed effects panels that follow neighborhoods over time may address problems of endogenous self-selection between neighborhoods if sorting takes place in the manner of the Tiebout hypothesis. We apply both points using three rounds of New Zealand census data to test whether volunteering rates are lowered by neighborhood heterogeneity by race/ethnicity, birthplace, income or language. We find boundaries matter, and that only ethnic/racial heterogeneity is robustly associated with lower volunteering. Key words: heterogeneity, neighborhood effects, volunteering JEL Classification: D13, D64, H31 a Author for correspondence. Department of Economics and Finance, University of Canterbury, Private Bag 4800, Christchurch, 8140 New Zealand. jeremy.clark@canterbury.ac.nz Phone: Fax: b Department of Economics, Seoul National University, 599 Gwanak-no, Gwanak-gu, Seoul , Korea. bgkim07@snu.ac.kr Fax:

3 1. Introduction Do individuals in communities that become more heterogeneous lose concern for the welfare of others? Support for this provocative claim has emerged in the past decade over various dimensions of heterogeneity and manifestations of concern for others. To address this question empirically, researchers have commonly tested for an effect of a concave neighborhood characteristic (heterogeneity), on people s propensity to give time or money to public goods, become members of organizations, return census forms, express trust in others, and so on. More generally, researchers have used a similar approach to test for the effect of various concave neighborhood characteristics on various outcomes of interest. Examples include the effect of neighborhood income inequality on mortality rates (Lynch et al. 1998, Mellor and Milyo 2001, Deaton and Lubotsky 2003, Lochner et al. 2001), self reported health (Blakely et al. 2002, Mellor and Milyo 2002), and homicide rates (Mellor and Milyo 2001), or the effect of neighborhood racial diversity on population and economic growth (Rappaport 1999, Alesina and La Ferrara 2005), and adolescent sexual activity (Brewster 1994). 1 In this paper we make two contributions. By using unusually fine geographical units and a panel of those units, we address two potential problems in previous studies: neighborhood size and endogenous neighborhood choice. Endogenous neighborhood choice may be characterized by a Tiebout-style hypothesis, where people self-select to live in areas with others who share their preferences regarding the optimal trade-off between private consumption and public goods provision (Tiebout 1956). Thus, if people who are less inclined to volunteer are attracted to locate in urban centres, which tend to be more heterogeneous, then the literature s estimated cross sectional effects of heterogeneity on volunteering may be exaggerated. Similarly, the size of neighborhood used to estimate the effect of a concave characteristic (like heterogeneity) on people s behavior may introduce bias if it is defined too broadly. This is because larger unit analysis is likely to incorrectly capture the effect of differences between constituent smaller units due to the concavity of the heterogeneity measure. In contrast, if the size of neighborhood is defined too narrowly to 1 There is a larger empirical literature testing for the effect of non-linear (but not necessarily concave) neighborhood characteristics on various outcomes, covering a wider range of applications (for surveys, see Deaton (2003) or Durlauf (2004)). While neighborhood boundary choice may affect results for all non-linear measures, we have focused here on the consequence of using concave neighborhood measures in particular. 1

4 include the effects of heterogeneity in adjacent areas, the lower bound of heterogeneity s total effect can still be estimated. We address both endogenous location and size effects to further the empirical investigation of the effects of neighborhood heterogeneity on people s contribution of time to public goods, or volunteering. We use a heretofore untapped data source that provides several advantages over preceding studies: New Zealand census data on volunteering rates at the unusually fine levels of meshblock (= 100 people) and area unit (= 2000 people) for 1996, 2001 and We test whether heterogeneity of ethnicity/race, languages spoken, birthplace, or household income affects New Zealander s likelihood of volunteering. Questions regarding volunteering were asked of all New Zealanders in 1996, 2001 and 2006, enabling us to construct both pooled cross section and neighborhood fixed effects regressions for the entire country. The New Zealand census releases an unusually comprehensive list of covariates for all three years, allowing our cross section regressions to better control for confounding neighborhood characteristics such as deprivation, crime, housing and employment status, that may be correlated with heterogeneity. By using two levels of neighborhood boundary, we address the sensitivity of our empirical findings to neighborhood size and the robustness of previous studies using broader units. Our fixed effects regressions, following neighborhoods over time, can go part way to addressing residual endogeneity in our cross section analysis, based on a Tiebout style argument that unobserved attitudes towards volunteering may remain stable within neighborhoods over time even as the people in them come and go. The rest of the paper will proceed as follows. In Section 2, we review the growing empirical literature investigating the effect of neighborhood diversity on public good contributions and trust. In Section 3, we present a simple model that highlights the importance of neighborhood size when estimating the effects of concave neighborhood characteristics on outcomes of interest. In Section 4 we present descriptive statistics regarding volunteering and various measures of heterogeneity in New Zealand, followed by our estimation methods and results. In Section 5 we provide a discussion and conclusions. 2. Diversity, public goods and trust In recent years researchers have examined the effects of increased neighborhood heterogeneity by race, ethnicity, education, income or first language, on an individual s propensity to volunteer, contribute to fundraisers, be a member of any organization, trust 2

5 others or support welfare programs. 2 While the bulk of empirical studies have been carried out using data from the United States, others have used surveys from Australia, Kenya, Sweden, and the United Kingdom (see below). The most common approach has been to regress individuals survey responses on individual and neighborhood level characteristics, with the latter separately taken from census data for the neighborhood or region in which the respondents live. A selective summary of this literature might suggest that there is indeed a robust negative relationship between heterogeneity and support for public goods and trust in others. Alesina and La Ferrara (2000, 2002), using pooled cross sectional data from multiple years of the U.S. General Social Survey, find that increased neighborhood heterogeneity of income or race lowers an individual s probability of reporting membership in any organization, or of agreeing that most people can be trusted. Costa and Kahn (2003a), using pooled cross section data from two years of the U.S. Current Population Survey (CPS), find that increased heterogeneity of income or birthplace lowers an individual s probability of membership in any organization or of volunteering. Costa and Kahn (2003b) using the CPS and the DDB Lifestyle Survey, find that increased racial heterogeneity lowers individuals probability of volunteering. Vigdor (2004) finds that U.S. census tracts that were more heterogeneous in race, age or educational attainment in 2000 had lower response rates of households mailing in completed census forms. Returning such forms can be seen as a local public good, because local public funding depends on enumerated census tract population. Putnam (2007), using responses from the U.S. Social Capital Community Benchmark Survey of 2000, finds that individuals in more racially heterogeneous census tracts were less likely to give to charity or volunteer, trust others (whether of their own or other races), register to vote, or be optimistic that others would cooperate in dilemmas of collective action. Finally, Luttmer (2001), again using pooled cross section data from multiple years of the General Social Survey, finds that support for government welfare spending is lower in more racially heterogeneous states, and that this effect is significant in explaining some of the variation in generosity of welfare across states. 2 Political scientists such as Robert Putnam (2007) have emphasized the effects of heterogeneity on social capital, or peoples beliefs and actions that contribute to social networks and the associated norms of reciprocity and trustworthiness. 3

6 While the bulk of the adverse findings regarding heterogeneity have come from the United States, a limited number of papers have found similar results elsewhere, particularly related to trust. Leigh (2006), using the 1997/98 Australian Community Survey and 1996 Australian census data, finds that increased neighborhood heterogeneity of country of birth or of language spoken at home lowers the probability of individuals trusting their neighbors. Letki (2008), using data from the British Home Office Citizenship Survey of 2001 and census, finds that increased ward level racial heterogeneity lowers individuals trust in their neighbors. Gustavsson and Jordahl (2008), using the 1994 and 1998 Swedish Election Studies Panel and county level census data, find that increased income inequality in the lower half of the income distribution, or in the proportion of a respondent s county that is foreign born, lowered reported trust in others. Finally, Miguel and Gugerty (2005), using an NGOfunded survey of schools in rural Kenya, find that local ethnic heterogeneity is associated with sharply lower voluntary school fundraiser contributions, resulting in lower quality primary schools. Theoretically, the negative effects of heterogeneity on people s trust of others or contributions to public goods has been attributed to their innate preference to interact with others like themselves, which can cause social networks and the trust they generate to atrophy as dissimilarities increase (Alesina and La Ferrara 2000, 2002, Putman 2007). People may be less likely to internalize the benefits they bestow on the community at large by contributing to public goods if they perceive less similarity between themselves and that community (Vigdor 2004). Linguistic heterogeneity in particular may increase the costs of communication and reduce of quality of information exchanged in networks, making investments in such networks less attractive (Leigh 2006). Ethnic or cultural heterogeneity may also reduce the ability of communities to impose negative social sanctions for free riding across ethnic lines (Miguel and Gugerty 2005). Some have tried to organize these various causal mechanisms via preferences, strategies, and production (Alesina and La Ferrara (2005), Habyarimana et al. (2007). While the above (selective) summary might suggest conclusive evidence that heterogeneity corrodes people s trust in others and their contributions towards public goods, a closer inspection of this literature show the results to be less robust, and more problematic than they first appear. Regarding robustness, papers testing for the effects of different kinds of heterogeneity often find that some kinds matter, but others do not, or that multiple kinds may matter when 4

7 tested individually, but not when tested jointly. And the type of heterogeneity that affects behavior or trust seems to vary from study to study. For example, Alesina and La Ferrara (2002) find that higher neighborhood racial heterogeneity (white, black, Asian etc) lowered trust in other people, but higher heterogeneity of ethnic origin did not, while higher income inequality lowered trust when racial heterogeneity was excluded, but had no significant effect when it was included. Similarly, Alesina and La Ferrara (2000) find that while income, racial and ethnic heterogeneity all lowered the probability of group membership when entered separately, only income heterogeneity mattered when all three measures were included (Alesina and La Ferrara 2002). And while Letki (2008) finds that higher racial heterogeneity lowers trust in the United Kingdom, it has no effect on people s likelihood of formal or informal volunteering, unlike in the United States (Putnam (2007) and Costa and Kahn (2003b)). Again, while higher birthplace heterogeneity lowers trust in others in Sweden or Australia (Gustavsson and Jordahl 2008, Leigh 2006), higher ethnic heterogeneity (i.e. birthplace of ancestors) in Sweden does not. Regarding problems, many of the studies on heterogeneity have been based on cross sectional data (Putnam 2007, Alesina et al. 1999, Letki 2008, Leigh (2006), Vigdor (2004), Miguel and Gugerty 2005). They thus cannot be sure that effects attributed to heterogeneity are not instead caused by omitted variables that are correlated with heterogeneity. Letki (2008) in particular argues that neighborhood deprivation, poverty and crime may correlate with ethnic diversity yet be inadequately captured is many preceding studies, making diversity wrongly appear responsible for social withdrawal. Cross sectional studies also cannot determine whether it is the level of heterogeneity or changes in the level of heterogeneity that is affecting people s behavior. 3 There remains a problem, however, which heterogeneity studies have not generally recognized: how is the (often constrained) choice of neighborhood size affecting results? The coarseness of neighborhoods used has varied enormously. Gustavsson and Jordahl (2008) use Swedish counties, which contains 200, ,000 or even over one million people. Alesina and La Ferrara (2000, 2002), Costa and Kahn (2003a) and Luttmer (2001) use a 3 Luttmer (2001), Costa and Kahn (2003a, 2003b), and Alesina and La Ferrara (2000, 2002), construct pseudo panels of cross sectional survey data, which rely for legitimacy on the representativeness of each wave of the survey. Poterba (1997) uses panel data at the state level. Of the papers we have identified, only Gustavsson and Jordahl (2008) use true panel data at the individual level, using survey data. 5

8 respondent s US Metropolitan or Primary Metropolitan Statistical Area (MSA/PMSA), which contain a urban core of at least 50,000 and surrounding suburbs and affiliated towns. Letki uses the U.K census level of ward, which contain anywhere from hundreds to over 30,000 people. Alesina et al. (1999) uses U.S. county. Leigh (2006) uses the Australian postal area, typically containing 20,000 people. Since heterogeneity of income, race etc. can vary dramatically in just a few city blocks, the heterogeneity people experience most intimately may vary widely between areas of these coarsely defined neighborhoods. At the smaller end, Vigdor (2004) and Putnam (2007) define neighborhood at the U.S. census tract level, which commonly involves three to five thousand people. We turn next to a simple model to illustrate why the boundary of neighborhood used in these studies matters. The issue of endogenous neighborhood choice is addressed subsequently in Section Who are the people in your neighborhood? A model In this section, we present a simple model that illustrates the problems that can arise when researchers use overly broad or narrow neighborhood boundaries when estimating the effect of a concave neighborhood characteristic (e.g. heterogeneity) on an outcome of interest (e.g. volunteering). Consider a society with heterogeneity defined in terms of ethnicity, and assume for simplicity that there are only ethnicities 1 and 2. Assume next that the society can be divided into a number of small neighborhoods, each of equal size. We will start by assuming that people s likelihood of volunteering is affected only by the heterogeneity they perceive in their immediate small neighborhood, and relax this assumption later. Each small neighborhood i can then be defined as one of n constituent parts of a large neighborhood j. Following the literature already cited, we shall assume that ethnic heterogeneity can be correctly captured using a fragmentation index, though our argument holds for any concave measure. 4 Ethnic fragmentation x ij can be constructed for each small neighborhood i within large neighborhood j, and expressed as the product of the two ethnicities shares: x [1 ] [1 (1 ) ] 2 (1 ). (1) ij 1ij 2ij 1ij 1ij ij ij 4 Other concave measures used in studies of neighborhood effects include the Gini coefficient, Theil entropy index, Atkinson deprivation index, coefficient of variation, or neighborhood entropy index (see Hansmann and Quigley 1982). We have repeated the empirical analysis reported in this paper using entropy in particular, and found very similar results to those that follow. 6

9 Here ij 1 ij is ethnicity 1 s share of the population in small neighborhood i of large neighborhood j. With just two ethnicities, the fragmentation index reaches its maximum value at ij 0.5. In the same way, we can also construct a fragmentation index x j for the large neighborhood. This will be the product of the two ethnicity s shares in the large neighborhood, but can be equivalently expressed as each ethnicity s average share over the n constituent small neighborhoods. x [1 ] [1 (1 ) ] 2 (1 ), (2) j 1 j 2 j 1 j 1 j j j 1 ij i 1 where j 1 j n n is ethnicity 1 s share in large neighborhood j. From the (strict) concavity of x ij and x j in (1) and (2), it follows from Jensen s inequality that the fragmentation of the large neighborhood will be (strictly) greater than the mean fragmentation of the n constituent small neighborhoods, or E( x ) E( f ( )) f ( E( )) f ( ) x. (3) ij ij ij j j That is, as long as fragmentation varies across small neighborhoods, heterogeneity will appear greater, the larger the neighborhood over which it is defined. Intuitively, total fragmentation in a large neighborhood comes from heterogeneity within each of its small neighborhoods, but also from differences in heterogeneity between them. More formally, we define between heterogeneity x B j as the residual difference between the fragmentation index of large neighborhood j, and the average fragmentation index of its i = 1,,n constituent small neighborhoods, the latter equivalent to within heterogeneity x W j : xij B i W x x E( x ) x x x. (4) j j ij j n j j This mathematical discrepancy may have an empirical consequence: a study that uses large neighborhood boundaries may bias the relationship it finds between heterogeneity (or indeed, any concave neighborhood characteristic) and people s behavior. To see this, consider a benchmark regression model that correctly recognizes that changes in small neighborhood heterogeneity x ij can affect small neighborhood volunteering rates yij directly, or indirectly via the effects on heterogeneity between small neighborhoods B x j : y x x u. (5) B ij W ij B j ij 7

10 Here W and B are the population regression coefficients of within and between neighborhood heterogeneity, respectively, and uij is a pure random error. Compare to this an empirical study that uses only small neighborhood boundaries, and regresses y ij on xij alone, yielding a population regression coefficient S. From comparison with (5), S may be thought of as a biased estimate of the within effect of heterogeneity, W because of the omission of x. Using a standard result from omitted variable bias, B j B cov( xij, xj ) S W B. (6) var( x ) ij If heterogeneity between small neighborhoods has no effect on volunteering within small neighborhoods, or B 0, the total effect of heterogeneity will come from the within effect, and be captured by without bias. More generally, if between heterogeneity does affect S volunteering, the total (within and between) effect will still be captured by without bias. To see this, note that a change in y ij caused by a unit change in heterogeneity x ij is the sum of the (direct) within effect, x, and the (indirect) between effect, ( x B x ) x. The W ij S B j ij ij B B total effect on y ij is thus ( W B( x j xij )) xij or ( W B ( cov( xij, x j ) var( xij ))) xij. From (6), the latter can be re-expressed as x. S ij In contrast, empirical studies that use only large neighborhood boundaries, and regress y j on x j, will have problems. Returning to our benchmark, a simple aggregation over the correct small neighborhood specification in (5) yields the correct large neighborhood specification: yij i W B y ( ) x x u. (7) j n W j B j j By adding and subtracting a common term, (7) can be re-expressed as: W B B y ( x x ) ( ) x u j W j j B W j j B ( x ) ( ) x u W j B W j j (8) From comparison with (8), we can see that a study using only large boundaries is missing a term involving x B j. The resulting estimated effect of large neighborhood heterogeneity x j on y,, can be expressed as j L B cov( xj, xj ) L W ( B W ). (9) var( x ) j 8

11 In the special case that between heterogeneity has no economic effect on volunteering B ( B 0 ), L will not capture the true effect of within heterogeneity. Instead, if cov( xj, x j ) < B 0 then from (9) it follows that L W, or if cov( xj, x j ) > 0 then L W. The concavity of the heterogeneity measure introduces a sort of measurement error into large boundary regressions. More generally, when B 0, large boundary regressions will continue to provide biased estimates of the within- and total effects of heterogeneity. 5 Problems of a different sort arise if boundaries are set too narrowly. To see this, we relax the assumption that people s likelihood of volunteering is affected only by the heterogeneity they perceive in their immediate small neighborhood. 6 than assuming that yij can be affected only directly by xij and indirectly by More formally, rather B x j, we allow that it can also be affected by xij, i i. In this case the total effect on y ij would be the B combined one of ( ( x x )) x plus the effect of x, i i. As a result, small W B j ij ij boundary regressions will themselves yield downward biased estimates B ( ( W B( x j xij )) xij = S x ij ) of heterogeneity s total effect. Unfortunately, bad news for overly-small boundary regressions does not translate into good news for larger boundary regressions. L will become favorable in the sense of better capturing this additional effect of within heterogeneity from surrounding small neighborhoods xij, i i on y ij, but it will still be biased because of the omission of a term involving ij B x j as in equations (8) and (9). For example, in the special case that between heterogeneity has no economic effect on volunteering ( B 0), small neighborhoods on y ij. L will remain biased in capturing the within effect of xij for all i 5 In the special case that volunteering is identically affected by within and between heterogeneity, or W B, equations (5) and (7) produce S L W and W B. It is only in this case that neighborhood size has no effect on the estimated effect of a concave characteristic. 6 The possibility of heterogeneity outside a small boundary affecting behavior within it also needs to be addressed when we consider endogenous neighborhood choice. We attempt to control for endogenous selection to small or large neighborhood boundaries in Section

12 Our argument might suggest that researchers should estimate concave neighborhood effects using the smallest possible neighborhood size consistent with fully capturing posited effects on outcomes of interest. But researchers are often constrained as to available neighborhood size, or else unsure ex ante as to which boundary is just large enough to capture all posited influences. In an additional attempt to capture the within effect of heterogeneity on volunteering, as well as the between- and possible adjacent neighborhood within effects, we will use two boundary levels simultaneously. We will regress small neighborhood volunteering on small and associated large neighborhood heterogeneity measures. With the within effect of small neighborhood heterogeneity on volunteering controlled for, the remaining effect of large neighborhood heterogeneity will be equivalent to the sum of between effects and additional within effects of x, i i. This hybrid regression allows us to capture the within and between effects of heterogeneity on volunteering without the measurement error introduced by the concavity of our fragmentation index. We will also present the results of purely small neighborhood specifications to provide a lower bound on heterogeneity s total effect, and the results of purely large specifications to see if the bias we identify in theory makes much difference in practice. By further adding fixed effect estimation of our specification, we will also address in part the issue of endogenous location choice in small (or large) neighborhoods. ij 4. Empirical Analysis 4.1 The case of New Zealand Common to other Western nations, New Zealand has experienced a marked increase in social diversity over the past 25 years. Starting as a British colony in the mid-nineteenth century, New Zealand s population was predominantly of British ancestry, with a significant indigenous Maori population (Phillips, 2008). Immigration from other European and Commonwealth countries increased from the time of the second World War, and from neighboring Pacific Island and South East Asian nations. Changes to the Immigration Act of 1987, and the introduction of an ethnicity-blind points system in 1991 was followed by a substantial further diversification of migrants from China, India, and North African and Middle Eastern countries (Phillips, 2008). For more detail about social diversity and volunteering in New Zealand, we turn to the data. 10

13 4.1. Data and descriptive statistics Our data comes from the New Zealand census rounds of 1996, 2001 and The New Zealand census collects data on an exhaustive list of individual and household characteristics including volunteering activities, ethnicity/race, languages spoken, birthplace and household income. These data are released by Statistics New Zealand at various levels of neighborhood aggregation, right down to the unusually small size of meshblock ( 100 people) and area unit ( 2000 people). Constant 2006-defined neighborhood geographic boundaries are used for all three rounds to ensure consistency. Our sample is restricted to all those neighborhoods without missing or censored explanatory variables. 7 Over the three years of the census, our pooled sample is 3,504 area units and 49,600 meshblocks in New Zealand. A description of the dependent and explanatory variables used is provided in Appendix 1, and corresponding descriptive statistics are provided in Appendix 2. To provide the reader with a description of contemporary New Zealand, Figure 1 illustrates the volunteering rate and ethnic, language, birthplace and nominal household income shares for 1996, 2001 and These shares are population-weighted mean values, based on those meshblocks providing complete observations for our analysis, or our common sample. 8 Figure 1 suggests that New Zealand is increasing in diversity along each dimension. Moving from shares to fragmentation measures (the equivalent of one minus the Herfindahl Index of concentration), Table 1 provides key descriptive statistics. The population-weighted average proportion of New Zealanders aged 15 or over who reported volunteering at least once 7 In general, we constructed share variables for each neighborhood so as to ensure they were weakly positive and summed to one. In the case of gender, for example, we constructed ShareFemale by dividing the frequency of Number Female by ( Number Female + Number Male ). This assumes that non respondents had the same gender composition as respondents. See Appendix II for details of each variable s construction. 8 Corresponding descriptive statistics using all meshblocks providing observations for a given variable (our maximum sample ), are provided in Appendix 3. 11

14 outside the household in the previous four weeks was 18.3% in 1996, and then, using a slightly different definition, 15.6% in 2001 and 14.7% in During this same period, {Figure 1 about here.} {Table 1 about here.} fragmentation by ethnicity/race, languages usually spoken, birthplace, and nominal household income increased. Regarding ethnicity/race, a fragmentation index could conceivably range between 0 and.8 for five categories. The population-weighted mean fragmentation across all meshblocks rose from.347 in 1996, to.352 in 2001, to.378 in A similar index for language fragmentation, which could range from 0 to.75 over four categories, rose from in 1996 to.254 in 2001 to.275 in The index for birthplace fragmentation, which could range from 0 to.5 over two categories (inside or outside of New Zealand), rose from.293 in 1996 to.300 in 2001 to.329 in Finally, the index for nominal household income fragmentation, which could range from 0 to.833 over six unadjusted nominal income bands, rose from.746, to.757, to Another way to illustrate potential correlation between volunteering and fragmentation is to graph 95% confidence intervals of the best fit polynomial relationship between each 9 For 1996 volunteering was defined as having Attended Committee Meeting etc Unpaid for Group, Church or Marae. For 2001 and 2006 the definition was changed to be defined as any Other Helping or Voluntary Work For or Through any Organisation, Group or Marae. For all three years our definition excludes those caring for a child or someone who was ill, elderly, or disabled outside the household. See Appendix Table 1I for more detail. Because of the change in volunteering question, we have repeated all the analysis to follow using only 2001 and 2006 data. The results concerning heterogeneity s effects are very similar to what we report here, with the exception that the evidence for income heterogeneity s (negative) effect is slightly greater, becoming significant even in fixed effects analysis. 10 Because the six household income bands were not adjusted for inflation between each census, we can only measure how the dispersion of unadjusted nominal incomes across bands has changed over time, not real household incomes. 12

15 meshblock s volunteering rate and each dimension of heterogeneity. This is done in Figure 2 using the pooled sample. A clear negative relationship appears between volunteering and ethnic, language and birthplace fragmentation, while a more diffuse inverted U relationship appears for nominal income fragmentation. The non-monotonic relationship between income fragmentation and volunteering might reflect that there exists a degree of inequality beyond which volunteering is depressed, or simply that a correlation does not exist. {Figure 2 about here.} Thus, consistent with the findings in the literature, the fall in volunteering rates in New Zealand coincided with increasing heterogeneity by several dimensions. Nonetheless, many other changes were taking place in New Zealand over these years which could have influenced people s decision to volunteer (via their tastes or opportunity costs), or organizations decisions to demand volunteers (via volunteers non-wage costs and productivity (Handy and Srinivasan (2005))). We construct measures for many of these confounding factors, which are described in Appendices 1 and 2, including real median household income, and the ethnicity, language, and birthplace shares that underlie our fragmentation measures. Among these variables, the average real median household income across meshblocks rose from NZ$ 37,800 in 1996, to $39,000 in 2001, to $45,000 in The mean share of females remained steady at 51%, while the mean percentage whose highest education was a bachelor s or honour s degree rose from 8% to 10% to 12%. At the same time, the mean percentage of those aged 15 or over not in the labor force fell from 34% to 33% to 31%. The mean percentage claiming Christian religious affiliation also fell from 67% to 62% to 56%, while the mean percentage claiming no religious affiliation rose from 28% to 31% to 36%. We will try to untangle the effects of these various changes on volunteering rates in the regression analysis that follows. Finally, returning to the issue of neighborhood boundary, Table 2 compares the mean and standard deviation of meshblock and area unit measures of fragmentation. 11 As predicted in Section 3, the means of all four types of heterogeneity appear greater over area units than over meshblocks. In addition, the standard deviation of neighborhood heterogeneity is consistently lower at the area unit level than at the meshblock level for every measure. This 11 The descriptive statistics in Tables 1 and 2 use the common sample used for subsequent regression analysis. 13

16 suggests that our choice of neighborhood size may indeed affect the empirical relationship we estimate between social heterogeneity and volunteering Estimation strategy and results In this section, we lay out and implement our strategy for estimating the cross-sectional and longitudinal empirical relationship between social heterogeneity and volunteering. Because we have a wide, shallow panel of many neighborhoods over just three census years, our pooled data contains substantial variation between neighborhoods at any point in time, {Table 2 about here.} but less variation within neighborhoods over time. In line with the vast majority of studies in our literature review, we shall begin by using pooled cross section OLS as our baseline specification. To address the problem of omitted variable bias that attends cross section analysis, and to test the robustness of our results, we shall then add two steps. First, we will repeat the baseline cross sectional analysis using various additional groups of control variables. Second, we will switch to fixed effects analysis to address omitted variable bias resulting from location choice under a Tiebout-style hypothesis. Throughout this process, we shall provide (small) meshblock regressions that provide either an unbiased or lower-bound estimate of heterogeneity s total effect on volunteering. We shall also provide larger area unit regressions for comparison, and finally hybrid regressions. Hybrid regression coefficients may be thought of as decomposing the total effects of heterogeneity into within and between effects if volunteering is affected only by heterogeneity in x ij. If volunteering is also affected by heterogeneity in x, i i, hybrid regressions may be thought of as providing ij an upper bound of heterogeneity s total effect, (assuming area units contain all relevant heterogeneity), given by the sum of within and between coefficients. In such cases, the lower bound of total effects is given by the single heterogeneity coefficient in the meshblock regressions. Beginning with our baseline cross sectional analysis, we run regressions of the form y X u (10) ijt ijt ijt where y ijt is meshblock i s volunteering rate in area unit j in year t. X ijt is a vector of neighborhood characteristics, year dummies, and social heterogeneity measures, while u ijt is 14

17 a random error. In each case, we regress volunteering rates on one type of heterogeneity at a time, along with its underlying level variables (e.g. ethnic shares for ethnic fragmentation, language shares for language fragmentation, median household income for income fragmentation, etc.). To these we add the baseline covariates of share female, median age, population density, mean household size, share married, shares of families comprised of couples with children, and couples without, and year dummies. 12 Table 3 provides the results. Column (1) shows our baseline estimate of meshblock ethnic heterogeneity s effect on meshblock volunteering rates with controls for (meshblock) ethnic affiliation shares and baseline covariates. The estimated coefficient on ethnic fragmentation (-.128) implies a relatively strong negative effect of this type of heterogeneity on volunteering. In particular, a 10 percentage point increase in meshblock ethnic fragmentation is estimated to decrease the (meshblock) volunteering rate by 1.3 percentage {Table 3 about here.} points. This effect is only moderately overstated when area units are used in place of meshblocks in column (2), where a 10 percent increase in area unit ethnic/racial fragmentation decreases the (area unit) volunteering rate by 1.4 percentage points. Finally, from our hybrid specification in column (3), a 10 percentage point increase in heterogeneity within meshblocks decreases (meshblock) volunteering by 1.1 percentage points. At the same time, a 10 percent increase in corresponding area unit heterogeneity decreases meshblock volunteering by.4 percentage points. This indicates the existence of between effects and/or the effect of heterogeneity in other meshblocks x, i i. 13 ij 12 We have also added ethnic share composition to the baseline covariates when examining the effects of birthplace or income fragmentation on volunteering. This is because of the clear effect that Maori ethnic affiliation has on volunteering rates. Ethnic shares remain omitted when examining the effects of language fragmentation, because ethnic and language shares are highly correlated. 13 The 1.05 ( 1.1) within, and.40 between effects of heterogeneity can be related to the 1.28 total effect as follows. If we assume heterogeneity in other meshblocks has no effect, then on average a ten percentage point increase in ethnic fragmentation within meshblocks raises ethnic fragmentation between meshblocks in the associated area unit by 5.75 percentage 15

18 The estimated effects of language, birthplace and household income heterogeneity are similarly provided in columns (4) (12) of Table 3. Language and birthplace heterogeneity are again strongly negatively associated with volunteering rates under the baseline specification. From column (4), language fragmentation s (lower bound of) total effect on volunteering is a 2.8 percentage point drop, though it appears as a 3.7 percentage point drop when larger area units are used. Similarly, from column (7), birthplace fragmentation s (lower bound of) total effect on volunteering is a 1.1 percentage point drop, which appears as a.3 percentage point drop using area unit neighborhoods. Finally, nominal household income band fragmentation s (lower bound) total effect on volunteering in column (10) is a slight.2 percentage point drop, which appears as a 1.1 percentage point drop using area units. Thus, our baseline cross section results might suggest that New Zealand s shifting immigration and tax policy has been responsible for a drop in New Zealander s tendency to contribute time towards public goods. These initial results also suggest that using broader neighborhood boundaries can either inflate or attenuate heterogeneity s estimated effects by as much as five times. While our meshblock estimates of the effects of each type of heterogeneity on volunteering are almost uniformly negative, this could simply reflect the omission of other influences on volunteering that are correlated with heterogeneity. Omitted factors could include variation in religious affiliation, neighborhood deprivation, labor force status, or education. We have also yet to test whether one type of heterogeneity affects volunteering when other dimensions of heterogeneity (and their underlying share variables) are controlled for. Thus, in Table 4 we extend our cross sectional analysis to include groups of other confounding variables one at a time. 14 These groups are: 1) religious affiliation: Christian, other, and no religion affiliation rates, 2) neighborhood deprivation: home ownership rates, median number of bedrooms, crime rates, and percentage of individuals receiving single parent domestic benefits, 3) employment status: shares in full time work, part time work, points, or B x x = cov( x, x ) var( x ) =.575. The total effect is then (.575) B j ij ij j ij If we assume heterogeneity in other meshblocks is relevant, 1.28 becomes the lower bound of total effects, including the effect of x, i i. 14 High degrees of correlation between various covariates precluded us from including all clusters simultaneously. ij 16

19 unemployed, and not in the labor force, 4) education levels: the share of individuals lacking minimum high school qualifications and the share with bachelor s or (additional year) honour s degrees, and 5) including all heterogeneity measures simultaneously, together with their underlying share variables. Note that care must be taken in evaluating the effects of each type of heterogeneity when all are included simultaneously in 5), because they (and their underlying share variables) may be highly correlated. While many of our baseline and added covariates explain variation in volunteering rates, 15 we focus in Table 4 on showing the direct (remaining) effect of each type of heterogeneity on volunteering. Column (1) of Table 4 shows the (lower bound of) total effect of meshblock ethnic heterogeneity on meshblock volunteering as each group of confounding variables is added to the baseline covariates of Table 3. In each case, ethnic/racial heterogeneity retains a significant, negative effect on volunteering of roughly 1 percentage point. Language heterogeneity similarly retains a robust and relatively strong negative effect as other covariates are included. From column (4), a 10 percentage point increase in {Table 4 about here.} meshblock language fragmentation results in a roughly 2 to 2.5 percentage point drop in the volunteering rate. Birthplace heterogeneity (column 7) similarly retains a negative effect as covariates are added, similar in magnitude to ethnic heterogeneity (1 percentage point), with the exception of the case where all types of heterogeneity and underlying share variables are entered simultaneously. There a 10 percentage point increase in meshblock birthplace fragmentation raises volunteering by.2 percentage points. Finally, nominal household income band 15 Those covariates consistently positively related to volunteering rates were share with Maori ethnic affiliation, Maori and Samoan language shares, median age, share married, share of families that had couple with kids, and less so families that had a couple without kids, share with Christian or other religious affiliation, share who owned own home, median number of bedrooms, share with bachelors or honours degrees, and share employed part time. Those covariates consistently negatively related to volunteering rates were share with Asian or MELAA ethnic affiliation, English or other language share, household size, population density, share with no religious affiliation, share of families that were single parent, and share employed full time. 17

20 heterogeneity is similar to birthplace heterogeneity in retaining a negative effect on volunteering as covariates are added, except when all types of heterogeneity are included. However, the magnitude of the effect of income fragmentation on volunteering is smaller than for other types of heterogeneity. It ranges between lowering volunteering by.4 percentage points, to raising it by.1 percentage point when all heterogeneity is considered simultaneously. While it is interesting that birthplace and income heterogeneity no longer lower volunteering rates when all types of heterogeneity are controlled for, the correlation that exists between various types of fragmentation mean that these findings must be treated with caution. Taken together, the baseline and extended cross sectional evidence so far points strongly to a negative effect of ethnic/racial and language heterogeneity on volunteering, and possibly to a negative effect of birthplace and income heterogeneity as well. Nevertheless, as with any cross section analysis, there may remain unobserved characteristics that are correlated with heterogeneity that are skewing its estimated effects. Even with the inclusion of extensive covariates, omitted variable bias remains a strong possibility because of the Tiebout hypothesis that people will self-select to live in an area with others who share their preferences regarding the optimal trade-off between private consumption and public goods provision (Tiebout 1956). Specifically, if people who are less inclined to volunteer are attracted to live in urban centres, which tend to be more heterogeneous, then the cross sectional effects of heterogeneity on volunteering will be exaggerated. We therefore move to fixed effects analysis in an attempt to better control for unobserved characteristics like people s attitude towards volunteering or towards heterogeneity. One problem with using fixed effects analysis here is that we are following neighborhoods rather than individuals over time, and the latter are free to change where they live. Is there any reason to expect that unobserved characteristics like attitude to volunteering would remain constant over time for given neighborhoods, even as the individuals in them come and go? Perhaps ironically, our defence of this proposition comes from the same Tiebout hypothesis that raises the problem of endogenous neighborhood choice in the first place. For with freedom of movement, individuals who come to differ with the local prevailing preferences between private and public good provision may leave, and those who share those preferences may enter. If the effects of heterogeneity found with fixed effects do not correspond with those found in cross section, this might suggest that the cross section effects are spuriously caused by omitted variable bias. 18

21 To proceed, we estimate the following volunteering equation using panel data on the meshblocks of New Zealand: y X. (11) ijt ijt ij ijt yijt is the volunteering rate in meshblock i within area unit j in year t, while X ijt contains our set of heterogeneity measures and other control variables previously defined. The ij are unobservable meshblock-specific fixed effects which may be correlated with Xijt, while ijt is a pure random error term. To control for the potential correlation between and X, we apply OLS to the mean-differenced equation y y ( X X ). (12) ijt ij ijt ij ijt ij Here, for any variable Z, Z Z /3. 16 ij t ijt Table 5 presents the results. The control variables included are identical to those used in baseline cross section (Table 3), and the effects are again presented using meshblock, area unit, and hybrid specifications. In general, evidence of an effect of heterogeneity on volunteering has weakened. As shown in column (1), ethnic fragmentation retains its negative effect on volunteering, but the magnitude of the (lower bound of) total effect has fallen from 1.1 percentage points in cross section to.5 percentage points under fixed effects. This suggests that neighborhoods that experience an increase in ethnic heterogeneity also experience a decrease in volunteering rates on average, but the effect is modest after controlling for people s choice of neighborhood location in the manner of the Tiebout hypothesis. In contrast, language fragmentation has lost any significant effect (column (4)). Birthplace fragmentation in {Table 5 about here.} column (7) retains a negative but reduced effect, with a (lower bound of) total effect falling from 1.1 to.3 percentage points. Finally, nominal household income fragmentation in ij ijt 16 We consider also the more general case that endogenous location choice may occur even for large neighborhoods by adding an j term in equation (11), or an unobservable area unitspecific fixed effect which may be correlated with specification will then control for both ij and j. 19 X jt. The fixed effects of our hybrid

22 column (10) is similar to language fragmentation in losing a significant effect. 17 findings are qualitatively summarized in Table 6. All of these Methodologically, it is interesting to note that the choice of a larger neighborhood boundary would have substantially affected our (baseline) fixed effects results, just as it did our (baseline) cross section analysis. If we use area units rather than meshblocks, ethnic fragmentation would inflate to a -1.1 rather than -.5 percentage point effect. Language fragmentation would inflate to a -2.1 percentage point effect from no effect. Birthplace fragmentation would attenuate to have no effect, rather than a -.3 percentage point effect. Only household income fragmentation would have a similar result no effect under either neighborhood boundary. 5. Discussion and conclusions This paper has attempted to make two contributions to the growing empirical literature in which researchers estimate the effects of concave neighborhood characteristics like heterogeneity on outcomes of interest, such as people s likelihood of volunteering. With access to unusually fine geographical units, and a panel of those units, we have addressed the problems posed by 1) the researcher s choice of neighborhood boundary, and 2) the researchees endogenous choice of neighborhood. Regarding boundaries, we show that for concave neighborhood characteristics like heterogeneity, using the smallest possible boundary that includes all posited effects will provide an unbiased estimate of total effect. Using overly-large boundaries will incorrectly capture the effect of differences in heterogeneity between constituent small neighborhoods on the outcome of {Table 6 about here.} interest. This can bias estimates (up or down) of heterogeneity s total effect even when such between heterogeneity has zero economic effect. Using overly-small boundaries will avoid this bias, but provide only a lower-bound estimate of heterogeneity s total effect by ignoring 17 This is the single case where restriction of the sample to the two years with an identical volunteering definition, 2001 and 2006, resulted in a different finding. Here in fixed effects analysis, income fragmentation retains a negative effect of.2 percentage points, significant at the one percent level. 20