Does Social Capital Reduce Crime?

Transcription

1 Does Social Capital Reduce Crime? Paolo Buonanno University of Bergamo Daniel Montolio University of Barcelona Paolo Vanin University of Padua Abstract We investigate the effects of civic norms and associational networks on crime rates. Civic norms may attach guilt and shame to criminal behavior, thus increasing its opportunity cost. Associational networks may increase returns to noncriminal activities and raise detection probabilities, but they may also work as communication channels for criminals and may offer official cover to criminal activities. The empirical assessment of these effects poses serious problems of endogeneity, omitted variables, measurement error, and spatial correlation. Italy s great variance in social and economic characteristics, its homogeneity in policies and institutions, and the availability of historical data on social capital in its regions allow us to minimize the first two problems. To tackle the last two problems, we use report-rate-adjusted crime rates and estimate a spatial lag model. We find that both civic norms and associational networks have a negative and significant effect on property crimes across Italian provinces. 1. Introduction Crime and the fear it generates are among the most important determinants of individual welfare and of the expected returns to many economic activities. It is both intuitive and confirmed by recent theoretical literature that individual choices of crime participation may be significantly affected by the presence of civic norms and associational networks. Civic norms may attach guilt and shame to criminal behavior, thus increasing its opportunity cost. Associational networks may increase returns to noncriminal activities and raise detection probabilities, but they may also work as communication channels for criminals and may offer official cover to criminal activities. Thus, it is not a priori clear whether the We are grateful to Dennis W. Carlton, Sam Peltzman, an anonymous referee, Antonio Cabrales, Antoni Calvó-Armengol, Antonio Ciccone, Federico Cingano, Giacomo Pasini, seminar participants at Università degli Studi di Bergamo, Università degli Studi di Padova, and Universidad Torcuato Di Tella (Buenos Aires) and conference participants at Università degli Studi di Bologna, Duke University, Università di Milano Bicocca (Verbania), and Università Ca Foscari di Venezia (Treviso) for very useful comments. We thank Robert Putnam for providing his data on social capital in Italy. We gratefully acknowledge financial support from the University of Padua (research grant CPDA071899). The usual disclaimers apply. [Journal of Law and Economics, vol. 52 (February 2009)] 2009 by The University of Chicago. All rights reserved /2009/ $

2 146 The Journal of LAW& ECONOMICS empirical correlation between crime rates and civic norms and associational networks should be expected to be positive or negative and whether it reflects a causal link, in what direction, and with what implications for anticrime policy. Yet such questions have received little attention, particularly by economists. This is especially surprising in light of the recent economic literature on the microeffects on crime of some forms of social interaction, such as peer effects, neighborhood effects, family background effects, and so on, and, more important, in light of the increasing relevance in the economic debate on the concept of social capital, often conceived in terms of social norms and networks that favor coordination and cooperation. Starting from the seminal work by Putnam (1993) on the role of social capital for well-functioning government, several contributions in this literature have focused on Italian data. 1 This is due to the fact that Italy displays large and persistent provincial disparities in social and economic characteristics in spite of having common policies, institutions, laws, justice system, and school system; having police forces organized at national level; and being ethnically and religiously quite homogeneous. Thus, changes in these factors are not responsible for socioeconomic differences across Italian provinces, and this in turn substantially reduces the omitted-variable problems affecting many cross-country studies. 2 In this paper we exploit provincial-level variations in civic norms and associational networks in Italy to investigate their effects on crime rates. We focus on property crimes because they are more likely than violent crimes to depend on economic motivations. Since social characteristics in Italy are often peculiar traits of single cities, provincial data are probably best suited to capture social capital in this country. We are not aware of any previous studies on the effect of social capital on crime in Italian provinces. Several studies, from Knack and Keefer (1997) to Bjørnskov (2006), find empirically that social capital is best described as a collection of three main dimensions, namely, generalized trust, civic norms, and associational networks, and that these dimensions have different effects on economic outcomes. Since our aim is to study the latter two dimensions, we separately consider provinciallevel measures of cultural and recreational associations, voluntary associations, voter turnout at referenda, and blood donations. To account for criminal networks, we also include a measure of criminal association. In our estimates we control for other major socioeconomic determinants of crime rate, such as income, unemployment rate, education, urbanization rate, and share of youth. 1 Among others, two recent examples are the investigations of Guiso, Sapienza, and Zingales (2004) and Peri (2004) of the role of social capital on financial and economic development. Together with Putnam, Coleman (1988, 1990) may be credited with the main responsibility for the academic diffusion of the concept of social capital. 2 For instance, in Lederman, Loayza, and Menéndez s (2002) cross-country investigation of the relationship between social capital and crime, it is difficult to disentangle what is due to differences in social capital from what is due to different institutional settings.

3 Social Capital and Crime 147 Moreover, to account for provincial differences in the efficiency of the national justice system and in the effectiveness of police forces, we also control for the length of judicial proceedings and for crime-specific clear-up rates, both of which exhibit great variability across provinces. Finally, to further control for unobserved heterogeneity, we include different sets of geographical dummies. Empirical work on crime and on social capital is typically affected by several methodological problems, the main of which are, besides omitted variables, measurement errors, endogeneity, and spatial correlation. Measured crime rates crucially depend on report rates, which not only vary significantly across crimes and space but, most important, appear to be positively related to social capital. Regressing measured crime rates on social capital, we would thus obtain upwardbiased estimates. Endogeneity problems arise from the fact that certain forms of crime might affect social capital because they constrain social interaction, as shown, among others, by Liska and Warner (1991). 3 Such a negative effect, due to reverse causation, would bias downward our estimates. 4 Spatial correlation may be due to criminals interprovince communication or mobility. If ignored, spatial correlation works as an omitted variable, rendering estimates biased and inconsistent. As far as endogeneity is concerned, we consider blood donations and referenda turnout as safely exogenous variables with respect to crime rates. Indeed, blood donations seem to be as exogenous a variable as possible, and the referenda we consider have never concerned issues either related to crime or of direct interest to criminal groups. To control for the possible endogeneity of association density, we exploit the fact that associational networks in Italy are to a significant extent a historical heritage. We then use Putnam s (1993) historical data on associations in Italy as an instrument for current associations, arguing that it is unlikely that such an instrument is correlated with current crime rates through other channels. Measurement errors represent a relevant problem because underreporting, that is, the fraction of crimes that are not reported to the police, is likely to be determined not only by random errors but also by specific and persistent characteristics of each province, among which is its level of social capital. High and homogeneous report rates for car thefts offer a first solution to this methodological issue, but we pursue a more general strategy. We correct measured crime rates by crime-specific report rates (available from victimization surveys at the regional level), and we thus construct true crime rates, which we then regress on their determinants. Spatial correlation is a crucial issue when evaluating social determinants of crime since, as shown by the recent literature reviewed below, various forms of localized social interaction may give rise to a spatial multiplier (see Anselin et 3 Liska and Warner analyze 26 large U.S. cities in the mid-1970s and show that some forms of crime (for example, robberies) generate fear and constrain social interaction, which, in turn, reduces opportunities for other forms of crime. 4 An upward bias would in turn arise if higher crime rates stimulate associational activity as a community reaction.

4 148 The Journal of LAW& ECONOMICS al. 2000) by which crime rates in a given province also depend on crime rates in neighboring provinces. To address this issue, we control for spatially lagged crime rates, which consist of weighted averages of crime rates in neighboring provinces. Our evidence, based on an original data set that merges existing sources with data collected by the authors, indicates that both civic norms and associational networks have a negative and significant effect on property crimes across Italian provinces. In our baseline specification, a 1-standard-deviation increase in blood donations is associated with significant reductions in common thefts, robberies, and car thefts by approximately 1/4, 1/7, and 1/6 of a standard deviation, respectively (such reductions amount to around 15 percent). An increase of 1 standard deviation in association density is associated with significant reductions in robberies and car thefts by approximately 1/7 and 1/4 of a standard deviation, respectively (amounting to around percent); association density is not significantly related to common thefts. Yet, when we control for endogeneity and use historical instruments, we find that association density is negatively and significantly related to all crime rates, with a coefficient that is higher, in absolute value, than in the baseline regressions. Finally, referenda turnout is never significant, which confirms that one should carefully distinguish between different aspects of social capital, which indeed have different effects. The remainder of the paper is organized as follows. In Section 2 we briefly review the main literature on the effects of social environment on crime. In Section 3 we present our data and empirical strategy. Section 4 contains the results, and Section 5 concludes. 2. Literature on Social Determinants of Crime 2.1. Economic Models Based on Multiple Equilibria Following Becker (1968), most economists consider individual decisions of crime participation as rational choices taken by comparison of the expected costs and benefits. At the core of several theoretical models, which extend the basic paradigm to include the social component of such costs and benefits, lies some form of strategic complementarity: the returns to becoming criminal, relative to not doing so, are higher as the number of other individuals who choose criminal behaviors increases. This gives rise to multiple equilibria, which help explain the variation of crime rates across regions with similar fundamentals. Several specific mechanisms that yield such results are explored in the literature, for instance, by Sah (1991), Rasmusen (1996), Silverman (2004), and Calvó-Armengol, Verdier, and Zenou (2007). In particular, Calvó-Armengol and Zenou (2004) show that denser social networks may increase aggregate crime levels by facilitating information sharing among criminals. Weibull and Villa (2005) argue that the effectiveness of social norms against crime, due to both guilt and shame, decreases

5 Social Capital and Crime 149 as crime rates increase, thus being yet another possible source of multiple equilibria Empirical Evidence on Social Determinants of Crime Since the early 1990s, economists have collected more and more empirical evidence on the social determinants of criminal behavior. The crime effects of neighborhood, family, peers, and networks have been widely studied by, among others, Case and Katz (1991), Glaeser, Sacerdote, and Scheinkman (1996), Ludwig, Duncan, and Hirschfield (2001), Kling, Ludwig, and Katz (2005), and Calvó- Armengol, Patacchini, and Zenou (2005). 5 While this literature is more concerned with the microeffects of local interaction patterns on crime rates than with the effect of more aggregate aspects of the social structure, such as widespread civic norms and associational networks, as we are in the present study, it confirms three aspects that are important for us: the general relevance of social determinants of crime, the endogeneity and spatial correlation problems that arise when one tries to evaluate such determinants, and the possibility that apparently desirable aspects of the social structure turn out to be conducive to crime Effects of Social Capital on Crime Several theories, developed by sociologists and criminologists, imply a negative effect of social capital on crime. 6 Yet the above-mentioned economic literature suggests a more complex and multifaceted relationship. Very few empirical studies investigate directly the link between social capital and crime. Among them, most attention has been devoted to violent rather than property crime, as in Rosenfeld, Messner, and Baumer (2001), Lederman, Loayza, and Menéndez (2002), and Messner, Baumer, and Rosenfeld (2004). Chamlin and Cochran (1997) and Heaton (2006) consider the effects on property and violent crime of social altruism, proxied by charitable donations, and religious participation, respectively. These contributions convey two main messages: at the methodological level, they confirm both the opportunity to disentangle the different dimensions of social capital and the relevance of reverse-causation problems; at the substantive level, they offer mixed evidence on the effect of different forms of social capital on crime and make clear that more and careful empirical analysis is needed. 5 Different and convincing evidence of the relevance of social and nonpecuniary factors for crime decisions is also provided by Levitt and Venkatesh (2000). Rubio (1997) points out that in Colombia, where incentives for crime are high, social capital may have the perverse effects of reinforcing crime choices. 6 For instance, Rosenfeld, Messner, and Baumer (2001) argue that theories of social disorganization, anomie, and strain all predict that civic engagement and social trust (which they refer to as social capital) should reduce crime because they increase formal and informal social control, strengthen the effectiveness of social norms, and provide resources for individual goal attainment.

6 150 The Journal of LAW& ECONOMICS 2.4. On the Effects of Social Capital in Italy There is a wide literature on the effects of social capital in Italy. Here we limit our attention to two studies: Putnam s (1993) seminal book and Guiso, Sapienza, and Zingales (2004). In a nutshell, Putnam shows that local governments are more efficient where civic engagement is stronger. He empirically relates civic engagement to measures of horizontal association networks, voter turnout at referenda, newspaper readership, and the diffusion of preference vote at political elections. 7 Moreover, he relates the geographic distribution of social capital to historical origins and uses it to explain the wide and persistent differences between southern and northern Italian regions. 8 In an empirical investigation of the effects of social capital on financial development in Italy, Guiso, Sapienza, and Zingales (2004) focus on two proxies of social capital, voter turnout at referenda and blood donations, and argue, as we do here, that both can be safely considered exogenous because they are determined not by economic or legal incentives but rather by civic and altruistic norms. In particular, they find greater financial development where social capital is higher. 3. Data and Empirical Strategy As mentioned above, in our analysis we disentangle various aspects of social capital, so we are able to separately discuss the positive and negative effects of associational networks, the importance of both civic and altruistic norms, and the relevance of historical heritage. This last aspect, besides being of interest in itself, also allows us to appropriately tackle endogeneity. Moreover, we control for a number of other variables that are usually considered determinants of crime, thus minimizing the risk that our results are driven by omitted variables. To further control for unobserved heterogeneity due, for instance, to differences in cultural tolerance toward organized crime, we introduce macroregional dummies. We consider the cross section of the 103 Italian provinces. Summary statistics and extreme values for all the variables used are reported in Table 1, while data sources and the exact definition of each variable are presented in Table A Measures of Social Capital The most complex and debatable issue is how to measure social capital, given its multidimensional and multifaceted nature. Putnam (2000) argues that the more general forms of social capital are trust and social participation. In particular, he subdivides social participation into political participation, civic par- 7 Several decades before Putnam, Banfield (1958) looked at the strength of family ties and at what he termed amoral familism to find the roots of backwardness and criminality in southern Italy. 8 Putnam emphasizes that the process of social capital accumulation takes centuries. Guiso, Sapienza, and Zingales (2007) show that current differences in social capital across Italy can be traced back to the free-city experience enjoyed by some cities and not by others at the turn of the first millennium.

7 Social Capital and Crime 151 Table 1 Summary Statistics Mean SD Min Max Reported car thefts per 1,000 inhabitants True car thefts per 1,000 inhabitants Clear-up rate for car thefts Car theft report rate Reported robberies per 1,000 inhabitants True robberies per 1,000 inhabitants Clear-up rate for robberies Robbery report rate Reported thefts per 1,000 inhabitants True thefts per 1,000 inhabitants Clear-up rate for thefts Theft report rate Urbanization rate Length of judicial proceeding (years) Young males ages (%) Population with high school education (%) Unemployment rate Gross domestic product per capita (EUR 1,000s) Referenda turnout Associations per 100,000 inhabitants Voluntary associations per 100,000 inhabitants Blood donations per 100,000 inhabitants Criminal networks per 100,000 inhabitants ticipation, religious participation, altruism, and volunteering. In order to account for this multidimensionality, we consider four different measures of social capital: recreational associations, voluntary associations, referenda turnout, and blood donation. Table 2 reports the correlations among these four measures of social capital. Despite the different nature of these variables, their correlations are high, except for the one between blood donation and associations. As Figures 1 3 show, social capital variables are mainly concentrated in northern and central regions, although they present a heterogeneous distribution. The variable for recreational associations (Association) measures the number of recreational, cultural, artistic, sports, environmental, and any kind of nonprofit associations per 100,000 inhabitants in To narrow the focus, the variable for voluntary associations (Voluntary Association) captures in an analogous way the propensity of citizens to volunteer. Referenda turnout (Turnout) is the average of electoral turnout in the referenda held in Italy between 1974 and 1999 (notice that voting at referenda is not mandatory in Italy). Finally, the variable for blood donation (Blood) measures the number of donations per 100,000 inhabitants. Unlike Guiso, Sapienza, and Zingales (2004), our measure of blood donation includes all provincial and regional associations of blood donors and is not restricted to Associazione Volontari Italiani Sangue (AVIS), the major Italian association of blood donors. Indeed, although AVIS blood donations represent 90 percent of all donations at national level, they

8 Table 2 Correlations ln(car Thefts) ln(robberies) ln(thefts) Association Voluntary Association Blood Turnout Robbery Report Rate Theft Report Rate Car Theft Report Rate ln(car Thefts) ln(robberies) (.0000) ln(thefts) (.0000) (.0000) Association (.0620) (.2510) (.0000) Voluntary Association (.0440) (.2813) (.0000) (.0000) Blood (.0007) (.4320) (.0465) (.0070) (.0055) Turnout (.0067) (.1704) (.0000) (.0000) (.0000) (.0000) Robbery report rate (.0006) (.0775) (.6761) (.0002) (.0002) (.0239) (.0035) Theft report rate (.0169) (.1891) (.0007) (.0011) (.0008) (.0000) (.0000) (.0469) Car theft report rate (.4009) (.7221) (.6870) (.7677) (.7377) (.1236) (.8387) (.0007) (.0655) Note. Significance levels are in parentheses.

9 Social Capital and Crime 153 Figure 1. Voluntary associations per 100,000 inhabitants account for 0 percent in Friuli Venezia Giulia, 40 percent in Puglia, 60 percent in Tuscany and Calabria, and 80 percent in Veneto, Emilia Romagna, and Sicily. Thus, restricting the measure to AVIS would underestimate social capital in these regions Crime Variables We consider thefts, robberies, and car thefts. Table 2 shows correlations among reported crime rates, crime-specific report rates, and social capital variables. 9 Figures 4 6 and 7 9 show the geographical distribution (by quintiles) of reported 9 Data on crime-specific report rates are available from victimization surveys at the regional level. In particular, our source is a National Institute of Statistics survey conducted every 5 years on about 60,000 households. We use data from the 2002 wave (ISTAT 2002).

10 154 The Journal of LAW& ECONOMICS Figure 2. Blood donations per 100,000 inhabitants and true crime rates, respectively. 10 In our regressions, the dependent variable is the logarithm of true crime rates (per 1,000 inhabitants). Report rates in Italy vary significantly across crimes and space. Thefts and robberies are characterized by a high degree of underreporting and high levels of geographic heterogeneity in report rates. The rate of report of thefts is 30.8 percent in the north and 21.3 percent in the south, and the robbery report rate is 46.6 percent in the north and 32.2 percent in the south. By contrast, car thefts do not suffer from underreporting (more than 94 percent of car thefts are reported) and, more important, the rate of report is very similar (almost identical) across space. If a higher fraction of crimes are reported to the police in provinces with greater social capital, regressions based on reported crime rates 10 True crime rates are obtained by dividing reported crime rates (at province level) by the corresponding report rate (at the regional level).

11 Social Capital and Crime 155 Figure 3. Referenda turnout yield an upward-biased estimate of the effects of social capital. Table 2 shows that, while report rates of car thefts are not significantly correlated with social capital, the report rates of both thefts and robberies display a positive and significant correlation with all our measures of social capital. We take this as prima facie evidence of the importance of using true (that is, report-rateadjusted) crime rates rather than reported crime rates, especially for thefts and robberies Socioeconomic, Demographic, and Geographic Controls Our data comprise a set of socioeconomic and demographic variables that are likely to be correlated with crime rates. The explanatory variables are separated into three groups: deterrence variables, demographic variables, and socioeconomic variables. Deterrence variables determine the expected returns to

12 156 The Journal of LAW& ECONOMICS Figure 4. Reported common thefts per 1,000 inhabitants crime. The deterrence variables used here are the length of the entire judicial process (Length) and the (crime-specific) clear-up rate (Clear Up), lagged one period in time to avoid endogeneity. We include two standard demographic variables: the percentage of men ages (Youth) and the share of population living in cities with more than 100,000 inhabitants (Urbanization). Young men are said to be more prone to engage in criminal activities than the rest of the population (Freeman 1991; Grogger 1998). Big cities show much higher crime rate than do small cities (Glaeser and Sacerdote 1999). We complete our data set by including a set of socioeconomic variables: gross domestic product (GDP) per capita, unemployment rate (Unemployment), and share of the population with a high school education (High School). Gross domestic product per capita and the unemployment rate proxy for the

13 Social Capital and Crime 157 Figure 5. Reported robberies per 1,000 inhabitants general level of prosperity in each province and thus for legitimate and illegitimate earning opportunities (Ehrlich 1973). Education may reduce crime over and above its effect through labor market opportunities (employment rate and wage rate). In particular, education may have a sort of civilization effect (Fajnzylber, Lederman, and Loayza 2002). In order to address the possibility that certain associational networks increase crime rates, we control for the presence of criminal association at the provincial level. We proxy the extent of criminal networks with the number of charges for criminal association per 100,000 inhabitants (Criminal Networks). 11 This control 11 Association among three or more people to commit several crimes constitutes a specific type of crime in the Italian justice system, called criminal association and regulated by the penal code (Codice penale, art. 416).

14 158 The Journal of LAW& ECONOMICS Figure 6. Reported car thefts per 1,000 inhabitants is particularly relevant in the Italian context, which presents a strong and pervasive presence of organized crime in the southern provinces. Despite the wide number of controls mentioned above and the various variables used for robustness checks (see below), there still remains the possibility that Italian provinces differ along some unobserved characteristics correlated with both social capital and crime rates. The standard solution of panel data estimations with province fixed effects is not viable for the present study both because of data limitations and because social capital typically varies very slowly over time, so we would not likely have enough variation to identify its effects. So we exploit cross-sectional variations and address unobserved heterogeneity by including macroregional dummies for areas that are likely to be homogeneous along relevant unobserved characteristics, namely, the north, center, and south of Italy. To go even further, we split southern Italy in three further macroregions,

15 Social Capital and Crime 159 Figure 7. True (report-rate-adjusted) common thefts per 1,000 inhabitants distinguishing between areas plagued by organized crime (Sicily, Calabria, Campania, and Puglia), mainland areas that are not (Abruzzo, Molise, and Basilicata), and Sardinia Spatial Analysis As mentioned above, localized interaction and mobility across the borders of different provinces may give rise to spatial correlation in provincial crime data (Zenou 2003). Anselin (1988) presents the two standard econometric models to deal with this issue: the spatial lag model and the spatial error model. While the latter focuses just on spatial correlation in errors, for instance, owing to the way data are collected and aggregated, the former takes the substantive view that crime in a province is correlated with crime in neighboring provinces owing, for instance, to localized interaction patterns. Ignoring spatial correlation in

16 160 The Journal of LAW& ECONOMICS Figure 8. True (report-rate-adjusted) robberies per 1,000 inhabitants errors, when it is present, just causes a loss of efficiency of the estimator, but ignoring spatial correlation in the dependent variable, if present, causes more serious problems since estimates become biased and inconsistent. The solution adopted in the spatial lag model, on which we focus, is to introduce the average crime rate in neighboring provinces, adequately weighted to account for distance, as an additional regressor. Formally, this implies using a weight matrix W to construct a spatially lagged dependent variable, Wy, which is included in the regression: y p rwy Xb, where y is the dependent variable (a vector), X is a matrix of regressors, b and r are coefficients (a vector and a scalar), and is the error term (assumed independent and identically distributed). A crucial issue is how to properly define neighboring provinces and therefore the spatial lag. Different choices of the W matrix may yield different results, and unfortunately there is not a single obvious way to construct such matrix. To be sure

17 Social Capital and Crime 161 Figure 9. True (report-rate-adjusted) car thefts per 1,000 inhabitants that our results are not driven by a particular choice of W, we construct three different weighting matrices: one based on the inverse of road travel distance between the capital cities in each province, one based on the inverse of Euclidean distance between their geographic coordinates, and one based on simple contiguity among provinces. The first two matrices construct a given province s spatially lagged crime rate as a weighted average of crime rates in all the other provinces in Italy, with weights equals to the inverse of distance. The contiguity matrix constructs the spatial lag as the average crime rate in bordering provinces. Thus, the first two matrices allow us to consider spillovers from all other provinces in Italy, with farther provinces exerting less intense spillovers, whereas the contiguity matrix considers only local spillovers from bordering provinces. We run all our regressions with these three alternative specifications Moreover, for each matrix W, we check whether row standardization makes any difference, and for the distance-based matrices, we check whether imposing different cutoff distances (that is, setting w ij equal to zero if the distance is above a certain cutoff value) changes our results.

18 162 The Journal of LAW& ECONOMICS 4. Results Tables 3 5 present our baseline regressions of crime rates on the various measures of social capital, controlling for deterrence, demographic, and socioeconomic variables as well as for macroregional dummies and spatial correlation. Among the controls, the degree of urbanization, the presence of criminal networks, lagged clear-up rates, and the length of judicial proceedings are the most significant explanatory variables of all crime rates, each with the expected sign. The main results of our estimations concern the effects of social capital. A 1-standard-deviation increase in blood donations leads to significant reductions in common thefts by 13 percent and in robberies and car thefts by 15 percent (around 1/4, 1/7, and 1/6 of the respective standard deviations). Associational density is negatively correlated with all crime rates, but it is significant only for car thefts and robberies and not for common thefts. 13 Referenda turnout is not significantly correlated to any crime rate. We have performed a number of robustness checks to control for the effects of business cycle, large cities, and family ties. None of them is significant, and their inclusion does not significantly alter the effects of either social capital or the other controls. In terms of the sign and significance of the coefficients, our spatial analysis broadly confirms simpler results obtained by ordinary least squares regressions. The coefficient of the spatial lag is always positive but is significant in only one of 12 regressions (and just at the 10 percent level). All of the above-mentioned results, and in particular the sign, magnitude, and significance of the coefficients of social capital variables, are extremely robust to alternative matrices of spatial weights: we find the same results using road travel distance, Euclidean distance, and province contiguity, both row standardized and not, and even with different cutoff values for distance-based matrices. One reason for caution with these results, especially as far as associations are concerned, is that they may be due to reverse causation. As we argued in the previous sections, other studies find evidence that higher crime rates constrain social interaction, and theoretical models warn us that reverse causation may be present in a number of settings. The negative relationship we find could then reflect the effect of crime on associations rather than the other way around. To control for this, we instrument voluntary associations with historical data on mutual aid societies, cooperatives, electoral turnout, and cultural and recreational associations founded before 1860, 1 decade before Italy s unification. Historical data come from the seminal work by Putnam (1993) on the role of social capital on government functioning in Italy. Association density today is indeed significantly correlated with our historical 13 A 1-standard-deviation increase in voluntary association density is associated with significant reductions in robberies by 17 percent and in car thefts by 23 percent (around 1/7 and 1/4 of the respective standard deviations) and with a reduction in common thefts by 3 percent, which is not statistically significant. Results for recreational associations are almost identical.

19 Association Voluntary Association Blood Turnout Social Capital and Crime 163 Table 3 Determinants of Theft (1) (2) (3) (4).002 (.002).002 (.002).006** (.002).004 (.009) Criminal Networks.317*.320** ** (.140) (.104) (.134) (.098) Length.068*.071**.066*.071** (.032) (.025) (.030) (.027) Youth (.147) (.114) (.093) (.100) High School (.023) (.018) (.018) (.018) Unemployment (.022) (.019) (.024) (.030) GDP (.040) (.030) (.031) (.046) Urbanization.007**.007**.006*.006** (.002) (.002) (.002) (.002) Clear Up t 1.039**.039**.048**.042** (.009) (.010) (.012) (.010) Spatial lag ( r ) (.437) (.449) (.421) (.435) Note. The dependent variable is the logarithm of report-rate-adjusted common thefts per 1,000 inhabitants. Bootstrapped (10,000 repetitions) and regionally clustered standard errors are in parentheses. Spatial lag model regressions are used, with a row-standardized weighting matrix based on the inverse of road-traveling distance between provinces. Significant at 10%. * Significant at 5%. ** Significant at 1%. instruments, which offer different measures of civic participation between the mid-nineteenth century and the advent of Fascism. We argue that these instruments are not likely to exert any direct effect on current crime rates, besides the indirect effect through current association density, at least once we control for the other determinants of crime, such as income, criminal networks, macroregional dummies, and the other covariates mentioned above. Although high levels of persistence in associational networks and crime rates might imply that our strategy does not entirely solve the endogeneity problem, on one side our use of four different historical instruments minimizes the risk that persistence carries reverse causation through time, and on the other side this is probably the best one can do. 14 Table 6 presents the results of a two-stage least squares regression 14 Notice that a Hausman test confirms that our instruments are exogenous in all regressions. In particular, the Wu-Hausman F-test is equal to.903 (p-value p.344) for the car theft first stage,.842 (p-value p.361) for the robbery first stage, and.040 (p-value p.843) for the theft first stage.

20 164 The Journal of LAW& ECONOMICS Association Voluntary Association Blood Table 4 Determinants of Robbery (1) (2) (3) (4).011 (.006).011 (.007).007* (.003) Turnout.025 (.018) Criminal Networks.967**.963**.869**.922** (.219) (.285) (.228) (.247) Length (.074) (.067) (.062) (.067) Youth (.161) (.152) (.122) (.143) High School.077*.079* (.032) (.039) (.040) (.035) Unemployment (.048) (.041) (.042) (.043) GDP (.057) (.060) (.055) (.063) Urbanization.014**.015**.011**.009** (.004) (.004) (.003) (.003) Clear Up t *.012*.009 (.006) (.005) (.006) (.005) Spatial lag ( r ) (.541) (.525) (.459) (.456) Note. The dependent variable is the logarithm of robberies per 1,000 inhabitants. Bootstrapped (10,000 repetitions) and regionally clustered standard errors are in parentheses. Spatial lag model regressions are used, with a row-standardized weighting matrix based on the inverse of road-traveling distance between provinces. Significant at 10%. * Significant at 5%. ** Significant at 1%. in which voluntary associations are first regressed against our historical instruments and then the predicted value is used (with the other covariates) in the crime regression. When instrumented with historical data, voluntary associations have a negative and significant effect on all crime rates, with a coefficient that is higher, in absolute value, than in the baseline regressions. 15 The same holds for cultural and recreational associations. 16 These qualitative results are robust to different specifications of the first-stage regression. A 1-standard-deviation increase in history-predicted association density leads to significant reductions in common thefts by around 20 percent and in robberies and car thefts by 15 Instrumental variables estimates show a negative and significant spatial correlation in common thefts, whereas the spatial lag is not significant for robberies and car thefts. 16 The results are identical and are available on request.

21 Association Voluntary Association Blood Turnout Social Capital and Crime 165 Table 5 Determinats of Car Theft (1) (2) (3) (4).013** (.005).015** (.005).007* (.003).009 (.014) Criminal Networks.775*.766**.680*.766** (.332) (.269) (.324) (.257) Length.127**.124**.110**.118** (.038) (.028) (.036) (.033) Youth (.179) (.131) (.136) (.151) High School (.028) (.029) (.028) (.031) Unemployment (.037) (.034) (.028) (.029) GDP (.045) (.053) (.049) (.052) Urbanization.014**.014**.009**.010** (.003) (.004) (.003) (.003) Clear Up t 1.069**.072**.078**.076** (.015) (.015) (.015) (.016) Spatial lag ( r ) (.353) (.443) (.423) (.437) Note. The dependent variable is the logarithm of car thefts per 1,000 inhabitants. Bootstrapped (10,000 repetitions) and regionally clustered standard errors are in parentheses. Spatial lag model regressions are used, with a row-standardized weighting matrix based on the inverse of road-traveling distance between provinces. Significant at 10%. * Significant at 5%. ** Significant at 1%. around 30 percent (that is, to respective reductions by 1/2, 1/5, and 1/3 of a standard deviation). 5. Concluding Remarks We present robust evidence that once we control for a variety of possible other determinants, as well as for endogeneity, underreporting, spatial correlation, and unobserved heterogeneity, civic and altruistic norms and associational networks significantly reduce crime rates. To assess this result, we exploit the great variation in social characteristics across Italian provinces, the availability of historical data, which we use as instruments, the high and geographically stable report rates for certain forms of property crime, crime-specific report-rate-adjusted figures of crime rates (constructed from reported crimes and report rates taken from victimization surveys), estimation of a spatial lag model, a number of control

22 166 The Journal of LAW& ECONOMICS Table 6 Instrumental Variables Results Thefts (1) Robberies (2) Car Thefts (3) First stage: Association (2.931) (2.869) (3.039) Cooperatives (12.305) (11.900) (11.909) Mutual Aid Societies * 3.442* (1.817) (1.828) (1.466) Electoral Turnout 6.344** 5.921** 6.235** (1.894) (1.858) (1.567) Second stage: Voluntary Association.018*.026**.027* (.008) (.015) (.013) Criminal Networks.322**.954**.626** (.083) (.178) (.226) Length.112** ** (.023) (.065) (.035) Youth.209** (.079) (.150) (.137) High School **.057** (.017) (.030) (.022) Unemployment *.007 (.015) (.026) (.021) GDP (.027) (.048) (.043) Urbanization.012**.020**.019** (.004) (.007) (.006) Clear Up t 1.045** ** (.017) (.006) (.012) Spatial lag ( r).537** (.141) (.535) (.514) Note. In the instrumental variables first stage, Voluntary Association is instrumented with four historical variables, kindly provided by Robert Putnam. Bootstrapped (10,000 repetitions) and regionally clustered standard errors are in parentheses. Ordinary least squares regressions are used in the first stage, and spatial lag model regressions are used in the second stage, with a row-standardized weighting matrix based on the inverse of road-traveling distance between provinces. Significant at 10%. * Significant at 5%. ** Significant at 1%. variables (some of which were directly collected by the authors), and macroregional dummies. In particular, we show that blood donations have a negative, significant, and sizeable effect on all property crime rates, referenda turnout has no significant effect whatsoever, and history-predicted association density significantly and sizably reduces all property crime rates. Our results imply that a policy of promotion of civic norms and associational life may have beneficial side effects in terms of crime reduction. Thus, sociocultural interventions may usefully complement traditional anticrime policies based on the threat of punishment. Indeed, as argued by Kugler, Verdier, and

23 Social Capital and Crime 167 Zenou (2005), in some cases greater threat of punishment induces higher levels of corruption rather than lower levels of criminality, as appears to be the case in some Italian provinces. Our empirical investigation supports the idea that in such cases it makes sense to fight the deep social roots of crime in addition to repressing its specific manifestations. Since we have neither modeled nor empirically investigated the determinants of civic norms and associational networks, more precise policy indications would be purely speculative. In any case, if the negative correlation we find reflects a stable relationship, then policy intervention might alter the underlying parameters, for instance, those governing the flows of creation and destruction of new and old associations. If, by contrast, it reflects coordination of different provinces on different equilibria, some with high crime and low civic participation rates and some with the opposite features, then the rationale for intervention would come from coordination failure. If there is a critical mass of individuals that, once coordinated on a different equilibrium strategy, is sufficient to induce the rest of the population to form expectations accordingly and therefore shift behavior, then the policy target should be the simultaneous coordination of this critical mass. 17 While the evidence we present does not allow us to discriminate between these two views, it definitely speaks in favor of global strategies of tackling crime. Appendix Table A1 Variable Descriptions and Data Sources Variable Description Source Association Number of cultural, recreational, artistic, ISTAT sports, environmental, and other nonprofit associations per 100,000 inhabitants at the province level in 2000 Voluntary Association Number of associations whose members ISTAT work voluntarily per 100,000 inhabitants at the province level in 1999 (Legge quadro sul volontariato, Law No. 266, Gazz. Uff. 196 [August 22, 1991]) Blood Number of blood donations per 100,000 inhabitants collected in 2000 by AVIS and AVIS, FRATRES, FIDAS, AFDS other blood donor associations that collect blood at the province level (FIDAS, FRATRES, AFDS) Turnout Voter turnout at the province level for all referenda held ; for each province, turnout data were averaged across time Ministry of Interior 17 Observe that social capital persistence over time is compatible with both views: on one side, stocks change slowly when entry and exit flows are approximately balanced; on the other side, they change seldom when substantial critical masses are needed to trigger a change.

24 Table A1 (Continued) Variable Description Source Thefts Logarithm of common thefts per 1,000 ISTAT inhabitants at the province level; average Robberies Logarithm of robberies per 1,000 inhabitants ISTAT at the province level; average Car thefts Logarithm of car thefts per 1,000 inhabitants ISTAT at the province level; average Thefts report rate Percentage of common thefts reported to the ISTAT police in 2002 Robberies report rate Percentage of robberies reported to the ISTAT police in 2002 Car thefts report rate Percentage of car thefts reported to the ISTAT police in 2002 GDP Gross domestic product in the province ISTAT (EUR 1,000s) divided by population in the province in 2001 Unemployment Total rate of unemployment at the province ISTAT level in 2001 High School Percentage of population over 6 years of age ISTAT who completed at least high school at the province level as of 2001 Youth Percentage of young males ages in the ISTAT 2001 population at the province level Urbanization Percentage of population living in cities with ISTAT more than 100,000 inhabitants at the province level in 2001 Length Average number of years to complete the Ministry of Justice first and second phases of trial by the courts located in a province; data are computed using court-level data and then averaging across courts located in the same province in 2001 Clear Up t 1 Ratio of the number of crimes cleared by the ISTAT police to the total number of crimes reported, for each province and crime category in 1999 Criminal Networks Number of incriminations for criminal ISTAT association per 100,000 inhabitants at province level; average Association 1860 Fraction of cultural and recreational Putnam (1993) associations founded before 1860 and still existing in 1982 Cooperatives Density of cooperatives in the population, Putnam (1993) averaged across 1889, 1901, 1910, and 1915 Mutual Aid Societies Participation in mutual aid societies, Putnam (1993) standardized by the population and averaged across 1873, 1878, 1885, 1895, and 1904 Electoral Turnout Average turnout in national elections in 1919 and 1921 and in local and provincial elections in 1920 (the only elections with universal male suffrage before Fascism) Putnam (1993) Note. ISTAT p Istituto Nazionale di Statistica; AVIS p Associazione Volontari Italiani Sangue; FRATRES p FRATRES Donatori Sangue (Consociazione nazionale dei gruppi donatori di sangue); FIDAS p Federazione Italiana Associazioni Donatori di Sangue; AFDS p Associazione Friulana Donatori Sangue.

25 Social Capital and Crime 169 References Anselin, Luc Spatial Econometrics: Methods and Models. Dordrecht: Kluwer. Anselin, Luc, Jaqueline Cohen, David Cook, Wilpen Gorr, and George Tita Spatial Analyses of Crime. Pp in Measurement and Analysis of Crime and Justice. Vol. 4ofCriminal Justice 2000, edited by David Duffee. Washington, D.C.: National Institute of Justice. Banfield, Edward C The Moral Basis of a Backward Society. New York: Free Press. Becker, Gary S Crime and Punishment: An Economic Approach. Journal of Political Economy 76: Bjørnskov, Christian The Multiple Facets of Social Capital. European Journal of Political Economy 22: Calvó-Armengol, Antoni, Eleonora Patacchini, and Yves Zenou Peer Effects and Social Networks in Education and Crime. Working Paper No Research Institute of Industrial Economics, Stockholm. Calvó-Armengol, Antoni, Thierry Verdier, and Yves Zenou Strong and Weak Ties in Employment and Crime. Journal of Public Economics 91: Calvó-Armengol, Antoni, and Yves Zenou Social Networks and Crime Decisions: The Role of Social Structure in Facilitating Delinquent Behavior. International Economic Review 45: Case, Anne C., and Lawrence F. Katz The Company You Keep: The Effects of Family and Neighborhood on Disadvantaged Youths. Working Paper No National Bureau of Economic Research, Cambridge, Mass. Chamlin, Mitchell B., and John K. Cochran Social Altruism and Crime. Criminology 35: Coleman, James Social Capital in the Creation of Human Capital. American Journal of Sociology 94:S95 S Social Capital. Pp in Foundations of Social Theory, edited by James Coleman. Cambridge, Mass., and London: Harvard University Press, Belknap Press. Ehrlich, Isaac Participation in Illegitimate Activities: A Theoretical and Empirical Investigation. Journal of Political Economy 81: Fajnzylber, Pablo, Daniel Lederman, and Norman Loayza Inequality and Violent Crime. Journal of Law and Economics 45:1 40. Freeman, Richard B Crime and the Employment of Disadvantaged Youths. Working Paper No National Bureau of Economic Research, Cambridge, Mass. Glaeser, Edward L., and Bruce Sacerdote Why Is There More Crime in Cities? Journal of Political Economy 107:S225 S228. Glaeser, Edward L., Bruce Sacerdote, and José A. Scheinkman Crime and Social Interactions. Quarterly Journal of Economics 111: Grogger, Jeffrey T Market Wages and Youth Crime. Journal of Labor Economics 16: Guiso, Luigi, Paola Sapienza, and Luigi Zingales The Role of Social Capital in Financial Development. American Economic Review 94: Long Term Persistence. Unpublished manuscript. Northwestern University, Department of Finance, Evanston, Ill. Heaton, Paul Does Religion Really Reduce Crime? Journal of Law and Economics 49: