Politico1 Geography, Vol. 14. No. 617. pp. 521-541, 1995 Copyright 0 1995 Rsevier Science Ltd Rimed in Great Britain. All rights reserved 0962-629fm $10.00 + 0.00 0!%2-6298(9s)ooo53-4 Crime, policing and the perception of neighborhood safety JEFF GROGGER Department of Economics AND M. Department STEPHEN WEATHERFORD of Political Science, University of California, Barbara, CA 93106, USA ABSTRK~. Crime is a powerful political abate crime pose difficult trade-offs policing and the multiplicity imperative that spending issue in urban America, for governments. of other pressing be directed feels most information concerned is available or threatened. but policies needs for scarce budgets, as possible: to it is governments areas and crimes where the Unfortunately, about how individuals little systematic form a conception of neighbor- hood safety, or how the demand for police services relates to the occurrence crime in their question. The marginal science techniques spatial combines ambiguities tract-level reports on an exploratory approaches from to pay for publicly provided analysis data), for police of voting specifications services). a more realistic formulation Whilst analysis economics police and (to estimate services), political geography of the relationship the application of the analytical of of that (using between GIS crime of GIS techniques question, it also reveals in both the theory and method used in earlier work. Using census data on crime rates and voting on an initiative to increase taxes to pay for police services, the abatement sensitive paper to test alternative and demand our analysis shows that the public is more willing to pay for of violent crime than property to assumptions concerning range This study willingness (the fosters area. Santa Given the high costs of as efficaciously wish, that is, to direct spending toward geographical public Santa Barbara, crime. But even this inference about how individuals gather and process the frequency is information and severity of crime in their local area We specify a of information-processing models, and draw inferences for further research. Introduction Crime is a pressing crime are plentiful and they compete social problem and debate on the about public and a powerful them intense, agenda with political issue. but abatement other pressing Strategies for mitigating strategies are expensive issues. For example, incarcerating one malefactor for a year is estimated to cost taxpayers $15 000-30 the city of Los Angeles spent an estimated $374 million for police protection 000, and in 1980. Crime, policing and the perception of neighborhood safety 522 Given such sizable pronounced. outlays, the challenge of wisely allocating scarce resources is Society would wish, that is, to reduce the danger posed by crime by the greatest amount for a given criminal justice budget. In a democratic crucial information needed by the government political system, the would define the public’s aversion to different types of crimes. Other things being equal, we would like to spend additional dollars on crime control to prevent those crimes that society deems most severe. Such a severity ranking is not one an outside observer can readily make. While it is easy to state that murder, for example, is more severe than joy riding, and to rank burglaries of empty houses based on the value of goods stolen, there are many cases where such a ranking is considerably more difficult to make. For instance, the line between assault and robbery is often blurred- is a savage beating by itself worse than a mugging that involves a loss of funds, but less severe bodily injury? While one might attempt a ranking based on the sum of funds stolen, entirely the considerable attempts to characterize reflect hospital costs and lost income, psychic costs imposed on victims. A social welfare function that the way victims and their fellow In this paper, we propose over different neighborhood referendum protection. a strategy to obtain types of crimes, that explicitly linked the form psychological images and its potential an ordinal ranking by examining increased of the impact on of social (dis-) the relationship between voting patterns on a Los Angeles property with taxes We follow a strategy similar to that used by economists additional city police to estimate marginal for publicly provided goods. Our analysis is frankly exploratory, about citizens of the crime problem crime rates and neighborhood willingness-to-pay would miss the public’s relative aversion for different crimes should, in short, magnitude and spatial distribution themselves. preferences such a scheme public’s beliefs, one and the exposition involving crime, interweaves two related themes the other geography. Our initial hypothesis, drawn from the literature, was that the vote on a measure to increase police protection should neighborhood. depend in a fairly direct way on the crime The analysis shows that the relationship more complex unanticipated than that hypothesis subtlety-specifically, not correspond implies, because the geographical in any simple way to the psychological between rate in the voter’s crime and voting is ‘neighborhood’ is a construct space environing of individuals may space relevant to their assessment of crime and personal safety. In a nutshell, we find that the measured effect of crime on voting depends neighborhood conceptions potentially on the definition of crime rates. The problem ‘neighborhood’ of melding that of space occurs in a wide variety of social science serious impacts on estimation and inference. one geographical uses to construct and psychological settings, and may have While earlier researchers made little progress beyond identifying the problem, the use of GIS technology have has made it feasible for us to pursue the issue analytically. A model of optimal crime control In this section we present a simple formal model of optimal crime control. The model points out the importance of knowing the social disutility ranking of various crime types for the formation of optimal criminal justice policy. The exposition is for the most part a simplification of that in Grogger (1987). There are two crime types, the levels of which are denoted at c1 and c,. The direct costs imposed on society in terms of lost and damaged property, time losses and psychic losses are given by the social disutility function U (c,, c2). Both partial derivatives of U (> are JEFF GR~GGER AND M. negative but decrease police expenditures at a decreasing 523 STEPHENWEATHERFORD rate for all values of c1 and c,. Society allocates to abating both types of crime, and it is assumed that expenditures can be at least partially targeted toward abating one type over the other. Expenditures devoted toward crime 1 are denoted by sl, those devoted toward type 2 denoted The amount expenditure, of each type of crime is related s,. abatement according to two aggregate crime functions. Drawing on previous research, these can be thought of as the aggregated who respond consumption to the corresponding to the level of police first-order conditions expenditures of potential criminals, and to legitimate employment and prospects so as to maximize their expected utility. These reaction functions, which serve as behavioral feasibility constraints to society, are given as c, = f(sJ, It is assumed that A’(s3 < 0 where J;’ denotes i = 1,2. the derivative of A. Finally, the budget constraint of the police is given by K (sl + s2) < i?, where K+ > 0. Society’s problem is therefore to ::‘g L = U(c,, CJ s.t. cI = d (s,) i = 1, 2 K(s, The first-order conditions + s,) < Ic for a minimum can be written .fi’ (s1) -=_ Y2 f2’ Yl (s2) where y, and y2 are the Lagrange multipliers associated with the first two constraints. Now, y, is the shadow value of reducing the level of crime i by one unit, expressed in social utility units. In other words, it is the social disutility weight attributed to the ith crime type. The more severe crime type will therefore condition can thus be interpreted the ratio of the marginal deterrent effect of expenditures deterrent effect of expenditures have the larger yi. The first-order as stating that resources should be allocated such that on crime type 1 to the marginal on crime type 2 should be in proportion to their relative disutility weights. The relative expenditure values of two factors are therefore seen to influence the optimal to control a specific type of crime: the severity of the crime as reflected by its social disutility weight y,, and its deterability (I;‘). If deterability were equal, then more should be spent on deterring the more severe crime. Several researchers have examined the deterability of different types of offenses (Witte, 1980; Myers, 1983; Grogger, 1991). In contrast, there has been no research directed toward estimating societal preference orderings over different types of crimes. In the next section, we discuss the data we use in our preliminary estimate of such an ordering. The data Data on the number of crimes of different types in the city of Los Angeles are reported by police reporting district (PRD) on a quarterly basis by the Los Angeles Police Department. We aggregate the quarterly data to provide annual counts for 1985. The PRDs by which the data are reported are small geographical units which in many cases coincide with census tracts (CTs). In most cases where the two units do not coincide, two or more PRDs nest cleanly within a CT. The data were aggregated crime counts analyzed below. by hand to provide the census tract Crime, policing and the perception of neighborhood safety 524 Data on police expenditures per capita within Los Angeles are reported on an annual basis in the LAPD Annual Report. These are provided by police reporting area (PRA), a larger geographical unit nesting many PRDs.There are eighteen PRAs; we use expenditure data from 1985. The voting data come from a referendum 1985. The referendum, improvements Proposition on the Los Angeles city election ballot in June 1, would have mandated to finance the employment a special tax on land and of up to 1000 additional police officers over the period 1985 to 1990. The language of the measure made it very clear that funds raised by the tax would be used solely for increasing police patrols and for related support costs. The measure failed with a ‘no’ vote of about 55 percent. Precinct-level voting data on Proposition 1 were obtained from the Los Angeles City Registrar of Voters. The Registrar’s Office also provided detailed maps of the voting precincts, and these were used to hand-match precincts to CTs. In general, CT overlays were used to determine the proportion of each precinct that fell in each tract. About 80 percent of all precincts fell within a single CT. Tract-level vote tallies were then computed for analysis. Finally, to control for factors known to be important determinants ballot measures, demographic and income variables from the 1980 Census were taken from the Census Summary Tape File 3. The temporal 1980 demographic measurement omitted mismatch stemming from use of data and 1985 ballot and crime data is doubtless undesirable, error induced by this temporal mismatch variable demographic of voting on such bias that would and economic occur is of lesser magnitude in the absence but the than the of any data on important control variables.’ The variables to be analyzed are defined and statistically summarized in Table 1. Due to collinearity among crime counts, TABLE 1. Variable Variable abbreviation PYES LOGPYES VCRIME PCRIME POLEXP PRLACK I’ASIAN POTHER POP DAYPOP POPDENS PCINC PPOV the data were aggregated definitions and summary Description Percentage ye5 votes out of total votes cast, 1985 Log-odds of PYES (i.e. log[PYES/ UOO-PYES)l) Violent crimes per 1000 resident population, 1985 Property crimes per 1000 resident population, 1985 Police expenditure per capita, 1985 Percentage of population that is black, 1980 Percentage of population that is Asian, 19x0 Percentage of population that is nonwhite, non-black and non-Asian 1980 1980 resident population Number of jobs located in tract in 1980 Resident population density 1980 1980 per capita income ($lOOs) Percent of families below poverty line, 1980 into two types: violent statistics Mean 45.3 -0.28 9.4 SD 17.4 0.77 14.7 117.5 307.3 125.9 69.9 17.6 29.9 6.8 8.4 12.4 14.8 4051.3 2 410.0 1726.4 4 788.7 10 677.8 87.4 13.6 7 445.0 54.5 11.7 JEFF GROCCER ANDM. STEPHEN 525 WEATHERFORD crimes and property crimes. Violent crimes include murder, rape, robbery and assault. while property crimes include burglary, larceny and auto theft.* Both crime variables are expressed as rates, giving the number of crimes per 1000 resident population.’ Two sets of explanatory variables merit particular discussion. that police mistreat members presumably would desire lower levels of police protection all else equal. We therefore tract as controls of minority groups; include variables measuring for residents’ preferences. Second, A frequent complaint in this case, minority is members than their major counterparts, the racial composition in each it may seem natural to include a measure of housing tenure in the model since the referendum involved a property tax. While property taxes are in fact levied on owners, much of the actual tax burden may be passed on to tenants. In this case, income will provide a better measure of willingness to pay for increased police services. Graphical descriptions of the vote proportions and crime rates are provided in Figures la- lc. In Figure la, for instance, we see that the referendum in Central and South Central Los Angeles, generally carried in tracts and failed in West Los Angeles and the San Fernando Valley. A relatively strong correlation would appear to exist between yes votes and the violent crime rate (Figure lb), with a weaker association between property crime rates (Figure Ic) and yes votes. Other graphical analyses not provided here indicated that areas with high proportions population densities, of yes votes also had larger black populations, higher and lower per capita incomes than areas with low yes votes. Estimation for police K5e demand services Estimation can be motivated by considering the individual’s demand function for police services, For the Ch individual in tract j, demand for police expenditures is assumed to take the form E ‘1* = Where zii u - uq (1) E,,* gives desired expenditures, neighborhood disturbance crime rates, income ZY is a vector of relevant attributes, and demographic variables, including and Up is a random term. While E,,* is not directly observed, the individual is assumed to vote in favor of the measure if E,; > A,, where A1 is actual expenditures measure otherwise. logistically distributed, With the additional assumption in tract j, and to vote against the that the u,{ are independently one obtains the relation (2) where p, is the fraction of yes votes in tract j, and Z/contains tract averages of the variables in ZIr Equation (2) is expressed in terms of observable tract-level variables, and is linear in the parameters 6. It can therefore be estimated by ordinary least squares (OLS). The coefficients of the crime rate variables are proportional to the average change in desired police expenditures for a unit increase in the corresponding crime rate, holding all other factors constant. Presumably, voters’ demands for increased police protection in response to equal changes in different types of offenses reflect the seriousness they attach to those offense types. That is, a unit increase in more serious offenses will lead to a greater Crime, policing and the perception of neighborhood safety 526 m 0.14 to <0.25 m 0.25 to <0.50 q n 0.50 to <0.75 0.75 to 0.91 n No data FIGURFLA. Proportion of yes votes JEFF GROGGER AND M. STEPHENWEATHERFORD m 0 to <13.6 q n n 13.6 to <25.6 25.6 to <53.2 53.2 to 951 q No data The violent crime data robbery and aggravated include murder, assault FIGIRF 1~. Violent crimes rape, per 1000 resident population 527 Crime, policing and the perception of neighborhood safe& c 35 to <lo6 @fj 106 to (146 Property and q 146 to (200 H 200 to 2850 q No data crimes include burglary, larceny autotheft FIGURElc. Property crimes per 1000 resident population. 529 JEFF GROCCERANI M. STEPHENWFATHERFORI~ increase in desired reason, the estimated of different types expenditures crime than a unit increase rate coefficients in less serious can be used to construct crimes. For this a severity ranking of offenses. Specifuing the concept of neighborhood Some conception of ‘neighborhood’ individuals’ judgments individual’s judgment of their or ‘local area’ own is organized psychologically meaningful, conceptualizing the geographical likelihood must by a conception but the lie at the victimized of the spatial literature bounds logically of being offers little by environment empirical of this psychological precincts or other comparable reporting units that have requirements spatial been bear space. crime defined rate’ 1982a, individual, for demographic his/her of those living catchment area for information localized within passively. More interpersonal use attributable individuals persuade each interactions, other. as a surrounding Some The mismatch and exchange use surrogate that constitutes the opinions and homogeneous are fostered stations tend elite to be share. We might think of that most residents that attitudes exchanges may occupations and patterns do little indicators, in the by active more than of information but many will be contacts or attempt no ‘absorb‘ and beliefs likely to be nurtured have of an individual’s by the social class, ethnic information typically sciences of all but a few radio residents or bias, is more these the environments class or racial composition, scientists and thus of necessity the aggregate of only patterns condition of backgrounds, social it is a relatively with its implication direction Social data that give a opinions of opinion news-and that neighborhood the influences Local information of television purposively the level of the tract or voting that areas to a neighborhood’s in which whose like the ‘relevant political not The environment circulation of opinion’, conversations. data are available. processes across but also the climate on local and regional the commonality relevant to tracts, voting agencies, or the measure media-the commonality local area carry an identifiable acknowledge for pointedly, only for geographical of concepts studies vote to and opinion. environment The ‘climate the it, to the extent broadcasting and to concentrate or other demographic the informational affiliation of the mass the effects environment. behaviors and the spatial component it is conventional traits and partisan newspapers of public and cultural in contextual social psychological by the economics convenience to the social 198217). In predicting instance, record that is as a neighborhood. over how to specify is encountered (Weatherford, for the identity public finds that data are available little or no relation area its particular The ambiguity areas for which typically of The guidance there is virtually no reason to believe that the images of ‘neighborhood’ citizens’ reckonings about personal safety are coterminous with Census At the same time, the analyst heart crime. direct to influence measures such as census data at the level of the precinct, in order or of these information at to typify some of traits of the area.” of available spatial data and social psychological reality poses problems for both theory and method. Theories about how people form opinions or make political choices clearly err by treating each person as a monadic individual; attitudes and behaviors are the products of encounters between separate beliefs but also the need to test the accuracy of their beliefs their opinions) and their social surroundings (composed less predictable encounters>. To test and develop theories individual and society, we need measures of the social individuals (with their own and to gain approbation for of some mix of self-selection and about the interaction between context. I!sing available data. Crime, policing and the perception of neighborhood safety 530 almost certainly measured on area1 units that lack much ‘psychological reality’, is preferable to omitting contextual variables altogether; but the practice may simply replace one specification problem with another, properties lead to errors-in-variables since problems inappropriately (Weatherford, measured we present the results of our analysis of GIS data to investigate systematically alternative specifications of the spatial context contextual 1982a). In the next section, that environs a range of citizens’ thinking about different types of crime. Results The results of estimating Equation (2) from tract-level data are presented Although these models account for a high proportion voting data, with an R2 of 0.73, their implications for model specification Model 1, including the control variables along with a measure census tract, reveals a sizable, properly signed coefficient likelihood of voting worrisomely for Proposition in Table 2. of the variation in tract-level are inconsistent. of violent crime in the for the effect of crime on the I, but the standard error of the coefficient is large. Model 2 substitutes a measure of property crime and shows a small but significant effect of crime on the vote. Taking these two models together, we might infer that violent crimes have about 10 times the impact of property crimes on individuals’ demand for police protection, but that the measurement It might are artificial, however, since both types of crime occur in the real world and individuals’ views of personal be objected (or our specifications influence is imprecise. crime of its causal both presumably impact) of violent specification that these safety in the neighborhood. Model 3 includes the two crime variables along with the control variables. The results for property crime in this specification appear plausible, but the sign of the coefficient violent crime is wrong and its standard error is excessive. either problems with the specification problems with the data, presumably of the relationship on These symptoms might signal between crime and the vote, or collinearity between the two crime variables.5 In this section and in Appendix A we investigate each possibility seriatim. Do different crimes take place in different ‘neighborhoods’? Our measurement procedure of the crime rate variables in the models above follows conventional in contextual effects studies, aggregating all the variables at the level of the census tract and then using the tract as the unit of analysis. This specification voters had in mind information about the occurrence TAHLF2. Regression Variable VCRIME PCRIME Adjusted @ Model 1 0.207 (0.142) 0.730 results: dependent implies that of crime in their census tract, when variable is LOGPYES Model 2 Model 3 0.018 (0.006) -1.69 (0.190) 0.023 (0.008) 0.733 0.733 Notes. * Coefficient (x100), standard error. Mohel 1 includes VCRIME and control variables, Model 2 includes PCRIME and controls, and Model 3 includes both VCRIME and PCRIME along with the control variables. The control variables include POLEXP, PASIAN, PBLACK, POTHER, POP, DAYPOP, PCINC, AND PPOV JEFF GROCCER AND M. STEPHEN WEATHERFORD they voted on the proposal 531 for higher taxes to fund more police protection. Intuitively, however, we have little reason to believe that the basis for assessing one’s exposure crime comes from the crime committed solely within one’s tract of residence. to This is particularly likely given that the average area of a census tract in Los Angeles is only 1.7 kma, the area corresponding to a circle with a radius of only 730 meters, or less than half a mile. This seems well below the typical distance one would travel, say, on routine shopping trips, and even children who walk to school (about whose safety many voters are likely to be acutely concerned) on average. Moreover, crime spending problems measures, in adjacent are likely to traverse a greater distance than half a mile voters’ expectations about crime, and hence their votes on anti- might reasonably be based on information about crime areas, since these may spill over into their own neighborhood. Individuals’ assessments of personal safety, in short, are unlikely to be bounded area as small as a census tract; it would be reasonable for their perceptions by an of the relevant crime rate to take in a wider area. To test this intuition empirically, we constructed larger areas, and then included each sequentially of prior information constructed, each crime rate measures over successively in the regression model. Given the lack as to the proper area to be included, a series of such measures was indicating the crime rate over an area whose radius was a given multiple of ~1,where p is the radius of a circle with area equal to the average area of a tract. From the discussion corresponding above, then, [J is roughly 730 meters. Specifically, to k’p for k = 1 to 5, in increments total number of crimes committed FIGIXE 2. Census tracts of 0.25, computing we defined areas crime rates as the in tracts whose centroids fell within a circle of radius and neighborhoods in South-Central Los Angeles Crime, policing and the perception of neighborhood safety 532 @I from the centroid of the target tract, divided by total resident population over the same area.’ Figure 2 illustrates these areas for a representative tract. Table 3 presents the results of estimating the model of Table 2 with various different measures of the crime rates. Note that the number of observations models varies across the specifications. whose centroids lay within /+Iof the city boundaries. deletion of the first Iz observations distributed lag, and the consequences The operation from a time-series model is analogous containing are identical: fewer observations The spatial definition of ‘neighborhood’ the results are generally (x100), using different definitions to the a kth-order for larger Iz. takes in roughly the same-sized for census tracts, and, as might be expected, TABLE3. Crime coefficients used to estimate the This is due to the necessity of deleting observations area for 1~ as similar to those of neighborhood Coefficient and standard error Neighborhood radius, in multiples of mean tract radius Model A(l) VCRiME 0.072 0.071 0.133 0.08 0.127 0.08 0.124 0.104 0.162 0.126 0.217 0.158 0.253 0.172 0.445 0.185 0.765 0.237 1.197 0.307 1.879 0.398 2.036 0.416 2.362 0.424 2.451 0.432 2.807 0.471 .?.29_? 0.476 3.554 0.504 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 3.25 3.50 3.75 4.00 4.25 4.50 4.75 5.00 Note; Coefficient Model A(l) PCRIME is in italics above, standard error below. N 0.016 0.007 0.106 0.008 0.015 0.008 0.043 0.017 0.178 0.04 0.142 0.042 0.139 0.044 0.182 0.047 0.275 0.059 0.391 0.074 0.574 0.075 0.654 0.102 0.705 0.103 0.791 0.105 0.852 0.112 0.975 0.116 1.042 0.122 Model specifications match those 571 552 532 504 481 456 442 429 406 389 376 365 351 343 324 313 270 in Table 2. JEFF GROGGER AND M. computed using census in Model differs the two specifications. across The specifications, remaining significance masked levels. relationship degree neighborhood tract. Models rate in a larger and In Appendix crime to take to affect cases. estimates and substantive the pattern interpret effect difficult the evidence A, we review to on the the evidence of the area of the ‘neighborhood’ a series of spatial contexts 1 and 2 provide area point on the analysis. between is increased of the in the data that makes rates and voting. relationship from zero in both from a coherent we present and of similar of the VCRIME coefficient in terms impact qf crime on the vote: defining of the census estimate we believe, section, its impact Model 2 and the PCRIME coefficient different both results, Recall that our goal in simulating pattern widely 533 for PCRIME is significant the point of collinearity crime and assess i%e ua ying vary In the next between collinearity While This variation statistically. effect it is insignificantly estimates by a certain isolate tract crime rates. Comparing 3, we can see that the estimated size across STEPHENWEATHERFORD rates in a larger and the and larger the clearest the vote, of varying and picture vote, sizes is to track the as the size area surrounding of the effect of allowing we focus on those of the the voter’s the crime specifications in this section. Note first that the summary expectations. series: the higher increasing the crime rate, comports the whole range, with the more crimes average based for PCRIME (0.411). The pattern moreover, to reveal additional The pattern insight of the impact of crime crimes from 0.07 to 3.6, while grows the relative the entire will fAvor magnitude of the of the crime: over virtually powerful on the vote than property effect (1.287) about of coefficients three times the across the simulations voters conceptualize their as the size of the area increases, The coefficient average appears. spatial the strength for the rate of violent crime rate increases are more readily seen graphically, and Figures _ia and .3h chart of the crime and we think significant coefficients difference against ~1. These figures in the pattern of effects of the two types of display an on voting. In Figure .?a we see that the coefficients violent crimes 2.50, increasing which in the area for the property patterns crime are in line with across the coefficient from 0.016 to 1.04. These intriguing Similarly, on the vote also increases. the magnitude that the vote into the way is unmistakable: specifications on the severity coefficient coefficient environment. likely have a more VCRIME by these on the vote is consistent protection. with expectations violent the revealed of crime taxes to pay for more police coefficients crimes. properties The sign of the impact voters it comes steadily over the remaining about their residential crimes, events citizens range.’ on the vote of the rate of This suggests the rate of violent of violence appear the impact of p, and then begin is not so much but rather the awareness to violent showing small for small values are responding neighborhood, information is quite at about ~1 = that the information to grow to crime in the immediate in the larger surrounding to define in a quite large area (within their personal a radius of about area, When safety in light of 2.5 miles) around block.’ Figure .?b shows that the pattern of estimated coefficients for the impact of the rate of property crimes on the vote differs from the pattern for violent crimes. Like the violent crime rate, the rate of property crimes in the immediate neighborhood has little effect on the vote (see Figure _?b), but the rise in the coefficient occurs earlier (/I = 1.75) and the pattern shows a notable spike at /I = 2.0, where the size of the effect levels off until about {! = 2.75. This suggests that voters may be responding to a psychologically coherent sense Crime, policing and the perception of neighborhood safety 1 t.25 I5 FIGURE 3~. 1.75 ' 2.25 2'5 2.75 ' 3.25 3'5 3.75 4 4.25 4'5 4.75 5 Coefficient of VCRIME as a function of neighborhood radius in Model (PCRIME omitted). 1 1.2 1 0.8 0.6 0.4 0.2 0 1.25 1'5 FIGURE 3~. 1.75 4 3 2 1 2.25 2'5 3.25 3'5 2.75 3.75 5 4.25 4'5 4.75 Coefficient of PCRIME as a function of neighborhood radius in Model 2 @CRIME omitted). of ‘neighborhood’ at about p = 2.0-2.75 (a circle with radius roughly a mile), in addition to taking account of information about property crimes in a larger area around where they live. The stylized pattern research literature of coefficients on information observed flow here is interpretable and political paradox in the findings of Table 3 is the suggestion opinionization. that the psychological more severe crime resonates across a larger ‘neighborhood’. in light of the The apparent impact of the With property crimes, that is, the greater their frequency in contiguous and nearby surrounding areas, the greater the citizen’s support for expenditures on policing. With violent crime, on the other hand, greater frequency in the local area appears policing, while greater frequency to make little difference to support for in a relatively large region surrounding the individual’s residential tract does translate into more votes for anti-crime spending. We can see the explanation for this pattern by returning to the notion of the neighborhood context as a catchment area in the flow of information. It is a truism to remark that individuals shape their opinions from what they know of reality, and that if subjective perceptions and objective circumstances differ, then it is the perceptions that influence attitudes and actions. The truism carries special weight in our case, for it 535 JEFFGROCCERAND M. STEPHENWEATHERFORII appears highly different sources reporting likely that individuals than for property on different obtain information about violent crime, and the different transmission area1 units, paint systematically different crime from channels, by pictures of the spatial surround relevant to assessing personal safety. The dramatic nature of violent crime draws television coverage, crimes. Television and television is a key source of citizens’ information personalizes information, bringing about violent into the living room stories that carry the immediacy of local events. But television skews the presentation of crime in two ways: it presents as ‘local news’ events drawn from a catchment area much larger than the census tract or the immediate neighborhood, and especially perception of violent crime (Sheley and it exaggerates and Ashkins, the frequency 1981). Thus, of crime the subjective of violent crime, on which the individual’s vote is based, correlates with data on actual crime rates for a much larger urban environment than the individual’s census tract. Property crimes, on the other hand, are both less exciting and less unusual than violent crimes (especially coverage. in a large metropolis), Thus the bulk of information from local newspapers communities and they garner genuinely little television and personal contact. Dramatic property crimes from surrounding are by no means excluded from the individual’s attention (and they are likely to arrive at the individual’s consciousness alongside relatively about property crimes comes to the individual local events about via television), which but they take their place information comes from personal contacts. Differences in the dominant source of information help to explain why the correlation between crime is significant when the relevant context the impact of the property about different crime types, then, the vote and a measure of actual property is as small as the census tract, and why crime rate is especially pronounced for an area just a bit larger than a mile wide (a size that strikes us as intuitively consistent with the traditional image of a neighborhood). Local media and personal conversations information. The impact of the property crime rate for increasingly the relevant spatial context transmission flow-television steadily rising effect is, conversely, probably and metropolitan of the violent crime attributable newspapers-that rate on the vote transmit this large definitions of to the same sort of accounts as the spatial for the area is widened.9 Conclusion and directions for further work Our analysis set out to construct a social preference ordering over two qualitatively different types of crimes, and the results strongly suggest that citizens are much more sensitive to violent crime than to property crime, as shown by their willingness additional taxes for police protection. our ability to express Both theoretical this as an unambiguous and statistical challenges conclusion, to pay hamper but the results have strong intuitive appeal and substantial statistical support. The theoretical puzzle concerns a problem of long standing in social scientists’ research on contextual effects: how to provide a valid spatial definition of the concept of ‘neighborhood’. arises from the procedure presence of collinearity of aggregating between information The statistical problem from multiple area1 units, in the key variables. To bring empirical evidence to bear on defining the notion of ‘neighborhood,’ we have employed GIS data and analysis resources to investigate the way in which individuals Crime, policing and the perception of neighborhood safety 536 map their ‘psychological space’ depending on the practical problem they are confronting. We begin from the intuition that ‘neighborhood’ is not a constant construct. For instance, the relevant notion of neighborhood would differ depending on whether one needed to borrow a cup of sugar or form an opinion on alternative locations for a nuclear waste facility. We test this intuition in a structured comparison, focusing on the public’s reaction to different types of crime. Even though we hold constant the general subject-matter, crime and law enforcement, and the outcome variable, voting on a ballot proposition, nevertheless a subtle pattern of differences emerges between the area1 extent of ‘neighborhood’ that individuals use in evaluating violent crime versus property crime. The observed pattern is, moreover, consistent with what we know from media and information use studies of the flow of news in urban areas, and thus we conjecture that information flow conditions the area1 extent of individuals’ psychological conception of locality. We address the methodological puzzle by utilizing our GIS data to simulate a succession of neighborhoods of varying size, and then to track and compare systematically the effects of aggregation across units and collinearity between variables. Our analysis, reported in Appendix A, supports our decision to rely on separate regressions for each crime type in evaluating the substantive question of the public’s relative sensitivity to different crimes. These results lead to the interpretation that both types of crime are important determinants of the public’s demand for police protection, but that violent crimes carry the greater causal weight. In future work, we wish to build on the ordinal-level distinctions revealed by this research, to refine the model in two directions. One will pursue more extensive experimentation directed toward establishing empirical criteria for the proper specification of the construct of ‘neighborhood’ that citizens use implicitly to give psychological meaning to crime rates. The other will extend the analysis by investigating whether spatial dependence in the model disturbances may hinder efforts to obtain precise estimates of the crime rate coefficients. Acknowledgements An earlier version of this paper was presented at the NCGIA Conference Models of Political Behavior, gratefully acknowledged. Buffalo, New York, 22-25 October Rusty Dodson has provided excellent on Spatial and Contextual 1992. Support from NCGIA is research assistance. Thanks are also due to Dick Berk for providing the crime data and to Munroe Eagles for his constructive as discussant comments at the conference. Notes 1. In future work, we plan to use the 1980 and 1990 Census data to interpolate 1985 values of these variables for a sample of comparable areas, in order to assess the seriousness of this mismatch. 2. We note that different offense types within these two categories concern among voters. In particular, concerned, reported in spite of its relative infrequency. We therefore estimated models below in which we used the murder rate as the crime variable. qualitatively therefore, may well differ in the level of murder is a crime about which the public seems acutely very similar we present to the models which included all violent only the models which include the broader such as those These models were crimes. crime rates. To save space, JEFF GROGGER AND M. STEPHENWEATHERFORD 3. Note that we can centered place only assess the effect on the area of residence. of employment; unavailable, 4. Sprague detailed one would data matching tracts vote of crime also like to measure of residence in neighborhoods crime with tracts around of work one’s places are however. and Huckfeldt, significant progress Huckfeldt in a major data collection at depicting and Sprague, 5. The magnitude pooling on the referendum Ideally, 537 measures operation-that project and a series of insightful of information and influence articles, in the local make area (cf. rules out 1987, 1988). and statistical the two flows voters significance into of the difference an omnibus do not distinguish ‘crime’ between between variable, violent the coefficients even crime if the implication and property of this crime-were acceptable. 6. While in principle the centroid, one might the unweighted of accounting for crime wish to down-weight crime occurring will see, even use of the simple for serious of model additional complexity 7. The coefficient does is consistently H. The radius beyond not reach significant could actually occurring constructed the boundaries unweighted consideration add needless crimes rate variables variables selection criteria. conventional of the individual’s poses problems The choice and is, therefore, at a greater distance here demonstrate levels of statistical own tract. of interpretation of weighting left for future from the importance As we that call schemes would work. significance until p = 2.75 and after that, be larger, but the data do not permit us to estimate models with larger areas. 9. An alternative concerned explanation of the greater not only with crime their workplace) and residence. this alternative While for neighborhoods IO. This clustering the larger larger from that of violent around crime of census than importance neighborhoods hypothesis 3~, it does for areas tract about properties Central South parts References reflects these is that citizens are of the city as well (e.g. areas the increase why the impact the distinguishes of the metropolis Valley and on the West Side. parts as well as the area in the crime of property of coefficients crime differ\ one mile in diameter. Here it essentially Fernando encompass not explain in Los Angeles. Central neighborhoods but in other may explain segregation and of larger their homes, from relatively poor high degree of residential and working-class the more affluent districts areas in the in the San Crime, policing and the perception of neighborhood safety 538 Appendix A: The Problem of Collinearity among Crime Rates By regressing the vote on each type of crime separately, Models 1 and 2 produce a coherent and plausible picture of the process through which citizens translate their information about crime into political demands for more police protection. Both these specifications are, however, liable to the accusation that they omit a correlated, and substantively relevant, independent variable (i.e. PCFUME from Model I and VCRIME from Model 21, and hence that their coefficient estimates may be biased. Model A(1) addresses this possibility, by including both crime rate variables along with the control variables. The estimates from this specification, presented in Table Al, tell a more ambiguous story about the relationship between crime and the vote. Whether the data for the regression are tract- TABLEAl. Crime coefficients (x100), using different definitions of neighborhood Coefficient and standard error Neighborhood radius, in multiples of mean tract radius Model A(lj VCHME -0.084 0.104 0.011 0.133 -0.001 0.137 -0.104 0.144 -0.925 0.0215 -1.118 0.328 -0.954 0.352 -0.829 0.384 -1.021 0.518 -0.63 0.581 0.083 0.627 -0.03 7 0.655 0.245 0.691 -0.222 0.704 0.089 0.757 0.688 0.732 0.906 0.758 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 3.25 3.50 3.75 4.00 4.25 4.50 4.75 5.00 Note Coefficient Model A(1 j PCIUME is in italics above, standard error below. N 0.022 0.01 0.015 0.013 0.015 0.013 0.054 0.024 0.425 0.069 0.407 0.088 0.356 0.091 0.368 0.098 0.505 0.131 0.523 0.142 0.558 0.152 0.661 0.164 0.657 0.17 0.836 0.177 0.834 0.185 0.84 0.184 0.867 0.19 Model specifications match 571 552 532 504 481 456 442 429 406 389 376 365 351 343 324 313 290 those in Table 2. JEFF GROGGER ANDM. STEPHEN WEATHERFORD 539 level measures (see Table 2) or built indicators summed over successively larger areas (see Table .?J the impression is the same: the coefficients for PCRIME are generally significant and consistent with the estimates derived from Model 2, while the coefficients for VCRIME are unstable (though typically indistinguishable from zero) and vary widely from Model 1. We suspect that the explanation for these perplexing results is a combination of measurement error and multicollinearity, and in this Appendix we investigate the symptoms for that diagnosis. Collinearity, the problem of excessively strong correlation among independent variables, is difficult to diagnose and nearly impossible to solve. We know, for instance, that occurrences of violent crime and property crime are correlated, but intuitively the two types of crime appear to have such different impacts on their victims that concatenating them would misconstrue the causal process. Statistical tests of collinearity attempt to give some precision to ambiguities like this one. To check for collinearity, we ran the series of diagnostic tests designed by Belsley et al. (1980). In pursuing this lead, our use of GIS data allows us to identify the occurrence of collinearity with some precision, We present the clues in two steps: the first follows the conventional path of contextual analysis in placing sole reliance on data aggregated at the level of census tracts, the second focuses on data aggregated up to larger units that appear to be more typical of the spatial area people identify with the concept of neighborhood. In Table Al, the results for Model A(I), a conventional analysis of tract-level data on crime rates and voting, seem to show that citizens are willing to pay higher taxes in order to combat property crime, but that the effect of violent crime on voting for the measure is indistinguishably different from zero. This unexpected result makes it logical to look more closely at measurement and specification, In the census tract crime data (N = 727), VCRIME and PCRIME are correlated at r = 0.708, a strong but not clearly fatal relationship. Dimensional analysis of the 11 independent variables reveals two dominant principal components: one typified by high rates of violent crime, high expenditures on police, low income and high poverty rates; the other by high rates of property crime and low population density.‘” Among collinearity diagnostics, the highest condition number is only 14.9, the progression of condition numbers is quite smooth, and neither VCRIME nor PCRIME load highly on the offending component. In sum, although the coefficient for VCRIME when Model A(l) is estimated from tract-level data is unexpectedly paltry, its lack of robustness does not seem to be attributable to collinearity. We suspect that the small effect is due to a relatively weak signa-tonoise ratio for the VCRIME variable (attributable to measurement error), although we are also aware that the result may indicate that voters connect violent crime to political expression in a more complicated way than our model envisions. In future research we hope to investigate both possibilities. The more elaborate estimates of Model A(1) shed additional light on this problem. Given the relatively strong correlation between VCRIME and PCRIME in the smallest units of aggregation (Y = 0.708 in the data for census tracts), if the mechanics of aggregating tracts into larger units were to increase this correlation, they could undermine the efficiency of our estimates of causal effects. We pursue a collinearity analysis for the case of the neighborhood radius set at p = 3.0. This case was selected in part because it is intermediate in the succession of variations on the definition of neighborhood-as-census-tract, and in part because the estimated results contradict intuitive expectations. The diagnostics clearly show that estimation is hampered by correlations among key independent variables. In this reduced sample (N = 4061, VCRIME and PCRIME are correlated at K = 0.88. Principal components analysis of the independent variables reveals one dominant component, typified by high rates of both violent crime and property crime and by high poverty rates, and a weaker second component clustering high expenditures on police with low population density. Among collinearity diagnostics, the highest condition number is 25.8. approaching the range of strong and potentially harmful collinearity. This indicator is also nearly double the next largest, and both VCRIME (0.904) and PCRIME (0.939) load highly on the offending component. In short, the surprising estimates for this level of aggregation (SuhStdntial positive effects for property crime, large negative effects for violent crime) are no ground for inference about voters’ reasoning or the optimal choice of public policy: they are apparently the aherrant result of excessive correlation between the two types of crime in the data analyzed. 540 Crime, policing and the perception of neighborhood safety Given the potential theoretical value of obtaining empirically validated estimates of individuals’ conceptions of neighborhood, it is worth investigating further the sources of variation in the degree of collinearity. Aggregation could worsen the problem of collinearity in either of two ways. One is a sample composition effect, the other an aggregation effect, The sample composition effect is caused by dropping observations that border on the city boundary, a problem that parallels the loss of initial and concluding observations in time series analysis. In both cases, the number of observations lost is a function of lag length (in our case, ~1. Recall that the simulations presented in Model A(1) would be based ideally on the same number of centroids as there are census tracts, but that data for key variables are not available on contiguous tracts outside the city boundary. This means that, as the area1 units become successively larger, more tracts must be dropped from the analysis for lack of data. Given the irregular shape traced by the city boundary, this effect produces a noticeable drop-off in sample size as p increases. The upshot is to introduce a confounding source of imprecision into the analysis: depending on the pattern of exclusions, it might turn out that the composition of the resulting sample disproportionately represents tracts where VCRIME and PCRIME have above-average correlations, thus exacerbating the problem of collinearity. The aggregation effect is a direct result of the procedure of summing across census tracts, Given that the sample is most heterogeneous in the smallest units, summing across census tracts will result in more homogeneous units, in much the same way that averaging diminishes variance. In addition, however, larger neighborhoods overlap, while census tracts are mutually exclusive, and this reduces heterogeneity because the crime rates in different neighborhoods are computed on the basis of partially shared data. It is important to note that this effect is essentially the result of faithfully operationalizing the concept of neighborhood in a densely populated urban area: individuals living in one census tract are exposed to events in contiguous UXtS both through personal experience and through the flow of information, and it is natural to suppose that this exposure might result in a shared sense of concern about personal safety. In general, changes in sample composition accompany changes in the definition of neighborhood, so that these two possible sources of collinearity cannot ordinarily be disentangled. However, it is possible to hold one fixed and vary the other, and in this way to distinguish between the two with a straightforward test, Table A2 presents the results of a re-estimation of Model A(1) in which the sample composition is varied while the definition of neighborhood is held constant. That is, we specify the model in terms of tract-level crime rates (as in Model 3, Table 21, and then estimate the model for the various subsamples resulting from increasing p (as in Model A(l), Table Al). The coefficients are in every case quite similar to those presented in Table 2. In other words, it is not primarily changes in sample composition, but rather the process of aggregating to produce different definitions of the relevant crime rate for different definitions of neighborhood, that leads to the observed increase in the level of collinearity. TABLEA2. Regression l!J 5 3P 5p -0.169 (0.89) 0.023 (2.97) 715 733 -0.201 (1.01) 0.022 (2.75) 570 0.720 -0.204 (1.13) 0.023 (2.62) 481 0.699 -0.237 (1.07) 0.034 (2.76) 406 0.678 -0.127 (0.49) 0.025 (1.95) 290 0.684 PCRIME N Adjusted Z? (a) Absolute (b) It is the Cc) Regressions variable is LOGPYES Census tract VCRIME Notes. results: dependent t-statistics, radius based of the also on circle include the OLS with standard area control errors, in parentheses. equal to average variables mted tract in Tuhle area, 2. ahlut 730mL. JEFFGROGCERAND M. STEPHENWFNHERFORI) 541 We read this result as supporting our decision to focus on Models 1 and 2, which specil;i separately the effects of different crime types on the vote, in order to interpret the policy implications of the public’s aversion to violent crime versus property crime. That is, if the level of correlation between VCRIME and PCRIME is high enough to be pathologically collinear, then relatively little independent information is lost by estimating vote functions including only one of the crime variables at a time, as in Models 1 and 2. Thus, the diagnosis of the data problems hampering sound estimation of a combined mode1 strengthens the underpinning of our interpretation on the basis of separate estimates.