SPATIAL ANALYSES OF HOMICIDE WITH AREAL DATA *

Steven F. Messner

University at Albany, SUNY, and NCOVR

and

Luc Anselin

University of Illinois at Urbana-Champaign, and NCOVR

*Support for this research was provided by a grant from the National Consortium on
Violence Research (NCOVR). NCOVR is supported under grant # SBR 9513040 from
the National Science Foundation. Support was also provided by a grant from NSF to the
Center for Spatially Integrated Social Science (BCS-9978058) and by grants to the Center
for Social and Demographic Analysis (SUNY-Albany) from NICHD (P30 HD32041) and
NSF (SBR-9512290). Any opinions, findings, conclusions or recommendations
expressed herein are those of the authors and do not necessarily reflect the views of the
funding agencies.
 SPATIAL ANALYSES OF HOMICIDE WITH AREAL DATA

Abstract

This chapter highlights the ways in which the application of recently developed

techniques for spatial analysis contributes to our understanding of homicide. We begin

with a brief historical review of the role of geographic space in the sociological study of

crime and then discuss generic methodological issues involved in the study of areal units.

The logic of important techniques for spatial analysis is explained and illustrated using

two empirical case studies of variation in homicide rates across U. S. counties. One case

study involves the use of techniques of Exploratory Spatial Data Analysis (ESDA), and

the other applies spatial regression modeling. The analyses yield suggestive evidence of

diffusion processes and also reveal the incompleteness of well-accepted baseline models

of homicide rates. The chapter concludes with a brief discussion of pressing issues for

future research on the spatial dynamics of crime.

 SPATIAL ANALYSES OF HOMICIDE WITH AREAL DATA

I. Introduction

In a widely cited article published at the beginning of the 1990s, Land, McCall,

and Cohen noted an important development in sociological studies of crime over the

previous two decades. There had been a remarkable upsurge in the volume of empirical

studies seeking to explain why different areal units – e.g., cities, metropolitan areas, and

states – have high or low homicide rates (Land et al., 1990, p. 922). Much of this

research was inspired by theoretical debates surrounding a distinctively geographic issue:

the historically high homicide rates in the South (Hawley and Messner, 1989). However,

with the exception of this focus on regional differences in overall levels of homicide, the

early work on the social structural covariates of homicide rates was largely insensitive to

spatial context. The field has changed dramatically in recent years, and criminologists

are increasingly applying formal tools of spatial analysis to describe and explain

variations in levels of homicide (and other crimes).

The purpose of the present chapter is to highlight the ways in which the

application of newly developed techniques for spatial analysis contributes to our

understanding of homicide. We begin with a brief historical review of the role of

geographic space in the sociological study of crime. This is followed by a discussion of

generic methodological issues involved in the study of areal units. We then explain the

logic of important techniques for spatial analysis, and illustrate their use in two empirical

case studies dealing with the analysis of variation in homicide rates across U. S. counties.

The chapter concludes with a brief discussion of pressing issues for future research.
 2

II. Background

An interest in the geographic dimension of crime can be traced to the very

beginnings of criminology as a scientific discipline. In the mid-19th century, the so-

called moral statisticians, Quetelet and Guery, used the newly published criminal

statistics in France to search for social forces that might underlie criminal behavior

(Radzinowicz, 1966, pp. 29-38). They constructed maps of crime totals for different

regions in France and observed considerable stability over time. These stable patterns

suggested to them that levels of crime are to some extent a reflection of the larger

environment (social and physical). They proceeded to examine other features of the

respective regions of France, searching for covariates that could account for the

geographic patterning of crime.

In the early years of the 20th century, members of the Chicago School of

sociology embraced and extended the geographic approach pioneered by Quetelet and

Guerry. The human ecology perspective of Burgess, Park, and McKenzie (1925) focused

attention squarely on the spatial context for social activities, and the empirical research of

Shaw and McKay (1931; 1942) documented striking similarities in the distribution of

officially recorded rates of juvenile delinquency across neighborhoods and the

distribution of various indicators of housing structure, economic status, and population

composition (Vold et al., 1998, pp. 144-145). The common thread underlying this early

work in the sociological study of crime was the premise that social phenomena, including

criminal behavior, are not randomly distributed across the physical terrain, and as a

result, inquiry into the spatial patterning of these phenomena can yield unique insight into

their causal dynamics.

 3

Interest in the geographic dimension of crime waned in the middle years of the

20th century. As Coleman (1986) explains, major innovations in survey methodology

fostered an “empirical behaviorism” in the field of sociology at large. This had two

important implications for sociological studies of crime. One, the analytic focus in

empirical research tended to shift away from social groups and territorially based social

aggregates (e.g., neighborhoods, cities) to individuals (Bursik and Grasmick, 1993, p. ix).

Rather than studying variation in crime rates across areal units, researchers increasingly

concentrated on explaining differences in individual involvement in criminal or

delinquent activities.

The survey paradigm also encouraged researchers to conceptualize their units of

analysis as independent observations, consistent with the survey model of random

sampling from a population (Coleman, 1986, p. 1316). This conceptualization extended

beyond studies based on actual surveys of individuals to social structural analyses

employing macro-level units. Thus, the upsurge in areal studies of homicide rates that

appeared in the latter decades of the 20th century (referred to above and reviewed by Land

et al., 1990; see also Parker et al., 1999) reflected an accumulation of studies based

almost exclusively on analytic and statistical methods that assume independence of

observations, such as regression models estimated by Ordinary Least Squares (OLS).

Independent observations are just that – they are unconnected in meaningful ways,

including that of location in geographic space.

In recent years, criminologists have rediscovered the importance of geographic

information in the study of crime. This is driven primarily by substantive motivations as

reflected in influential place-based theories of crime such as routine activities theory

 4

(Cohen and Felson, 1979; Felson, 1998) and crime pattern theory (Brantinham and

Brantingham, 1993; see also Anselin et al., 2000). Similarly, geographic space plays a

central role in the burgeoning research on crime “hot spots” (Roncek and Maier, 1991;

Sherman et al., 1989; Sherman and Weisburd, 1995), the diffusion of violence (Cohen

and Tita, 1999; Messner et al., 1999; Morenoff and Sampson, 1997; Rosenfeld et al.,

1999; Smith et al., 2000), and neighborhood collective efficacy (Sampson et al., 1997).

The non-spatial modeling characteristic of earlier macro-level studies of crime has also

been increasingly replaced by approaches that are keenly sensitive to spatial dynamics

(e.g., Baller et al., 2001; Morenoff et al., 2001). This resurgence of interest in space has

clearly been facilitated by the increased availability of georeferenced information on

crime events and the dissemination of methods of spatial analysis to the practice of

empirical research in criminology. Thus, at the turn of the 21st century, geographic space

is returning to the forefront of criminological inquiry (Abbot, 1997).

III. Generic Methodological Issues in the Analysis of Areal Data

Areal analyses in criminology are typically based on data aggregated for

administrative spatial units of observation. The use of such units raises a number of

fundamental methodological concerns. We will limit our focus here to three that are

particularly important for spatial analysis: the relevance of scale and the associated

problem of ecological inference; the intrinsic heterogeneity of rates; and the substantive

interpretation of spatial autocorrelation in models estimated for aggregate areal units of

observation.
 5

Ecological inference and the associated issue of the ecological fallacy pertain to

problems that may occur when phenomena observed at an aggregate level are explained

in terms of an individual causal mechanism (King, 1997). Unless rather extreme

conditions of homogeneity are satisfied, such inference is misleading or spurious.

Aggregate analyses, such as the study of homicide rates at the county or census tract

level, are sometimes dismissed as naturally suffering from the ecological fallacy problem.

Indeed, in a naïve interpretation of aggregate results, this may well be the case. However,

there are many research questions (especially those with a public policy implication) that

are legitimately carried out at the scale of the administrative units to which the policies

will pertain. Specifically, when dealing with homicide rates, the main interest often lies in

explaining patterns and correlates for underlying risk. This risk pertains to a population at

risk which can legitimately be taken as the population of an areal unit of observation.

Accordingly, it is instructive to relate the levels of risk to other characteristics of

populations. However, extending such a model to the interpretation of the causal

dynamics underlying individual behavior would be spurious. We argue that in

sociological studies of criminal violence, where the interest lies in macro conditions or

context, an areal (or ecological) perspective using administrative units is legitimate.

Moreover, it allows for a much richer empirical context in terms of the range of

information available on covariates.

Scale is also important due to the potential mismatch between the scale and

spatial extent of units of observation for the data and the scale and spatial extent of the

process at hand. This mismatch will tend to result in a statistical problem wherein error
 6

structures are spatially correlated and/or heteroskedastic. Fortunately, this statistical

problem can be addressed by applying proper spatial econometric methods.

A second methodological issue encountered in the use of areal data is the intrinsic

heterogeneity of rates computed for varying populations at risk. Unless the areal units all

have the same population, the variance of the rate (as an estimate for the parameter in an

underlying binomial random variable) will not be constant, but instead is inversely

related to the size of the population (larger populations yield more precise estimates).

This is mostly a problem for “small area” estimation and is especially pronounced when

the risk pertains to relatively rare events, such as homicides in rural counties. There are

three important implications of this variance instability for the spatial analysis of

homicide rates. First, the visualization of rates for areal units such as counties or census

tracts, unless properly corrected (or smoothed), may yield spurious outliers. More

precisely, since the estimates of underlying risk in areas with sparse populations will have

a much greater variance, “spikes” are likely to be observed, even though the areas are

subject to the same risk as other, more densely populated ones. This will affect the

interpretation of results in an exploratory spatial data analysis (ESDA), as outlined in the

next section. Secondly, the variance instability may yield spurious indications of

significant spatial autocorrelation when using traditional statistics. All familiar test

statistics (such as Moran’s I and Geary’s c) are based on an assumption of spatial

stationarity, which includes a requirement for the variance of the process to be constant.

The potential of extreme variance instability for rates of small areas violates this

assumption, again possibly yielding misleading inference. Finally, the non-constant

variance will be intrinsic to the error term in any regression model (aside from other
 7

sources of heteroskedasticity), and inference must be properly adjusted to take this into

account.

A third issue pertains to the substantive interpretation of empirical results that

suggest the presence of spatial autocorrelation in areal data, especially when the data are

limited to a single cross-section. In this situation, it is impossible to distinguish true

contagion from apparent contagion without additional information (such as space-time

data). Any spatial cluster of similar values may be the result of a process of spillover,

contagion, or spatial externalities, or, instead, follow from structural differences due to an

intervening (unobserved) variable specific to the locations in question. It is important to

keep this limitation in mind when interpreting evidence of spatial autocorrelation.

Similarly, spatial autocorrelation can be embedded in regression models in a variety of

ways, each of which imply different types of spatial externalities. A common

characteristic of these externalities is that even though they are modeled as spatial

relations between the dependent variable of the model (e.g., homicide rates), they

ultimately result from spatial relations among the explanatory variables and/or the error

terms in the model (Anselin, 2003). Caution must be used in cross-sectional settings not

to view the models as conditional upon observed neighbors. This easily leads to the trap

of ecological fallacy. It also suggests an interpretation that implies the availability of

observations over time as well as across space. However, in a pure cross-sectional setting,

this is not the case. Instead, equilibrium must be assumed, and the nature of the dynamic

processes (such as contagion or diffusion) that yielded the equilibrium cannot be inferred

without further information.

 8

IV. Spatial Analyses of U. S. County Homicide Rates

Spatial analyses are potentially useful for addressing a variety of substantive

issues in criminology. For present purposes, we focus on two questions that have been of

keen interest in the homicide literature. The first deals with the potential impact of

homicide in one place on the likelihood of homicide in nearby locations, i.e., is the spatial

patterning of homicide consistent with some kind of diffusion process? Any

phenomenon that is generated by diffusion should exhibit a distinctive geographic pattern

that is non-random: higher incidences of the phenomenon should be observed near the

location where the initial incidents occurred (the point of origin). The theoretical

possibility that criminal violence might spread in such a manner from one place to

another has been raised prominently in the public health literature (Hollinger et al., 1987;

Kellerman, 1996). Loftin, for example, proposes that criminal assaults can be usefully

regarded as “analogous to disease,” capable of “contagious transmission” (1986, p. 550).

With the aide of formal techniques of spatial analysis, it is possible to search for spatial

imprints of homicide that are suggestive of diffusion or contagion.

A second important question in the literature concerns the possible interaction of

structural factors with features of the geographic context. Are the structural determinants

of homicide rates invariant across space (cf. Land et al., 1990), or do structural conditions

exert differing effects on homicide levels in various sub-regions of the geography? The

tools of spatial analysis facilitate the identification of spatial regimes that are likely to

exhibit distinct causal processes. These tools also provide formal tests of differences in

the effects of predictor variables across regimes.

 9

We illustrate the application of spatial analysis to address these substantive

questions by referring to the results from two recent studies of county-level variation in

homicide rates.1 The first entails an exploratory spatial data analysis (ESDA) applied to a

case study of homicides in the St. Louis region. The second study applies the techniques

of spatial econometrics to a model with structural covariates for homicides, based on data

for all counties in the continental U. S.

The Diffusion of Homicide in the St. Louis Region: An Application of ESDA

Despite widespread speculation about possible diffusion processes in homicide,

our understanding of the nature of such processes is primitive at best. Given such limited

knowledge, an exploratory data analysis (EDA) approach is an extremely valuable

starting point for systematic inquiry.

EDA consists of descriptive and graphical statistical tools intended to discover

patterns in data and suggest hypotheses by imposing as little prior structure as possible

(Tukey, 1977). Contemporary EDA methods emphasize the interaction between human

cognition and computation in the form of dynamically linked statistical graphics that

allow the user to manipulate directly various views of the data (e.g. Cleveland, 1993;

Buja et al., 1996). ESDA extends standard EDA by focusing on techniques to describe

and visualize spatial distributions, identify atypical locations or spatial outliers, discover

patterns of spatial association, clusters or hot spots, and suggest spatial regimes or other

forms of spatial heterogeneity (Anselin, 1994, 1998, 1999a; Haining 1990; Bailey and

Gatrell, 1995).

1
These studies are reported in Messner et al. (1999) and Baller et al. (2001). The discussion below draws
upon these sources.
 10

ESDA is especially useful in the study of possible diffusion processes because it

has at its core a formal treatment of the notion of spatial autocorrelation, i.e., the

phenomenon where locational similarity (observations in spatial proximity) is matched by

value similarity (attribute correlation) (see Cliff and Ord, 1981, and Upton and Fingleton,

1985, for extensive treatments). The particular ESDA techniques used for our

illustrations focus on the detection of local patterns of spatial autocorrelation through the

implementation of so-called LISA statistics (Anselin, 1995) in dynamically linked

graphic windows that visualize the location, magnitude and pattern in the data, such as

box maps, Moran scatterplots and LISA maps (Anselin, 1996, Anselin and Smirnov,

1999, Anselin et al., 2002).

The areal units for our case-study investigation of diffusion processes are counties

in the St. Louis Metropolitan Statistical Area (MSA) and additional counties within three

layers of adjacency to the MSA.2 For each of the counties in the region, rates were based

on homicide counts aggregated by the decendents’ county of residence (per 100,000). To

compensate for the rate instability in areas with small populations, we smoothed the data

by computing averages for two periods, 1984-88 and 1988-93. The earlier period is one

of relative stability in homicide, while the latter is a period of generally increasing

homicides.

We illustrate two distinctive aspects of ESDA. First, we consider selected maps

and graphical statistics for the two time periods under investigation to bring out overall

trends and to identify possible outliers (see Figure 1).3 The specific graphics are: (1) Box

Maps which show the location (quartile) of every county within the overall distribution of

2
See Messner et al. (1999) for a detailed description of data sources and variable definitions.
 11

homicide rates for the period, and (2) Box Plots which show graphically the variation of

homicide rates (see Anselin, 1999a). In both the maps and plots outlier counties are

identified.

-----

Figure 1 about here

-----

Visual inspection of the Box Map for the period of stable homicide rates reveals

that the counties tend to fall into two general regimes, defined by a diagonal line running

from the upper left of the region to the lower right. Those counties falling below the

diagonal show generally higher homicide rates (third and fourth quartiles). Between

1984 and 1988 only two areas had homicide rates so extremely high that they can be

considered outliers — St. Louis City and St. Clair County, IL. These basic patterns in the

geographic distribution of homicide might be viewed as evidence of a spatial clustering

of homicide rates; however, as will be demonstrated below, that conclusion would be

premature.

The geographic distribution of homicide appears to have shifted between the

periods of stable and increasing homicide. The “southwest vs. northeast” division

deteriorates somewhat over time, with more high homicide counties becoming evident in

the northeast quadrant. Indeed, an especially salient characteristic of the Box Map for the

latter (increasing) period is the emergence of four additional outliers – Reynolds County,

MO; Bond County, IL; Macon County, IL; and Cumberland County, IL. However, the

substantive importance of these four new outliers should be interpreted cautiously. Their

3
Note that in a software implementation of ESDA, these graphs would be linked in real-time, to allow for
interaction with the various views of the data, see Anselin et al. (2002).
 12

homicide rates are far more modest than those for the two outliers that appear in both Box

Maps, and closer examination reveals that three of the new outliers (all except Macon

County, IL) may have achieved that status because of variance instability.

The Box Maps and Box Plots are useful for describing the general characteristics

of the distribution of homicide throughout the 78-county area under study, and for

revealing specific areas with exceptionally high levels of homicide. However, they are

limited in their ability to identify any significant spatial clustering of homicide rates. To

take into account the spatial arrangement of the homicide values, we make use of

measures of spatial autocorrelation, and specifically local spatial autocorrelation.

Local indicators of spatial association (or LISA, an acronym coined in Anselin,

1995) assess a null hypothesis of spatial randomness by comparing the values in each

specific location with values in neighboring locations. Several LISA statistics can be

considered, but a local version of Moran’s I is particularly useful in that it allows for the

decomposition of the pattern of spatial association into four categories, corresponding

with four quadrants in the Moran Scatterplot (Anselin, 1996). Two of these categories

imply positive spatial association, namely, when an above-average value in a location is

surrounded by neighbors whose values are above average (high-high) or when a below-

average value is surrounded by neighbors with below average values (low-low). By

contrast, negative spatial association is implied when a high (above average) value is

surrounded by low neighbors and vice versa. Both of these instances are labeled spatial

outliers when the matching LISA statistics are significant. Each of the quadrants matches

a different color in the so-called LISA Map, a map that shows both the locations with
 13

significant LISA statistics (i.e., a rejection of the null hypothesis of spatial randomness)

as well as the category of spatial association.

The classification into four categories of spatial association is illustrated by the

Maps in Figure 2. They match a Moran scatterplot (Figure 3), where the horizontal axis

is expressed in standard deviational units for the homicide rate. The vertical axis

represents the standardized spatial weighted average (average of the neighbors) for the

homicide rates.4 The slope of the linear regression through the scatterplot is the Moran’s

I coefficient, as shown in Anselin (1996). This allows for an easy interpretation of

changes in global spatial association (the slope) as well as a focus on local spatial

association (the quadrant). The Scatterplot presented in Figure 3 also report the Global

Moran’s I statistic described above as a measure of spatial autocorrelation among all

counties in the analysis.

-----

Figure 2 about here

-----

Beginning with the LISA Map for the period of stable homicide, we find evidence

of spatial grouping. A cluster of counties with high homicide rates, as well as neighbors

with high homicide rates, is apparent in the area around St. Louis City. This urban core

of high homicide is also implicated in the three surrounding counties with low homicide

rates, but high-homicide neighbors. These are suburban residential counties near the city

of St. Louis. In addition, two clusters of low-homicide counties, surrounded by other

counties with low homicide rates, can be seen in the northern and eastern fringes of the
 14

region. Consistent with a possible diffusion process, a distinct “hot spot” appears in the

data centered on the St. Louis urban core, while “cool spots” are also detected in areas

geographically removed from St. Louis.

Turning to the subsequent period during which homicide rates increased

throughout the region, the same general profile of spatial clustering appears. However,

one notable difference emerges during the later period when both Sangamon and Morgan

Counties made the transition to high-homicide counties, surrounded by low- homicide

counties. In conjunction with Macon County, they now formed a string of counties (see

the northern section of the map), suggesting the possible east-to-west diffusion of

homicide out of Macon County, as homicide in general increased throughout the region.

A final piece of evidence about the spatial clustering of homicide can be gleaned

from the Moran Scatterplots. The overall clustering suggested by significant (global)

Moran’s I statistics is not due simply to the disproportionate influence of the two,

neighboring, counties with the highest homicide rates in the region – St. Louis City and

St. Clair County. In fact, the degree of spatial clustering increases for both time periods

when those two counties are omitted from the analysis, as shown in Figure 3 for the later

period (for further details, see Messner et al. 1999).

-----

Figure 3 about here

-----

The application of ESDA to the St. Louis region provides suggestive evidence of

a diffusion process for homicide. The static comparisons of clustering in the two time

4
The spatial correlation statistics were based on a spatial weights matrix that labeled counties as neighbors
when their centroids (centers of gravity) were within 31.7 miles from each other, for details, see Messner et
 15

periods reveal a spatial imprint to be expected in the presence of past diffusion, and a

dynamic comparison across time periods indicates changes consistent with a westward

spread of homicides in a selected portion of the geography. We emphasize, however, that

ESDA is a starting point for analysis – it is explicitly exploratory. It permits rejection of

the null hypothesis of spatial randomness, but processes other than diffusion can produce

spatial non-randomness. To illustrate the application of tools of spatial analysis to detect

some of these processes, we turn to our study based on county-level data for the U.S. as a

whole.

Modeling Homicide Rates Across the U. S.: An Application of Spatial Econometrics

As noted at the beginning of this chapter, much of the initial interest in areal

studies of homicide in the U.S. focused on the issue of regional differences and, more

specifically, the high homicide rate in the South. A large number of quantitative studies

were conducted to determine whether the Southern effect on homicide rates persists after

various social structural variables are taken into account in multivariate models.

Although the results of this research are inconsistent (for a review, see Parker et al.,

1999), it is hard to imagine that any serious quantitative study of homicide based on areal

units for the nation at large would not take into account potential regional differences.

This interest in the geography of homicide has been largely limited to the level

effect of region, i.e., whether homicide rates are higher in the South, even net of relevant

control variables. Much less attention has been devoted to the possibility that geographic

context might condition or moderate the effect of predictor variables (for an exception,

see Messner, 1983). The tools of spatial econometric modeling are especially well suited

to address such an issue.

al (1999).
 16

To explore the possibility of differential effects of homicide predictors across

regions, we examined homicide rates calculated for each of the US counties, smoothed

over the three-year period centered on the Census years of 1960 to 1990 (i.e.,

observations for four time points). Our selection of independent variables was guided by

the seminal work of Land et al. (1990). These variables include measures of resource

deprivation/affluence, population structure, median age, unemployment, the percent

divorced, and a dummy variable indicating Southern location (based on Census

definitions). The measures of resource deprivation/affluence and population structure

were constructed as composite indexes derived from principal components analysis.5

Before turning to spatial econometric models, consider a conventional, non-

spatial model of homicide for the sample of U.S. counties. Table 1 presents the results

of an OLS regression of homicide rates on the baseline model of predictors.

Similar to the findings in Land et al., resource deprivation/affluence, population structure,

and divorce are positively related to homicide rates. The negative effect of median age is

also unsurprising, indicating that counties with younger populations have higher

homicide rates. The negative coefficient for the percent unemployed is perhaps

counterintuitive but consistent with the results of Land et al. (1990). They suggest that

unemployment indicates reduced opportunity for violence (less social activity) once

resource deprivation is controlled for (see also Cantor and Land, 1985; Land et al., 1995).

The significantly positive coefficient for the dummy variable for South reveals the often

observed level effect of region – comparatively high homicide rates in the South. In

general, these non-spatial results for counties mirror those found in empirical studies for

other common areal units in the U. S. (cities, MSAs, and states).

5
See Baller et al. (2001), for a detailed description of data definitions and data sources.
 17

-----

Table 1 about here

-----

To assess the extent of spatial effects, we carried out the usual spatial econometric

battery of diagnostics for heteroskedasticity, spatial regimes as well as spatial lag and

spatial error dependence (Anselin 1988).6 Both heteroskedasticity as well as spatial

dependence were significantly present in the residuals of the OLS model.

First, consider whether there is evidence to go beyond the simple level effect for

the South. A test for structural stability of the regression coefficients across regions (a

spatial Chow test; see Anselin, 1990) permits a formal assessment of this (see Table 2).

The null hypothesis of coefficient stability is clearly rejected, suggesting that the

assumption of an identical (or stable) pattern of effects across regions is implausible.

Moreover, an examination of the tests of individual coefficients reveals that several of the

predictor variables exhibit significantly different effects in the South in comparison with

the non-South. These results clearly run counter to the assumption that the same causal

processes operate throughout the geography under investigation, an assumption that is

implicit in non-spatial OLS analysis. Our analyses thus reveal not only that the South

exhibits comparatively high homicide rates, even adjusting for social structural variables;

in addition, the social structural variables commonly used in research affect homicide

rates differently in the South than in other regions.

-----

Table 2 about here

-----
 18

Furthermore, the estimates also indicate a larger residual variance for the model in

the Southern counties, suggesting a poorer fit in this region. All of this suggests that the

invariance of the Land et al. (1990) baseline model of homicide may have been

overstated. In addition, these results highlight the inadequacy of reducing spatial

heterogeneity to a dummy variable for the South and the need to model regional variation

in the effects of covariates explicitly.

Given the strong evidence of distinct spatial regimes in the South and non-South,

we pursued a disaggregated modeling strategy by estimating separate models for each of

the regions, examining the residuals for possible spatial effects and implementing a

spatial regression model where appropriate. The spatial regression model takes the form

of either a spatial lag model or a spatial error model. A spatial lag model implies that the

geographic clustering of homicide is due to the influence of homicide in one place on

homicide in another. This model is consistent with some kind of diffusion process. A

spatial error model indicates that clustering reflects the influence of unmeasured

variables. For the South, there was strong evidence of the need for a spatial lag

specification in each of the years considered. In contrast, the results for the non-South

suggested a lag model in 1960, but a spatial error model in the subsequent years.7 Tables

3 and 4 summarize results for these spatial models in the Southern and non-Southern

counties respectively.

Beginning with the results for the South (Table 3), the signs of the coefficients for

structural covariates are generally consistent with those observed in non-spatial analyses

for the full sample of counties. However, there are interesting changes in magnitudes

6
All computations were carried out by means of the SpaceStat software package (Anselin 1999b).
7
For detailed results, see Baller et al (2001).
 19

(and significance) over time. The resource deprivation component is positively related to

homicide rates throughout the period, but the strength of the effect steadily increases over

time. The population structure variable exhibits non-significant effects in 1960 and 1970.

It is only in the latter years (1980 and 90) that the expected positive effects emerge.

Divorce rates are significantly related to Southern homicide rates throughout the period

but the effect is noticeably weaker in 1990. Unemployment is negatively related to

homicide rates in all years except 1960, while median age exhibits significantly negative

effects sporadically.

-----

Table 3 about here

-----

Table 3 also indicates that the effects of the Southern spatial lags of homicide are

positive and statistically significant in all time periods. These findings support the claim

that homicides in Southern counties influence homicides in other counties, consistent

with a diffusion interpretation. Note also, however, that the effects of the spatial lags

generally weaken overtime. An examination of the betas indicates that the spatial lags

are the strongest predictors of Southern homicide in 1960 and 1970 but are eclipsed by

the structural predictors in 1980 and 1990.

Turning to the non-South (Table 4), the results for the structural covariates are

quite similar to those for all counties in the non-spatial analyses, with the exception of the

unemployment variable. Resource deprivation, population structure, and divorce exhibit

significantly positive effects on homicide rates, while median age yields significantly
 20

negative effects. The only significant effect for unemployment is in 1970 and it is

positive, contrary to the general pattern.

-----

Table 4 about here

-----

With respect to spatial dependence, in every year except for 1960 the spatial error

model provides a better fit in the non-South than does the spatial lag model.

Substantively, this implies that, for the most part, the residual spatial autocorrelation in

the non-South can be adequately accounted for in terms of unmeasured predictor

variables. A diffusion process thus seems unlikely in non-Southern counties over recent

decades.

In sum, our application of techniques of spatial modeling in the case study of U.S.

counties demonstrated striking spatial patterns of homicide. Homicide researchers should

attend to these patterns for at least two very important reasons: spatial dependence needs

to be modeled properly to estimate the effects of non-spatial variables, and spatial

dependence directs attention to potentially interesting substantive processes, such as

diffusion.

V. Conclusion

The two case studies reviewed above illustrate the utility of exploiting an explicit

spatial perspective in the analysis of homicide. Apart from purely methodological reasons

to pursue such an approach, such as the inefficiency and biases in coefficient estimates

that may result when spatial effects are ignored, substantive insights were gained as well.
 21

Importantly, the explicit spatial approach allowed us to identify the

incompleteness of well-accepted baseline models. It provided a way to move beyond

simple dummy variable proxies for geography and permitted the assessment of the degree

of regional heterogeneity in much greater detail. This points to the need for extending the

baseline model of homicide with variables that can capture the suggested regional effects.

A spatial approach also allows one to shed some light on the presence of potential

diffusion processes. Most importantly, the distinction between different spatial

econometric specifications provides grounds to dismiss diffusion, or, alternatively, to

identify the types of spatial externalities or spillovers that may be instrumental in

generating the observed spatial patterns.

Our approach is only the beginning, however. Powerful models of the dynamics

of the space-time evolution of homicide and other violent crime still largely remain to be

formulated. Recently developed methods for the spatial econometrics of panel data are

very promising in this respect. The procedures for identifying clusters, outliers, and

spatial regimes discussed in this chapter are only initial steps in the understanding of

these patterns. Richer models and additional specifications need to be considered, and

the underlying vectors of transmission yielding the spatial patterns or outliers need to be

identified. Successful efforts along these lines are likely to require more sophisticated

theorizing about the determinants of levels of homicide risk across varying populations.

Much remains to be done, but the current methods of exploratory spatial data

analysis and spatial econometrics provide a solid base for further inquiry. It is our hope

that the examples provided here will stimulate future work of both an inductive and
 22

deductive nature to enhance our understanding of homicide and to promote a spatially

informed criminology more generally.

 23

REFERENCES

Abbot, A. 1997. “Of Time and Space: The Contemporary Relevance of the Chicago

School,” Social Forces 75: 1149-1182.

Anselin, L. 1988. Spatial Econometrics: Methods and Models. Boston: Kluwer

Academic.

Anselin, L. 1990. “Spatial Dependence and Spatial Structural Instability in Applied

Regression Analysis.” Journal of Regional Science 30:185-207.

Anselin L. 1994. “Exploratory Spatial Data Analysis and Geographic Information

Systems.” Pp. 45-54 in New Tools for Spatial Analysis, edited by M. Painho.

Luxembourg: EuroStat.

Anselin, L. 1995. “Local Indicators of Spatial Association – LISA.” Geographical

Analysis 27:93-115.

Anselin, L. 1996. “The Moran Scatterplot as an ESDA Tool to Assess Local Instability in

Spatial Association.” Pp. 111-125 in Spatial Analytical Perspectives on GIS in

Environmental and Socio-Economic Sciences. London: Taylor and Francis.

Anselin, L. 1998. “Exploratory Spatial Data Analysis in a Geocomputational

Environment.” Pp. 77-94 in Geocomputation, A Primer, edited by P.A. Longley,

S. Brooks, B. Macmillan and R. McDonnell. New York: John Wiley.

Anselin, L. 1999a. “Interactive Techniques and Exploratory Spatial Data Analysis.” Pp.

251-264 in Geographic Information Systems: Principles, Techniques,

Management and Applications, edited by P.A. Longley, M.F. Goodchild, D.J.

Maguire and D.W. Rhind. New York: John Wiley.

 24

Anselin, L. 1999b. SpaceStat Software Program for Spatial Data Analysis, Version 1.90.

Ann Arbor, MI: BioMedware Inc.

Anselin, L. 2003. “Spatial Externalities, Spatial Multipliers and Spatial Econometrics.”

International Regional Science Review (forthcoming).

Anselin, L., J. Cohen, D. Cook, W. Gorr, and G. Tita. 2000. “Spatial Analyses of Crime.” Pp.

213-262 in Criminal Justice 2000, Volume 4, Measurement and Analysis of Crime and

Justice, edited by David Duffee. Washington, DC: National Institute of Justice.

Anselin, L. and O. Smirnov. 1999. The DynESDA Extension for ArcView 3.0. Bruton Center,

University of Texas at Dallas. Richardson, TX.

Anselin, L., I. Syabri, O. Smirnov and Y. Ren. 2002. “Visualizing Spatial Autocorrelation with

Dynamically Linked Windows.” Computing Science and Statistics 33 (forthcoming).

Bailey, T.C. and A.C. Gatrell. 1995. Interactive Spatial Data Analysis. New York: John Wiley.

Baller, R. D., L. Anselin, S. F. Messner, G. Deane, and D. F. Hawkins. 2001. “Structural

Covariates of U. S. County Homicide Rates: Incorporating Spatial Effects.” Criminology

39:201-232.

Brantingham, P. L. and P. J. Brantingham. 1993. “Environment, Routine and Situation: Toward

A Pattern Theory of Crime.” Pp. 259-294 in Routine Activity and Rational Choice:

Advances in Criminological Theory, 5, edited by R. V. Clarke and M. Felson. New

Brunswick, NJ: Transaction Books.

Buja, A., D. Cook and D. Swayne. 1996. “Interactive High Dimensional Data Visualization.”

Journal of Computational and Graphical Statistics 5:78-99.

Burgess, E., R. E. Park, and R. D. McKenzie. 1925. The City. Chicago: University of

Chicago Press.
 25

Bursik, R. J., Jr. and H. G. Grasmick. 1993. Neighborhoods and Crime: The Dimensions

of Effective Community Control. New York: Lexington Books.

Cantor, D. and K. C. Land. 1985. “Unemployment and Crime Rates in the Post- World

War II United States: A Theoretical and Empirical Analysis.” American

Sociological Review 50:317-332.

Cleveland, W. S. 1993. Visualizing Data. Summit, NJ: Hobart Press.

Cliff, A. and J.K. Ord. 1981. Spatial Processes: Models and Applications. London: Pion.

Cohen, J. and G. Tita. 1999. “Diffusion in Homicide: Exploring a General Model for

Detecting Spatial Diffusion Processes.” Journal of Quantitative Criminology

15:451-493.

Cohen, L. E. and M. Felson. 1979. “Social Change and Crime Rate Trends: A Routine

Activities Approach.” American Sociological Review 44:588-608.

Coleman, J. S. 1986. “Social Theory, Social Research, and a Theory of Action.”

American Journal of Sociology 91:1309-1335.

Felson, M. 1998. Crime and Everyday Life, 2nd Edition. Thousand Oaks, CA: Pine

Forge Press.

Haining, R. F. 1990. Spatial Data Analysis in the Social and Environmental Sciences.

Cambridge: Cambridge University Press.

Hawley, F. F. and S. F. Messner. 1989. “The Southern Violence Construct: A Review of

Arguments, Evidence, and the Normative Context.” Justice Quarterly 6:481-511.

Hollinger, P. C., D. Offer, and E. Ostrov. 1987. “An Epidemiologic Study of Violent Death,

Population Changes, and the Potential for Prediction.” American Journal of Psychiatry

144:215-219.
 26

Kellerman, A. 1996. Understanding and Preventing Violence: A Public Health Perspective.

National Institute of Justice Review. Washington, DC: US Government Printing Office.

King, G. 1997. A Solution to the Ecological Inference Problem. Princeton: Princeton

University Press.

Land, K. C., D. Cantor, and S. T. Russell. 1995. “Unemployment and Crime Rate Fluctuations

in the Post-World War II United States: Statistical Time-Series Properties and Alternative

Models.” Pp. 55-79 in Crime and Inequality, edited by J. Hagan and R. D. Peterson.

Stanford, CA: Stanford University Press.

Land, K. C., P. L. McCall, and L. E. Cohen. 1990. “Structural Covariates of Homicides

Rates: Are There any Invariances Across Time and American Journal of

Sociology 96:1441-1463.

Loftin, C. 1986. “Assaultive Violence as a Contagious Process.” Bulletin of New York

Academy of Medicine 62:550-555.

Messner, S. F. 1983. “Regional Differences in the Economic Correlates of the Urban

Homicide Rate: Some Evidence on the Importance of the Cultural Context.”

Criminology 21:477-488.

Messner, S. F., L. Anselin, R. D. Baller, D. F. Hawkins, G. Deane and S. E. Tolnay. 1999. “The

Spatial Patterning of County Homicide Rates: An Application of Exploratory Spatial

Data Analysis.” Journal of Quantitative Criminology 15: 423-450.

Morenoff, J. D. and R. J. Sampson. 1997. “Violent Crime and the Spatial Dynamics of

Neighborhood Transition: Chicago, 1970-1990.” Social Forces 76:31-64.

 27

Morenoff, J. D., R. J. Sampson, and S. W. Raudenbush. 2001. “Neighborhood

Inequality, Collective Efficacy, and the Spatial Dynamics of Urban Violence.”

Criminology 39:517-560.

Parker, K. F., P. L. McCall, and K. C. Land. 1999. “Determining Social-Structural

Predictors of Homicide: Units of Analysis and Related Methodological

Concerns.” Pp. 107-124 in Homicide Studies: A Sourcebook of Social Research,

edited by M. D. Smith and M. A. Zahn. Thousand Oaks, CA: Sage.

Radzinowicz, L. 1966. Ideology and Crime. New York: Columbia University Press.

Roncek, D. W. and P. A. Maier. 1991. “Bars, Blocks, and Crimes Revisited: Linking the

Theory of Routine Activities to the Empiricism of Hot Spots.” Criminology

29:725-755.

Rosenfeld, R., T. M. Bray, and A. Egley. 1999. “Facilitating Violence: A Comparison of

Gang-Motivated, Gang-Affiliated, and Nongang Youth Homicide.” Journal of

Quantitative Criminology 15:495-516.

Sampson, R. J., S. W. Raudenbush, and F. Earls. 1997. “Neighborhoods and Violent

Crime: A Multilevel Study of Collective Efficacy.” Science 277:918-924.

Shaw, C. R. and H. D. McKay. 1931. Social Factors in Delinquency. Chicago:

University of Chicago Press.

_____. 1942. Juvenile Delinquency and Urban Areas. Chicago: University of Chicago

Press.

Sherman, L. W. and D. Weisburd. 1995. “General Deterrent Effects of Police Patrol in

Crime ‘Hot Spots’: A Randomized, Controlled Trial.” Justice Quarterly 12:625-

648.
 28

Sherman, L. W., P. R. Gartin, and M. E. Buerger. 1989. “Hot Spots of Predatory Crime:

Routine Activities and the Criminology of Place.” Criminology 27:27-55.

Smith, W. R., S. G. Frazee, and E. L. Davison. 2000. “Furthering the Integration of

Routine Activity and Social Disorganization Theories: Small Units of Analysis

and the Study of Street Robbery as a Diffusion Process.” Criminology 38:489-

523.

Tukey, J. 1977. Exploratory Data Analysis. Reading: Addison Wesley.

Upton, G. and B. Fingleton. 1985. Spatial Data Analysis by Example. New York: John

Wiley.

Vold, G. B., T. J. Bernard, and J. B. Snipes. 1998. Theoretical Criminology, Fourth

Edition. New York: Oxford University Press.

Table 1. Ordinary Least Squares Regression of County Homicide Rates 1960-1990

Independent Variables 1960 1970 1980 1990

Resource 1.798** 2.913** 3.412** 3.872**

Dep/Aff Component [.318] [0.396] [0.500] [0.583]
(14.571) (19.511) (28.268) (27.133)

Pop. Structure .359** 0.812** 0.747** 1.353**

Component [.064] [0.111] [0.109] [0.204]
(3.892) (6.959) (7.315) (13.491)

Median Age -0.231** -0.191** -0.242** -0.101**

[-0.192] [-0.130] [-0.137] [-0.055]
(-11.931) (-8.394) (-9.671) (-3.691)

Divorce 1.160** 1.264** 1.250** 0.583**

[0.205] [0.184] [0.266] [0.152]
(12.233) (12.109) (18.586) (10.690)

Unemployment -0.062 -0.278** -0.122** -0.306**

[-0.028] [-0.087] [-0.059] [-0.141]
(-1.762) (-5.562) (-3.965) (-7.472)

South 2.639** 3.589** 2.113** 2.194**

[0.233] [0.243] [0.154] [0.165]
(11.312) (12.557) (9.129) (9.952)

Intercept 8.126** 8.653** 8.541** 6.517**

(12.804) (11.275) (9.720) (6.364)

Adj. R-Squared 0.295 0.360 0.431 0.435

N 3085 3085 3085 3085

Unstandardized regression coefficients are reported (standardized regression

coefficients in brackets, t-ratios in parentheses).

** P < .01 * P < .05 (two-tailed tests)

 Table 2. Stability of Regression Coefficient by Spatial Regime -- County

Homicide Rates 1960-1990

1960 1970 1980 1990

Spatial Chow Test Overall stabilitya

150.527** 227.468** 162.712** 168.438**

Stability of Individual Coefficients (Non-South versus South)b

Res. Dep. Component 0.135 0.868 7.303** 36.065**

Pop. Struc. Component 0.118 0.286 32.490** 18.758**

Median Age 3.480 0.036 7.352** 0.982

Divorce 0.057 11.088** 15.822** 0.641

Unemployment 24.849** 45.870** 12.922** 28.150**

Heteroskedastic Coefficients

Non-South 9.776 16.016 21.750 16.209

South 36.930 54.544 30.451 34.204

Test on Heteroskedasticityb

360.392** 328.375** 40.296** 164.284**

N (N of South)

3085 (1412) 3085 (1412) 3085 (1412) 3085 (1412)

** P < .01 * P < .05 (two-tailed tests)

a
distributed as χ2 with 6 degrees of freedom

b
distributed as χ2 with 1 degree of freedom
 Table 3. Spatial Lag Models of Southern Homicide Rates 1960-1990

Independent Variables 1960 1970 1980 1990

Resource 0.832** 1.792** 3.026** 4.028**

Dep/Aff Component [0.121] [0.218] [0.478] [0.602]
(3.386) (5.820) (13.994) (14.814)

Pop. Structure -0.057 0.401 1.551** 1.747**

Component [-0.007] [0.041] [0.198] [0.209]
(-0.265) (1.497) (7.637) (8.247)

Median Age -0.129** -0.060 -0.150** -0.018

[-0.099] [-0.039] [-0.093] [-0.009]
(-2.942) (-1.378) (-3.736) (-0.368)

Divorce 0.786** 0.642** 0.775** 0.482**

[0.092] [0.075] [0.149] [0.097]
(3.241) (3.060) (6.302) (4.251)

Unemployment -0.070 -0.353** -0.244** -0.438**

[-0.026] [-0.092] [-0.108] [-0.191]
(-0.897) (-3.023) (-4.145) (-5.928)

ρ)
Spatial Lag (ρ 0.713** 0.651** 0.182* 0.230**
[0.379] [0.359] [0.100] [0.125]
(6.005) (6.905) (2.431) (3.261)

Intercept 4.108* 4.153* 9.101** 5.249*

(2.207) (2.042) (5.364) (2.513)

Sq. Corr. 0.178 0.239 0.311 0.333

N 1412 1412 1412 1412

IV estimation. Unstandardized regression coefficients are reported

(standardized regression coefficients in brackets, t-ratios in parentheses)

** P < .01 * P < .05 (two-tailed tests)

 Table 4. Spatial Regression Models of Non-Southern Homicide Rates

Independent Variables 1960 1970 1980 1990

Resource 1.571** 3.007** 4.143** 2.875**

Dep/Aff Component [0.275] [0.389] [0.467] [0.405]
(9.395) (14.626) (19.837) (13.435)

Pop. Structure 0.386** 0.859** 0.290* 0.962**

Component [0.126] [0.211] [0.056] [0.229]
(5.011) (7.795) (2.132) (8.299)

Median Age -0.156** -0.157** -0.304** -0.066*

[-0.191] [-0.163] [-0.197] [-0.050]
(-7.336) (-6.452) (-8.607) (-2.034)

Divorce 0.833** 1.403** 1.318** 0.572**

[0.276] [0.359] [0.366] [0.239]
(8.552) (13.980) (14.560) (9.156)

Unemployment 0.079** -0.024 0.008 -0.045

[0.061] [-0.013] [0.005] [-0.029]
(2.622) (-0.502) (0.196) (-0.888)

ρ)
Spatial Lag (ρ 0.415** NI NI NI
[.197]
(4.645)

λ)
Spatial Error (λ NI 0.243** 0.329** 0.268**
Intercept 4.832** 6.164** 9.622** 3.261**
(6.544) (7.309) (7.588) (2.621)
Sq. Corr. 0.199 0.234 0.348 0.258

N 1673 1673 1673 1673

Instrumental Variables estimation in 1960, Generalized Moments in 1970-1990.

Unstandardized regression coefficients are reported (standardized regression

coefficients in brackets, t-ratios in parentheses). No significance is reported

for the λ parameter estimated by means of the Generalized Moments technique.

Significance is based on rejection of the null hypothesis in the OLS regression.

** P < .01 * P < .05 (two-tailed tests)

 HomRate8488

first quartile

second quartile

third quartile

fourth quartile

upper outliers

HomRate8893

first quartile

second quartile

third quartile

fourth quartile

upper outliers

Figure 1. Box Map (Outlier Map) and Box Plot (Outlier Plot) for Homicide Rates 78-84
and 88-93 in the St Louis Region. The outliers in the Box Plot match the outlier locations
in the Box Map. In a dynamically linked windows environment, these outliers are linked.
 HomRate

not significant
High-High

Low-Low

High-Low

Low-High

Figure 2. LISA Maps for St. Louis Region Homicide Rates, 1984-88 (left) and 1988-93
(right). Counties with significant Local Moran statistics are highlighted by the type of
spatial association.
Figure 3. Moran Scatterplot for Homicide Rate in St. Louis Region, 1988-93. Steepest
slope is with two central counties excluded. The slope of each line is Moran’s I.