Change in the Spatial Clustering of Poor Neighborhoods within U.S. Counties and Its Impact on Homicide: An Analysis of Metropolitan Counties, 1980-2010

ABSTRACT

Recent scholarship has examined changes in the geographic distribution of poor persons in America, but it remains unclear whether high- and low- poverty neighborhoods have become more, or less, spatially clustered over the past several decades. Additionally, while many have argued that growth in both high-poverty spatial clusters and high-low poverty spatial clusters could yield conditions that are conducive to increases in homicide, previous research has not considered that possibility. We contribute to knowledge by examining whether there have been important shifts in the spatial clustering of poverty in America between 1980 and 2010, and if so, whether those shifts were related to changes in homicide during the period. The descriptive results of our study reveal that there were notable changes in population exposure to both high- poverty and high-low poverty spatial clusters between 1980 and 2010. Fixed-effects negative binomial regression models yield limited support for the idea that changes in spatial inequality, as measured by the clustering of high- and low-poverty neighborhoods, are associated with changes in homicide rates. In contrast, the results indicate a significant positive association between changes in exposure to very high-poverty spatial clusters and homicide trends. The findings affirm the importance of considering the spatial dynamics of demographic conditions when explaining changes in violence across communities.

KEYWORDS:

• Poverty
• spatial clustering
• homicide

Disclosure Statement

No potential conflict of interest was reported by the authors.

Notes

1. On the other hand, gentrification also can create spatial clusters of adjacent high- and low-poverty neighborhoods if its impact is limited to one or more neighborhoods that compose a larger high-poverty spatial cluster.

2. Relatively few non-metro counties were “tracted” in 1980. Additionally, imposing 2010 uniform county and census tract boundaries, which is important for isolating changes in spatial organization that are unrelated to boundary changes, reveals that some metro counties were not fully tracted in 1980. We include counties in which 90% or more of its 2010 tracts can be identified (either directly or through interpolation) in the 1980 census tract data.

3. We retained counties that had valid data on homicide and the other variables considered for at least three of the four time points considered. After doing so, less than one percent of the county-year observations had missing data on the variables included in the main analysis.

4. While superior to equivalent products, such as the Neighborhood Change Database, it is important to recognize that the LTDB interpolates data to 2010 tracts based on land area proportion. This makes a critical assumption that the population is distributed evenly across land area, which in many instances may not be the case (Schroder 2017). The implications of this for our analysis are unclear, but it could yield an underestimate of the degree of clustering for poor neighborhoods within counties in 1980, 1990, and 2000, while overstating the degree to which clustering increased between those years and 2010.

5. The weights matrices for each year are constructed separately on the subset of tracts for which percent poverty is defined in each year. Moran’s I is not a valid measure for counties with very small numbers of tracts. Clustering is not defined for a single tract, and the approach we employ requires there to be at least five neighbors, meaning that a county must have a minimum of six tracts for inclusion in the analysis. Moreover, Moran’s I is highly susceptible to outliers with low tract numbers in a particular county. To limit this type of bias, we focus on counties with a minimum of 20 census tracts, as noted above.

6. If a county has multiple high-poverty spatial clusters, the average poverty rate across clusters is used. Similarly, if a county has multiple “high-low/low-high” clusters, the average difference across clusters is used.

7. For this study, we obtained restricted-use mortality data from the NCHS, which provides data on all counties and identifies the county in which recorded deaths occurred. Prior research has shown that the NCHS data reveal temporal variation in homicide over time and spatial variation across large counties that closely parallel patterns reflected in the other national-level source of information about homicide, the Uniform Crime Reporting (UCR) data (Baller et al. 2002; U.S. Department of Justice 2014; Wiersema, Loftin, and McDowall 2000). The NCHS data are the preferred source of homicide counts for our study, however, because they offer the most comprehensive coverage for estimating county homicide rates over the period and maximize the sample available for analysis. The UCR data and the Supplementary Homicide Reports (SHR), the sub-component the UCR that provides details of homicide incidents, are less optimal because they contain significant missing data due to selective nonparticipation among law enforcement agencies. Missing data in the UCR and SHR grew over the study period, resulting in considerable missing data for many agencies within counties and, in a few instances, missing data for all counties within nonparticipating states (Maltz 2006).

8. Specifically, we averaged homicide counts for 1980-81, 1990-91, 2000-2001, and 2009-2010. We used 2009-2010 for the latter period, rather than 2010-2011, because the restricted-use NCHS data obtained for the study contains data only through 2010.

9. We excluded other potentially important county attributes because they contained significant missing data for our sample (i.e., police force size, arrest rates, county prison admissions rates), but below we report supplementary models based on smaller county-year samples that incorporated these measures.

10. Given the potential bias that can arise with the conditional fixed-effects negative binomial estimator (e.g., Allison and Waterman 2002), we estimated unconditional negative binomial models with county and period fixed-effects specified as dummy variables.

11. We replicated the models after adding other control variables that were available for a smaller subset of county-years (e.g., police per capita, drug arrest rates, and county imprisonment rates). The results of these specifications yielded conclusions that are substantively identical to those reported in Table 3.

Additional information

Notes on contributors

Eric P. Baumer

Eric P. Baumer is Professor of Sociology and Criminology at Pennsylvania State University and Faculty Affiliate of the Population Research Center.  His research explores demographic, temporal, and spatial patterns of violence, the mobilization of law, and the application of criminal justice sanctions.

Christopher Fowler

Christopher Fowler is Associate Professor of Geography and Demography at Penn State University. His research examines the relationship between geographic contexts like neighborhoods and outcomes for individuals with a particular focus on racial and economic inequality.

Steven F. Messner

Steven F. Messner is Distinguished Teaching Professor of Sociology at the University at Albany, State University of New York.  His research focuses on social institutions and crime, understanding cross-national variation in levels of crime, and crime and social control in contemporary urban China.

Richard Rosenfeld

Richard Rosenfeld is the Curators’ Distinguished Emeritus Professor of Criminology and Criminal Justice at the University of Missouri - St. Louis. His current research focuses on crime trends and public policy.