«Urban Problems and sPatial methods VolUme 17, nUmber 1 • 2015 U.S. Department of Housing and Urban Development | Office of Policy Development and ...»
Social capital factors were obtained from the Speak to Your Health! Survey (STYHS) disseminated to Flint residents in 2007. To ensure that the survey respondents represented all geographic regions of the city of Flint and Genesee County, random samples of households were drawn from Genesee County census tracts. At least 20 residents were obtained for each of the 39 residential census tracts in Flint. The frequency of positive responses per CBG was derived by dividing the number of “yes” responses by the total number of responses because of the ordinal scale of the data.
Public health variables in this study included body mass index (BMI) and nonmotorized mobility safety. BMI was obtained from the STYHS survey, and the average was joined to each CBG. Travel risk factors were obtained from the State of Michigan Department of Transportation for the year 2010.
Previous research has also posited that environmental stressors such as lead can have a long-lasting effect on criminal activity (Shaker, Rybarczyk, and Eno, 2009; Wright et al., 2008). Therefore, the average blood lead level (BLL) concentration from 9,000 adolescent people within each census tract was obtained from the State of Michigan Department of Community Health, Childhood Lead Poisoning Prevention Program, for the year 2010. An areal interpolation method was used to join
these data to the CBGs. The interpolation method is used when the spatial units are not congruent (Flowerdew, Green, and Kehris, 1991). The open and closed leaking underground storage tank, or LUST, dataset served as a proxy for brownfields in this study. The data were obtained from the State of Michigan, Department of Environmental Quality, for the year 2007.
Several neighborhood contextual variables were obtained from the City of Flint, Planning and Engineering Departments. The most recent data were obtained and consisted of transportation routes (bus, bicycle, railroad, and sidewalks), schools, parks, land use, housing type (owner- or renter-occupied), and vacant parcels. Food outlet data were retrieved from the ReferenceUSA database for the year 2010 (http://www.referenceusa.com/). Food outlets were construed as an indicator of neighborhood quality, land use diversity, and opportunistic locations for criminal activity (Gruenewald et al., 2006).
Each explanatory factor in this research was aggregated and then normalized by the area (square feet) of each CBG. This process was used to bring all factors into the same resolution for further spatial analysis and minimize errors associated with the modifiable areal unit problem, or MAUP.
Modeling Approach To examine relationships among crime, spatial, and aspatial factors within CBGs, we advanced a comprehensive statistical and spatial modeling approach. After cursory descriptive and spatial analysis procedures were conducted, two models were calibrated. The first model, OLS, was developed to detect global crime correlates. A second model, GWR, was then enlisted to highlight significant localized crime explanatory variables.
Exploratory Analysis Exploratory data analysis (EDA) and exploratory spatial data analysis (ESDA) were conducted to detect for statistical or spatial relationships among crime and potential predictors. The EDA consisted of a Pearson’s correlation analysis using SPSS software, version 19 (International Business Machines Corporation, or IBM). The analysis was used to test for linear relations among crime and potential predictors. To further refine potential variable selection, the variance inflation factor (VIF) index was also used to detect for multicolinearity; a threshold of 5 was established as a cutoff value based on previous statistical research (Kutner, Nachtsheim, and Neter, 2004). Exhibit 2 shows the final selection of independent model predictors and their descriptors.
ESDA was imparted in this research to examine spatial associations among all of the factors and to aid with model development. A classic spatial autocorrelation index, Moran’s I, was implemented to explore spatial nonstationary effects on crime (Anselin et al., 2000). Exhibit 3 demonstrates that elevated high-high (HH) and low-low (LL) CBG crime densities are clustered throughout space, suggesting the need for a model that accounts for spatial relationships.
Model Development A global regression model was developed using SPSS to determine the generalized causal associations among crime and the explanatory variables. The outputs from this model consist of global predictions of the dependent variable, using several independent variables. The OLS model’s strength, coefficients, and residuals served as a comparison with the GWR model. To determine if the OLS model residuals were spatially autocorrelated, a global Moran’s I was implemented.
Spatially clustered residuals is an indicator of spatial nonstationarity.
A GWR model was developed to measure the magnitude, directionality, and geography of crime predictors for each CBG. The mathematical expression for GWR is similar to the OLS in that local parameters take the place of global parameters, while accounting for distance (Fotheringham, Charlton, and Brunsdon, 2002).
The GWR equation can be expressed as, (1) where yi is the dependent variable at location i, b0(ui, ni) is the intercept at location i, bk(ui, ni) is the estimated kth parameter at location i, cik is the independent variable of the kth parameter at location i, and i is the error term at location i. The GWR model assumes that the error term is independent and identically distributed (Zhao and Park, 2004).
The GWR model was developed using GWR4 software, developed by Fotheringham, Charlton, and Brunsdon (2002). The GWR model produces parameter estimates for each CBG based on the kernel and the bandwidth selection, producing a continuous surface in return. In this research, a fixed Gaussian weighting scheme was used. The scheme is based on a distance-decay function following a Gaussian curve. The function is adjusted by the bandwidth setting, which dictates the distance that neighborhood parameters from the centroid i will count toward the estimate (Mennis, 2006). The bandwidth setting was set to automatically obtain optimal values to minimize the Akaike Information Criterion (AIC). The setting was chosen to account for the variation in size and quantity of the CBGs. Moreover, the AIC optimization technique assures a robust model, signified by an ideal goodness-of-fit and reduced degrees of freedom coefficients (Graif and Sampson, 2009). The outputs from the GWR model included parameter estimates, R2 values, and t-values for each CBG.
The directionality, significance, and degree of spatial variability (nonstationarity) of the parameter estimates were assessed using ArcGIS software. The parameter estimates and diagnostics were assessed in accordance with the Mennis (2006) study. Significant relationships between each explanatory variable and dependent variable were obtained by querying t-values at the 90-percent significance level (z-scores ± 1.6565) for each unstandardized parameter estimate. Using a Jenks Natural Breaks sequential classification scheme, an area-class map was produced that grouped the significant estimates into five classes. Those parameter estimates that fell outside the significance threshold were displayed in white.
Exhibit 7 uniquely depicts the magnitude, directionality, and geography of crime correlates (unstandardized parameter estimates, b). Exhibits 7a through 7d indicate that SES estimates marginally affect crime. Renter-occupied housing (exhibit 7a) positively affects crime east of the central business district (CBD), while English-speaking residents in the same vicinity have a negative influence (exhibit 7b). Similarly, significant parameter estimates for non-White residents and households living below the poverty level appeared to affect crime in the same general area (exhibits 7c and 7d). The result suggests that these factors have a compounding effect on crime, requiring specific crime-reduction strategies. The SES estimate for educational attainment positively affects crime overwhelmingly in two CBGs south of the CBD (exhibit 7e). Interestingly, the relationship appears counter to previous research that has shown that reduced educational attainment increases crime (Kruger et al., 2007). We can infer from exhibit 7e, however, that criminal activity may be diffusing into this area from nearby locations because the neighborhood is of high SES status.
Exhibit 7f depicts a positive association between crime densities and BMI in several CBGs east of the CBD. The result supports previous research that has shown a link between obesity rates and the probability for criminal arrests (Kalist and Siahaan, 2013). Conversely, self-reported general health shows no significant statistical influence on crime densities (exhibit 7g). The factor is indirectly linked to the fear of crime, which has been purported to affect actual crime. The outcome in this research, however, contains no such linkages (Chiricos, Padgett, and Gertz, 2000). The concentration of BLL appears to negatively affect crime densities among several CBGs in Flint (exhibit 7h). It can be inferred from this result that areas with increased environmental stress do not statistically affect crime, despite prior research indicating otherwise (Needleman et al., 1979).
The neighborhood contextual factors that negatively altered crime densities included recreational trails (exhibit 7i) and railroads (exhibit 7k). The density of trails in the southern portion of Flint appeared to reduce crime. Strong spatial and statistical dependencies between railroads and crime were discovered along the western boundary of Flint (exhibit 7k). One possible inference from this result is that the population density is low in this area, thereby reducing opportunities for victimization. The neighborhood contextual factors that displayed a positive effect on crime included the density of food outlets (fast-food restaurants, liquor stores, and convenience stores) and sidewalks.
As evidenced in exhibit 7l, the density of poor-quality food outlets moderately affects crime, with the strongest relationship evidenced in the southeast portion of the city. The positive relationship found in this study is supported by previous research. For example, Gruenewald et al. (2006) found that alcohol establishments had the propensity to accelerate criminal activity. Exhibit 7m exhibits the spatial patterning of significant sidewalk-density estimates on crime, albeit with marginal
influence. It is clear from the spatial pattern that sidewalk density may be facilitating criminal activity throughout a large portion of Flint. This result may partly be because sidewalks are vectors for criminal activity. In other words, the increase in pedestrian mobility and exposure may be increasing crime opportunities. The finding here is substantiated by earlier research conducted by Doyle et al. (2006), who found a positive correlation between neighborhood walkability and crime.
Conclusion The key goal of this article was to critically examine the utility of a spatially explicit model and refined cartographic technique to uncover detailed relationships among crime and important socioeconomic conditions, public health, social capital, environmental stress, and neighborhood contextual variables in Flint, Michigan. The objective was reached using a strategic modeling strategy that consisted of EDA, ESDA, global, and localized modeling approaches. The strength and performance of the GWR model were reasonably good in comparison with the OLS model, exhibiting an adjusted R² of 0.536. The GWR model provided local coefficients for each CBG, which were integrated into a robust visualization strategy using GIS. The results included several nuanced cartographic outputs that displayed statistically significant relationships between crime and the explanatory factors, which were not evidenced from the OLS model. More importantly, the visualization of the significant GWR coefficients suggest that targeted enforcement strategies should vary based on localized geography
Cityscape 31Rybarczyk, Maguffee, and Kruger
and contextual conditions. With a better understanding of how, and to what extent, various factors influence crime at a microscopic scale, law enforcement officials, community planners, and citizenry can develop more insightful crime-prevention/reduction strategies.
Acknowledgments The authors thank the Office of Research at the University of Michigan-Flint for sponsoring this research through the Research and Creative Activity Fund.
Authors Greg Rybarczyk is an assistant professor in the Department of Earth and Resource Science at the University of Michigan-Flint.
Alex Maguffee is a research assistant in the Department of Earth and Resource Science at the University of Michigan-Flint.
Daniel Kruger is a research assistant professor in the School of Public Health at the University of Michigan.
References Anselin, Luc, Jacqueline Cohen, David Cook, Wilpen Gorr, and George Tita. 2000. “Spatial Analyses of Crime,” Criminal Justice 4: 213–262.
Cahill, Meagan, and Gordon Mulligan. 2007. “Using Geographically Weighted Regression To Explore Local Crime Patterns,” Social Science Computer Review 25 (2): 174–193.
Chiricos, Ted, Kathy Padgett, and Marc Gertz. 2000. “Fear, TV News, and the Reality of Crime,” Criminology 38 (3): 755–786.
Doyle, Scott, Alexia Kelly-Schwartz, Marc Schlossberg, and Jean Stockard. 2006. “Active Community Environments and Health: The Relationship of Walkable and Safe Communities to Individual Health,” Journal of the American Planning Association 72: 19–31.
Flowerdew, Robin, Mick Green, and Evangelos Kehris. 1991. “Using Areal Interpolation Methods in Geographic Information Systems,” Papers in Regional Science 70 (3): 303–315.
Fotheringham, A. Stewart, Chris Brunsdon, and Martin Charlton. 2002. Geographically Weighted Regression: The Analysis of Spatially Varying Relationships. London, United Kingdom: John Wiley & Sons.
Graif, Corina, and Robert J. Sampson. 2009. “Spatial Heterogeneity in the Effects of Immigration and Diversity on Neighborhood Homicide Rates,” Homicide Studies 13 (3): 242–260.
Gruenewald, Paul J., Bridget Freisthler, Lillian Remer, Elizabeth A. LaScala, and Andrew Treno.