Neighborhood Effects on Crime in San Francisco: An Examination of Residential, Nonresidential, and “Mixed” Land Uses

ABSTRACT

Communities and crime research presumes that purely nonresidential land uses are associated with more crime, whereas purely residential land use is associated with less crime. However, few studies have evaluated this prevailing assumption in tandem with the notion of “mixed” land use. Drawing on a sample of San Francisco census blocks, this study examines the neighborhood effects of several land use measures on crime counts using negative binomial regression. The results provide moderate support for the prevailing assumption; however, we also find that the effects of mixed land use are nuanced based on differing measurement approaches. Whereas two types of mixed-use buildings have differential effects on crime (i.e. commercial mixed-use and residential mixed-use) we determine that the amount of heterogeneity among eight specific land uses, next to or in direct proximity of each other, is consistently related to more crime. The implications of these findings for criminology and public policy are discussed.

Notes

¹ A related, but distinct, area of research has determined that certain land use characteristics spatially collocate in high-crime areas; however, these studies examining crime concentration generally do not systematically account for socioeconomic differences across geographic units, nor do they use regression analysis (e.g. Kinney et al. 2008; Sherman, Gartin, and Buerger 1989). Therefore, it is somewhat difficult to draw causal inferences from these studies.

² The San Francisco Police department does not provide geospatial information on murders.

³ The 2010 Census is a short-form questionnaire thereby it does not provide these variables at the block level (except for single-parent households). Thus, we imputed the remaining variables using the synthetic estimation approach by Boessen and Hipp (2015).

⁴ We started estimating models using crime lags derived from a first-order queen contiguity matrix. However, some models indicated large Moran’s I values of the residuals (e.g. >.10 Moran’s I value index), which suggests that substantive spatial autocorrelation remains in the model. We therefore turned to estimating models using crime lags derived from a second order (cumulative) queen contiguity matrix, and they produced small Moran’s I values (i.e. <.05 Moran’s I value index). These small values provide evidence that the models have adequately accounted for most of the spatial autocorrelation.

⁵ An alternative approach is to construct and include spatially lagged variables of the independent variables. However, this approach did not prove to be effective in removing substantive spatial autocorrelation from models. Not surprisingly, two famous studies examining land uses and neighborhood crime similarly only include spatially lagged crime variables (Browning et al. 2010; Stucky and Ottensmann 2009). This highlights the need to methodologically control for spillover effects of crime (Anselin 1988).

⁶ Blocks on Treasure Island (an island in the San Francisco Bay) have been dropped.

⁷ The variation inflation factor scores (VIFS) for the first series of models did not exceed a cutoff of 4. For the second series of models, when using quadratic terms derived from mean centered variables, the VIFS did not exceed a cutoff of 5. The highest pairwise (absolute value) correlation among the land uses was .51, and the average (absolute value) pairwise correlation was .10. We assessed the possibility of influential cases by estimating ancillary models that dropped blocks with the most extreme 1% of Hadi values (Hadi 1994). This did not change the pattern of results.

⁸ Commercial and residential mixed-use buildings, respectively, maintain their positive and negative relationships from the first set of models. Residential mixed-use exhibits a slowing negative relationship with most crime types (not shown). Commercial mixed-use is linearly and positively related with most crime types. The second set of models effectively allow us to assess proposition 5, while at the same time, reaffirming the conclusions we made from the first set of models regarding proposition 3 and proposition 4.

Additional information

Notes on contributors

James C. Wo

James C. Wo is an Assistant Professor in the Department of Sociology and Criminology at the University of Iowa, and a Research Fellow at the University of Iowa’s Public Policy Center. His research interests include neighborhoods and crime, local institutions and organizations, land uses, and quantitative research methods.

Young-An Kim

Young-An Kim is an Assistant Professor in the College of Criminology and Criminal Justice at Florida State University. His research interests focus on various areas such as neighborhoods and crime, criminology of place, geo-spatial analysis, immigration and crime, and quantitative research methods.