Home Foreclosures and Neighborhood Crime Dynamics

Abstract

We advance scholarship related to home foreclosures and neighborhood crime by employing Granger causality tests and multilevel growth modeling with annual data from Chicago neighborhoods over the period 1998–2009. We find that completed foreclosures temporally lead property crime and not vice versa. More completed foreclosures during a year both increase the level of property crime and slow its decline subsequently. This relationship is strongest in higher income, predominantly renter-occupied neighborhoods, contrary to the conventional wisdom. We did not find unambiguous, unidirectional causation in the case of violent crime and when filed foreclosures were analyzed.

Keywords::

• Foreclosures
• crime
• neighborhoods
• Granger causality
• multilevel growth models

Acknowledgments

This research was supported from a grant from the MacArthur Foundation to MDRC related to the evaluation of the New Communities Program. The opinions expressed in this paper are the authors and do not necessarily represent those of MDRC, the Foundation, or their respective Boards of Trustees.

Notes

¹ However, when Cui (2010) replicates a cross-sectional model like Immergluck and Smith's she finds a variety of implausible relationships. Moreover, she finds that crime rate strongly predicts foreclosures 3 years later, suggesting that endogeneity is indeed worrisome.

² Cui (2010) and Ellen et al. (2013) explore the causality issue in a conceptually similar way by testing if the temporal leading value of foreclosures predicts crime.

³ A map showing these 80 neighborhoods is available from the authors. Note that some of the traditional 77 Community Areas of Chicago were split for our analysis because parts were outside NCP boundaries. With these few exceptions, our 80 neighborhoods follow the encompassing boundaries of their constituent Census tracts.

⁴ The Census tract designations used for these transformations are the definitions created after the 2000 Decennial Census; data collected or assembled using earlier designations was transformed to the 2000-era designations using the relational matrices published by the U.S. Census Bureau. For most of the neighborhoods, the definitional boundaries align with tract boundaries such that the neighborhood-level measure is the aggregation of the tract-level measure. In cases where this is not true, the tract values were apportioned between multiple neighborhoods based on the distribution of the tract's population between the multiple neighborhoods.

⁵ Cluster analysis was used to create the neighborhood grouping used in this paper. This technique is widely used to sort cases (people, things, events, neighborhoods, and so on) into groups, or clusters, so that the degree of association is stronger among members of the same cluster and weaker among members of different clusters. The procedure was applied to classify all Chicago neighborhoods, and more than 20 variables were used to group these neighborhoods, including measures of economic context, housing market dynamics, and racial/ethnic diversity of neighborhoods; see Table A1. These measures were assembled for the period 2000–2005, in order to capture the ‘starting context’ for the NCP initiative. In general, a good cluster solution is one in which each cluster is very different from other clusters (between-cluster heterogeneity) and in which units in each cluster are as similar as possible (within-cluster homogeneity). We used the Ward clustering method. The statistical diagnostics available for cluster analysis were calculated and examined, and they confirm that the resulting five clusters of neighborhoods largely differ from one another. In addition, tests were conducted to assess the ‘goodness of fit’ of neighborhoods with their groups (including an examination of the extent to which neighborhoods differ from their cluster, on average, as well as sensitivity tests comparing findings before and after the exclusion of a potentially ‘outlier’ neighborhood). These tests provide further confidence in the five clusters.

⁶ Both filed and completions data exclude ownership transfers that occur as the result of financial distress (short sales or deed-in-lieu-of-foreclosure transactions).

⁷ With multiple incidents the report is classified in the UCR category of the most serious crime (generally, the crime with the highest potential penalty). Note that these are police reports and do not reflect later adjudication of the incident (e.g., an assault recorded on the initial report as a criminal act later adjudicated as justifiable self-defense is still included).

⁸ While the LOESS function is parametric within a single ‘neighborhood’ of the data, the overall function—the compilation of the local regression results for each point—fit to the trend is nonparametric. Weighted moving averages are a simple example of a LOESS function, where the local polynomial regression function has a degree of 0. For the LOESS functions estimated for the indicators, a higher degree polynomial was used so that shifts in trend direction would be more readily identified.

⁹ The use of the causal in the name of the test and throughout this section is a reference to the particular statistical relationship, not to be mistaken for conceptual causality and/or causal mechanisms.

¹⁰ The value of k is not determined empirically as Granger causality is a ‘brute force’ method. In general, the value of k is set high enough to ensure that the autodependence of the outcome is accounted for within the limits of the input data.

¹¹ Note that the Granger causality test relies on overfitting to ensure that all autodependence is removed from the data. This overfitting is harmless in regards to bias, meaning that it does not imperil the validity of the statistical significance tests. However, overfitting causes estimator inefficiency and so results in uninformative coefficient estimates. In other words, Granger causality tests whether there is a relationship (i.e., its existence) but is uninformative in regards to the strength or direction of effects of X on Y.

¹² Results are robust to k = 1 as well, though we did not have adequate observations to estimate a model with k = 3.

¹³ The multilevel model random effects are different from the random effects commonly used in econometric time series models (sometimes referred to as fixed effects models). In the multilevel formulation, the random effects are equivalent to main effects, with the descriptor ‘random’ referring to the nature of the neighborhoods (i.e., they are theoretically drawn at random from some larger population). In contrast, the econometric random effects describes the effect of individual differences (i.e., random disturbance), which is necessary to account for to generate unbiased estimates for the main effects. For an introduction to multilevel models, see Kreft & de Leeuw (1998).

¹⁴ Modeling the foreclosure–crime relationship using counts rather than rates allows coefficients to be interpreted as the increase/decrease in the number of crimes due to changes in the number of foreclosures.

¹⁵ Since there are only 3 or 4 years (observations) in each period, the model specification used only the linear model of change described above.

¹⁶ The control variables, population, commercial land area, and single-family owner-occupied housing units, were centered prior to entering into the model. This means that the estimated effects for these variables correspond to the effect of deviations from the average neighborhood. For example, the results indicate that neighborhoods with population that is greater than average have higher levels of property crime. It was not possible to estimate the effect of the control variables on the rate of change in the number of crimes as annual data for these variables were not available.

¹⁷ This model also includes random effects to represent neighborhood variation in the property crime trend; inclusion of the neighborhood group effects reduced level of variation in the random effects and the correlation between the random starting level and annual change parameters slightly (correlation: − 0.646).