No abstract

No abstract

No abstract

For decades, social scientists have been trying to answer causal questions about the effectiveness of certain programs or policies. The conventional methodology for answering such causal questions relies on the "no interference between different units" assumption; that is, a unit's outcome depends solely on the treatment that the unit is assigned to (or exposed to) and does not depend on the treatment assignment (or exposure) of other units in the population. However, this assumption is likely to be violated in spatial settings because of various spillover, diffusion, and displacement effects. In this paper, we present a case study evaluating the causal effects of Chicago's community policing program (a community-wide intervention) on neighborhoods' crime rates. We use semiannual crime data from Chicago to evaluate (1) whether community policing is more effective at decreasing rates of reported personal crime when implemented everywhere versus nowhere in a city; (2) whether community policing is effective if implemented in a single local area, holding constant policing in adjacent areas; (3) whether implementing community policing in surrounding areas affects crime in a focal area; and (4) whether community policing's impact on crime in surrounding areas depends on whether community policing is also implemented in the focal area. To answer these questions, we relax the no-interference assumption. Our approach allows the potential outcomes in any local area to depend on a function of the treatment assignments in all other units. We define causal effects and evaluate assumptions that make the framework tractable within the framework of a generalized linear model with spatially auto-correlated random effects.present two approaches to data analysis. The first, a naive approach, is appropriate if confounding is solely attributable to the pre-treatment crime rate. The second approach relaxes the ignorability assumption by controlling for district-varying confounding and identifies spatially auto-correlated random effects in the multilevel model for personal crime rates. Finally, in section 4, we discuss conclusions and future work.

No abstract