Regression to the Mean in Place-Based Crime Prevention

Most place-based crime prevention interventions, such as hot spots policing, vacant lot remediation, lighting upgrades, CCTV deployment, business improvement districts, and focused-deterrence at micro-places, share a common deployment logic: target the locations with the highest recent crime. Across these strategies, quasi-experimental evaluations consistently report larger effects than randomized trials. The gap is acknowledged but rarely explained. This paper formalizes a mechanical explanation that applies to any place-based intervention with threshold-style selection. Locations enter the top of a recent crime ranking partly because of persistent features and partly because of transitory shocks that mean-revert in expectation. Standard panel estimators cannot separate that mechanical reversion from a treatment effect. We first derive an expression for this bias, consisting of entirely estimable quantities. Through simulations and a placebo application to real crime data in which the threshold-selected treatment corresponds to no actual intervention and so has zero effect by construction, we demonstrate that standard estimators generate a statistically significant ``effect'' of the sign and magnitude the analysis predicts---comparable to reductions reported in the published literature. We conclude with concrete guidelines on how to estimate this bias prior to an intervention, and how to design interventions to mitigate the size of the bias.