A novel approach to creating pseudo-control groups applied to the evaluation of Hotspot Policing

• In Event: Advances in Experimental Methods Panel 1 : Experimental Frontiers in Policing and Threat Assessment

Abstract

From 2021-2024 the UK Home Office funded the Grip programme which used pulse police patrols to reduce serious violent crime in the highest crime hot-spots across 20 police forces in England and Wales. The initial evaluation utilised a cross-over design which, while controlling for within force effects, has its limitations. Criticisms of this approach are: i) criminals may learn the patrol schedule – increasing crime on non-patrol days, overestimating effectiveness ii) patrols may have a residual impact on non-patrol days - decreasing crime, underestimating effectiveness. We have developed an innovative quasi-experimental design methodology which overcome these limitations. We created pseudo-control groups that resemble the treatment group in terms of both crime trends and geographical profiles.

We ran LASSO regression to identify which geographical points of interest (POIs) are most predictive of crime (numbers of pubs, fast-food restaurants etc..). We layered a tessellated hexagonal grid across each force and translated this grid multiple times to build the initial pool of control areas. We carried out a Greedy search to determine which hexagons had the highest amount of serious violent crime, removing hexagons that overlapped with the treated areas.

For each treated force we then used Genetic Matching to build a pseudo-control group, matching using the POI that were predictive of crime as well as the crime history. Determining when a match is sufficient is a general problem. To counter this we leveraged the stochastic nature of the genetic matching algorithm to build a collection of matches for each treated police force.
For each collection of controls, we ran a Difference-in-Difference Poisson regression model. This gives a distribution of estimated effect-size of the intervention on the treated areas. We took the median effect-size to be the result for that force and then pooled these force level results in a meta-analysis.

Authors

• Peter Dawson, Home Office

• Timothy Eales, Home Office UK