Heterogeneous Associations Between Policing Perceptions and Residents’ Sense of Safety: A Causal Forest with Double Machine Learning Approach

This study examines whether the association between residents’ perceptions of policing and their sense of safety varies across individual characteristics and neighbourhood contexts. Using survey data from 1738 residents of the Humberside region of England, linked to small-area deprivation and police-recorded crime data, we apply a causal forest with double machine learning to estimate this conditional association flexibly while adjusting for demographics, experience of crime and neighbourhood conditions. Because the data are cross-sectional, the estimates are interpreted as adjusted associations rather than causal effects. We find a positive average association between policing perceptions and sense of safety that is present across all groups examined. A calibration test indicates some variation in the strength of this association, but differences between pre-defined demographic and neighbourhood groups are modest, their confidence intervals largely overlap, and much of the variation lies within rather than between groups. The clearest between-group pattern concerns neighbourhood deprivation: the association tends to be somewhat weaker among residents of more deprived neighbourhoods. Differences by ethnicity are suggestive but based on small subgroups. These patterns generate hypotheses for more targeted policing and neighbourhood strategies that should be tested with longitudinal or experimental designs.