Working Paper: NBER ID: w26224
Authors: Abhijit Banerjee; Esther Duflo; Daniel Keniston; Nina Singh
Abstract: Should police activity be narrowly focused and high force, or widely-dispersed but of moderate intensity? Critics of intense “hot spot” policing argue it primarily displaces, not reduces, crime. But if learning about enforcement takes time, the police may take advantage of this period to intervene intensively in the most productive location. We propose a multi-armed bandit model of criminal learning and structurally estimate its parameters using data from a randomized controlled experiment on an anti-drunken driving campaign in Rajasthan, India. In each police station, sobriety checkpoints were either rotated among 3 locations or fixed in the best location, and the intensity of the crackdown was cross-randomized. Rotating checkpoints reduced night accidents by 17%, and night deaths by 25%, while fixed checkpoints had no significant effects. In structural estimation, we show clear evidence of driver learning and strategic responses. We use these parameters to simulate environment-specific optimal enforcement policies.
Keywords: drunk driving; police resources; randomized controlled trial; learning behavior; public health
JEL Codes: D83; K42; O18
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| Driver learning and strategic responses to police behavior (K49) | influence on outcomes (I24) |
| Effectiveness of the crackdown persists after intervention ends (E65) | lasting impact on driver behavior (R48) |
| Rotating sobriety checkpoints (R48) | reduced night accidents (R41) |
| Rotating sobriety checkpoints (R48) | reduced night deaths (I14) |
| Fixed checkpoints (Y10) | no significant effects on night accidents (R41) |