Abstract
We study how a regulator can best allocate its limited inspection resources. We direct our analysis to a US Occupational Safety and Health Administration (OSHA) inspection program that targeted dangerous establishments and allocated some inspections via random assignment. We find that inspections reduced serious injuries by an average of 9% over the following five years. We use new machine learning methods to estimate the effects of counterfactual targeting rules OSHA could have deployed. OSHA could have averted over twice as many injuries if its inspections had targeted the establishments where we predict inspections would avert the most injuries. The agency could have averted nearly as many additional injuries by targeting the establishments predicted to have the most injuries. Both of these targeting regimes would have generated over $1 billion in social value over the decade we examine. Our results demonstrate the promise, and limitations, of using machine learning to improve resource allocation. JEL Classifications: I18; L51; J38; J8