Optimal Policy Targeting with Machine Learning: Evidence from Pakistani Audits

• In Event: Fiscal Institutions and Capacity

Abstract

This paper bridges welfare economics and machine learning econometrics to develop empirically implementable algorithms for optimal audit targeting. We formulate a structural model in which sufficient statistics for welfare -- specifically the welfare-weighted marginal value of public funds (WMVPF) and the welfare-weighted net social benefit (WNSB) -- identify optimal audit policies under a range of government objectives and constraints. These statistics depend on four individualized causal effects: the immediate revenue recouped from an audit, the causal effect of an audit on long-run tax revenue, the marginal administrative cost of an audit, and the marginal compliance cost of an audit. We estimate these effects using generalized random forests applied to the universe of Pakistani income tax returns, exploiting a year in which audits were assigned completely at random. We then use these estimates to implement our targeting algorithms in the following year, when audits were conducted semi-randomly among likely evaders. We show that our counterfactual policies reduce the expected welfare cost per dollar of revenue raised by 40–57% over the real-world policy, or alternatively, more than double expected revenue under a constant welfare cost. This framework offers a general approach to empirical welfare maximization using machine learning in resource-constrained policy settings.

Authors

• Nicholas Lacoste, Tulane University

• Zehra Farooq, Pakistan Federal Board of Revenue