Website CV Paper
Fields
Environmental & Energy, Econometrics, Innovation
Job market paper
Targeted Interpretability for Estimating Policy Parameters: Evidence from U.S. Gasoline Demand
Economists usually settle the trade-off between interpretability and flexibility for the whole model at once. I argue for settling it part by part, an approach I call targeted interpretability: keep interpretable only the parameter that policy uses, and let the data determine the rest, including nonlinearities and interactions, as the partially linear model estimated by double/debiased machine learning allows. Interpretability where policy does not need it can be costly: linear controls bias the estimate of that parameter, even in large samples, whenever their approximation errors in the instrument and outcome equations are correlated, and this bias is testable. In a new weekly U.S. state panel, the approach yields a gasoline price elasticity of −0.155, against −0.512 with linear controls. Under the Ramsey rule, −0.512 implies a gasoline tax of 28 to 70 percent of the price, whereas −0.155 implies no upper limit, so the estimates call for different policies.