Orthogonal Machine Learning for Demand Estimation: High Dimensional Causal Inference in Dynamic Panels (with Victor Chernozhukov, Vira Semenova, and Matt Taddy)

We provide estimation and inference tools for a high-dimensional sparse treatment effect in a partially linear model in a panel setting. First, we estimate the expectation of treatment and outcome variables conditionally on lagged realizations of treatment, outcome and any other exogenous predictors using modern machine learning (ML) methods. Second, we construct a point estimate of the treatment effect using Lasso with the dependent variable and the covariates equal to the residuals from the previous regressions, respectively. We construct confidence intervals for single and many coefficients using a version of debiased Lasso. We apply these methods to estimation of heterogeneous demand elasticities -- a challenge frequently faced while optimizing pricing for a wide catalog of products -- and validate our estimates using experimental variation in prices.