Scalable Demand Estimation with High-Dimensional Controls: A DML-SNFP Approach

We develop a double/debiased machine learning stochastic nested fixed-point (DML--SNFP) estimator for random-coefficients logit demand models with many markets and high-dimensional controls. DML--SNFP uses a pilot sample to estimate the nuisance functions associated with the high-dimensional controls. These nuisance estimates are held fixed during the online stage, which performs market-level demand inversion and updates only the low-dimensional structural parameters through stochastic approximation. We establish consistency and large-$T$ asymptotic normality and show that, under appropriate rate conditions, the effect of nuisance estimation on the final estimator is asymptotically negligible. In Monte Carlo experiments, DML--SNFP reduces bias and root mean squared error relative to updating the full parameter vector during the online stage while remaining close to the oracle-nuisance benchmark. An application to Medicare Advantage demand illustrates the implementation of DML--SNFP in a large-scale demand setting with many controls.

Stochastic GMM for Large-Scale Demand Estimation (with Zhentong Lu, Myung Hwan Seo, and Youngki Shin) [arXiv]

Dyadic Data with Ordered Outcome Variables (with Chris Muris and Cavit Pakel) [arXiv]