Abstract

We present a novel framework for estimation and inference for the broad classof universal approximators. Estimation is based on the decomposition of modelpredictions into Shapley values. Inference relies on analyzing the bias andvariance properties of individual Shapley components. We show that Shapleyvalue estimation is asymptotically unbiased, and we introduce Shapleyregressions as a tool to uncover the true data generating process from noisydata alone. The well-known case of the linear regression is the special case inour framework if the model is linear in parameters. We present theoretical,numerical, and empirical results for the estimation of heterogeneous treatmenteffects as our guiding example.

 

Quick Read (beta)