Scalable inference for individual treatment effect

Discovery of the discrepant treatment effect across each individual is important for precision decision making. Different from the conditional average treatment effect (CATE) which is sensitive to the uncertain environments, the developed individual treatment effect (ITE) is more scalable in term of generating average treatment effect, group-specific treatment effects or CATE and more robust to the model uncertainty, because the proposed heterogeneous model includes partially linear and high-dimensional regression, etc. Under the potential outcome framework, we use multiple imputation techniques first to recover the conditional expectation of unobserved potential outcome. Ridge fused penalty-based quadratic loss function is developed to estimate all population and individual parameters and we deduce the explicit expression for individual estimators, avoiding the computational burden caused by the massive data. Theoretical results argue the asymptotic distribution of individual parameters for statistical inference of individual treatment effect. Simulation studies pose supportive evidence that the proposed unified structure of individual analysis including estimation, inference and prediction performs well with finite samples, and a real data example is provided for illustrating the flexibility of ITE in finding the individual information and forming all sorts of treatment effects.

Keywords

heterogeneous analysis, treatment effect, fused penalty

Full Text (PDF format)

Received 30 November 2022

Accepted 29 January 2023

Published 19 July 2024