Primary Submission Category: Causal Inference and SUTVA/Consistencies Violations
On the Causal Effects of Long-Term Treatments
Authors: Jinglong Zhao, Shan Huang, Chen Wang, Yuan Yuan,
Presenting Author: Jinglong Zhao*
One lingering challenge of randomized controlled trials is to estimate the unobserved long-term treatment effects with limited short-term experimental data. Existing literature concerns more about the long-term effects of short-term treatments. In this paper, we focus on the long-term treatments, which would be repeatedly assigned to users once the intervention is rolled out. We propose a mathematical framework, which we refer to as a longitudinal surrogate model, to study the long-term treatment effects with historical and short-term experimental data. We show that under standard assumptions, the long-term treatment effects can be estimated by an iterative expectation expression conditional on short-term surrogates and treatment assignments. Detailed instruction on the practical estimation process and required assumptions is discussed. We verify the efficacy of our approach with empirical large-scale holdout experiments conducted on the WeChat platform. Considering the accumulated short-term effect as the benchmark, we evaluate the estimated long-term effect generated by our framework and show the validity of our approach.