Estimating Individual Treatment Effects under Uncertainty in Randomized Trials

In this paper, we propose a method for estimating individual causal effects in randomized experiments. We develop an approach taking into consideration an appearance of uncertainty in estimates in analyzing causal inference between treatment and outcome. An introduced approach includes the construction of confidence intervals for treatment effects within subpopulations of customers. The purpose of the study is to create an approach, which identifies an uncertainty of heterogeneous treatment effects. Besides, we try to assess the quality of modeling taking into consideration an uncertainty of model assessment using logarithmic likelihood estimates. We apply econometric and machine learning models to get individual treatment effects: logistic regression with a full set of interactions, a transformed outcome decision tree, and a transformed outcome random forest. Besides, the study is based on the experiment carried out by the Russian restaurant company and examines the effect of scoring on the probability of purchase. So, the findings of the study could be used by companies to improve marketing strategies related to the retention and attraction of customers and by investigators to be sure of the reliability of individual treatment effect estimates.