Title: Uplift modeling with multitreatment for observational pretest-posttest data
Authors: Hiroaki Naito - Doshisha University (Japan) [presenting]
Hisayuki Hara - Doshisha University (Japan)
Abstract: Uplift modeling is used to optimize the effect of intervention by predicting the causal effect for each unit from its covariates. Uplift refers to the causal effect for each value of covariates. Uplift modeling has been developed in a randomized controlled trial like A/B test of direct marketing. Recently, uplift modeling is also extended to the nonrandomized study, where only observational data are available. For observational data, a switch doubly robust method (SDRM) has been proposed recently. SDRM is based on doubly robust (DR) estimator and avoids instability of DR estimator due to extreme propensity scores. SDRM considers the case where the treatment assignment is binary (whether or not treatment has been received). Currently, SDRM is the state of the art of uplift modeling for observational cross-sectional data. We will extend SDRM for observational pretest-posttest data. Pretest-posttest data consists of outcomes before and after the intervention and are often used to evaluate causal effects qualitatively. In addition, we will extend the treatment assignment to multi-treatment. This extension allows for more complex intervention strategies. We will perform some simulation studies and apply the proposed method to real data example to confirm the usefulness of the proposed methods.