Heterogeneous Causal Effect Analysis and Application Based on Meta-Learners

Meta-learners have been widely used to estimate heterogeneous causal effects, especially in the fields of psychology and economics. In this paper, we investigate the performance of several kinds of meta-learners through simulation design. We find that the extreme propensity scores may lead to outcomes off-target, making the DR-learner estimation unstable. We propose a new method to improve DR-learner, which we call DRI-learner. In this method, we do not weight directly by the reciprocal of the estimated propensity score. We use the symmetric kernel function form of propensity score as the new weight. The propensity score is smoothed for greater stability, as the weight for a particular subject now depends on all estimated propensity scores and their absolute difference from the subject's propensity score, rather than just the propensity score for a single subject. The proposed method does not suffer from the instability of the DR-learner when the propensity scores are extreme values. Finally, we apply different meta-learners to actual data analysis. Whether in simulation design or actual data analysis, the performance of DR 1-learner is better than that of DR-learner.