Avoiding Simpson's Paradox In Heterogeneous Treatment Effect Analysis

Heterogeneous Treatment Effect Analysis is a method used to identify the treatment effect of heterogeneous subpopulation in causal inference.It has great potentials in the fields of medical,Internet industry and intelligent marketing.However,popular heterogeneous treatment effect analysis methods,such as causal tree,are based on the hypothesis of unconfoundedness,which usually does not hold for real-world high-dimensional data.The presence of confoundedness greatly weakens the power of those causal estimation methods.The purpose of this paper is to give a practical guide on how to appropriately apply causal tree to real-world data,put forward the causal modeling idea of diagnosing and avoiding Simpson's paradox,so as to mine the feature of heterogeneous population legitimately.Before modeling,we should search Simpson's paradox in the dataset.If a confounder is found,it is necessary to diagnose the causal tree that we grow subsequently,observing whether each leaf node contains the confounder.Confounding adjustment is a must if there is a leaf node that does not contain the confounder.This paper proposes two adjustment methods,one is to adjust growth parameters of the causal tree,and the other is to conduct backdoor adjustment or propensity score weighting to correct causal effect estimates according to the characteristics of the confounder.The random simulation is carried out under three different data generation mechanisms.The simulation experiment verifies the effect of causal tree on conditional average treatment effect estimation.The results of simulation show that the causal tree identified and adjusted by Simp-son's paradox has better estimation effect and escorts that the causal effect estimate obtained from the adjusted causal tree are appropriate for unconfoundedness assumption.In terms of marketing data,the customer cluster is divided by causal tree,which is a useful guideline for the follow-up marketing plan.