Treatment Effect Identification as An Out-of-Distribution Generalization Method for Semi-Markovian Causal Model

Evaluating treatment effect plays a vital role in individual medicine in which the interpretability of the prediction model is critical due to unobservable confounders. The challenge is to guarantee consistency when generalizing over unknown distributions. However, current researches mainly focus on treatment effect estimation on specific hypotheses. Identification from structure hypothesis, as the base of estimation, has not been emphasized and integrated into the treatment effect estimation framework. In this paper, we introduce graphical identification methods to create predictors automatically and transform the data distribution specification task into prediction task which was well handled by deep learning. It will take the unknown common cause variables and hidden mechanisms into consideration without modeling them directly. Then we propose a treatment effect estimation algorithm based on identifiable semi-Markovian causal model. The estimator created by identification outperformed the traditional estimator in our linear out-of-distribution testing. The experiment results show the potential ability of complete identification methods to generalize over unknown distribution.

Paper