題目：Causal inference integrating existing propensity score and
outcome regressions with heterogeneous and incomplete covariates
講者：中央研究院統計科學研究所程毅豪研究員
時間：111年6 月10 日 14：10-15：40
摘要：
The use of propensity score (PS) is a common method to make an adequate causal
inference on the treatment effect in observational studies. With the development of
big data, databases in many areas of applications have been established. However,
some covariate variables may be unobserved in some of the existing datasets. We
propose a new PS‐based causal inference framework integrating results of the PS
and outcome regressions based on possibly heterogeneous and incomplete covariate
information from existing studies. The new proposal uses a reference sample
containing data on a complete covariate set facilitating an unbiased causal inference
on treatment effects. We obtain consistent estimates for the parameters of the full
PS and outcome regression models, and by which we further obtain unbiased causal
inferences on the average treatment effect. Asymptotic theory for the proposed
regression parameter estimators is established, and simulation results reveal that
the proposed method performs well for drawing causal inferences from various PS
and outcome regression analyses with incomplete covariates. An empirical study on
the causal effect of waist circumference on hypertension based on two existing
studies with different sets of covariates, as well as a dataset with complete
covariates, is provided for an illustration.