Simple and robust estimation of average treatment effects for the overlap population using model averaging

Abstract

The propensity score plays a central role in estimating causal effects in observational studies. Recently, considerable research has been conducted on specific causal estimands that represent the target population for causal effect estimation. In this manuscript, our focus is on the average treatment effects for the overlap population (ATO), where the estimand relies on the ‘true’ propensity score. Consequently, any misspecification of the propensity score model can introduce significant bias to the estimation of the causal effect. To address this issue, we present a novel approach for estimating propensity scores using a model-averaging method that ensures more robust causal effect estimators than a single propensity score model even with the misspecification of some propensity score models. The proposed procedure is user-friendly, as it only involves the use of generalized linear models in freely available software packages, such as ‘glm’ in R and the ‘GENMOD’ procedure in SAS, among others. Also, we propose a novel augmented inverse probability weighting estimator for the ATO that has boundedness; the estimator has the same support as the interested outcome. Through extensive simulation and real data analysis, we demonstrated that the proposed method yields stable and efficient causal estimators for the ATO estimand.

Keywords:

• Causal inference
• model averaging
• overlap weight
• propensity score
• robust estimation

Acknowledgments

We would like to express our gratitude to the editors for their useful comments, which have contributed greatly to focusing and improving the manuscript. We would like to thank Editage (www.editage.com) for the English language editing. All the authors accept responsibility for the content of this manuscript, approved its submission.

Data availability statement

The data used in the real data analysis cannot be made available for general use. The simulation analysis programs are available at the following URL: https://github.com/SOrihara/Simple-and-Robust-Estimation-of-ATO.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This study was supported by the JSPS KAKENHI Grant Number JP21K02905.

Notes on contributors

Shunichiro Orihara

Dr. Shunichiro Orihara earned his Bachelor's degree in Engineering from Keio University in 2015 and then obtained a Master's in Engineering from Osaka University in 2017. He was awarded his Ph.D. in Data Science from Yokohama City University in 2023. Dr. Orihara's career includes a significant tenure at Kyowa Kirin Co., Ltd. from 2017 to 2022. Subsequently, he joined the Department of Health Data Science at Tokyo Medical University. An active member of both the Biometric Society of Japan and the Society of Clinical Epidemiology, Dr. Orihara is renowned for his contributions to causal inference. His research focuses on adjusting for confounding factors, particularly using instrumental variables that account for both observed and unobserved confounders. Notable among his publications are studies on the double robust estimator in general treatment regimes and determining the optimal number of strata for propensity score subclassification. Dr. Orihara's exemplary work has been acknowledged with the Best Presentation Award for Young Scientists from the Biometric Society of Japan in both 2022 and 2023.

Yurika Amamoto

Ms. Yurika Amamoto earned her Bachelor's degree in Economics from Tokyo University of Science in 2020 and then obtained a Master's in Data Science from Yokohama City University in 2022. She began her notable tenure at Teijin Pharma, Ltd. in 2022. Ms. Amamoto has a keen interest in causal inference and has delivered presentations at multiple domestic conferences in Japan.

Masataka Taguri

Dr. Masataka Taguri After receiving a Ph.D. in Health Science from the University of Tokyo, he was appointed as a visiting scholar at the University of California, San Francisco (UCSF) in the United States. After returning to Japan, he became a faculty member at the Yokohama City University Graduate School of Medicine and was appointed to a professor in the School of Data Science in 2020. He is also a visiting faculty member at the Institute of Statistical Mathematics of the Research Organization of Information and Systems and was appointed as a visiting professor at the Institute in April 2021. He has been active in both research and teaching since April 2022 as a professor in the field of health data science at Tokyo Medical University. He is particularly interested in methods for performing causal inference from observational research data, and actively conducts research on methodologies while contributing to many clinical studies as a statistician. He has received many awards, including the Honorable Mention of the Biometric Society of Japan and the Excellence Report Award of the Japanese Joint Statistical Meeting.