Causal Inference With Survey Data: A Robust Framework for Propensity Score Weighting in Probability and...

Area of Science:

• Statistics
• Epidemiology
• Causal Inference

Background:

• Confounding and selection bias are significant challenges in observational causal inference.
• Existing methods often fail to address both biases simultaneously or assume data representativeness.
• Selection bias, introduced during data collection, is frequently overlooked.

Purpose of the Study:

• To propose a unified weighting framework for simultaneously addressing confounding and selection bias.
• To develop a robust inferential procedure for population-weighted average treatment effects.
• To extend the framework to non-probability data using auxiliary information from external probability samples.

Main Methods:

• Developed a survey-weighted propensity score weighting framework.
• Proposed a doubly robust inferential procedure.
• Extended the method for non-probability data with partially observed confounders in external samples.
• Investigated the role of key variables in external data for treatment effect heterogeneity and selection mechanisms.
• Explored combining auxiliary information from multiple probability samples.

Main Results:

• The proposed survey-weighted propensity score weighting framework effectively addresses both confounding and selection bias.
• The method provides a doubly robust estimation for population-weighted average treatment effects.
• Extensions successfully handle non-probability data using auxiliary information, even with partially observed confounders.
• Identified crucial external variables related to treatment effect heterogeneity and selection.
• Simulations and real-world application confirmed the superiority over standard propensity score weighting.

Conclusions:

• The unified weighting framework offers a powerful approach to mitigate bias in observational causal inference.
• The method enhances the reliability of findings from both probability and non-probability survey samples.
• This work provides practical solutions for complex data scenarios in causal inference research.