Estimating the propensity score when some confounders are partially observed
Causal Inference & Missing Data and Measurement Error Themes
How should the propensity score be estimated when some confounders are partially observed?
Abstract: Propensity score (PS)-based approaches are very popular to estimate causal effects in observational studies at risk of confounding bias. A major issue when estimating the PS is the presence of partially observed covariates. In this talk, we will investigate 3 different strategies to handle missing data for PS analysis: complete case analysis (CC), the missingness pattern approach (MP) and multiple imputation (MI).
We will look at the assumptions required for each method and think about when each should be used in practice. For multiple imputation, there are important questions regarding its implementation in the PS context (e.g. should we apply Rubin’s rules to the treatment effect estimates or to the PS estimates themselves? Does the outcome have to be included in the imputation model?). This talk will give an overview of key questions and outline some solutions from our recent research.