CMStatistics 2017: Start Registration
View Submission - CMStatistics
Title: Propensity score analysis with partially observed confounders: Multiple imputation and the missingness pattern approach Authors:  Clemence Leyrat - London School of Hygiene and Tropical Medicine (United Kingdom) [presenting]
James Carpenter - London School of Hygiene and Tropical Medicine (United Kingdom)
Elizabeth Williamson - London School of Hygiene and Tropical Medicine (United Kingdom)
Abstract: Propensity scores (PS) estimate the probability of an individual being treated given their characteristics. They are commonly used to address confounding bias in observational studies. One popular method to achieve covariate balance between exposure groups is to re-weight individuals by the inverse of their PS value (Inverse Probability of Treatment Weighting). In applications, a major issue is how to estimate the PS when confounders are partially observed. Multiple imputation (MI) is a natural, and widely used tool. However, in the PS context a number of key issues remain unresolved, included (i) how to build the imputation model to ensure compatibility between the substantial, the imputation and the PS models, (ii) how to apply Rubins rules in the PS context, and (iii) how to extend MI to address the question of time-varying treatment and exposure. An alternative strategy, the missingness pattern approach (MPA), has been proposed: although simpler to implement than MI, the assumptions required for its used are unclear, so we investigate this question further. We explore the performance of MI and MPA both theoretically and using a simulation study. We thus provide guidelines to perform a PS analysis in presence of missing data.