Primary Submission Category: Propensity Scores
Handling covariate missingness in propensity score weighting with clustered data: A comparison of multiple imputation and complete-case analysis
Authors: Xiao Liu,
Presenting Author: Xiao Liu*
Propensity scores (PS) are commonly used for causal inference of treatment effects. Two issues can complicate PS analysis: clustered data structure and missing data. Methods for handling either issue alone in PS analysis have been studied. Methods for PS analysis when both issues exist have been under-evaluated. This study considers PS weighting for average treatment effect estimation with clustered data where individual-level covariates contain missingness and (fully) unobserved cluster-level confounders exist. Simulation studies were conducted to compare different missing data methods (complete-case, single- or multi-level multiple imputation), PS models (random- or fixed-effects models), weighting (marginal, clustered), and outcome models (nonparametric, regression on treatment and random or fixed cluster intercepts; sandwich standard errors were used for methods with the nonparametric and fixed-effect outcome models). When the missingness was due to moderators (which was the unobserved cluster-level confounder in our simulation), complete-case analysis produced biased estimates. Single-level imputation performed well in some conditions but did not in a majority of conditions, particularly when dependence among variables used in the imputation was low. Overall, multilevel imputation, fixed-effect PS, clusterered weighting, and outcome regression on treatment and random cluster intercepts appeared perform well in bias, RMSE, and coverage across the simulation conditions.