Abstract:
|
Missing data in cluster randomized trials are often handled with parametric multiple imputation (MI), assuming multivariate normality and using random effects to incorporate clustering. Since data do not always satisfy this assumption, a nonparametric approach to MI is desirable. Predictive mean matching (PMM) is a nonparametric approach where missing outcomes are imputed with observed outcomes in the data from donors that are similar to the missing cases. It is not clear how best to extend PMM to multilevel data. Two possibilities are to ignore clustering in the imputation model or to include fixed effects for clusters. In parametric MI, ignoring clustering in the imputation model leads to underestimation of the MI variance, while including fixed effects for clusters tends to overestimate the variance. A mixed effects imputation model can be used as the basis for matching, but this is computationally intensive and increases reliance on distributional assumptions. To simplify computation and reduce bias in the estimated variance, we investigate a weighted PMM approach that incorporates both the fixed effects imputation model and the imputation model that ignores clustering.
|