512 – Various Flavors of Missing-Data Problems
Towards Multiple-Imputation-Proper Predictive Mean Matching
Philipp Gaffert
GfK SE
Florian Meinfelder
Otto-Friedrich-University
Volker Bosch
GfK SE
Rubin 1987 introduced multiple imputation in a parametric Bayesian framework and considers it proper if the uncertainty of the imputation fully propagates. Augmenting it with a semiparametric concept like predictive mean matching (Rubin 1986, Little 1988) promises both, valid inferences and robustness against some model misspecifications. Although numerous multiple imputation predictive mean matching algorithms exist their theoretical properties remain largely unexplored. In this talk, we show why all of these algorithms are improper, but the one by Siddique & Belin 2008. On this exception we build a new algorithm and demonstrate its superiority in terms of coverages of frequentist confidence intervals within a comparative simulation study.