How robust are multiple imputation based inferences in multilevel models?

Conference of the European Survey Research Association (ESRA), Lisbon, Portugal

Abstract

Predictive mean matching (PMM) is a state-of-the-art hot deck multiple imputation (MI) procedure. The quality of its results depends, inter alia, on the availability of suitable donor cases. Applying PMM in small sample scenarios often found in psychological or medical research could be problematic, as there might not be many (or any) suitable donor cases in the data set. So far, there has not been any systematic research that examined the performance of PMM, when sample size is small. The present study evaluated PMM in various multiple regression scenarios, where sample size, missing data percentages, the size of the regression coefficients, and PMM’s donor selection strategy were systematically varied. Results show that PMM could be used in most scenarios, however results depended on the donor selection strategy: overall, PMM using either automatic distance-aided selection of donors (Gaffert, Meinfelder, & Bosch, 2016) or using the nearest neighbor produced the best results.