Tutorials in Quantitative Methods for Psychology (Sep 2014)
Handling missing data in cluster randomized trials: A demonstration of multiple imputation with PAN through SAS
Abstract
The purpose of this study is to demonstrate a way of dealing with missing data in clustered randomized trials by doing multiple imputation (MI) with the PAN package in R through SAS. The procedure for doing MI with PAN through SAS is demonstrated in detail in order for researchers to be able to use this procedure with their own data. An illustration of the technique with empirical data was also included. In this illustration thePAN results were compared with pairwise deletion and three types of MI: (1) Normal Model (NM)-MI ignoring the cluster structure; (2) NM-MI with dummy-coded cluster variables (fixed cluster structure); and (3) a hybrid NM-MI which imputes half the time ignoring the cluster structure, and the other half including the dummy-coded cluster variables. The empirical analysis showed that using PAN and the other strategies produced comparable parameter estimates. However, the dummy-coded MI overestimated the intraclass correlation, whereas MI ignoring the cluster structure and the hybrid MI underestimated the intraclass correlation. When compared with PAN, the p-value and standard error for the treatment effect were higher with dummy-coded MI, and lower with MI ignoring the clusterstructure, the hybrid MI approach, and pairwise deletion. Previous studies have shown that NM-MI is not appropriate for handling missing data in clustered randomized trials. This approach, in addition to the pairwise deletion approach, leads to a biased intraclass correlation and faultystatistical conclusions. Imputation in clustered randomized trials should be performed with PAN. We have demonstrated an easy way for using PAN through SAS.
