Multiple imputation compared with some informative dropout procedures in the estimation and comparison of rates of change in longitudinal clinical trials with dropouts.

Statistical analysis based on multiple imputation (MI) of missing data when analyzing data with missing observations is gaining popularity among statisticians because of availability of computing softwares; it might be tempting to use MI whenever data is missing. An important assumption behind MI is the "ignorability of missingness." In this paper, we demonstrate the use of MI in conjunction with random effects models and several other methods that are devised to handle nonignorable missingness (informative dropouts). We then compare the results to assess sensitivity to underlying assumptions. Our focus is primarily to estimate and compare rates of change (of a primary variable). The application dataset has a high dropout rate and has features to suggest informativeness of the dropout process. The estimates obtained under random effects modeling with multiple imputation were found to differ substantially from those obtained by methods devised to handle informative dropouts.