Missing radiographic data handling in randomized clinical trials in rheumatoid arthritis.

In recent years, there has been increasing interest in compounds that have potential to slow down the structural joint damage in rheumatoid arthritis (RA) patients. Radiographs are instrumental in assessing structure damage in RA. Radiographic analyses results have become essential in establishing a "delay in structural progression" claim in newly developed agents for the treatment of RA. It is well known that the radiographic progression data generally follow a nonnormal distribution that is loaded with excessive zeros. A special concern about the radiographic data analyses is the handling of the seemingly high rate of missing values due to dropout or unreadable images. There are no uniform ways to handle missing radiographic data, and such data usually show considerable sensitivity to the imputation method chosen under the complexity of the nonnormal data and the unique missing mechanism. In this research, we proposed both an innovative multiple-imputation algorithm and a novel method called the mean rank imputation method under the nonparametric framework for sensitivity analyses. A simulation study was designed using rank analysis of covariance (ANCOVA) to extensively assess and compare the finite performance of these two new methods along with four other missing data handling methods previously used in the RA trials, namely, linear extrapolation, last observation carried-forward (LOCF), median quartile bin imputation, and median imputation under various settings. Our simulation results suggest that the multiple-imputation algorithm, providing an mITT analysis population, yields an inflated type I error and artificially good power. The proposed mean rank imputation method, following a true ITT principle, both is powerful and maintains type I error at the nominal level.