Missing data imputation in two phase III trials treating HIV1 infection

In most longitudinal clinical trials, some patients drop out before the end of the planned follow-up, and, in order to allow an all-patient intent-to-treat analysis to be performed, it is common practice to use some method of imputation to estimate values for missing data. However, different imputation methods may provide different results, and it is essential to investigate the sensitivity of the analysis using different imputation rules. In our analysis of two trials of the new HIV1 fusion inhibitor enfuvirtide, we compared some standard methods of imputing and analyzing HIV1-RNA data with two novel alternatives, to check the robustness of the primary endpoint results. The standard methods were: (1) last-observation-carried-forward, (2) baseline carried forward, and (3) multiple imputation. These were compared with a nearest-neighbour hot-deck method, specifically proposed for imputation of missing HIV1-RNA data, and with a heuristic approach: censored regression analysis of the last-observation-carried-forward. To supplement this analysis of real clinical trial data, we investigated the performance of the same imputation methods on simulated datasets designed to cover a broader range of missing data patterns.