Visit-driven endpoints in randomized HIV/AIDS clinical trials: impact of missing data on treatment difference measured on summary statistics

Abstract

In randomized HIV/AIDS clinical trials, CD4 lymphocyte counts and plasma HIV-1 RNA measurements are often used as endpoints. The comparison between treatment groups is mainly based on a summary measure of outcome, so-called summary statistic. Such analyses are often complicated by missing data occurring as drop-outs. For the most currently used summary statistics in these trials, we examined the impact of missing data occurring as drop-outs on test size, in order to help choosing between these statistics. A simulation of missing-data patterns was performed, using HIV-1 plasma RNA measurements as the main endpoint, to compare the effect of three plausible informative patterns, depending on treatment group, and on baseline or current plasma viral load, on eight different summary statistics. Missing data resulted in test sizes over the nominal value for the area under the curve minus baseline, the least-squares slope, the slope estimated with use of a mixed effects linear model, assuming a linear trend over the entire study, the difference between baseline and nadir, and the difference between baseline and week 24. The difference between baseline and week 8 was an acceptable summary with respect to the test size, but did not reflect accurately the durability of the effect of treatment. Two criteria appeared as the best summary statistics: the slope estimated by a mixed effects model, with a change of slope after two weeks of treatment, and to a lesser degree, the area under the curve after carrying forward the last observation.

CD4 T cell recovery is slower in patients experiencing viral load rebounds during HAART

“…Slopes of changes over time, mean changes at week 72 and confidence intervals were estimated from the models, using all available data except baseline. Piecewise linear models (two phases) were proposed when the improvement in data fitting was significant [14,15]. According to the published studies [8], we explored three possible junction points: W8, W24 and W36.…”

Enfuvirtide, an HIV-1 Fusion Inhibitor, for Drug-Resistant HIV Infection in North and South America

“…Although the last-observation-carried-forward method can overestimate the individual drug effect if the HIV-1 RNA level rebounds quickly after virologic failure, it was chosen to provide a consistent rule for handling patients in either group who discontinued treatment or had virologic failure, as well as for those who switched to enfuvirtide. 18,19 The robustness of the primary result was confirmed by three stringent sensitivity analyses that clearly showed that the magnitude of the estimate of the effect of enfuvirtide treatment was not determined primarily by the method of imputation.…”

Optimizing the design of clinical trials where the outcome is a rate. Can estimating a baseline rate in a run‐in period increase efficiency?

“…In practice, dropouts in the run-in period will carry a cost penalty and detailed comparisons might wish to take this into account. An algorithm was used to compute the necessary sample size for all possible designs involving up to five scans with scans potentially at baseline, 3,6,9,12,15,18,21,24,27,30,33 and 36 months. The 'optimal' design of a particular type is defined to be the one requiring the fewest subjects and this is presented assuming 10, 20 and 30 per cent annual dropout rates post-randomization.…”

“…Hence, the rate of brain loss is a natural outcome measure to use in a randomized controlled clinical trial in AD. Other settings in which a rate of change is a natural outcome include applications in cardiovascular disease [2] and AIDS [3] or indeed any situation where a linearly divergent treatment effect might be anticipated [4].…”