Clinical Trials

❘❙❚■ Chapter 30 | Missing DataTable 2. Comparison of estimates of treatment effect for different missing data scenarios with three differentanalysis strategies. aMissing schemes Datasets Strategy for dealing with missing dataAll available data Completers only LOCFNo missing data I 10.4 (0.45) 10.4 (0.45) 10.4 (0.45)MCAR II 10.4 (0.58) 10.7 (0.63) 8.9 (0.51)MAR III 10.0 (0.58) 7.4 (0.56) 17.3 (0.60)MNAR IV 5.6 (0.61) 3.6 (0.59) 18.3 (0.59)aThe treatment effect is defined as the difference in mean response between groups A and B at the fourth visitas estimated from the mixed model analysis. The figures in brackets stand for the standard error of the treatmenteffect. LOCF = last observation carried forward; MAR = missing at random; MCAR = missing completely at random;MNAR = missing not at random.Analysis of all available data at the final visit in isolation would not yield unbiasedestimates. Analysis of completers only, whether through the repeated measurementsmodel or through simple comparison of means at the final visit (Table 1),underestimates the treatment effect. The reason for this is that participants with‘low’ levels tend to be lost to follow-up, and there are more such losses ingroup A. In contrast, LOCF overestimates the treatment effect because LOCFtakes no account of the general increase in levels with time, with this effect havinggreater impact in group A, where there are more losses to follow-up, than ingroup B.Not surprisingly, when there is a large amount of MNAR data (Dataset IV),none of the methods provides a reliable estimate of the treatment effect.ConclusionWhen some values are missing in a randomized clinical trial there are implicationsfor both efficiency and bias. The extent of the bias in the analysis depends verymuch on the cause of missing data in the trial in question. In the absence ofspecific knowledge about the reasons why data are missing, a graphicalexploration of the pattern of missing data is likely to give the analyst some hints.There is a large amount of statistical literature on the analysis of studies withmissing data. There is a growing consensus that ad hoc imputation methods suchas LOCF should be avoided. When data can reasonably be assumed to be MCARor MAR, unbiased methods of analysis do exist. In the latter case, it is usuallynecessary to carry out an appropriate analysis of all the available data at all follow-350