Testing for differences in changes in the presence of censoring: parametric and non-parametric methods.

Some commonly used parametric and non-parametric methods for analysing repeated measures with incomplete observations are briefly reviewed. The performances of these methods in the presence of completely random, as well as informative censoring are compared in simulated experiments generated under the linear random effects model with parameter values derived from realistic examples. The effects of some moderate model deviations are also compared. The results indicate that in the presence of informative censoring, the usual parametric and nonparametric methods derived under the assumption of random censoring could either suffer severe loss of power or provide false positive results. The conditional linear model for informative censoring when used in conjunction with the bootstrap variance estimation procedure performed well under both random and informative censoring mechanisms. The non-parametric procedure obtained by ranking the individual summary statistics, although not as efficient as the conditional linear model with robust variance, also performed relatively well in most situations. Therefore, in situations in which informative censoring is likely to occur it is important to select the proper method of analysis to test for the informativeness of censoring and to account for its effects.