Intention-to-treat with drop-out.
Collaborative Research Center 386, Discussion Paper 199
We consider the problem of an ITT analysis in a randomized clinical trial. Due to (study) drop-outs, standard methods are not applicable and simple imputation methods like LOCF (last observation carried forward) may lead to biased results. Since a patient who drops out of the study often will also change or drop the assigned treatment, an "ignorable" analysis in the sense of Rubin (1976) assuming MAR (missing at random), as e.g. a propensity weighted analysis or a likelihood based MAR-analysis is not valid. This is due to the fact that information is missing about outcomes as well as the covariate treatment after drop-out. That is, even if the drop-out process itself is ignorable, we can not treat the problem as ignorable because of the missing covariate information. We follow the path given by Little and Yau (1996), who created multiple imputations under various assumptions about the actual treatment after drop-out, and conduct a simulation study on the alpha-error and power of simple endpoint tests. This should also shed light onto the problem whether the true treatment effect can be sensibly bracketed by assumptions like zero dose or continuing dose after drop-out.