Abstract:

Often in a clinical trial there are a number of patients whose outcomes are missing. If the missing patients differ in important ways from the patients whose outcomes are observed then results based on the observed patients could be misleading. If the outcomes are binary then a simple "tipping point" plot of the potential results of the trial as a function of the outcomes of the missing patients can be informative. In simple situations where large numbers of multiple imputations can be computed we can produce an estimate of the predictive distribution of the missing data, conditional on the model and the observed data. This distribution can be used to enhance the tipping point plot in several ways. In addition, by varying the model we can investigate the sensitivity of the results to the choice of model, and by averaging over many models we can expand the sensitivity analysis beyond the usual comparison of MCAR versus MAR versus specific MNAR models. All of this can be displayed on the tipping point plot in a natural way. An R package is available to carry out the imputations, obtain the predictive distribution, and create the tipping point plots for clinical trials with binary
