Keywords: tipping point analysis
It is critical to evaluate the sensitivity of conclusions from a clinical trial to potential violations in the missing data assumptions of the statistical analysis. Sensitivity analyses should not consist of a few methods that might have been reasonable alternatives to the chosen analysis method, nor should they explore only a limited space of violations in the assumptions of the analysis. Instead, sensitivity analyses should systematically and comprehensively explore the space of possible assumptions to evaluate whether the key conclusions hold up under all plausible scenarios. In a randomized, controlled trial, this can be achieved by tipping point analyses that vary assumptions about missing outcomes on the experimental and control arms in order to identify and discuss the plausibility of scenarios under which there is no longer evidence of a treatment effect. In this talk we introduce a simple, novel tipping point approach in which inference on the treatment effect is based on the observed data and two sensitivity parameters, with minimal assumptions and no need for imputation. The sensitivity parameters to be varied are the mean differences between outcomes in dropouts and outcomes in completers on each of the two treatment arms. We note that the proposed test statistic has good asymptotic properties, and illustrate the utility of such an approach with two examples of drug reviews in which the methodology was utilized to inform regulatory decision-making.