ABSTRACT
A common logistical problem with sequentially outcome-adaptive clinical trial designs arises from the fact that the outcomes of treated patients must be observed soon enough to apply a design's decision rules to choose treatments for new patients. In phase I–II trials, the late-onset outcome problem arises if either Toxicity or Efficacy is not scored sufficiently soon, relative to the accrual rate. The data structure assumed by most phase I–II dose-finding methods consists of assigned doses and two binary outcomes. A central concept for the statistical analysis of missing data is the missing-data mechanism, which determines what type of method is appropriate to use. In terms of robustness, the missing-data approach is preferable to the alternative approach that directly treats Toxicity/Efficacy as time-to-event outcomes and handles unobserved delayed outcomes as censored observations.
