Hui Xie

In many clinical trials there is a possibility that some subjects will cross over between treatment arms. To evaluate the effect of such crossover on key inferences, one can model this phenomenon as a missing-data problem, where for subjects who cross over one treats the unobserved value of the outcome in the original randomization group as the missing data (see Heitjan 1999 Controlled Clinical Trials 20:309; Heitjan 1999 Statistics in Medicine 18:2421). The as-treated estimate of the treatment effect is biased if the crossover is nonignorable, in the sense that the crossovers represent a non-random sample of the randomized subjects. A recent area of general methodologic interest in biostatistics is the development of methods for measuring the sensitivity of inferences to nonignorability in the missing-data mechanism; one such approach is the method of Troxel, Ma and Heitjan (2002, unpublished MS). In this paper we apply their method to the problem of measuring sensitivity to nonignorable crossover in randomized trials. A drawback of their approach is that it assumes a single nonignorable selection model that applies equally to all patients, whereas in clinical trials we might expect each arm to have its own nonignorable selection mechanism. Accordingly, we extend the Troxel method to account for the possibility of multiple nonignorability mechanisms. We conclude that although the as-randomized analysis is always preferable, with our method one can delineate circumstances under which the as-treated analysis may be more or less sensitive to nonignorable crossover. We apply our method to the evaluation of sensitivity to nonignorable crossover in a clinical trial of chemotherapy for the treatment of multiple sclerosis. This is a joint work with Dr. Daniel F. Heitjan, Department of Biostatistics & Epidemiology, University of Pennsylvania.