ABSTRACT
In clinical trials, it is common practice to categorize subjects as responders and non-responders on the basis of one or more clinical measurements under pre-specified rules. Such a responder analysis is often criticized for the loss of information in dichotomizing one or more continuous or ordinal variables. It is worth noting that a responder analysis can be performed without dichotomization, because the proportion of responders for each treatment can be derived from a model for the original clinical variables (used to define a responder) and estimated by substituting maximum likelihood estimators of model parameters. This model-based approach can be considerably more efficient and more effective for dealing with missing data than the usual approach based on dichotomization. For parameter estimation, the model-based approach generally requires correct specification of the model for the original variables. However, under the sharp null hypothesis, the model-based approach remains unbiased for estimating the treatment difference even if the model is misspecified. We elaborate on these points and illustrate them with a series of simulation studies mimicking a study of Parkinson’s disease, which involves longitudinal continuous data in the definition of a responder.
Acknowledgements
We thank an anonymous reviewer for constructive comments that helped improve the manuscript. The views expressed in this article do not represent the official position of the U.S. Food and Drug Administration or the U.S. Department of Defense.