Intent-to-treat analysis for clinical trials: use of data collected after termination of treatment protocol.

Following patients for a period of time after termination of treatment protocols is a common practice in clinical trials of drug treatments. After termination of the protocol treatment, patients are usually provided with medical care as deemed appropriate, e.g. different doses of the same drug, augmentation with other drugs, or different drugs. Data collected on patients when they are not in the treatment protocol are termed "off-treatment" data to be differentiated from "on-treatment" data that are collected while patients are in the treatment protocol. The purpose of the present paper is to describe some recent statistical methodological advances in the use of "off-treatment" data for intent-to-treat (IT) analysis using mixture models [Biometrics 52 (1996) 1002]. Two-piece spline models conditional on the protocol treatment dropout time are developed first. The two pieces of the spline model represent the on-treatment and off-treatment segments of data and are joined at the dropout time. The weighted average of the two-piece conditional models across realizations of the dropout times provides the mixture model for the pragmatic IT analysis. The weights are the estimates of the probabilities of the dropout times. The mixture model is amenable to an explanatory analysis which assumes that the patients remain on their assigned treatments. The model allows the parameters of the splines to depend on the dropout time. Statistical inference is based on bootstrap sampling procedures. We have illustrated this methodology using data from a drug trial comparing nortriptyline and paroxetine in the treatment of major depression in older patients.