Basing Intention-to-Treat on Cause and Effect Criteria

Abstract

Statisticians often employ the intention-to-treat principle in clinical trials to address any possible bias if subjects are dropped from the analysis data set because of events that might be caused by treatment, for example, protocol violations or missing data. However, there are issues that occur in actual trials that lead to concern about and confusion in identifying a full analysis set for an intention-to-treat analysis. An approach clarifying the problems and suggesting solutions is to consider the finite population induced by randomization. Achieving a balance of observed data from a specific randomization and the unobserved data (counterfactuals) from other possible randomizations is the main principle. In some cases, subsets from the population of all treatment-subject cases can be identified as reduced populations, retaining the cause and effect structure. Corresponding reduced analysis sets, which are subsets of all randomized subjects, can be selected for inference to reduced populations. With a reduced analysis set, unbiased statistics for the reduced population's parameters may be constructed. A subset of all randomized subjects can be arguably superior for the analysis of cause and effect, since it is more consistent with the intended population and its scientific procedures and measurements.