Abstract:
|
Recent interest in the topic of estimands is motivated by heightened realization that in familiar ITT approaches, failure to account for post-randomization events can disconnect analyses from issues of main importance. As a simple example, a treatment inducing a high degree of rescue medication could potentially result, in effect, in comparison versus the rescue and thus misalign with a study's true objective. At the same time accounting for post-randomization events can introduce bias if mechanisms are not fully understood, or if implemented inappropriately. We illustrate relevant concepts in the setting of trial design for a particular condition causing pain. Data spans a multi-week period, and a number of post-randomization events could potentially confound questions of interest: intake of rescue therapy or prohibited medication; changed standard of care administration; treatment or study discontinuation, for lack of benefit or AEs. Missing data is expected and proper methods must be applied. In this setting we discuss estimand definitions that could be considered, corresponding analytic approaches and strengths and cautions regarding decisions which might be taken.
|