Estimands: what they are and why we should use them
Brennan C Kahan, Declan Devane
Abstract
In clinical trials, postrandomization events, such as treatment discontinuation or the use of rescue medication, can complicate the interpretation of results. An estimand is a precise description of the treatment effect that investigators wish to estimate. Estimands facilitate more straightforward interpretation of trial results by explicitly defining how postrandomization "intercurrent" events are incorporated into the research question. This article introduces the five key attributes of estimands (population, treatment conditions, endpoint, summary measure, and strategies for intercurrent events) and explains the five main strategies for managing intercurrent events (treatment policy, composite, while on treatment, hypothetical, and principal stratum). Using a practical example of a trial comparing cognitive behavioral therapy vs medication for mild anxiety, we demonstrate how different estimand choices lead to varying study designs, analyses, and interpretations. Understanding estimands helps researchers design better trials and enables stakeholders to determine if the results are relevant to their situation. We also explain how sensitivity analyses can be used to check the reliability of results by assessing how results change under different statistical assumptions.