In this paper, the authors discuss a more fundamental issue that of quantifying the treatment difference for benefit/harm. In particular, they emphasized that an appropriate summary of the treatment difference preferably has the following features:
- The summary is clinically interpretable, ideally in layperson’s terms, and is accompanied by an appropriate summary of the endpoint in each treatment group.
- The summary of the treatment difference does not have modeling constraints, and the corresponding inference procedures are robust and model-free for drawing valid conclusions.
For feature 1, an intuitive and clinically interpretable treatment summary is essential for allowing clinicians and patients to make better treatment selection decisions at the end of the study. For feature 2, if the treatment difference is defined via a model, but at the interim or final analysis the model does not fit the data well, then it is unclear what conclusions to draw from the study, or how. If a poorly fitting model suggests a treatment difference, is that due to evidence for a treatment difference or due to lack of model fit? Moreover, there are no satisfactory analytical procedures for assessing the adequacy of the assumed model. If possible, one should avoid modeling at the design stage. For the same reasons, at the analysis stage, the inference procedures used to draw formal conclusions about the treatment difference should be model-free.
For illustration, the authors present three examples from recent clinical studies for treating patients with COVID-19, cancer, and heart failure. These examples discuss issues with model-based quantification of the treatment difference, and present simple, model-free alternatives that are easier to interpret clinically. The authors encourage the investigators to consider the principles of validity, interpretability and robustness when designing and analyzing future studies.More Details