Is it common or even methodologically correct (or when?) to report demographic characteristics separately for responders and non-responders (assuming they are clearly defined), in addition to those for the overall study population?
By the usual definition of “responder”, any use of responder analysis is highly discouraged as discussed in detail https://discourse.datamethods.org/t/responder-analysis-loser-x-4.
This is particularly problematic when a table is constructed that reverses the roles of the dependent and the independent variables. When stratifying on the outcome (into columns of the table) one is not conditioning in the correct time order.
Thanks a lot. Apart from the challenge to make a convincing argument with the clinician scientists, most statisticians would probably agree that any sort of artificial categorization cannot be a good idea (will keep the link to the loser x 4 for future reference!)
My question in this case was more related to the second point you make about reversing the order, and actually for the case of observational studies rather than RCTs (though this is not so relevant in this case). Assuming we actually have a binary response (mortality following a surgical operation is probably the closest example I can think of…), I often see the sort of table you mention stratifying on the outcome, which however seems to be asking the question in the wrong direction (here a couple of examples: https://pubmed.ncbi.nlm.nih.gov/16473714/ ; https://academic.oup.com/jpubhealth/article/20/3/275/1499774). Should those tables always be avoided then, what should one report instead, a multivariable model of the outcome? and are there any references out there which may be useful for communication with the “clinician scientists”?
Great question. To me the answer is to show information that is both new and useful: how does each baseline variable relate to the outcome, with no binning. A proposal for doing this is here: https://discourse.datamethods.org/t/should-we-ignore-covariate-imbalance-and-stop-presenting-a-stratified-table-one-for-randomized-trials
This applies equally to randomized as well as to observational cohort studies. It is a common error in the literature and leads to thinking that is not so clear. There is only one sort of study design for which stratifying by the outcome (if it’s categorical to begin with): a case-control study.