Helmut ★★★ Vienna, Austria, 20210408 19:09 (1285 d 17:32 ago) Posting: # 22305 Views: 26,875 

Dear all, we are collecting data for a publication about the annoying ‘GroupbyTreatment interaction’ (for some thoughts see this article). Below what we have so far in a metaanalysis of multigroup studies: To make a long story short: As expected, significant GroupbyTreatment interactions were detected at approximately the level of the test (0.05) and below the upper 95% significance limit of the binomial test. Hence, based on our observations in wellcontrolled studies likely they are mere ‘statistical artifacts’, i.e., false positives. The Kolmogorov–Smirnov tests were not significant, accepting the expected standard uniform distribution of \(\small{p(G\times T)}\). At a recent BE workshop^{1} I had a conversation with a prominent member of the FDA. She told that the agency tests the GroupbyTreatment interaction at the 0.05 level, thus reducing the false positive rate observed at the previously used 0.1 level. But then she also said »We often see not only pvalues just below 0.05, but also some with 0.0001…« It’s a common fallacy to regard the pvalue as the probability that the null hypothesis is true (or the alternative hypothesis is false).^{2} The outcome of a level \(\small{\alpha}\)test is dichotomous: Hypotheses are considered true or false, not something that can be represented with a probability.Any type of comparative BA study (BE, foodeffect, DDI, doseproportionality) is welcome. Of course, data will be treated strictly confidential and not published . The preferred data format is CSV (though xls(x), ODS, SAS XPT or Phoenix project files would serve as well). Columns (any order is fine):
No cherrypicking, otherwise we will fall into the trap of selection bias and the outcome will be useless. Hence, if you decide to provide data, please do so irrespective of whether you ‘detected’ a significant GroupbyTreatment interaction or not. We are primarily working on 2×2×2 crossover designs. However, if you have data of replicate designs, fine as well. In HigherOrder crossover designs indicate which of the treatments is the test and the reference. If possible, give the analyte. Once we have enough data sets, we will perform subgroup analyses. So far we have only data of one multisite study. If you could share some data, great. THX in advance!
— Diftor heh smusma 🖖🏼 Довге життя Україна! _{} Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes 