Hi, Dr Mircea Zloteanu! Let me one question. It seems like that all those adaptive designs that allow early completion of the study on an interim analysis due to proven effectiveness are very questionable? Moreover, for intermediate analyses, we usually use an alpha spending function, and so the alpha threshold is set more strictly than usual. So, in the case of "success" in the interim analysis, are we really catch only some completely wild variations and outliers?
Interm analyses (and sequential designs more broadly) are fine if you follow what the analysis plan requires. Those alpha functions operate on the assumption you collect sufficient data at each stage. If as in my examples you don't, then they do not guarantee the error rates you claim. What i would suggest if sample size is an issue (for financial or logistic reasons) to consider other approaches, e.g. e-values, bayesian, likelihoodist, or a compromise analysis (alpha/beta best ratio) if you need to be NHST.
Hi, Dr Mircea Zloteanu! Let me one question. It seems like that all those adaptive designs that allow early completion of the study on an interim analysis due to proven effectiveness are very questionable? Moreover, for intermediate analyses, we usually use an alpha spending function, and so the alpha threshold is set more strictly than usual. So, in the case of "success" in the interim analysis, are we really catch only some completely wild variations and outliers?
Interm analyses (and sequential designs more broadly) are fine if you follow what the analysis plan requires. Those alpha functions operate on the assumption you collect sufficient data at each stage. If as in my examples you don't, then they do not guarantee the error rates you claim. What i would suggest if sample size is an issue (for financial or logistic reasons) to consider other approaches, e.g. e-values, bayesian, likelihoodist, or a compromise analysis (alpha/beta best ratio) if you need to be NHST.
Dr Mircea Zloteanu, thank you very much!