On 15 Jan 2003 06:04:49 -0800, [EMAIL PROTECTED] (Lizzy) wrote: > With a very large sample size (n=15000), what is the validity of > secondary outcomes if sample size calculations were based on the > primary outcome only(death)and the primary outcome itself does not > show differences between treatment arms?
I can think of "secondary outcomes" as denoting (say) symptoms of the disease. In that case, the failure of the "primary" question probably undermines every secondary conclusion. Outcomes would have to be "exploratory" and "speculative" even if the p-levels were impressively small. I have also thought of "secondary outcomes" as more independent than that. The hypotheses were secondary, perhaps, in terms of justifying the funding for the study, but where rather extrinsic to the other ideas. - That is relevant, I think, in justifying how one splits the overall error. We do want to protect a 5% error level on an experiment, for instance. On the other hand, we do want to save a passel of money, by tossing in an extra data-collection form or two, and testing an extra hypothesis or two -- instead of re-running the experiment, at a cost of millions, in order to preserve this theoretical virtue. You have to figure this out. > > One to four out of 13 secondary outcomes/ outcome subgroups showed > significant difference between treatment arms. Can one state that > these data are valid, "hard" data, taking into account the very large > sample size? No adjustments for cut off point for p-values where done. If these were 13 parallel, equal hypotheses, then you do *need* to control for multiple tests, or else, label it as "exploratory", that whole portion of the followup. Huge sample size has a drawback, in that it can elevate tiny biases into being 'significant' effects. That is especially important if this huge N was observational, and not a randomized trial; for convenience samples, 'significant' might indicate that, yes, the outcome was *interesting*; now it is time to explain away all the other possible causes. (That is very difficult when the literal effect is small.) > This regards the clinical interpretation of data from one large drug > trial. - Large trials usually have large planning, and that should have included statements about the philosophy of dealing with multiple outcomes. Epistemology of the epidemiology... -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
