<<<Iwfal on this board and Clarksterh on the I-Village board have posted that a Cox regression analysis performed on a trial whose arms are either well balanced or slightly favoring the placebo arm improves the p value by much, much more than one would expect.>>>
I've seen this, and I have a comment and then I have 2 questions.
My comment would be that going from 0.01 to 0.002 is not a quantum leap (sorry mr. quantumdot if you're reading). After all, it's the threshold that counts.
Now my first question is, seeing as the log-rank and cox-R are 2 totally different methods, is Iwfal sure his analysis is universal generally to survival trials, and general to all kinds of analysis ?
I mean you have big trials and little trials. But you also have trials like ALTH's where the patients die very fast, so that they precede many censored patients some of whom will live to durations 20 times longer. Then you have others where most patients live long (dorb's for example) and very, very few die early.
I would suspect (but it's a long while since I did a cox-R, and it had nothing to do with biotech) that if the cox-R was universally favorable, then all the well-paid biostatisticians would know that as a fundamental, and would push for it every time.
Finally, log-rank v cox-R for trials like 9901 & 9902a, what happens to the HR in these circumstances (i.e. does it also generally get better even if balanced) ? Because obviously the HR is more important than the p-value, once you're significant.
"....on the biotech battle-field, you need some élan...."
