Replies to post #19754 on Biotech Values

[iwfal]

If Cox lowers the p-value from 0.33 to 0.02, then the trial is ipso facto not intelligently-designed and well-balanced.

Ok, I think we are reaching the essence. If there existed a tool that incontrovertibly compensated for imbalance then there is no need to run a 'well-balanced trial'. Cox Regression is, to a large extent, that tool. Your position is analogeous to saying you must design a transmitter system with S/N of twice Shannon's limit despite having Turbo Coding. I assert that it's passe. (Yes, I am definitely exagerating. Apologies. Just trying to make it clear.)

A measure of this is exactly what I said before - look at a correlation of size of Cox Regression to correctness of corrected HR. I'll bet they are not highly correlated.

(Another measure is to see how much Cox Regression overcorrects - simulate a trial, add Cox Regression noise, perform Cox Regression. I think that there is a good chance that in general Cox Regression produces a better HR than the original data before Cox Regression noise was added.)

[rkcrules2001]

DNDN -- Question for FDA mavens...

Dew's views (empahsis mine):
<<Has there ever been a trial accepted by the FDA as legitimate support for a marketing application where Cox lowered the p-value from 0.33 to 0.02? I don’t think there has been and I doubt there ever will be.>>

And:
<<If I were a reviewer at the FDA or another regulatory body, I would tell any company that comes to me with a raw p-value of 0.33 and a Cox-adjusted p-value of 0.02 to redo the trial and come back with clean data.>>

Are there examples of FDA rejecting such a trial (ie, one relying so heavily on Cox analysis) for a life-threatening disease?

And I suppose more accurately, examples of FDA rejection due to reliance on pre-specified Cox analysis? (In the last DNDN SEC filing, "Cox" analysis is discussed four times, and every time it is described as "pre-specified Cox multivariate regression analysis", presumably emphasizing that it was not post hoc.)