Replies to post #19746 on Biotech Values

[iwfal]

In regards to your final paragraph, walldiver raised an important point a few days ago: Cox can in some cases result in a p-value that is worse than the “raw” number. When that happens, companies are of course apt to tout the “raw” number as the one that really matters.

I agree absolutely that you need to prespecify the analysis method. That, however, is not an indictment of Cox Regression - on the contrary, a prespecified Cox Regression is probably generally a better indicator of efficacy than a raw log rank. (BTW - I believe that a Cox Regression that turns out to be worse than the Log Rank absolutely always means an imbalance significantly favoring the treated group. I.e. the raw log rank is probably significantly wrong.)

My contention is that Cox should be thought of like the fine-tuning knob on your old-fashioned radio. From a regulatory standpoint (as opposed to a theoretical one), it’s OK to run a pre-specified Cox analysis that improves the p-value modestly.

However, when Cox improves a p-value from 0.33 to 0.02, as it did in the case of DNDN’s 9902a trial, the data set is fatally flawed.

You are sounding an awful lot like that company you just dissed that didn't like the Cox Regression, so it touted the raw. -g- Either the methodology of Cox Regression is valid or it isn't regardless of the amount of change in p value. It might be true that Cox Regression, as a semiparametric method, is worthy of some(!!!) penalty, but it probably won't be based on p value size. And it is clearly not correct to throw out a prespecified Cox Regression entirely just because it generally reduces noise - in the comm world if that were done we'd be back in the late 1800's.

Final note, given that I agree that it is possible that a prespecified Cox Regression might deserve a penalty, it is highly unlikely that it would be sufficient to invalidate 9902a as supporting. I wouldn't want to submit 9902a as the primary trial, but as supporting evidence I think it is reasonable (only because they used the same covariates found in 9901).