Replies to post #19442 on Biotech Values

[DewDiligence]

> But if I remember iwfal's comment correctly, even correcting for factors that were not well correlated to survival could reduce the "noise" in the trial and improve the P value. That sounds to me like you could always clean up the P value a bit by doing the Cox analysis (?)<

First, I think iwfal overstates the “noise cleanup” thesis, especially in a trial with as many as 500 patients. Second, picking achievable endpoints and analyses is as much a part of the negotiation process with the FDA as is setting the trial size, eligibility requirements, and so forth. Companies don’t get extra p-value for picking dumb endpoints, and they don’t forfeit p-value for picking smart ones.

[walldiver]

<<But if I remember iwfal's comment correctly, even correcting for factors that were not well correlated to survival could reduce the "noise" in the trial and improve the P value.

That sounds to me like you could always clean up the P value a bit by doing the Cox analysis (?)>>

I think iwfal's point was, if the trial was well balanced, Cox analysis would improve the p value every time. I'm not so sure about that point, but it did seem that DNDN's 9901 Phase III was fairly well balanced but slightly biased against the treatment arm. The p value went from log rank 0.01 to Cox 0.002, quite a large improvement for a slight imbalance.

However, I believe the overall survival p value for the ITT group in the tesmilfene Phase III went from 0.02 log rank to 0.03 using Cox, so it does go the other way sometimes.

As for 9902A, it was extremely biased against the treatment arm, as interpreted from the 14.4-fold improvement from log rank to Cox regression (0.332 to 0.023).

Dew, lol, I'm not a DNDN insider, nor do I have a magic whisper pipeline to the execs. I just made a few predictions that came true several months later, such as 9901 turning out to be stat sig on survival for the ITT group and DNDN opening up 9902B to all Gleasons. I also missed on some predictions, i.e. by a large margin on 9902A's log rank p value. I was also wrong inre DNDN trying to preserve TTP as the first chronological endpoint in 9902B. We know now that the primary endpoint will be survival and the unblindings will be event-driven, and no, I don't know how the p value endpoints will be allocated and I don't know the # of events that will trigger the unblindings. I will take a guess that unblinding #1 will be after 250 deaths and unblinding #2 will be after 375 deaths.

I'm also not very concerned about Provenge not achieving stat sig in the IMPACT (9902B) trial. The Cox p value for 9901/9902A was 0.0006 or 0.0007 for only 225 patients. There will be 500 patients in IMPACT.