Biotech Values

[iwfal]

First, I think iwfal overstates the “noise cleanup” thesis, especially in a trial with as many as 500 patients.

Maybe. I haven't done all the parametric analysis (got it all set up, but got distracted before I finished debug). But common sense says you are wrong. Noise is noise. Slicing the noise in half should always have a large effect on p value. Exactly how much? I'd guess if you have 2 covariates that have an HR greater than or equal to the HR you are trying to measure you get >2x improvement in p value. And that factor grows substantially as the covariates HR grows.

But, again, I didn't finish the analysis, so this is somewhat speculation based on some early calculations and analogy with the Chi Square example:

Imagine you have a PAD trial. 100 patients in each arm and at the time of measurement 40 had improved in the treated arm, but 30 had improved in the placebo arm (strong placebo response as usual in PAD). Chi Square p value of 0.138. Now regression fixes the numbers and makes it 20 improved in treated arm and 10 in placebo arm. P value improves to 0.048. A factor of 2.9. According to Dew this factor should get smaller when the trial size is increased. It does not. In fact it gets bigger. I.e with the same response rates in trial with 300 patients in each arm the uncorrected p value is 0.01 and the corrected p value is 0.006 or a factor of 17.

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.

I agree this seems most likely. If the analysis they are doing is not a data mining exercise. I.e. if they picked 3 (or 5) covariates that are highly acknowledged predictors of outcome and state that they will correct for all of them in a one pass multivariate Regression (i.e. as they did in 9902a). This is born out looking at Regressions accepted by the FDA in the past. No penalty was assessed. OTOH if they do the same data mining type Cox Regression that they did in 9901 (starting with 20 covariates, find which matter, and then doing the Regression) it seems much more likely that they would pay some kind of penalty. But not clear how that would be sized.

Despite their PR I think it unlikely that they will be doing what they did in 9901. It would be a huge change in policy for the FDA. I suspect that, at best, they get to do what they did for 9902a - correct for 5 (probably fewer since I've never seen an approved Cox Regression with more than 3) prespecified covariates.