Replies to post #4795 on Dendreon Corp fka DNDNQ

[nerdseeksblonde]

Many people seem to think the FDA is
nitpicking on stats. However, the issue with taxotere is sill fishing- post hoc cherry picking if you will. There is no "alpha" criterion if you can pick stuff you like. My optimisim earlier was that they could allocate the variance- noise or deviation- to specific causes in a robust and
credible way ( make a causal argument to supplement the statistics) and eliminate many of the stats-only concerns about post hoc analysis. Taxotere synergy could be a fluke, or maybe not, but beating the data to death, not trying to gloss over stuff, was the only way to convince anyone. Why do we need to guess about the survival benefit in provenge-only patients? Aren't they advertising it? Scatter plots, clustering, etc any way to separate drug effects in any specific patients, was the way to go. With the placebo group, they had a way to validate their model beyond some simple summary statistics.

AFAIK, the FDA is looking at the overall evidence, of course in a rather subjective way, and would and did consider evidence beyond statistical hardlines.

[iwfal]

As a result, an interim look with a substantially lower number of events is not expected to be statistically significant, but if the sponsor desires an interim look (for example, to allow a cost avoiding termination of a trial for futility) and continues the trial, an alpha allocation is deducted from the normal statistical significance to account for having a second look, even though bias, post the interim look, should not be a problem in a survival trial.

This is part of the correct answer. The missing part is that the alpha is the thing they need to meet for the interim. It isn't about bias but about getting two bites at the apple.

Lets say you are testing a coin to see if it 'fair'. So you flip the coin 100 times and see if the number of heads is under p=0.05. What is the chance that a fair coin will produce such a p value? 5%.

Now for each flip use the same p calc method and calc p after each flip. What are the odds that a fair coin would at some point produce p<0.05? A lot higher than 5%. Because you had 90+ bites at the apple.

As for your questions about post hoc data mining - and the Petrylak's presentation was indisputably that - I don't really understand the question. But the chance that the FDA would give any credence to such a complicated post hoc analysis is zilch.

Clark

[p3analyze]

" why would the FDA have any problem allowing earlier termination and the filing of a BLA /NDA due to an unexpectedly efficacious treatment effect?"

If I understand correctly, FDA agreed with the pixantrone EXTEND adaptive trial design which allows the sponsor the potential to not only terminate early due to exceptional efficacy (should any of the 2 interim analyses have been wildly significant), but also the added flexibility of adjusting sample size (or size of the primary analysis set) for the final analysis. Today's release indicate that CTIC had good interim efficacy data, while not sufficient for declaration of statistical significance, likely due to low power (because only 100 subjects were analyzed?), but did allow them to reduce the final analysis set from 320 to a smaller number, so they can get the final result slightly earlier than if all 320 subjects had to be followed to become eligible for analysis of the primary endpoint - response rate.

Regarding the second question:
" how would a statistician decide whether the standalone use of Provenge in asymptomatic AIPC/HRPC would be approved by the FDA vs. the combination therapy of Provenge plus Taxotere, or both or neither? Could this have been one of the reasons for the FDA approvable letter".

FDA seems to be perfectly happy with the fact that secondary treatment with docetaxel was relatively balanced between the two treatment arms from 9901. So no, this did not contribute to the approvable letter.

If 9902b is successful and provenge is approved as monotherapy, then a possible design would be to compare provenge vs provenge+taxotere vs taxotere would answer the above question. If 9902b fails, then further provenge prostate trial almost has no choice but to do either a head to head provenge vs docetaxel (like vital-1), or a combo study provenge+docetaxel vs docetaxel alone (like vital-2).