Replies to post #1425 on Dendreon Corp fka DNDNQ

[p3analyze]

Thanks Rancherho, I just responded to your IV post.

Now, what Gametheory on this board said make a lot more sense, confirmed by io_io that one interim analysis will be done in 2007. But when and what's the likelihood of success?

Most likely it will be done after FDA approval. Assuming what I projected correct (if my assumptions are wrong, the numbers could be wrong), but for the sake of argument, there seems to be concensus that the target event goal is probably in the low 60%, although I think it could be lower. Nonetheless, suppose there would be 200 events at end of June 2007, the information fraction would be 0.67, teh OBF 2-sided alpha would be 0.012, at a hazard ratio of 1.8 (per the pooled cox regression), the probability of terminating early is 95% and declare success.

I have to slap myself to remind me not to get ahead of myself. But wow, that's scary.

[poorgradstudent]

>In the 9901/9902a trials, treatment effect was the most statistically significant.<

If this is true, then it's very impressive that Provenge beat out mets... usually that is the most statistically significant covariate for survival in cancer trials.

[iwfal]

The statistical significance of 9902b is to be determined by the Cox regression analysis where we know the pooled 225 patients 9901/9902a p value was 0.0006 vs. the Kaplan Meier p value of 0.011, in spite of a placebo crossover rate somewhat in excess of 70% It would seem to me that if one assumes the same treatment effect and event rate for 9902b as for the pooled 9901 and 9902a enrollees, the target p value threshold of 0.05 would be reached with substantially fewer events than if a Kaplan Meier derived p value was needed. Would this be true?

It is absolutely true that Cox Regression increases the power of the trial. But it isn't a huge amount. My guesstimate(!!!) is by the equivalent of about 20% more patients. (I.e. in this case the 9902b power would be about the same with 600 patients without Cox Regression). Note also that while, on average, Cox Regression is more likely to help your trial than hurt it, there will certainly be instances where it causes a good p value to pop over 0.05.



trials were stopped short of full planned enrollment due to unexpectedly high treatment effects, a caveat being that statistical data can have spikes during the trial period and stopping a trial short of full enrollment could be precipitated by such a spike.

This would almost always (I know of no exceptions - but won't swear there are none) be only per pre specified trial protocol. I.e. Interim looks by the DMSB can only stop a trial for bad events unless it is pre-specified by the company that they are willing to spend some p on the interim look.

[iwfal]

In the 9901/9902a trials, treatment effect was the most statistically significant.

This is not correct. PGS is correct that in 9901 and in 9902a bone mets was the strongest prognostic (tied with age in 9901). (I have no idea about 9901 combined with 9902a - but I'd be surprised if bone mets wasn't again the strongest prognostic. It probably has to be mathematically.)

[io_io]

rancherho:

"<<<It would seem to me that if one assumes the same treatment effect and event rate for 9902b as for the pooled 9901 and 9902a enrollees, the target p value threshold of 0.05 would be reached with substantially fewer events than if a Kaplan Meier derived p value was needed. Would this be true?>>>"


Seeing as no-one replied to this point, might I add that even more obviously you could assume a much lower treatment effect, on the order of 9902a's (log-rank) would be generally sufficient in a trial this size.

Of course another way of saying this is that the target (assumed) HR for success has been lowered.