>> The p-values you cite for 9901 and 9902a are based on cherry-picked Cox analyses. <<
True. My intuitive analysis technique is not without flaws , I'll admit. Still , even just looking at the integrated log-rank P=.011 , obtained with fewer events , and assuming at least some plumping of the data by Cox analysis , it seems like a good bet.
Another consideration , I suppose , is whether one believes Provenge works. If Provenge works , then the extra events will be meaningful in increasing the odds of success at the interim look.