Can you elaborate on this some ?
P value vs trial size follows (approximately) an exponential curve. So if 9902a had had the same HR as 9901 it would have had a p value in the neighborhood of 0.018 to 0.025. But it had a p value of 0.33. I.e. the trial size was a miniscule contributor to the failure of 9902a.
Say a large trial achieved the same 3-yr. survival figures as 02A , 32% vs 21%. That's a 52% relative improvement in 3-yr. survival. Surely that would yield high statistical significance at some reasonable trial size , at least by landmark-type tests , no ?
Yes, but the 9902b trial is not really all that big since, to first order, the trial size is a function of the smallest arm. In the case of 180 deaths you'd expect perhaps 75 deaths in the placebo arm. Pretty small by comparison to other cancer trials. Also true that the 3 yr landmark p value will better than the HR p value because the separation only happened at the end - example: assume an HR test for the two following scenarios:
1) Curves separate immediately and dramatically with 79% of placebo patients dying within 12 months and then no one dying at all to 36 months - vs treated where no one dies until 35 months and then 68% die. At 36 months there would 32% vs 21%.
2) Curves are exactly the same until month 35 and then separate to 32% vs 21%.
In a landmark analysis the two would produce the same p value. But in an HR analysis the p value for scenario 1 would be a LOT lower than the second. And I believe that the HR p value for scenario 1 would be a lot lower than the landmark p - and for scenario 2 would be inverse would be true.
9901/2a is a lot more like #2 than #1. So the p value should be substantially higher than the 36 mo landmark p value.
Clark
