The more talented a biostatistician is in the practice of mathematics, the less likely they are to understand the clinical impacts of their theories.
All technical professions have this issue. Statistics is, sadly, just one example. Hence the classic stereotype of the genius geek (a geek being someone with no common sense).
But there are exceptions - e.g. many physics Nobel Laureates are people with mathematical ability AND common sense. And others with this dual capability, even if they aren't nobel class, are often the people promoted rapidly in Corporate America (my universe).
The really sad thing here is that companies could save a lot of time and money if they understood these things - proper stratification, Cox Regression, ... . I would guess that proper use of the tools can decrease the cost by 25% over no Cox Regression, no stratification. Baysean methodologies might save another 50% (WAG).
The triangle test is a prime example where the biostatistician who created it never stopped to think how it would compare to a log rank system if the trial enrolled much faster than exoected.
Given that: if you give me trial size, number of events and HR I can tell you approximately the p value then the Triangle Test is really no different than any trial with predefined interm alphas and futility limits. The triangle test is different in only two ways - it pre-specifies the alphas assigned to each interim and it prespecifies the futility limits. (If the FDA allowed Baysean trials then tests like this COULD assign more alpha if they lower the limits at which futility is triggered. But it see my below suspicions.)
I suspect that in total YMI actually got less than a total of 0.05 alpha across all their interims. I.e. they got screwed, but because it was in HR space, instead of "p value" space they didn't know it. They would have been better off just using standard alpha assignment.
I will convert the YMI HRs and event numbers into approximate p values and let you know.
A restatement because of known communication issues: If you give me a trial size, a number of events and the resulting HR I will tell you the p value with fairly tight tollerances. So a triangle test (with predefined HRs vs events) is not conceptually different than assigning alphas in the traditional way. It just provides guidance on how to do it.
