There's a whole statistical literature on survival bias. The basic equations all require "independent and identically distributed" data loss. As soon as that's

There's a whole statistical literature on survival bias. The basic equations all require "independent and identically distributed" data loss. As soon as that's not the case (e.g., more placebo patients drop out than treatment patients), you have to compensate for that. That being said, I suspect this analysis is operating at a higher level where survivor bias isn't as relevant, because it's looking across many studies at once.