Dew - The meaning of my post might have been lost on you because you couldn't get past its flippancy. I'll try to be more serious.
I do agree with you that program-survival bias is a good filter for taking a first impression of any new dataset that seems more impressive than it should be. You could, in fact, argue that program-survival bias is a corollary of the "Regression to the mean" argument that Thomas Fleming brought up in his FDA letter against the D9901 and D9902a data in the BLA.
In your case, you would say that the dominant mode (ie, the mean) for the universe of possible trials is that the null hypothesis takes place. As such, a phase-3 study designed from weak and possibly spurious data in a phase-2 trial will be susceptible to a regression to the mean, ie, showing the null hypothesis. There, that is a slightly more rigorous argument for your principle from Fleming's argument.
But as I said, general principles for filtering or taking first impression do not necessarily apply to a particular dataset once you've decided to take a closer examination of it. There, you need to understand the semantics of the domain of study where the data come from and taking in all corroborative data to see if the whole picture makes sense. That too should be self-evident to anyone who ever does data analysis. And, I can assure you that I've looked at data far larger than you could ever imagine.
