"do you have any intrinsic definition of "program-survival bias" other than being just a fancy name for "phase-2's don't reliably predict phase-3's""
I'm not sure what he means, but perhaps he refers to the clinical trial equivalent of "survivorship bias." Meaning, a successful trial, like a successful hedge fund record, is more likely attributable to random luck than one might have thought. given that unsuccessful trials, like unsuccessful hedge fund managers, often quietly disappear.
Put another way, if the p-value is .04, the probability is .04 that random variation could prduce the result, but that does not mean that there is a 96 percent proablility that the result is due to skill/successful drug. And, where the trial fails various prespecified endpoints, but meets a restrospectively defined endpoint , arguably that presents a similar issue.