The short answer is program-survival bias. Since a huge number of drugs and indications are tested in phase-2, by sheer chance many drugs with no true efficacy will appear to show efficacy according to conventional statistical metrics.
I agree that Program Survival Bias is a factor. But I'd say that a much much bigger factor is that phase 2 trials are generally not a filter for efficacy. At best they are a filter for severe safety issues (e.g. side effects or negative efficacy):
I'd guess significantly more than 1/2 of biotech ph ii's are single arm - historical comparisons - and they often use patients that are in better health than in the general population. The resulting fidelity inre a measurement of efficacy positively sucks.
And the remaining, randomized, ph ii trials are almost always small. Ultimately powered nowhere near what they would need to be to detect the 20 or 30% benefit that the later ph iii is powered for. And most continue to ph iii if they see either an ITT efficacy on the right side of HR=1.0 or in some important subgroup.
Market Data 
Markets 
AI
