There is a lot of literature with highly visible peaks just below 0.05. And the easiest way to get that is to p-hack under the null bc that's the scenario with the most density in the neighborhood of 0.04.

Before writing this I had somehow assumed if you had lowish power you'd be more likely to get 0.04+delta than 0.01 + delta. But the p-curves, increase as you move towards zero. So a lot of p-hacking practices will show up in ways that aren't particularly visible--e.g. 0.01.

This work illustrates how large-scale text analysis, combined with extensions of p-curve methods to variable-power settings, can determine how efforts to improve research rigor are influencing the quality and credibility of large bodies of the scientific literature.

The fraction of borderline significant P values has modestly declined. But this appears to reflect increased study power, not reductions in P hacking or publication bias.

This link might be better!

These methods should be replicated in trials with greater clinical decline in placebo groups and larger treatment effects (e.g., lecanemab, donanemab) to assess whether amyloid truly works as a surrogate.

... and does not find evidence that amyloid is a good surrogate marker.

But: A4 had minimal progression and modest amyloid removal.

FDA accelerated approvals for Alzheimer’s drugs have relied on amyloid reduction as a surrogate marker, but does it actually lead to cognitive benefit?

This preprint uses individual-level A4 trial data to ask that question rigorously ...

Thanks!

This would seem to contradict that, at least for the most recent death. Looking into the other 2.

bsky.app/profile/kaka...