The file-drawer problem isn't the real issue with mindset research

Most people assume the gap between meta-analyses and preregistered replications is about publication bias—failed studies sitting in drawers while positive ones get published. That's part of it, but the weirder problem is that the original studies often weren't designed to fail in any honest way. The early mindset work (Dweck's stuff, the Blackwell intervention in the mid-2000s) typically had small samples, no preregistration, and lots of flexibility in how outcomes got coded and analyzed. A teacher notices girls in the intervention seem more persistent on Task B but not Task A? Task B becomes the focal outcome. The effect sizes look huge because the researchers weren't trying to fool anyone—they were just following the data they found interesting. Meta-analyses inherit this mess. They aggregate studies that were each individually honest but collectively cherry-picked in different ways. Then the preregistered replications come along and lock everything down first: specific sample size, specific measure, specific analysis plan. No surprises. The effects shrink. This isn't corruption in either direction; it's what happens when you go from exploratory work to confirmatory work. The frustrating part is how long we collectively pretended the original effect sizes were the default expectation rather than the optimistic endpoint of a messy discovery process. We're still teaching mindset interventions to schools on the basis of estimates that probably never should've been trusted as effect sizes in the first place.