The file-drawer problem with growth mindset research

I watched a colleague spend three months designing a replication of Dweck's classic mindset study with her ninth-graders. Same praise protocol, same measures, same population. The effect was barely there. She was genuinely confused—the meta-analyses said this was robust. The answer isn't that mindset doesn't matter. It's that published meta-analyses are inherently describing a different phenomenon than preregistered replications. Meta-analyses aggregate studies where researchers had discretion—in how they framed the intervention, who they recruited, which outcomes they emphasized, when they stopped collecting data. That discretion gets baked into the summary effect size. Preregistered replications lock down those choices beforehand, and the effect shrinks. This isn't unique to mindset work. It's what happens when a finding gets popular enough that it attracts both true believers (who tweak things until they find support) and skeptics (who try to falsify it under strict conditions). The meta-analyses are mostly pulling from the first group. When you actually run a large, prespecified study in a real classroom with real constraints—the kind you can't fudge because reviewers are watching—you get something smaller. Sometimes much smaller. The uncomfortable part is that this gap doesn't mean mindset interventions are worthless. It means the published literature systematically overstates them. We keep acting surprised by this cycle. We shouldn't.