Publication bias is doing the heavy lifting

I watched a colleague spend two years on a large, careful replication of a mindset intervention in her district. Null result. The study was tight—preregistered, good sample size, real classrooms. When she presented it at a conference, someone in the audience asked why she was "attacking Carol Dweck's work." She wasn't. She was just reporting what happened. Here's what I think is actually going on, and it's not that Dweck is wrong or that the replication folks are out to get her. Meta-analyses aggregate whatever made it into journals. And what made it in skews hard toward "it works." Studies showing small effects or null results from underfunded labs, unpublished dissertations, failed classroom pilots—those live in filing cabinets. The meta-analyses then average effect sizes that aren't a random sample of reality; they're a sample of publishable findings. You get 0.4, 0.35, 0.3 averaged together, and you report a meaningful effect. The preregistered replications are different. They're funded specifically to test whether something holds at scale. They're not trying to publish a surprising finding; they're trying to publish the truth. And they keep finding that when you actually control for publication strategy and prior beliefs, the effect shrinks to something like 0.05 or 0.10—teachable, maybe, but not the engine of achievement we were told it was. It's not that meta-analyses are bad methodology. It's that they're only as honest as the literature they're analyzing. Right now, the literature has a built-in tilt toward positive findings. You'd need the earlier papers themselves to have been preregistered and well-powered to trust the aggregate. Most weren't.