The replication crisis — the discovery that a significant portion of published findings in psychology, medicine, nutrition, and economics cannot be reproduced by independent researchers — has been a recurring story for over a decade. It keeps happening because the incentive structures that produced the original problem have not fundamentally changed.
The academic publishing system rewards novelty. New findings get published; null results and replications rarely do. This creates systematic publication bias: the studies that make it into journals are disproportionately the ones that found something, which means the published literature systematically overstates effect sizes and understates the failure rate of hypotheses. Researchers know this. The journals know this. The system continues producing it because the incentives for individual researchers point in one direction.
Statistical practice compounds the problem. The p-value threshold of 0.05 — used as a binary cutoff for significance across most fields — is a convention, not a law of nature, and it is poorly understood even by the researchers applying it. Flexible analysis choices — which variables to include, when to stop collecting data, how to handle outliers — can be made in ways that move a result across the significance threshold without rising to the level of fraud. This practice, sometimes called p-hacking, is common enough to explain a substantial portion of findings that fail to replicate.
The fixes are known: pre-registration of study designs, mandatory disclosure of null results, larger sample sizes, and greater weight on replication in publication decisions. Progress on all of these is real but slow, because the people who would implement them are evaluated under the current system and have limited incentive to change it.
Science as a method is self-correcting in the long run. The incentive structure is not. That gap is where the crises live.