Winning the auction is the bad news.

Picture a jar of coins worth exactly some amount, passed around a room. Everyone estimates; the estimates scatter around the truth, some high, some low. Now auction the jar to the highest bidder. The winner is not the wisest person in the room — it is whoever was most over-optimistic. The very act of winning selects for an inflated estimate, which is why the winner so reliably overpays. Economists call it the winner's curse: the prize for being highest is the near-certainty that you were too high.

This isn't a quirk of greedy bidders; it survives even when everyone is honest and unbiased on average. The trouble is that an auction doesn't show you the average bidder — it shows you the maximum. And the maximum of a pile of noisy estimates lands in the right tail, further out the more estimates you take. The slider proves it: each extra bidder makes the winning bid worse, because you are sampling the extreme of a larger draw. The curse is a property of selection, not of stupidity.

Now swap the jar of coins for a real effect in the world — a drug's benefit, a psychological nudge, a genetic association — and swap the bidders for studies. Each study estimates the true effect with noise. And there is a filter that decides which studies the world gets to see: statistical significance. A study tends to be publishable when its estimated effect is large enough to clear the significance bar.

That filter is an auction, and significance is the winning bid. When the true effect is small and studies are noisy — the normal condition in much of social and biomedical science — a study can only reach significance if luck pushes its estimate upward. So the studies that pass the filter are exactly the ones that overestimated, and the published effect size is inflated for precisely the reason the auction winner overpays.

The winning bid grows with the number of bidders; the published effect inflates as power falls. Underpowered fields — small samples chasing small effects — suffer the worst curse, because in those fields significance is almost unreachable without a generous helping of luck.

Two filters stack on top of each other. First, within each study, significance selects for inflated estimates — the winner's curse. Second, across the whole literature, non-significant studies are quietly never written up, never submitted, or never accepted — the file-drawer problem, the classic form of publication bias. The result is a published record that is a biased sample of the research actually done: the positives are visible, the nulls invisible, and any reader or meta-analysis that trusts the visible record overstates how real, and how large, the effect is.

The funnel is diagnostic precisely because honesty has a signature. Small studies should scatter widely on both sides of the truth. When the left side of their scatter is missing — the small studies that found little or nothing — the asymmetry betrays a literature that has been filtered for good news.

If published effects are inflated, a clear prediction follows: try to reproduce them and they should shrink. That is exactly what happened. In 2015 the Open Science Collaboration's Reproducibility Project repeated 100 prominent psychology studies under high-powered, pre-agreed designs. In the originals, 97% had reported a statistically significant result. On replication, only 36% did — and the replication effects came in at roughly half the size of the originals. Not fraud; the winner's curse and the file drawer, cashing out on schedule.

The same story has played out beyond psychology. A large replication effort in cancer biology found the effects it could reproduce were, on average, dramatically smaller than first reported. This is sometimes called the decline effect — the tendency of exciting findings to fade as they are re-examined — and much of it is not mysterious at all. It is regression to the mean wearing a lab coat: an estimate selected for being extreme is, on a second look, less extreme.

The cures are structural. The reforms that work all attack the filter rather than the researcher. Pre-registration commits to the analysis before the data, so a result can't be quietly reshaped into significance. Registered reports go further: a journal accepts the study on the strength of its design, before the results exist, so null findings get published too and the drawer is emptied. Reporting effect sizes and intervals, not just a significance verdict, lets readers see the uncertainty the winner's curse hides. And honest meta-analysis models the missing studies instead of averaging only the survivors.

For the rest of us. You will rarely run a funnel plot over your morning news. But the pocket question does most of the work: is what I'm seeing all the attempts, or only the lucky ones? A single dramatic study is the auction winner — treat its number as a ceiling, not an estimate. A finding that has survived large, pre-registered replication is worth far more than a surprising first result in a small sample, however exciting the headline.

This is the selection illusion in its purest, most consequential form: the same mechanism as survivorship bias, but operating on knowledge itself. What reaches you has passed through a gate that prefers the surprising and the significant, and the gate keeps the overestimates. Read accordingly — and trust the boring, replicated number over the brilliant, lonely one. The rest of the compendium is full of honest numbers misleading us; this is the one that quietly shaped the textbooks.