Do Not Trust Statistics You Did Not Fake Yourself

Four hundred and fifty-four points of honest noise, correlation minus zero point one — a finding so absent you could publish the absence. Then you put a magnifying glass over one corner of it, and through the glass: five points, a straight line, R² = 0.99. A relationship. A result. A paper. This is the whole replication crisis in one instrument. The five points were always there, sitting in the cloud at exactly the same size as everything else, completely invisible until somebody went looking for them. Nobody lied. Every number here is real: the cloud is a genuine draw from a bivariate normal, the r in the fine print is the actual sample correlation of the whole sample, and the R² is the actual fit to the five that got picked. Even the optics are honest — a lens at that magnification really would show you only that much of the plate, and the points it hides are still counted in the n. The only dishonest step was where you chose to point it, which is, of course, always the only dishonest step. A riff on the line usually pinned on Churchill: "I only believe in statistics that I doctored myself." For statisticians, quants, data scientists, and anyone who has ever watched a p-value improve on the fourth attempt. Printed in white on dark or black on light, in a chunky monospace that looks like it escaped from a terminal.