HARKing
Hypothesizing After the Results are Known, then pretending you predicted it.
What it means
HARKing — Hypothesizing After the Results are Known — is the questionable research practice of presenting a hypothesis generated after inspecting the data as though it had been specified in advance of data collection. Named by Norbert Kerr, it corrupts the logic of confirmatory hypothesis testing, which assumes the prediction was made independently of the data being used to test it; by retrofitting the hypothesis to patterns that may be pure chance, HARKing dresses up noise as theory-confirming signal. The mechanism that makes it so misleading is capitalization on chance: in any rich dataset some apparent effects will arise by luck, and selecting the ones that 'worked' and narrating them as predicted dramatically inflates the false-positive rate while hiding the multiple comparisons that produced them. It is closely allied with p-hacking and is one of the practices implicated in the replication crisis, since HARKed findings, built on flukes, fail to reproduce. A nuance worth keeping is that exploratory, post-hoc hypothesis generation is a legitimate and valuable part of science — the problem is specifically misrepresenting it as confirmatory — and the standard remedy, pre-registration, exists to keep the two clearly separated.
Examples
Running an experiment, noticing an unplanned gender difference in the data, and then writing the paper as though that difference was the hypothesis you set out to test.
An A/B test shows no overall lift, but the new checkout wins among Android users on Tuesdays; the deck presents that slice as the hypothesis all along, and rollout finds nothing.
A trial's main outcome fails, so the write-up leads with a benefit spotted in one age band — a result the team first met in the data, not before collecting it.
First described in Norbert Kerr (1998).