Short answer: p-hacking is the practice of trying many analyses, variables, exclusions, or comparisons until a statistically significant result appears, then presenting that result as if it were the planned test all along.
Not all p-hacking is deliberate fraud. Sometimes it happens because researchers are exploring data and gradually convince themselves that the most interesting-looking result was the real result. The problem is that repeated searching increases the chance of finding a result that looks meaningful by accident.
How p-hacking happens
A researcher might test several outcomes but report only the significant one. They might remove “outliers” in a way that changes the result. They might stop collecting data once significance appears. They might try different statistical models, subgroups, or time windows until one crosses the threshold. Each choice may look defensible in isolation, but together they can turn noise into a publishable story.
This matters because many fields have strong incentives to publish surprising, positive, or clean findings. A messy null result may be harder to publish than a neat significant result. If the system rewards significant findings too heavily, researchers may be pushed toward flexible analysis even without intending to mislead.
How to reduce it
Good practice includes preregistration, where researchers state their hypotheses and analysis plans before looking at the data. Sharing data and code can also help others check whether the result depends on hidden decisions. Journals can reduce p-hacking by valuing careful null results, replication studies, and transparent methods.
The key distinction is between exploration and confirmation. Exploring data is valuable when labelled honestly. The problem begins when exploratory results are presented as if they were planned confirmatory tests. Readers should ask whether a study clearly separates what it intended to test from what it found along the way.