P-hacking

The mechanics are usually mundane rather than fraudulent. A researcher might drop outliers under one rule and not another, add or remove covariates, test several outcome variables, split the sample into subgroups, or stop collecting data once the p-value dips below 0.05. Each individual decision can seem defensible, but exploring the garden of forking paths inflates the real false-positive rate far above the nominal 5 percent.

For peer reviewers, p-hacking is rarely visible on the surface because the paper reports only the surviving analysis. Warning signs include a p-value suspiciously close to 0.05, a primary outcome that shifts between the registration and the manuscript, unexplained covariate choices, and a flurry of subgroup findings without correction for multiple testing. Comparing the methods against any preregistration is the single most effective check.

The countermeasures are procedural. Preregistration and registered reports fix the analysis plan before the data arrive, and reporting every outcome tested, significant or not, lets readers judge the true evidential weight. ManuscriptMind flags mismatches between a stated hypothesis and the analyses actually run, which is where p-hacking most often leaves a trace.