Statistics Glossary → Jackknife
Jackknifeconcept
A resampling method that recomputes an estimate leaving out one observation (or one cluster) at a time and uses those replicates to estimate the standard error and bias.
With n observations, the estimate is recomputed n times, each time with one observation removed. The spread of these leave-one-out estimates gives the standard error: SE² = ((n − 1)/n) Σ (θ̂₍ᵢ₎ − θ̄)². The difference between the replicate mean and the full-sample estimate also gives an estimate of bias. In complex survey designs, a primary sampling unit (PSU) is dropped instead of a single observation.
The jackknife predates the bootstrap and is deterministic: the same data always give the same answer. It works well for smooth statistics such as means, proportions and regression coefficients, but it can estimate the standard error of non-smooth statistics such as the median or other percentiles inconsistently; the bootstrap is better suited there.
An everyday example
A class average is computed for 10 students. The average is then recomputed 10 times, leaving out one student each time. If the 10 averages are very close, the class average is a stable estimate; if one of them drops noticeably (when the top-scoring student is left out), the result hinges on a single person and the standard error comes out large.
In MerQur
In Partial Least Squares Regression, 'Extra output: Jackknife CI' gives coefficient variances from the cross-validation segments; in Survey-PHREG, the 'Jackknife deviation centre' option sets whether deviations are taken around the replicate mean or the full-sample estimate. It is offered as Jackknife (JK1 / JKn) under the 'Variance method' option of the Weighted Mean, Frequency, Total, Linear and Logistic Regression analyses, and it is used to compute the acceleration of the BCa interval in Bootstrap Confidence Interval, Correlation Analysis, Mediation Analysis and PLS-SEM.
Analyses where this concept appears
Bootstrap CICorrelation AnalysisWeighted Mean (Survey Means)Weighted Frequency (Survey Freq)Weighted Total (Survey Total)Weighted Linear Regression (Survey Regression)Weighted Logistic Regression (Survey Logistic Regression)Partial Least Squares Regression (PLS)Mediation AnalysisSurvey-PHREG (Weighted Cox)PLS-SEM (Partial Least Squares Path Modeling)
Related concepts
Bootstrap (resampling)Taylor linearization and replication variance methods