Jackknife
The jackknife estimates statistical bias or uncertainty by recalculating an estimator after systematically deleting observations or groups.
In the ordinary delete-one form, remove each of n observations in turn and recompute the same statistic. With leave-one-out estimates theta_(i) and their average theta_bar, the variance estimate is (n - 1)/n times sum((theta_(i) - theta_bar)^2); its square root is the standard error. The basic independent, identically distributed setup does not establish validity for every estimator. [1]
Deleting a whole cluster is different from deleting one row. Daily records from the same drive are not automatically independent units; the deletion scheme must match the sampling design and dependence. Cluster deletion does not itself remove dependence between clusters. [2]
For a completely observed fixed-horizon binary endpoint sample, every delete-one event fraction is zero when no event is observed. Its jackknife standard error is therefore zero. This is an algebraic property of the observed sample, not proof of zero population risk. With 100 independent trials sharing one event probability, zero events instead give a one-sided 95% exact binomial upper bound of about 2.9513%. This invented endpoint example is not AFR from unequal drive-time exposure. [4] [5]
Relations
Limitations
- The ordinary delete-one jackknife can fail for nonsmooth statistics such as the sample median; computing a spread does not guarantee a calibrated confidence interval. [3]
Citations
- Efron and Stein (1981): The Jackknife Estimate of Variance, equations 1.1-1.3
- Stata: jackknife, observation and cluster deletion, Methods and formulas
- Efron (1979): Bootstrap Methods: Another Look at the Jackknife, section 3
- NIST Dataplot: exact binomial confidence limits, one-sided and two-sided bounds
- Hesela: jackknife spread and zero-event bounds, synthetic reproducible example