In statistical or mathematical terms, the jackknife is a method of determining the behaviour of a model or a distribution through simulation. It is one of the Monte Carlo methods of simulation, along with permutation and bootstrapping.

Using a real data series, jackknifing allows a researcher to create a large number of hypothetical cases by resampling the data by eliminating one (or more) case at a time. It is used principally to determine the effect outliers or abberrant cases have on the results of the study. The fundamental procedure is as follows:

  1. Identify the data series of interest. This data series will have n observations.
  2. Calculate the parameter(s) of interest on the original data.
  3. Eliminate the 1st observation from the data series.
  4. Calculate the parameter(s) of interest on the reduced data set.
  5. Repeat steps 3-4 for the 2nd to nth observation.
  6. If desired, repeat steps 3-5, eliminating more than one case at a time.
This method of simulation is particularly useful when the data being studied do not conform to typical distributions, or may contain a lot of uncertainty.

Update: ariels tells me that this is called leave-one-out cross-validation by computer scientists, but that the procedure is not quite. Leave-one-out cross validation is used to estimate the error probabilites, while jackknifing can be used to estimate many different parameters or even distributions.