Contents
- Index
Bootstrapping
Method of analysis that involves the generation of bootstrap samples. In the case of a change-point analysis, a bootstrap sample is a reordering of the original data. For each reordering of the data, a CUSUM chart is constructed and the maximum difference calculated. Since each bootstrap represents a random order, the bootstraps estimate the distribution of the maximum difference if no change has taken place.
The actual value of the maximum difference can then be compared with the bootstrap estimates. If the actual value of the maximum difference is large relative to the bootstrap estimates, there is strong evidence that a change occurred. The confidence level is determined by the percentage of bootstrap estimates that are less than or equal to the observed value of the maximum difference. For example, if 1000 bootstrap samples are selected and for 992 the maximum difference is less than or equal to the maximum difference calculated from the original data, the confidence level that a change occurred is 99.2%.
A similar type of analysis is used to calculate the confidence intervals for the time of the change. One weakness of a bootstrap analysis is that, if repeated, slightly different results might be obtained. This weakness is minimized by using a large number of bootstrap samples. A minimum of 1000 is recommended which is the default value. A larger number of bootstrap samples may be specified using the Advanced Options dialog box.
Bootstrapping is a distribution free (nonparametric) method of analysis that has only one assumption, that of independent errors.