|
The straightforward way to double bootstrap
a statistic, t is to simply draw B samples
from the data and then for each sample, draw an additional S
samples for a total of B*S samples. An example would be when t is
a p-value and the distribution of the p-value is desired.
I have a problem where the total of B*S
samples is prohibitive because the calculation of t is an involved
and expensive process.
I seem to recall that Bradley Efron
and others have a trick for double bootstrapping that only involves B + S
samples. Does anyone have some code or an algorithm for such a routine
?
Paul H. Lasky
P & B Consulting
|