[R] bootstrap: definition of original statistic

[Gad Abraham]

> Gad Abraham wrote:
>> Hi,
>> In the boot package, the original statistic is simply the statistic
function evaluated on the original data (called t0).
>> However, in Harrell et al 1996 "Multivariable prognostic models..."
Stats Med vol 15, pp. 361--387, it is different (p. 372):
>> The statistic function evaluated on the original data is called "D_app"
(apparent statistic), whereas "D_orig" (original statistic) is derived
by first fitting the model to a bootstrap sample, then freezing the
model and applying it to the original data. The
>> bootstrap statistic D_boot is derived by applying the function to the
bootstrap sample. Then optimism O = average(D_boot - D_orig), and the
bias-corrected estimate is D_app - O.
>> Can someone explain this difference, and especially why boot
>> doesn't evaluate the frozen model again on the original data?
>> Thanks,
>> Gad
>
> The optimism bootstrap is a different kind of bootstrap to estimate bias
from overfitting.  boot is for the regular bootstrap.

Thanks.

Is there a general-purpose optimism bootstrap function in CRAN? It seems
that Design:::validate does the optimism bootstrap but only for cases like
lrm and cph; I'd like to validate the mean-square error of a time series
method like arima or forecast:::ets.