Wild bootstrap

Actually the wild bootstrap just samples the residuals, following different
schemes:
- 1 and -1 ("rademacher" bootstrap)
-normal distribution
-other schemes... see
http://www.econ.queensu.ca/working_papers/papers/qed_wp_1028.pdf

I have implemented such a bootstrap for VAR models (allowing to include HC
estimators from pkg sandwich), see:
library(vars)
?causality

Note that as only the residuals are bootstraped, and not the regressors, a
lot of computation can be saved as you don't need to compute every time the
(X'X)-1 matrix. Because you want to do it many times, it can be useful not
to reuse lm() at each step

I explored:
update(lm)
and finally used something like:
Ynew<-pred+res*rnorm(n=obs, mean=0, sd=x)
PI.boot<-solve(cross, crossprod(Zmlm,Ynew))

But this sure could be more efficient...

Matthieu



Patrick Burns a ?crit :

It is good if the original poster explains what
a term means that they are searching for. ?That
has at least two benefits:

* it avoids misunderstandings if the same or a
similar term has a different meaning.

* it educates subscribers (like me) who don't
know the term.

If not the original poster, then it's nice if a
response contains it. ?So here goes:

The wild bootstrap is bootstrapping residuals and
then randomly multiplying by 1 or -1. ?Hence ensuring
the distribution is symmetric around zero.

I've done it -- I just didn't know it had a name.

The bootstrapping tutorial on www.burns-stat.com
would tell you how to do it yourself (minus the last
step of multiplying by -1 or 1).


On 29/03/2010 22:08, Brian G. Peterson wrote:

On 03/29/2010 04:00 PM, Paulo Grahl wrote:

I am wondering whether it is possible to use 'boot' package (or any
other) in order to use the wild bootstrap method.
Any help, examples, would be appreciated.

There are many implementations of the wild bootstap in R.

Perhaps you should do a little searching, try a few things, and refine
your question with a code example.

Regards,

- Brian