Greg wrote:
> "David Jones" <[EMAIL PROTECTED]> wrote in message
> news:<[EMAIL PROTECTED]>...
>> Greg wrote:
>
>
>
>> Yes, and possibly overlooking all the subtleties of bootstrapping
>> itself.
>
>   Can you say more about this?

I tried to imply that I don't follow all of bootstrapping. If you look
at a good book on bootstrapping you will find many different versions
of ways for constructing confidence regions ... these often give
rather different answers, but the subtleties of why any given approach
might be good or bad are beyond me, except that I prefere the approach
of constructing confidence intervals out of significance tests, not
the other way round.

>
>> In your case you would need to find an
>> adjustment of the data so that g(x)=0 does hold. Can you do this?
>
>   It is my impression that adjusting the resampling distribution to
>   reflect the null needs to be done when testing hypotheses (see eg.
>   Hall, P. & Wilson, S. R. (1991). Two guidelines for bootstrap
>   hypothesis testing. Biometrics, 47, 757-762.) but is not done when
>   calculating confidence intervals.)

 But I think you said you were trying to do a significance test. The
"confidence interval" you mentioned is simply a reflection of the
sampling uncertainty of the test-statistic you are using ... the
underlying procedure estimates the distribution of possible
test-statistic values given that the "true" population test-statistic
exactly equals the test-statistic value from your sample. In turning
this sampling distribution into a confidence region in the way that is
done, you are making two assumptions : symmetry of the distribution
and a purely additive effect when the "true" value of the population
statistic changes. For some reason, even if these assumptions are
invalid, this does not prevent the method sometimes providing good
confidence regions ... at least according to the performance measures
for particular examples that you will find in books/papers on
bootstrapping

David Jones


.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to