| |
 |
|
|
Science Forum Index » Space - Consult Forum » Bootstrap estimates and bias...
Page 1 of 1
|
| Author |
Message |
| John Uebersax... |
 Posted: Wed Jun 18, 2008 11:46 pm |
|
|
|
Guest
|
I wrote a simple bootstrap estimation program, and, as a test, used it
to get the bootstrap estimate of the SEM for nine data values.
Using the standard bootstrap method (e.g., Efron and Tibshirani, An
Introduction to the Bootstrap, p. 47), however, what I seem to get is
a very accurate estimate of the 'view sample as population' SEM (i.e.,
based on n, not n-1 in the variance denominator). This, of course, is
a biased estimate of the population SEM.
Efron and Tibshirani (p. 43) seem to treat this issue lightly,
suggesting that "for most purposes" either method (n or n-1 in the
denominator) is just as good as the other.
I don't understand this. It seems in my simple example above that if
I multiply the bootstrap estimate by:
sqrt[n/(n-1)]
I get the usual, unbiased estimate of the population SEM; and with n=9
it makes a noticeable difference. If getting an unbiased estimate
were difficult, then I could understand E&T's comments. But in this
case it appears, at least empirically, that multiplying by the
constant above produces an unbiased estimate of the SEM.
Is this something unique to using the bootstrap to estimate an SEM, or
does it apply more generally? Can a case be made for always
multiplying a bootstrap-derived standard error estimate times sqrt[n/
(n-1)]?
John Uebersax PhD
Rixensaart |
|
|
| Back to top |
|
| |
|
Page 1 of 1
All times are GMT - 5 Hours
The time now is Fri Dec 05, 2008 9:15 am
|
|