Joseph - Thanks for your suggestion, but I don't think it's -bs-. I tested to see what would happen if I didn't resample at all, but kept estimating and testing on the original data - same problem. In this test program, I also took out all the -lincoms- and -test- commands and just redid the estimation. There was still a very small increase in execution time with iteration number, but nowhere near what it was before.
Al
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Joseph Coveney
Sent: Monday, December 28, 2009 9:12 PM
To: [email protected]
Subject: st: Re: increasing time burden during resampling
Al Feiveson wrote:
Hi - I am running Stata 11 on Windows XP and I am implementing a form of
multiple testing using a resampling method as described in Westfall & Young's
text. Basically, for each iteration the method is:
1)resample the data modified to a joint null situation (I use the Stata -bs-
command for this)
2) fit a model to the resampled modified data
3) do a bunch of tests using -test-
4) save the test results.
These operations are identical for each iteration, yet the time per iteration
increases roughly linearly until it becomes prohibitive to continue.
If I stop after a given number of iterations, I must close out all Stata
processes before starting again, else the bogged down state still holds.
Previous versions of this that used the estimated coefficients and standard
errors without -test- didn't seem to have this problem. So I suspect it has
something to do with repeated use of the -test- command. Maybe something needs
to be reset or cleared after each iteration?
I tried increasing the memory allocated to Stata, but that didn't seem to help.
I would appreciate any suggestions for improving efficiency here.
--------------------------------------------------------------------------------
Could the problem lie with -bs-? I think that -bs- keeps track of the original
configuration of the dataset with pointers or hidden variables or something.
If, between iterations, you're not allowing -bs- to restore the original
configuration, then these pointers, hidden variables or whatever will cumulate
linearly with each iteration and tie up resources.
It looks like you've already done the Westfall-Young "maxT" (previous versions
using only coefficients and their standard errors) and are now going for their
"minP" (current version using -test- results). Is adding -test- to the routine
the only change, or could there also be some difference in how -bs- is used,
e.g., by making changes to the dataset (saving model results in the dataset via,
say, -estimates-) before -bs- has had the chance to restore the dataset and
release resources used in keeping track of its original configuration?
Joseph Coveney
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
