Statalist


[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: After non-convergence with xtmixed


From   Maarten buis <[email protected]>
To   [email protected]
Subject   Re: st: After non-convergence with xtmixed
Date   Wed, 2 Jan 2008 14:33:57 +0000 (GMT)

--- Clyde Schechter <[email protected]> wrote:
> When estimating a model with xtmixed, if the gradient-based 
> optimization fails to converge (as is painfully often the case!), 
> Stata reports results based only on the iterated EM 
> optimization.  Why is that?  Despite non-convergence, the log 
> restricted-likelihood at the point where gradient-based optimization 
> quits is, in my experience, typically substantially higher than what 
> prevailed at the end of iterated EM.  So why not report the results 
> of the incomplete gradient-based calculations instead?  Is there some
> reason to believe that when the gradient-based method fails it 
> actually provides worse estimates?  Or is it simply a matter of the 
> details of programming that prevents that?
> 
> If the incomplete gradient based results are, as I'm inclined to 
> believe, better than the final EM iteration, is there any option or 
> work-around that will make Stata report them?  After all, those 
> calculations sometimes represent many hours of extra calculation, and
> they seem to just go to waste.

If -xtmixed- doesn't converge, you should fixed the model till it does
converge. You should not report any model that does not converge.

The EM result might help you diagnose the problem: e.g. are any
parameters very large/small compared to the other parameters, in which
case you might want to rescale that variable. 

Also think about whether any of the variables are redundant (like a
dummy for male and a dummy for female), in which case you need to drop
one of the variables. Alternatively, two or more variables could be
very highly correlated, in which case you can drop one of those
variable (since they are so highly correlated they do not contain much
independent information anyhow.), combine those variables into a single
variable (high correlation could mean that they are actually different
measures the same concept), or use -orthog- on those variables.

Hope this helps,
Maarten

-----------------------------------------
Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands

visiting address:
Buitenveldertselaan 3 (Metropolitan), room Z434

+31 20 5986715

http://home.fsw.vu.nl/m.buis/
-----------------------------------------


      __________________________________________________________
Sent from Yahoo! Mail - a smarter inbox http://uk.mail.yahoo.com

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/



© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index