Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down at the end of May, and its replacement, statalist.org is already up and running.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: RE: save iterations of xtmelogit - need to break/pause


From   Jeph Herrin <stata@spandrel.net>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: RE: save iterations of xtmelogit - need to break/pause
Date   Thu, 09 May 2013 10:37:56 -0400

To the other good tips offered here, I would add another. If you try to estimate your model as a linear probability model (that is, use -xtmixed- instead of -xtmelogit-) you will likely get a quick sense of whether you have too many levels and which ones they are. -xtmixed- typically converges much more quickly, and will indicate whether some of your random effects are essentially zero.

Generally, while it is frustrating to abandon 6 days of calculations, the slow convergence likely indicates that you headed down the wrong path.

hth,
Jeph


On 5/9/2013 8:17 AM, JVerkuilen (Gmail) wrote:
On Thu, May 9, 2013 at 1:11 AM, tshmak <tshmak@hku.hk> wrote:

Hi Ann,

Despite your 160,000 individuals, I am rather unsure you have enough data
to fit a model with 4 random effects, if that's what you mean by 4 levels.
Since it's a logistic regression, perhaps a better measure of how much data
you have is the smaller of the number of 1's and 0's. There's probably huge
uncertainty over the variance of the random effects, and Stata is having a
hard time converging.

100% agree. Everyone always says "I have X thousands of data
points!?!?!?" but in a multilevel model the Ns at each level matter.
With logistic it's even worse, just as Tim noted, because it's
essentially the minimal number of events or non-events that matter.
These can become quite small.


My suggestion is:
1. Try and fit a simpler model. Perhaps some of the random effects have
only a handful number of different values. Perhaps consider treating them as
fixed-effects?

Yes, start with the random intercept model and build upwards towards
the model you want.



2. Consider doing it in WinBUGS/OpenBUGS. If you use informative priors,
you can pretty much fit anything. You can also use it to tell you whether
your model should be fit in Stata. If using different uninformative priors
gives you very different results, you probably shouldn't use Stata for your
model, and your best bet is either simplify it or use a subjective Bayesian
approach.

Agreed, but that's no light undertaking. I suspect the amount of
effort it would take would be substantial. An alternative would be to
look at MCMCglmm in R, which is a nice program but has some quirks in
the interface.

Jay
--
JVVerkuilen, PhD
jvverkuilen@gmail.com

"They were careless people, Tom and Daisy - they smashed up things and
creatures and then retreated back into their money of their vast
carelessness, or whatever it was that kept them together, and let other
people clean up the mess they had made." -- F. Scott Fitzgerald
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/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/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/


© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index