Stata The Stata listserver
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

RE:st: HSHAZ iteration stopped at number 290.


From   "Stephen P. Jenkins" <stephenj@essex.ac.uk>
To   <statalist@hsphsun2.harvard.edu>
Subject   RE:st: HSHAZ iteration stopped at number 290.
Date   Thu, 2 Jun 2005 09:00:34 +0100

> Date: Wed, 1 Jun 2005 21:54:30 -0400
> From: Cody Knox <codyknox@gmail.com>
>
> I have been running HSHAZ for a large dataset with a number of
> covariates. After going through 280 iterations, STATA (hshaz) showed
> virtually the same log likelihood for about 10 iterations and then
> stopped to move into the next iteration - what a disaster!
> 
> If it had sent an error message, I would have had no choice but to
> give up. Bottom line is STATA is still running using most of the CPU
> capacity - and about a day and a half has passed.
> 
> Does anyone have similar experience (and got it through and got the
> estimated results later)? Any tips/help are welcome. - if I stop here,
> I should run the program for another 20 days... please help me.

You have a large dataset and a large number of covariates (how many
precisely?). -hshaz- uses -ml- method d0 (numerical derivatives, and so
it is therefore common in this situation for estimation to take a  l  o
n  g  time. 

Slow convergence is also common when the likelihood is in (almost)
non-concave regions. You should get signals of this with messages in the
iteration log.  Trying to fit "too many" mass points may also
exascerbate this problem (you don't say how many you specified).  So can
poorly specified models. (Suggestion: spend a lot of time making sure
that you are confident about your no-frailty model specification before
exploring the effects of frailty using -hshaz- or -pgmhaz8- or
-xtcloglog-.)

You can get more information about convergence by specifying the -trace-
and -gradient- options when using -hshaz- (see maximize_options).

You have not said whether you have experimented with using the
-technique()- or -difficult- options, either. Or different starting
values for the mass points and associated probabilities. (These are
mentioned in the help file.)

Overall, don't expect these programs to provide some sort of magic
bullet for estimation. Fitting models with unobserved heterogeneity is
an intrinsically tricky business.  (The reason for what you label a
"disaster" may be a result of your over-opimistic expectations.)

I also suggest that you include key extracts from your log file in any
future posts so that readers can see exactly what you typed, and exactly
what Stata reported.

Stephen (author of -hshaz- and -pgmhaz8-)
-------------------------------------------------------------
Professor Stephen P. Jenkins <stephenj@essex.ac.uk>
Institute for Social and Economic Research
University of Essex, Colchester CO4 3SQ, U.K.
Tel: +44 1206 873374.  Fax: +44 1206 873151.
http://www.iser.essex.ac.uk   


*
*   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 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index