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

RE: st: No constant with XTMELOGIT

From   "Verkuilen, Jay" <[email protected]>
To   <[email protected]>
Subject   RE: st: No constant with XTMELOGIT
Date   Mon, 28 Jan 2008 15:47:23 -0500

>>I have been hesitant to use laplace because of the potential bias.
I noted earlier  I have over 56,000 observations so each run can take
several days even using the difficult command.  If I can't report the
difference in the models using laplace and not using laplace then I
don't suspect I should use laplace.  Would that be a safe assumption?<< 

Be wary when comparing across models estimated with different quadrature

You may want to look at the HLM 6.0 software, which has a higher-order
Laplace method that is very accurate (not as accurate as Gauss-Hermite
quadrature, though) built into it. Hopefully the higher-order Laplace
method will become more commonly available. Alternatively, you could use
MCMC, which replaces the quadratures with Monte Carlo integration. This
would most likely require switching software. winBUGS, for instance, is
nice, but it's not going to be happy with a dataset the size of yours;
there are a few other options around. Unfortunately, many random effects
simply require a lot from your computer. Gauss-Hermite quadrature is
very accurate, but it's EXPENSIVE and if you have more than, oh, three
random effects, is is often intractable. 

>>Here is example model not using diff: you can see there are so
many valleys that I wasn't getting anywhere.<<

You might want to fit a simpler model to get good starting values for
your "final" model. This often speeds convergence and lets you avoid .
However, convergence difficulties like the valleys you mention are often
a sign of a model that's too complex for the data. Bayes estimation
often succeeds here because it smooths the valleys out. 

J. Verkuilen
Assistant Professor of Educational Psychology
City University of New York-Graduate Center
365 Fifth Ave.
New York, NY 10016
Email: [email protected]
Office: (212) 817-8286 
FAX: (212) 817-1516
Cell: (217) 390-4609

*   For searches and help try:

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