Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.


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

Re: st: GLLAMM for 2-level latent factors not converging


From   Nirup Menon <[email protected]>
To   [email protected]
Subject   Re: st: GLLAMM for 2-level latent factors not converging
Date   Mon, 11 Mar 2013 11:15:01 -0400

Thank you for the prompt advice.
The indicators are likert scale ranging 1-7.
BTW, in an earlier model, where I had constrained one of the regression weights from the level-2 latent to the level-3 latent, converged fine. I used the initial values from those results to run this model. I will try other constraints, as you suggested, to look for identification issues.
Nirup

On 3/11/2013 10:22 AM, JVerkuilen (Gmail) wrote:
This sounds like either an identification problem or a Heywood case to
me. Do you have reason to suppose the model is identified? (i.e., try
it with simulated data, or you know otherwise). You may have empirical
unidentification due to needing a correlation to be nonzero but having
a value too close to 0, too. A Heywood case may be in the offing, too,
which is created by a seriously misspecified model. Often the best
diagnosis for this is to look at the observed correlations. If one of
them is extraordinarily high compared to the rest that is often a
sign.

Another trick is to really constrain the model a lot by fixing
loadings to 1 or some other meaningful value and slowly freeing them
to see where it blows up.
What is the nature of the indicators? Are they binary?

In order I'd check:

Identification. I suspect you may not have an identified model.
Empirical underidentification
Heywood case


On Mon, Mar 11, 2013 at 10:05 AM, Nirup Menon <[email protected]> wrote:
Hi,
I am running gllamm on a data file in which "scene" is nested inside "idno"
(please see code below). Each idno has responded to 5 or less scenarios.

At idno level, I have 2 latent (reflective) factors made up of 2 and 3
indicator (observed) items respectively.

At the scene level, I have 1 latent (reflective) factor made up of 2
indicator (observed) items.

The scene level latent factor is regressed on the two idno level latent
factors.

I have omitted the random effect from idno to scene for now. The code is:

gen scene=_n

reshape long y, i(idno scene) j(item)

tab item, gen(d)

eq load1: d1 d2

eq load2: d3 d4 d5

eq load3: d6 d7

matrix bm1 =(0,1,1\0,0,0\0,0,0)

eq f1: x1 x2 x3

gllamm y d2-d7, i(scene idno) l(oprob) bmat(bm1) nrf(1 2) eqs(load1 load2
load3) geqs(f1) trace nip(5) nocorrel

The problem is that the lambdas for d4 and d5 are large positive numbers
(around 200), while the regression weight of the factor made up of d3, d4,
and d5 on the scene factor is becoming a large negative number (around
-300). The variance of the both factors are reasonable. Is this a
convergence problem? The "adapt" option is not helping either. Any
suggestions?
Thanks
Nirup





*
*   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 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index