Dear Martin,
I have used the gllamm manual quite a lot for multilevel multinomial and
ordinal logit models and linear multilevel growth model with latent classes
for the unobserved heterogeneity. I have, however, not found a presentation
of how to model a nested multinomial logit model, hence this thread on the
statalist.
(For clarification: Yes, you get it right ? the idea was to model the latent
class membership. How specifically to do this I haven't decided on yet.
Sorry about being unclear about the identification issue. My point here was
a mere indication that a more realistic model with transition varying
covariates would be "better", i.e. not depend critically on distributional
assumptions of the errors. I am not an expert in identification theory, so
I'd better read some more about before I mention it again ;) The problem
about intervening variables is also unclear in my study; I'll have to look
at this more thoroughly).
Nonetheless, my problem here is technical. Maybe I need to go through the
gllamm manual once again (and this time read it more in depth). However, if
anyone can give me a helping hand with the estimation of the nested
multinomial logit model in -gllamm- stated in the top of this thread, I
would be thankful.
All the best,
Kristian
-----Oprindelig meddelelse-----
Fra: [email protected]
[mailto:[email protected]] På vegne af Maarten buis
Sendt: 18. november 2008 23:06
Til: [email protected]
Emne: Re: st: GLAMM and nested logit/probit
--- Kristian Karlson <[email protected]> wrote:
> The main point of the study is to delve into what the unobserved
> components in educational decisions consist of, i.e. describing the
> latent classes that account for unobserved heterogeneity with
> variables that indicate different forms of motivation.
Just for clarification: You have observed variables that are indicators
of a latent variable, and you are interested what part of the
unobserved heterogeneity can be captured by this latent variable?
Have you already found www.gllamm.org, which has the manual and lots of
example code?
> The main point is that the more complex a model (i.e. the more
> realistic model of the educational system), the better the
> identification of the unobserved heterogeneity influencing
> educational decisions net of respondent background characteristics
> (family background, gender, ability, etc.).
I am not sure I buy this point.
First, I always think of identification in terms of where the
information you are using in your estimation comes from: is the data,
the design, or your assumptions. Better identification for me means
that you are making better use of the observed data or the study
design, rather than assumptions. More complex models typically mean
that you are making more use of assumptions.
Second, controlling for variables is not an aim in it's own right, but
a means of controlling confounding variables. So, a clear (theoretical)
idea about which variables are confounders and which variables are
intervening variables is crucial. So controlling for background
characteristic does _not_ necessarily lead to better estimates, it
depends on the place of these variables in the causal chain.
-- 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 N515
+31 20 5986715
http://home.fsw.vu.nl/m.buis/
-----------------------------------------
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/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/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
