The thresholds are applied to the latent response y*
underlying the observed response y.
y* can take on any value from - infinity to + infinity.
If the lowest observed response category, say 1, happens
very rarely, the lowest threshold may have to be a large
negative number because Pr(y=1) = Pr(y*<threshold)
and similarly for the largest category.
I am not sure what predictions you have considered.
If you are referring to 'xb' (the linear predictor),
then this is the mean of the latent response y* which
doesn't need to lie on the range of y.
In gllamm, you can obtain the predicted (population
averaged) cumulative probabilities, e.g., Pr(Y>1) using
gllapred probgr1, mu above(1) marg
and similarly for the other categories.
You can also get predicted probabilities for particular
values of the random effects or posterior mean probabilities.
Sophia
Erik Melander wrote:
I have a panel dataset with an ordinal dependent variable, judgmentally
coded from 0 to 4. There is considerable inertia in the dependent variable
and I thus want to include a lagged dependent variable (actually a set of 4
dummies since it is ordinal scale) to control for autocorrelation. I have
tried to run random effects ordinal probit and logit for panel data, using
for example the stata commands below:
reoprob dependentvar dependentvar1t-1 dependentvar2t-1 dependentvar3t-1
dependentvar4t-1 indepvarA indepvarB indepvarC indepvarD, i (panelunit)
gllamm dependentvar dependentvar1t-1 dependentvar2t-1 dependentvar3t-1
dependentvar4t-1 indepvarA indepvarB indepvarC indepvarD, link(oprobit)
i(panelunit)
gllamm dependentvar dependentvar1t-1 dependentvar2t-1 dependentvar3t-1
dependentvar4t-1 indepvarA indepvarB indepvarC indepvarD, link(ologit)
i(panelunit)
In the output, one thing that seems a little weird is that the
cuts/thresholds give a broader range than the dependent variable itself. The
range of the coefficients of the categories of the lagged variable is only
about half the range of the cuts/thresholds.
Most disturbing of all is that the resulting models give rise to predictions
that are outside the range of the dependent variable. Why is this so, and is
there anything I can do in order to arrive at models with more reasonable
predictions?
Thanks for your attention.
Erik Melander
*
* 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/
--
Sophia Rabe-Hesketh, Professor
Educational Statistics
Graduate School of Education
3659 Tolman Hall
University of California, Berkeley
Berkeley, CA 94720-1670
WWW: http://www.gllamm.org/sophia.html
*
* 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/
