Statalist


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

RE: st: RE: Multinomial logit for panel data?


From   "Nick Cox" <n.j.cox@durham.ac.uk>
To   <statalist@hsphsun2.harvard.edu>
Subject   RE: st: RE: Multinomial logit for panel data?
Date   Thu, 9 Jul 2009 16:22:20 +0100

Thanks for this information. 

-markov- certainly calculates no standard errors. 

Years ago in teaching I used Stata's matrix language to get limiting distributions from transition probabilities for first-order, discrete-state, discrete-time Markov chains. However, I don't know how that kind of calculation carries over to quintile-based classifications.  

Nick 
n.j.cox@durham.ac.uk 

Austin Nichols

Nick et al.--
I think Markov-style models are still used by implication, for example
in the economic mobility literature, where quintile transition
matrices are popular, but probably not justified. One reason: the
product of two transition matrices estimated for t-1 to t and for t-2
to t-1 often bears no resemblance to a matrix estimated from t-2 to t,
contradicting assumptions required for a Markov process.  If Dmitriy
Krichevskiy believe he has a Markov process, he can estimate on pooled
data by duplicating and reshaping data, though this seems like a bad
idea--better to estimate for each pair of times. Probably he should
use svy:tab to get standard errors on transition probabilities, which
do not seem to be estimated by -markov- on SSC.  In any case, the
limiting distribution can be derived via a little linear algebra, and
probably it is not too much work to get some standard errors for that
distribution as well.

On Thu, Jul 9, 2009 at 10:58 AM, Nick Cox<n.j.cox@durham.ac.uk> wrote:
> I don't think so; that is, it does not seem that multinomial *it will be
> of much use to you. As I understand it, you want a kind of Markov model
> in which you estimate transition probabilities. Such models were much
> discussed a few decades ago but in literature I see appear rather
> unfashionable now, for no good reason obvious to me.
>
> I can confirm that -markov-, which I wrote in 1998 and which remains on
> SSC, is no use for your purpose. I get occasional mail about it, some of
> it rather indignant or at least surprised that it does not do what it
> does not claim to do, namely support panels. This attitude sometimes
> appears connected to the paradox that those claiming to be doing
> cutting-edge research often expect that canned software is always
> available already in precisely the form they need.
>
> Yvonne Bishop, Stephen Fienberg and Paul Holland in their text "Discrete
> multivariate analysis" (MIT Press 1975; reissued more recently by
> Springer) have a chapter relating Markov and log-linear modelling. So,
> and this is wild surmise, -glm- or -poisson- may be of some use to you.
>
> Nick
> n.j.cox@durham.ac.uk
>
> Dmitriy Krichevskiy
>
> I have a large panel dataset (about 14000 individuals over 84 periods
> - unbalanced). I want to get to the limiting distribution of
> distribution dynamics for 2 separate groups. Is anyone aware of the
> way to analyze this?
>  -markov- does not work on panel data and -xttrans- produces a single
> transition matrix by lumping all of the transitions into '1 old to 1
> new' transition. Ideally I would want a dynamic panel data model but
> the dependant variable is a quintile so it seems to me that
> multinomial logit or probit is in order. Thanks for all of your help
> again.
>
>
> *
> *   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/

*
*   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/



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