Re: st: asmprobit, problems with correlation patterns

["Barbara Hanel " <Barbara.Hanel@wiso.uni-erlangen.de>]

Hallo Rich,
In the first run I tried to estimate the model just with the a-s- Variable "eink" - it doesn�t work too. In fact the model converged, but estimated a correlation coefficient of exactly -1. Furthermore, I got extremly large standard errors for the variance parameter. (It was a specification with only 3 destination states, so there was only one covariance parameter beside the variance parameter.)
I�t seems to me to be a similar problem like in the model with 4 destination states and a correlation coefficient of _exactly_ -0.5 But I just don�t know what it in fact could be.
That problem even persists only with the as-variable.

I defined the restricted pattern primarily because I hoped it would improve the estimation compared to the full model, because there are less parameters to identify. So I have even less hope to find a solution with the default option.



I hope I don�t waste too much of your time...

Thanks a lot,
Barbara


Am 23 Apr 2007, um 8:09 hat Richard Gates geschrieben:

Barbara,

If the correlation problem persists after dropping the
c-s variable from the indepvar variable list let me
know.

Try the default differenced covariance, it should give
you the optimal likelihood.  It will produce a 3x3
covariance, estimating 5 Cholesky transformed covariance
parameters.  These estimates are not interpretable, but
generally the covariance parameters are considered a nuisance
parameters.  It will also give you the full model to compare
with your restricted (in the covariance structure) model using
-lrtest-

I believe it is the alternative-specific (a-s) variables that
identify the covariance parameters since they vary within
each case.  I have tried running models with and without a-s
variables using the differenced covariance structure and as
soon as I drop the a-s variables -asmprobit- does not converge.
So I think your eink variable will be key in identifying the
covariance parameters.

My guess is that once you drop AB from indepvars estimation
will improve.

Best Regards,

-Rich

On Sun, 2007-04-22 at 15:39 +0200, Barbara Hanel wrote:
> Hallo,
>
> That sounds good. (I thought I first had to define all explanatories
> and than wich of them is case-specific.)
> I will try as you suggested.
>
> Of course I would still be appreciated if there is anybody with an
> idea concerning the correlation-pattern problem...
>
> Thanks a lot, Rich!
> Barbara
>
> Am 20 Apr 2007, um 12:48 hat Richard Gates geschrieben:
>
> Barbara inquired about the collinear message she is receiving
> from -asmprobit-:
>
> > 2) If I include more than one explanatory variable in the model I get
> > the message:
> > "Note: two or more of the variables are collinear; convergence may
> > not be achieved"
> > That's the case even if I include:
> > * one variable varying across alternatives and cases and
> > * only one other variable that is constant across the alternatives,
> > varies only by "i".
> >
> > I don´t understand how two such variables could be collinear?
>
> I will include a snippet of her output below
>
>
> > ------------------------------------------------------------------------------
> >         wahl |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
> > -------------+----------------------------------------------------------------
> > altern       |
> >         eink |    .000631   .0001332     4.74   0.000       .00037    .0008921
> >           AB |  (dropped)
> > -------------+----------------------------------------------------------------
> > ET           |  (base alternative)
> > -------------+----------------------------------------------------------------
> > AL           |
> >           AB |   6.498919   2.957141     2.20   0.028     .7030285    12.29481
> >        _cons |   -1.80445   .1309075   -13.78   0.000    -2.061024   -1.547876
> > -------------+----------------------------------------------------------------
> > R            |
> >           AB |  -4.308268   2.407709    -1.79   0.074     -9.02729    .4107547
> >        _cons |  -1.236563   .0867427   -14.26   0.000    -1.406576   -1.066551
> > -------------+----------------------------------------------------------------
> > KS           |
> >           AB |   15.49664   1.622816     9.55   0.000     12.31598     18.6773
> >        _cons |  -.1677057   .0752585    -2.23   0.026    -.3152096   -.0202019
> > -------------+----------------------------------------------------------------
> >   /lnsigmaP1 |   .0504696          .        .       .            .           .
>
> The call to -asmprobit- is
>
> . asmprobit wahl eink AB, case(z1) alternatives(altern) casevars(AB)
> > correlation(pattern B)
>
> Barbara is including a case-specific variable, AB, in two places
> in her model specification:  the -indepvars- variable list and
> the -casevars()- variable list.  This really should not be allowed
> and -asmprobit- should have terminated with an error.
>
> Since AB is case-specific Barbara will have to drop it from the
> -indepvars- variable list.
>
> . asmprobit wahl eink, case(z1) alternatives(altern) casevars(AB)
> > correlation(pattern B)
>
>
> I hope this helps.
>
> -Rich
> rgates@stata.com
>
>
> *
> *   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/
> -- -- -- -- -- -- -- -- -- -- -- -- -- --
> Dipl.-Volksw. Barbara Hanel
> FAU Erlangen Nürnberg
> Lehrstuhl für Statistik und empirische Wirtschaftsforschung; Prof. Riphahn,
> PhD
> 0911/5302-258
> Lange Gasse 20
> 90403 Nürnberg
> barbara.hanel@wiso.uni-erlangen.de
--