[Limdep Nlogit List] random parameters logit

[William Greene]

Dear Marjolein. Evidently the log likelihood has more than one mode.
Also, I have seen other cases in which the starting values affect
whether the estimator is able to converge. This can happen when
there is more than mode or the likelihood function is complex, as it
is for a mixed logit model.  Notice that the function value in your
second model is lower than the first. This is the negative of the
log likelihood, so the second estimates are noticeably better than
the first. I would trust the second set of results. One remaining
possibility is to try other starting values using ;Start=list of
values (look in the final results for the correct ordering).
/Bill Greene

----- Original Message -----
From: "Marjolein Maria Juliette Harmsen-van Hout" <Marjolein.harmsen at rwth-aachen.de>
To: "Limdep and Nlogit Mailing List" <limdep at limdep.itls.usyd.edu.au>
Sent: Friday, September 17, 2010 11:08:43 AM GMT -05:00 US/Canada Eastern
Subject: Re: [Limdep Nlogit List] random parameters logit


Dear prof. Greene,

Thanks for your reply. Even estimation with 500 Halton draws provides the same difference. Moreover, the program claims normal exit for both models: 

1st model: "* Converged Normal exit from iterations. Exit status=0. Function=  .11994470974D+04, at entry,  .11904739292D+04 at exit Error  1027: Models - estimated variance matrix of estimates is singular"

2nd model: "* Converged Normal exit from iterations. Exit status=0. Function=  .11906631459D+04, at entry,  .11817440931D+04 at exit +---------------------------------------------+ | Random Parameters Logit Model etcetera"

In case of either of your suggested possibilities, in how far can I trust the estimations found with my 2nd model?

Best regards,
Marjolein Harmsen - van Hout


On 17/09/10, William Greene  <wgreene at stern.nyu.edu> wrote:

> Dear Marjolein. It does appear that the two commands should produce
> the same model and, using Halton draws, the same results.  
> I can see two possibilities:
> (1) Although the models are the same, the starting values for the
> mixed logit estimates are not the same in the two cases. In the first
> case, the starting value for the interaction term would be zero. In 
> the second, it would be the MNL coefficient.  It is possible that the
> mixed logit has the kind of ill behaved log likelihood that would 
> produce this result.  
> (2) A less likely possibility is that rounding error in the computations 
> in the first case is producing the problem. This would imply that the 
> data set is very badly conditioned.
> Theory (1) sound more likely to me.  The second set of results are
> obviously to be preferred. I would add ;OUTPUT=3 to your command to
> display the derivatives during the solution, and verify that you really
> did get a normal exit from the optimizer with the second specification.
> I would also suggest increasing the number of draws to, say, 100. This 
> may help.
> /Bill Greene
> 
> ----- Original Message -----
> From: "Marjolein Maria Juliette Harmsen-van Hout" <Marjolein.harmsen at rwth-aachen.de>
> To: limdep at limdep.itls.usyd.edu.au
> Sent: Friday, September 17, 2010 4:24:05 AM GMT -05:00 Colombia
> Subject: Re: [Limdep Nlogit List] random parameters logit
> 
> 
> Dear all,
> As I did not receive any reply to my previous message, I hereby send it once more. If you need more information, please notify me. Thanks!
> 
> On 25/08/10, "Marjolein Maria Juliette Harmsen-van Hout"  <Marjolein.harmsen at rwth-aachen.de> wrote:
> 
> > > Dear all,
> > > 
> > > I encountered the following puzzle with my dataset:
> > > 
> > >  NLOGIT;
> > >  Lhs=CHOICE,CSET,ALT;
> > >  Rhs=VAR1;
> > >  Pds=PER;
> > >  Rpl=VAR2;
> > >  Fcn=VAR1(N);
> > >  Halton;Pts=20$	
> > > 
> > > leads to Error 1027: Models - estimated variance matrix of estimates is singular, whereas:
> > > 
> > >  CREA;VAR1_VAR2=VAR1*VAR2$
> > >  NLOGIT;
> > >  Lhs=CHOICE,CSET,ALT;
> > >  Rhs=VAR1,VAR1_VAR2;
> > >  Pds=PER;
> > >  Rpl;
> > >  Fcn=VAR1(N);
> > >  Halton;Pts=20$        	
> > > 
> > > provides normal estimations. Does anybody have an idea how this could be possible?
> > > 
> > > Thanks in advance for your reaction!
> > > 
> > > Kind regards, 
> > > ____________________________________________________________
> > > 
> > > Dr. drs. Marjolein Harmsen - van Hout
> > > Post-doctoral researcher
> > > 
> > > Tel.: +49 241 80 49835
> > > Fax: +49 241 80 49829
> > > E-mail: mharmsen at eonerc.rwth-aachen.de
> > > Website: www.eonerc.rwth-aachen.de
> > > 
> > > Lehrstuhl für Wirtschaftswissenschaften, insb. Energieökonomik
> > > Institute for Future Energy Consumer Needs and Behavior (FCN)
> > > E.ON Energy Research Center (E.ON ERC)
> > > RWTH Aachen University
> > > Mathieustraße 6, D-52074 Aachen
> > > Germany
> > > ____________________________________________________________
> -- 
> ____________________________________________________________
> 
> Dr. drs. Marjolein Harmsen - van Hout
> Post-doctoral researcher
> 
> Tel.: +49 241 80 49835
> Fax: +49 241 80 49829
> E-mail: mharmsen at eonerc.rwth-aachen.de
> Website: www.eonerc.rwth-aachen.de
> 
> Lehrstuhl für Wirtschaftswissenschaften, insb. Energieökonomik
> Institute for Future Energy Consumer Needs and Behavior (FCN)
> E.ON Energy Research Center (E.ON ERC)
> RWTH Aachen University
> Mathieustraße 6, D-52074 Aachen
> Germany
> ____________________________________________________________
> _______________________________________________
> Limdep site list
> Limdep at limdep.itls.usyd.edu.au
> http://limdep.itls.usyd.edu.au
> _______________________________________________
> Limdep site list
> Limdep at limdep.itls.usyd.edu.au
> http://limdep.itls.usyd.edu.au
-- 
____________________________________________________________

Dr. drs. Marjolein Harmsen - van Hout
Post-doctoral researcher

Tel.: +49 241 80 49835
Fax: +49 241 80 49829
E-mail: mharmsen at eonerc.rwth-aachen.de
Website: www.eonerc.rwth-aachen.de

Lehrstuhl für Wirtschaftswissenschaften, insb. Energieökonomik
Institute for Future Energy Consumer Needs and Behavior (FCN)
E.ON Energy Research Center (E.ON ERC)
RWTH Aachen University
Mathieustraße 6, D-52074 Aachen
Germany
____________________________________________________________
_______________________________________________
Limdep site list
Limdep at limdep.itls.usyd.edu.au
http://limdep.itls.usyd.edu.au