Date:         Fri, 9 Jan 2009 04:45:49 -0800
Reply-To:     Oliver.Kuss@MEDIZIN.UNI-HALLE.DE
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Oliver.Kuss@MEDIZIN.UNI-HALLE.DE
Organization: http://groups.google.com
Subject:      Re: proc nlmixed for multinomial logit with random effects
Comments: To: sas-l@uga.edu
Content-Type: text/plain; charset=ISO-8859-1

On 8 Jan., 19:59, osbur...@BLS.GOV (Jane Osburn) wrote:
> I want to use proc nlmixed in the following problem so that I can later
> include other components in the likelihood. I am trying to estimate a
> multinomial logit with random effects similar to the one below, (Dale
> McLerrin helped that person) except I do not want to expand the dataset to
> include a separate observation for every level one unit, because it would
> make the dataset too large. I would instead like to have my likelihood
> include n, the numbers of workers in occ1, occ2, or occ3 ;
>
> if occ = 1 then then loglike= n* log(prob1);  else
> if occ = 2 then then loglike= n* log(prob2);  else
> if occ = 3 then then loglike= n* log(prob3);
>
> proc nlmixed data=cars;
>   parms b1_1=0 b2_1=0 b3_1=0
>         b1_2=0 b2_2=0 b3_2=0
>         sig11=0.10 sig12=-0.05 sig22=0.10;
>   prob1 = exp(b1_1 + gamma1 + b2_1*age + b3_1*gender) /
>           (1 + exp(b1_1 + gamma1 + b2_1*age + b3_1*gender)
>              + exp(b1_2 + gamma2 + b2_2*age + b3_2*gender));
>   prob2 = exp(b1_2 + gamma2 + b2_2*age + b3_2*gender) /
>           (1 + exp(b1_1 + gamma1 + b2_1*age + b3_1*gender)
>              + exp(b1_2 + gamma2 + b2_2*age + b3_2*gender));
>   prob3 = 1 /
>           (1 + exp(b1_1 + gamma1 + b2_1*age + b3_1*gender)
>              + exp(b1_2 + gamma2 + b2_2*age + b3_2*gender));
>
>   if manufact='American' then loglike=log(prob1);  else
>   if manufact='Asian'    then loglike=log(prob2);  else
>   if manufact='European' then loglike=log(prob3);
>
>   model manufact ~ general(loglike);
>   random b1 b2 ~ normal([0,0], [sig11, sig12, sig22]) subject=family;
> run;
>
> Thanks in advance for any help.
>
> Jane Osburn
> Bureau of Labor Statistics
> 202 691-6504

Dear Jane,
sorry for the self-promotion, but together with Dale I wrote a paper
on the topic (Kuss O, McLerran D. A note on the estimation of the
multinomial logistic model with correlated responses in SAS. Comput
Methods Programs Biomed. 2007 Sep;87(3):262-9.) which hopefully would
answer most of your questions on the model you have.
I would be happy to send you the paper if you have no direct access.

Hope that helps,
Oliver


and