Date:         Thu, 6 Jan 2005 14:10:52 -0800
Reply-To:     Dale McLerran <stringplayer_2@YAHOO.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
Comments:     DomainKeys? See http://antispam.yahoo.com/domainkeys
From:         Dale McLerran <stringplayer_2@YAHOO.COM>
Subject:      Re: Clustering with conditional logit model -- NLMIXED
In-Reply-To:  <200501052346.j05Nkv8t013339@listserv.cc.uga.edu>
Content-Type: text/plain; charset=us-ascii

Noreen,

I have no expertise with the procedure MDC.  I took a brief look
at the description of the MDC procedure and noticed that for
conditional logit models, the MDC procedure includes a random
error term in eta{ij} with Gumbel distribution.  Could this be
why the results from the MDC procedure and the procedure NLMIXED
do not match each other?

Your NLMIXED code looks like code which I have seen for brand
choice modeling.  I believe that the random effects model which
you specify is overparameterized in that you have included a
random intercept term for each of the three choices.  Note that
you only have fixed effect intercept terms for walking and
taking the bus.  You don't include a fixed effect intercept for
driving because that would result in overparameterization, right?
Well, by the same token, including a random effect intercept
term for each of the choices would also result in
overparameterization.  Thus, I believe that you want to fit the
model

  proc nlmixed data=mnlplay;
    parms beta 1=-.1 beta2=-.1 beta3=-.1 busdum=-1 walkdum=2
          /*starting values are from proc mdc results*/;
    eta1 = beta1*timeauto;
    eta2 = busdum + beta2*timebus + u_bus;
    eta3 = walkdum + betae*timewalk + u_walk;
    prob1 = eta1 / (eta1 + eta2 + eta3);
    prob2 = eta2 / (eta1 + eta2 + eta3);
    prob3 = eta3 / (eta1 + eta2 + eta3);

    if mode=1 then loglike=log(prob1);  else
    if mode=2  then loglike=log(prob2);  else
    if mode=3  then loglike=log(prob3);
    model mode ~ general(loglike);
    random u_bus u_walk ~
           normal([0,0], [exp(2*s_bus),
                          exp(s_bus+s_walk),
                          exp(2*s_walk)]) subject=houseID;
  run;


Dale


>


--- Noreen McDonald <noreen@BERKELEY.EDU> wrote:

> I am developing a relatively standard multinomial model to predict
> travel
> mode (auto, walk, bus).  I've used proc mdc to estimate the basic
> conditional logit model and get reasonable results (see below for
> code).
> My problem is that I sample multiple members of the same household
> and
> want to account for this clustering in order to get appropriate std
> errors.  There appears to be no simple 'repeated' statement available
> in
> proc mdc.  I've been looking at proc nlmixed, but have been utterly
> flummoxed by the syntax.
>
> I would like to estimate the probabilities of choosing each mode with
> a
> random effect based on household membership(ie Ua=b1*timeauto+u1;
> Ub=a1*busdum+b2*timebus+u2; Uw=a2*walkdum+b3*timewalk+u3, where u1-u3
> are
> re with subject=houseid).  I have tried to play around with the
> nlmixed
> syntax simply to replicate my proc mdc results and have been
> unsuccessful.  I'd really appreciate help on:
> 1.  how to formulate multinomial logit models for nlmixed (see my
> very
> weak attempts below),
> 2.  how to introduced a random effect based on household id, and
> 3.  how the data should be set up for nlmixed (ie is it the same as
> for
> proc mdc with a conditional logit model)
>
> Thanks so much !!! Noreen McDonald
>
> /*working conditional logit code*/
> proc mdc data=logit.mnl3std;
>      model decision = busdum walkdum  timeauto  timebus timewalk
>      / type=clogit optmethod=qn covest=hess    choice=(mode 1 2 3);
>       id pid;
>      title 'MDC Simple: Car Bus Walk';
> run;
>
> /*non-working attempts to generate nlmixed model which replicates
> proc mdc
> results*/
> proc nlmixed data=mnlplay;
>   parms beta 1=-.1 beta2=-.1 beta3=-.1
>  busdum=-1 walkdum=2/*starting values are from proc mdc results*/;
>    prob1 = exp(beta1*timeauto) /
>           (exp(beta1*timeauto) +
>            exp(busdum +beta2*timebus) +
>            exp(walkdum + beta3*timewalk) );
>   prob2 = exp(busdum+beta2*timebus ) /
>           (exp(beta1*timeauto) +
>            exp(busdum+beta2*timebus ) +
>            exp(walkdum + beta3*timewalk) );
>   prob3 = exp(walkdum + beta3*timewalk) /
>           (exp(beta1*timeauto) +
>            exp(busdum +beta2*timebus) +
>            exp(walkdum + beta3*timewalk) );
>
>   if mode=1 then loglike=log(prob1);  else
>   if mode=2  then loglike=log(prob2);  else
>   if mode=3  then loglike=log(prob3);
>   model mode ~ general(loglike);
> run;
>


=====
---------------------------------------
Dale McLerran
Fred Hutchinson Cancer Research Center
mailto: dmclerra@NO_SPAMfhcrc.org
Ph:  (206) 667-2926
Fax: (206) 667-5977
---------------------------------------

__________________________________________________
Do You Yahoo!?
Tired of spam?  Yahoo! Mail has the best spam protection around
http://mail.yahoo.com