Re: st: Using mixlogit as a substitute of xtlogit.

[Arne Risa Hole <arnehole@gmail.com>]

As Maarten says it is more relevant to compare -mixlogit- with
-xtmelogit- since they both allow for random intercepts *and* slope
parameters while -xtlogit- allows for a random intercept only. In
addition -mixlogit- can handle cases with more than two discrete
outcomes, although this is not relevant for ippab's application. The
other main difference between -mixlogit- and -xtmelogit- is that the
integrals in the likelihood function are approximated by using
simulation in the former case (-mixlogit-) and quadrature in the
latter (-xtmelogit-). As pointed out by Maarten the results should be
similar, but estimation times are likely to be faster with -mixlogit-
when there are several random parameters in the model.

I agree with Nick - a familiar line by now to those who followed the
recent UK election - that it's always a good idea to point out when a
command is user-written.

Arne (author of -mixlogit-)

On 4 June 2010 09:31, Maarten buis <maartenbuis@yahoo.co.uk> wrote:
> --- On Fri, 4/6/10, ippab ippab wrote:
>> I am wondering if there is any benefit of using mixlogit
>> for a binary dependent variable in a panel data.
>> Basically, to use mixlogit, I will have to create two
>> alternatives (which are complementary), an alternative
>> specific constant, and interactions between independent
>> variables and the alternative specific constant.  Is
>> this is a wise thing to do?
>
> No, you should get very similar results, and the differences
> are due to particularities of the maximizing algorithms, so
> have no substantive interpretation.
>
>> My vague understanding is that mixlogit allows for more
>> heterogeneity than xtlogit.
>
> Not true, the value of -mixlogit- is that it allows one to
> estimate a random effects model within a multinomial logit.
>
>> But, I am confused about interpreting a population average
>> beta obtained from xtlogit in light of estimates from
>> mixlogit.  Just to give an example, if the estimate
>> for x1 is 2.05 from xtlogit, the interpretation would be
>> that increasing x1 increases the the likelihood of y=1.
>> Now, getting a similar estimate from mixlogit, e.g., mean
>> is 2.5 for x1 with std 3.10, makes the interpretation
>> complicated.  This means for about 20.9% of the sample,
>> increasing x1 does not increase the likelihood of
>> y=1.  Am I understanding these correctly?
>
> Yup, and now it is more clear to me why you are comparing
> with -mixlogit-. I guess that you were interested in making
> some of the coefficients random as well. If that is the case
> then you should have used -xtmelogit- instead.
>
>> The more critical question is, if more than 20% or 30% of
>> the sample have different preference, what does it mean to
>> have a positive significant coefficient (mean)?
>
> It means that people behave differently to the same stimulus,
> but that on average the effect is positive. A possible next
> step would be to try to explain this variation by adding
> interactions between higher level variables and the lower
> level variable x1.
>
> Hope this helps,
> Maarten
>
> --------------------------
> Maarten L. Buis
> Institut fuer Soziologie
> Universitaet Tuebingen
> Wilhelmstrasse 36
> 72074 Tuebingen
> Germany
>
> http://www.maartenbuis.nl
> --------------------------
>
>
>
>
> *
> *   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/