[Limdep Nlogit List] Should mixed logit (ML) be able to replicate base choice shares and should MLs be used for forecasts?

William Greene wgreene at stern.nyu.edu
Thu Mar 9 05:57:09 AEDT 2017


Richard.
1.  Mixed logit with ASCs will not exactly replicate the base choice shares.
That result is a specific algebraic result for the MNL model. Other models,
like the mixed logit will approximate the result, but only the MNL hits it
exactly.
2.  Not clear what you mean by "accurate forecasts for the population."
But, the "individual" parameters are estimates of the conditional means of
the parameter distribution.  Averaged across individuals, you get an
estimate
of the unconditional means.  If you are forecasting the sample totals, I
would use the estimated population averages not the sum of the individual
results.
/Bill greene

On Wed, Mar 8, 2017 at 10:48 AM, Richard Turner <
richard.turner at imarketresearch.com> wrote:

> Greetings,
>
> I have a couple of questions. I would greatly appreciate some commentary on
> them. Thanks in advance for any answers.
>
>    1. Should mixed logit with alternative specific constants and estimated
>    individual parameters be able to replicate the "base" choice shares
> (i.e.
>    choice shares in the data) within the data? Alternative specific
> constants
>    have this affect in multinomial logit choice models and I was wondering
> if
>    the same applied to mixed logit. *In a mixed logit model that I
>    estimated, I calculated shares using the fixed coefficients and the
>    individual level parameters, but the shares were a bit off from the
> actual
>    shares in the data*. I'm not sure if this would make a difference but,
>    about 70% of my variables in the multinomial logit model to create my
>    starting values were significant, but none of the parameters in my ML
> were.
>    The ML parameters did not seem strange in magnitude though.
>    2. Should ML with estimated individual parameters be used for forecasts?
>    I've read that they should not, but I am not sure why. Shouldn't the
>    individual parameters, if accurate, create accurate forecasts for the
>    population? If not, why should one trust that the distribution of
>    parameters was estimated correctly in the first place.
>
> Regards,
>
> Richard
> _______________________________________________
> Limdep site list
> Limdep at limdep.itls.usyd.edu.au
> http://limdep.itls.usyd.edu.au
>
>


-- 
William Greene
Department of Economics
Stern School of Business, New York University
44 West 4 St., 7-90
New York, NY, 10012
URL: http://people.stern.nyu.edu/wgreene
Email: wgreene at stern.nyu.edu
Ph. +1.212.998.0876


More information about the Limdep mailing list