[Limdep Nlogit List] (no subject)

[Fred Feinberg]

The problem seems more severe than would be addressed by reformatting
the data.  It's not clear from your description whether you're running a
(McFadden type) conditional logit model -- iin which the characteristics
of the choices themselves would be changing from choice occasion to
choice occasion -- or a multinomial logit (where you would obtain
coefficients for characteristics of the decision-maker).  It sounds like
you have elements of both, which is fine.  [Note that programs like SPSS
and SAS, for example, will run multinomial models, but not conditional
ones, and have to be tricked into doing so by, for example, Cox
regression.]

Anyway, if you are trying to get some set of coefficients for the
*attributes* of the choices themselves, as common in transportation and
ubiquitous in marketing (e.g., price changes), then not having those
attributes for the options that were not chosen is a BIG problem. One
possibility is to do some type of (perhaps multiple) imputation.  This
is complicated by your need to assume that the he covariates that were
NOT observed are drawn from the same distribution of those that were.
For example, if we don't observe the price of something you didn't buy,
can we assume it's drawn from the same distribution of the prices that
WERE observed?  That's rarely true.  Usually, you'll need a separate
model for the unobserved outcomes, which is where Heckman-type models
can be useful.

Joe Ibrahim (http://www.hsph.harvard.edu/facres/ibrhm.html) has done a
lot of work on this and related problems, primarily from a Bayesian
perspective, which handles missing data in a natural fashion.  He may
have written something on your specific problem.  A good general
reference is:

Ibrahim JG, Chen MH, and Lipsitz SR. Missing Responses in Generalized
Linear Mixed Models When The Missing Data Mechanism is Nonignorable,
Biometrika, 2001; 88:551-564.

FF


"Koetse, M.J.J. (Mark)" wrote:

> L.S.
>
>
>
> We are doing some research in which we want to estimate a multinomial
> logit model on transport mode choices by individuals. We only have
> information on the choices that individuals make, some characteristics
>
> of the transportation mode chosen, and on individual characteristics.
> We
> do not have any information about the choices that are not selected.
>
>
>
> As far as we can see, the data set-up in Limdep requires information
> about the selected choice as well as the non-selected choices. This
> implies that we need to convert our single line observations (in which
>
> one line represents one individual) to multiple line observations (in
> which one individual is represented by a number of lines that is equal
>
> to the number of alternatives in the choice set). However, since we
> have
> no specific information on these possible non-selected alternatives,
> this would imply that these extra rows contain identical information
> as
> the original row.
>
>
>
> Since our dataset is very large, this procedure would take a lot of
> time. Our question is basically whether it is possible to estimate a
> multinomial logit without applying this data transformation procedure?
>
>
>
>
> Any help is greatly appreciated.
>
>
>
> Kind regards,
>
> Mark Koetse and Muhammad Sabir
>
>
>
> Department of Spatial Economics
>
> Vrije Universiteit Amsterdam
>
> De Boelelaan 1105
>
> 1081 HV Amsterdam
>
>
>
> _______________________________________________
> Limdep site list
> Limdep at limdep.itls.usyd.edu.au
> http://limdep.itls.usyd.edu.au