[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: RE: RE: out-of-sample predictions with GLM

From   Maarten buis <>
Subject   Re: st: RE: RE: out-of-sample predictions with GLM
Date   Thu, 26 Jul 2007 15:30:41 +0100 (BST)

--- wrote:
> In my case, my unit of observation is the share of a new car model as
> a percent of the total market for new cars in a given year. I have
> this share over several years. Because there are hundreds of car
> models in my data, I don't think it would be feasible to use a
> multinomial logit. None of my observation are zero or one. The x's
> are mostly car attributes for the particular car models, many of
> which change over time.
> I pool all the data and estimate with:
> glm share_i_year_t x_i_t, fam(binomial) link(logit) robust
> Company dummies are included as fixed effects. The predicted shares
> from the model sum to 1 for a given year, which is gratifying.
> Out-of-sample predictions do not sum to one, which is not gratifying
> but also not unexpected. Maarten's comment makes me wonder if there
> is a better approach.

The constraint that the proportions add up to 1 in a given year is true
in your data (if you have all models (or a model "other") as is
apperently true in your case). This constraint is however not enforced
in your model. The fact that when you add the predicted proportions up
you get a number close to 1 is primarily the result of your data and is
certainly not a characteristic of your model. With a fractional mlogit
model you could enforce that constaint, though I agree with you that in
your case that is probably unpractical. 

A different remark is that you are ignoring the panel structure of your
data. The panel equivalent of -glm- with the -robust- option is 
-xtgee-. So you might try:

xtgee share_i_year x_i_t, fam(binomial) link(logit) i(model) t(year)

(assuming that the variable indicating the model is called model, and
the variable indicating the year is called year.)

Hope this helps,

Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands

visiting address:
Buitenveldertselaan 3 (Metropolitan), room Z434

+31 20 5986715

Yahoo! Answers - Got a question? Someone out there knows the answer. Try it
*   For searches and help try:

© Copyright 1996–2017 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index