Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Suest with different estimators (logit & mlogit)

From   Maarten buis <[email protected]>
To   [email protected]
Subject   Re: st: Suest with different estimators (logit & mlogit)
Date   Fri, 11 Feb 2011 11:07:46 +0000 (GMT)

--- On Fri, 11/2/11, carola.degroot wrote:
> I would love to use the suest command to test the
> differences between the coefficients of two different
> models. Although the sample of the models differ,
> they partly overlap (the sample used in the second
> model is a subselection of the sample used in the
> first model). The first model is a logistic
> regression model (logit), the second model is a
> multinomial regression model (mlogit). So although
> the models are completely the same with respect to
> the independent variables, the dependent differs. 
> I've read that Suest allows for different estimators,
> but my supervisor is worried that suest won't provide
> "reliable" ("fair") results because it might not be
> fair to compare the coefficients of a model with an
> estimator with only 2 categories with a model with an
> estimator with 3 categories.  Unfortunately, I'm not
> really an expert with the methodology behind the
> suest command, nor is my supervisor. Therefore I'm
> not sure whether this concern is indeed justified,
> and whether it's necessary to change the dependent
> of the second model into a binomial variable in
> order to run a fair suest for the cross-model
> comparison.

This is not a matter of fairness and it is also not a
matter of -suest-, it is a matter of comparing 
comparable stuff. -suest- will let you make that 
comparisson, question is does it mean anything?

Lets say your mlogit model has for its dependent
variable categories A, B, and C, and that A is
the reference category. In that case the parameters
in the first panel of the output have a dependent
variable that is the logarithm of the expected
number of Bs per A, while in the second panel the
dependent variable is the logarithm of the 
expected number of Cs per A.

Now assume your logit model only distinguishes
between (A or B) and C. In that case the dependent
variable is the logarithm of the expected number 
of Cs per (A or B). The parameters are obviously 
very different beasts, so it is hard to see how 
such a comparison can be meaningful.

I see two options: often, you can, with a bit of 
creative derivation, find comparable contrasts.
In that case you can use -suest- with -logit- and
-mlogit- models, but your subsequent test commands
will be complicated. Alternatively, you can turn
your -mlogit- in a -logit- model by recoding the
dependent variable such that they have the same

A good introduction in -mlogit- and -logit- models
is: J. Scott Long and Jeremy Freese  (2006) "Regression 
Models for Categorical Dependent Variables Using 
Stata, 2nd Edition". College Station, TX: Stata Press.

Hope this helps,

Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen


*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index