Re: st: Chow test quadratic effect

[David Hoaglin <dchoaglin@gmail.com>]

Hi, Natalia.

The Chow test applies to ordinary linear regression models.  It is
worth noting that Greg Chow's paper appeared in 1960, when computing
resources were much more limited than they are today (think
electromechanical desk calculators).  I see no need to use it now.  I
would set up the models with the appropriate interactions and use a
likelihood-ratio test.  That approach is more flexible.  And,
importantly, it generalizes to other types of regression models that
involve a likelihood, such as your multinomial logit model.

As I read your message, it is not quite true that "interactions would
not help."  They would be tedious, but you should be able to reduce
the effort by using factor variables in Stata.

It would make sense to focus first on agesq, since its presence in the
model requires that age also be included.

David Hoaglin

On Tue, Apr 2, 2013 at 11:34 AM, natalia malancu
<natalia.malancu@gmail.com> wrote:
> Hi David,
>
> thank you for your answer.
> The example I gave in my message is an invented one. Since I am
> interested in the stata command, I thought the easiest way to receive
> any suggestions was to use an already existing dataset (in this case a
> stata online one) and generate a super simple example.
> That being said, I fully agree with both your remarks. My actual model
> refers to two differente policy periods and intereactions would not
> help since my assumption is that each IV contributes differently to
> the outcome depending on the period we are talking (i would basically
> have to interact each with the dummy for period). My nonlineraty does
> actually have a quadratic form.
> Any suggestion on how the syntax for age +agesq looks like when
> performing a chow test for a multiniomial logit?
>
> Thanks,
> Natalia
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/