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

# Re: st: Interpreting 3 way dummy interaction with margins

 From Maarten Buis To statalist@hsphsun2.harvard.edu Subject Re: st: Interpreting 3 way dummy interaction with margins Date Wed, 8 Feb 2012 11:43:28 +0100

On Tue, Feb 7, 2012 at 5:14 PM, Lauren Beresford wrote:
> I have a rather long list of controls and even when I take those out, using your tip produces too much collinearity and most of the model is dropped.  I think this is because I am including all lower order interactions in addition to the higher order interactions.

That suggests that you are not using the factor variable notation (see
-help fvvarlist-). So my first suggestion is to use those instead of
creating your indicator variables and interactions manually. You can
do so, but it is easy to do it wrong. If you have Stata 10 or less
than you cannot use factor variables, and you'll have to manually
create and choose which variables to add. For that case I translated
the example in the Stata tip to Stata 10 or less:
<http://www.maartenbuis.nl/publications/ref_cat10.html>. A quick check
is that the number of variables in the different models should not
change (don't forget the constant when you haven't excluded it with
the -nocons- option).

In your case the fact that you include all higher and lower order
interaction variables is likely a mistake: the point of that trick is
to create the mathematically equivalent model that does not include
the higher order interactions, thus making your results easier to
interpret. Say your interaction term says how the effect of race is
different between men and women, than the trick is to instead create a
model that gives you the effect of race for men and the effect of race
for women. That way the models are mathematically equivalent but you
did not include the higher order interaction term. This idea scales up
to higher order interactions as well.

Another point is that if you think that the controls belong in your
model than you should not take them out. You do need to make sure that
they all have meaningful values for zero, so center them at some
meaningful value within the range of your data. That is almost always
a good idea anyhow.

Hope this helps,
Maarten

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

http://www.maartenbuis.nl
--------------------------

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/