Re: st: RE: Plotting interactions

[David Hoaglin <dchoaglin@gmail.com>]

Kostas,

It may be desirable to hold constant the variables that are not of
primary interest, but whether one can do so in a particular situation
is an empirical question.  One cannot assume that those variables can
be held constant.  Those variables may vary in the data, despite the
analyst's desire to hold them constant, so that acting as if they are
constant produces an extrapolation beyond the data.  Those .do files
provide an intuitive approach, but not necessarily a realistic one.

I make a point of not referring to the other variables in the model as
"control variables," because some people will interpret the
coefficient of X as summarizing how Y changes with X when the other
predictors are held constant.  The proper general interpretation of
the coefficient of X is that it summarizes the change in Y per unit
increase in X after adjusting for simultaneous linear change in the
other predictors in the data at hand.  It is straightforward to show,
mathematically, that that is what multiple regression does.

David Hoaglin

On Mon, Sep 30, 2013 at 8:01 AM,  <k.gemenis@utwente.nl> wrote:
> This is indeed an assumption of these .do files for the marginal effects plots. For many applications (not only in political science), computing marginal effects by holding control variables constant makes sense. In many instances one wants to see the marginal effect of X on Y across the levels of modifying variable Z assuming a host of controls that are not the primary interest of the investigated hypothesis (age, gender and what not) constant. This is definitely a limitation because if one wants to vary the controls they would have to compute multiple plots. Yet the provided .do files provide a pragmatic and intuitive approach in plotting interactions.
>
> Best,
> Kostas

*
*   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/