Re: st: Prediction at the 'mean' of a categorical- xi3?

[David Hoaglin <dchoaglin@gmail.com>]

Andrew,

I think of Jay's suggestion as equivalent to taking a mixture of the 6
categories as the basis for the centering. In that mixture, the weight
for each category would be its proportion in your data or in a
relevant population.  Approached in that way, the contribution to the
constant term should not be too difficult to interpret.

If you are setting various (continuous) predictors at their means for
a prediction, you still need to check that your data support such a
prediction.  Many sets of data do, but not all.

David Hoaglin

On Wed, Mar 13, 2013 at 11:59 PM,  <graupel75@gmail.com> wrote:
> JVerkuilen writes:
>
>>Nothing in the math says you can't center dummy variables; I do it all
>>the time using the usual Stata commands to make the intercept a
>>meaningful number. Effects coding in some sense does this in the
>>classic ANOVA setup but won't in a general regression due to the lack
>>of orthogonality.
>
> I've used both -xi3- center option and generated the mean-centered var
> by hand, and run the regression with this new var, but the log lik and
> AIC/BIC are quite different indicating that more than just the ref
> category has changed...
> so I must be confused.
>
> Let's say I have a categorical (levels 1 to 6), do I then estimate
> with the mean-centered value as a continuous var, and predict at 3.5?
> This doesn't make so much sense as it would assume the steps between
> levels are equivalent; I've looked in both Agressi and Long's Stata
> book without success-
>
> thanks!
> Andrew
*
*   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/