Re: st: why regress drops dummy variables

["Michael S. Hanson" <mshanson@wesleyan.edu>]

On Jan 2, 2007, at 9:20 AM, Rajesh Tharyan wrote:

> char sex[omit]m
> xi: regress car mcap no  i.director i.sex  i.director*i.sex
>
>
>          car |      Coef.   Std. Err.      t    P>|t|     [95% Conf.
> Interval]
> ------------- 
> +--------------------------------------------------------------
> --
>         mcap |  -.0031127    .000147   -21.17   0.000    -.0034009
> -.0028245
>           no |   .0030767   .0001176    26.17   0.000     .0028463
> .0033072
> _Idirector_2 |    -.00387   .0006057    -6.39   0.000    -.0050571
> -.0026829
>      _Isex_1 |   .0004936   .0023094     0.21   0.831    -.0040328
> .00502
> _Idirector_2 |  (dropped)
>      _Isex_1 |  (dropped)
> _IdirXsex_~1 |   .0007891   .0031731     0.25   0.804    -.0054303
> .0070085
>        _cons |   .0035362   .0015175     2.33   0.020     .0005619
> .0065106
>
> I do not understand why the dummy variables are dropped (are they  
> dropped?).
	Yes, they are surely dropped as Stata says they are dropped!


> Is there a solution to this?
	Yes, use the command:

xi: regress car mcap no  i.director*i.sex

As you will note from -help xi-, the above interaction term "Creates  
dummies for categorical variables ... and all interactions and main  
effects."  That is, by putting "i.director*i.sex" in your -xi: regress-  
command, Stata automatically expanded that expression to include  
"i.director" and "i.sex" individually (the "main effects") as well as  
the interaction term you had intended.  Thus, because you already  
separately included "i.director" and "i.sex" as regressors, they  
appeared in the regression twice and Stata had to drop one of their  
occurrences each.  This is clearly indicated in your -regress- output,  
quoted above.

	Why does Stata do this?  I suspect it is to protect less sophisticated  
users from including only interaction terms without the main terms...  
unless they really know what they are doing.[*]  As your example  
demonstrates, it is relatively uncommon for someone to want to do that.  
 Hope this helps.

                                        -- Mike

[*] As far as I can tell, it is not possible to include only the  
interaction term and not the main terms for two categorical  
(dichotomous) variables using the -regress...- command with the -xi:-  
prefix.  Rather, one must first create the interaction terms with -xi-  
as a command, then include only the interaction term(s) in the  
specification of -regress-.  It might be possible to accomplish the  
inclusion of *just* the interaction terms with the following syntax:

xi: regress car mcap no  i.director|sex

but as I do not have access to a database on which to test such a  
specification, that remains a conjecture (and exercise for the reader).

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/