Re: Re: st: Interaction terms

[maarten buis <maartenlbuis@googlemail.com>]

On Wed, May 4, 2011 at 12:00 PM, lreine ycenna wrote:
>  I tried the factorial method, however, I have very large
> coefficients,high SE, especially when the group dummy G1 is interacted
> with all the variables and on sub-sample (e.g. if year < 2000 and >
> 1995).so my results are mostly insignificant.

If you say that both the coefficient and the SE are large than that does not
tell us much, it is the ratio of the two that determines whether or not you
can reject the hypothesis that the parameter equals 0 (i.e. = "significant").

> (1) I wonder if it's because the second half of my variables are the
> bi-products of first half, even though I'm meant to treate these
> bi-product variables as individual variables. As I gradually add more
> variables, I also have more bi-product variables.
>
> e.g. (a) regress y ov edu wealth gd eduxwealth eduxgd weathxgd ovxedu
> ovxwealth ovxgd. ovxeduxwealth ovxeduxgd ovxwealthxgd.
>
> In this case, would it bias my result to include so many bi-product
> variables?

No, if you believe that you need to add all those interactions, than you
would bias your results be leaving them out. However, I probably would
not believe that it was necessary to include all those interactions unless I
had a pretty strong theory and substantive interest in these interactions.
However, it is your research, so you alone decide what should be added
to your model not anyone else on this list.

> If so, does it make sense to run all the bi-products
> separately on a single regression? e.g. (b) regress y ov eduxwealth
> eduxgd weathxgd ovxeduxwealth ovxeduxgd ovxwealthxgd. And then compare
> the coefficients in (b) with (a).

There are special situations where such a comparison can make sense,
but it is not the lack of significance of your coefficients. Low power is a
natural consequence of adding that many interaction terms. If you believe
you need them, than you must live with the consequences.

> (2) I notice that "Regress y ov edu wealth gd ovxedu ovxwealth ovxgd
> if G1==1" produces different/ smaller coefficients and SE from that of
> the ==G1 command. Would it be incorrect to use the if G1==1 method
> instead of regress y i.G1##c.(ov edu wealth gd ovxedu ovxwealth
> ovxgd)? I don't quite understand the difference.

That means that G1 can take other values than 0 or 1. Whether or not you
want to interact with a dummy for G1== 1 or treat it as a categorical variable
depends on what that variable is supposed to contain and what you want
your model to say.

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/