Re: st: RE: interaction without main effects

[Steven Samuels <sjsamuels@gmail.com>]

I'm strongly share Nick's opinion. In addition, how you report the  
interaction depends on how you "found" it. If you had an advance  
hypothesis about that specific interaction, then you can present it  
with the given p-value.  But if you found it after a search, then you  
must correct the p-value for multiple testing. See the -help- for - 
_mtest- and look at Roger Newson's -smileplot-, available via --ssc-.

Steve
sjsamuels@gmail.com

On Dec 14, 2010, at 2:37 PM, Nick Cox wrote:

Setting aside Cox models: this issue has been aired often on this list.

If you're fitting a term in x * z without fitting terms in x and z you  
are implying that you know that a and b in

a x + b z + c x z

must be 0, which is why you are fitting without x and z. How come? If  
there's theory for that, well and good. But I don't think being  
uninterested in something adds up to a theory that something is  
negligible.

On the whole my prejudice is that interactions without main effects  
just make life more difficult and focusing on individual significance  
tests implies that the terms are separate, which they aren't really.  
There is a brief but compelling discussion in McCullagh and Nelder's  
book (reference in [R] glm).

Nick
n.j.cox@durham.ac.uk

Grethe Søndergaard

I have a question about interactions.

I am analysing my data using a stratified cox regression analysis. In
one of my analyses I have found one interaction between two variables:
xi: stcox i.birthyear i.education i.birthyear#i.education,  
strata(strata)

After having found this interaction I run the same analysis again -
but this time without the main effects of education, since I am not
interested in presenting these:
xi: stcox i.birthyear i.birthyear#i.education, strata(strata).
This model runs without any problems.


In another analysis I find two significant interactions:
xi: stcox i.birthyear i.sex i.education i.birthyear#i.education
i.sex#i.education, strata(strata)

However, once I try to run this analysis without the main effects of
education, an error message occurs:
xi: stcox i.birthyear i.sex i.birthyear#i.education i.sex#i.education,
strata(strata)
error: 3.birthyear#1. education omitted because of linearity
error: 3.birthyear#2. education omitted because of linearity
error: 3.birthyear#3. education omitted because of linearity
(birthyear is categorised in three categories and education is
categorised in four categories)

Is it okay to include two interactions in a model, where not all main
effects are included?
And is it possible that collinearity only occurs in the model without
the main effects – and not in the model where the main effects are
included (or am I doing something wrong)?





I am analysing my data using a stratified cox-regression analysis. I
find two signficant interactions, when I write my model like this:
xi: stcox i.sex i.birthyear i.education i.sex#i.education
i.birthyear#i.education
This model runs without any problems.

However, I dont want to include the main effects of education, so I
have tried to write my model like this:
xi: stcox i.sex i.birthyear i.sex#i.education i.birtheyear#i.education

But. When I write my model like this, I get an error message saying
that the interaction between




I have found two significant interactions and I now want to a

*
*   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/


*
*   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/