Hi Erik -
Trying to look at 20*19 = 380/2=190 two-way interactions seems a bit like
fishing . .
Have you thought of doing an omnibus test? (I believe it's
possible/applicable in PHreg, someone correct me if I'm wrong?)
ie run the model with NO interactions and then run the model with all 2-way
The difference between the -2 log likelihoods is distributed as chi-square
with the degrees of freedom equal to the difference in df in the 2 models.
If that's significant, then it's telling you that you do have significant
interactions you should investigate. If not, it's telling you that all the
interactions just add noise.
That said, if you're sure you need all 20 variables and sure you need to
look at all possible 2-way interactions:
If it were me, I think I'd do a macro loop to run the PHreg, output the
variable estimates and p values to a file which is stacked onto one overall
file each time, then at the end print out just that file and look through
it. Otherwise you'll have overwhelming amounts of output.
You could even loop your macro calls . .
ie if you havea macro called phregloop that calls in var1 & var2, instead of
writing out 190 different versions of %phregloop(var1=a, var2=b);
then you might do a loop to call those.
Hope this helps
> From: Erik Christiansen[SMTP:ec@SELVMORDSFORSKNING.DK]
> Reply To: Erik Christiansen
> Sent: Friday, June 25, 2004 6:20 AM
> To: SAS-L@LISTSERV.UGA.EDU
> Subject: phreg and var1*var2 (interaction variables)
> I am working with proc phreg in a case/control design. I would like to
> analyse interaction between to variabls. I have the total of 20 variables
> my analyse, and this will result i many interaction variables. One varibel
> can interact with the rest 19 varibles. I am only interested in
> interaction variables. Is there an easy way to define the interaction
> variables? It is possible to define them in a data statment as
> but this includes a lot of work. Is it possible to do this in the proc
> statment, maby like:
> proc phreg data=XXX;
> strats ID;
> model Y = VAR1 VAR2 VAR3 (and all significant interaction variables with
> degree of 2);
> thanks Erik, Denmark