You're right; fairly easy to test. Here are the results:
I tried a smaller model with three dependent variables, call them A B &
C. fitint put in all 3 interaction terms (i.A*i.B, i.A*i.C, i.B*i.C)
along with main effects. It then removed each interaction term one at a
time and calculated the lrtest comparing the model with 3 interaction
terms and the nested model with 2.
My question is now whether or not this is a reasonable way to look for
potential two-way interactions, ie., testing each in the face of all
other possibilities? Using a dataset of 1495 obs, I created a model with
8 dichotomous predictors. Therefore main effects plus two-way
interactions creates a model with 8 main effects & 28 interactions. Is
it reasonable to test each interaction in the presence of all possible
other interactions or should each interaction be tested only with the
main effects?
If it is, here is my proposed approach:
1) use fitint to look for potential candidates for a model, e.g., select
interaction terms whose p-values are lower than some probability, say
<0.1 or 0.2.
2) Create new model with main effects plus the interactions from step 1)
3) Test each interaction term's contribution to the model using lrtest
on the model with and without that term.
4) Keep those interaction terms where lrtest p-value < 0.05.
Does that seem like a reasonable way to look for interaction terms? I've
also thought about a purely hypothesis-driven approach to creating and
testing interaction terms but I must say, this 'fishing expedition'
approach did uncover some sensible interactions that I would not have
thought of looking for.
John
[email protected] wrote:
John, this seems to be a question that you can answer by test which
specifies a few interaction terms. Be sure to report back your
findings.
-Steve
On Tue, May 19, 2009 at 12:20 PM, Nick Cox <[email protected]> wrote:
-fitint- is a user-written program from SSC. Please remember to specify
where user-written stuff you refer to comes from.
In this case, I don't recognise either of the authors as contributors to
the list. Either way, if you get no good answers to this question, you
may need to look at the code -- which appears to be well structured and
commented -- and/or to contact the program authors directly.
Nick
[email protected]
John LeBlanc
I have an -ologit- model with 8 predictor variables and fortunately a
large dataset. I'm using -fitint- to see if there are any interesting
two-way interactions among the 28 possibilities. Does -fitint- put all
28
in simultaneously or does it (I hope) test them one by one?
by the way, what is the purpose of the -factors- option? At this point,
I'm simply copying all dependent variables into both -factors()- &
-twoway()-. I can't figure this out from the .hlp file.
*
* 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/
--
Regards, John LeBlanc
_________________________________________________________
John C. LeBlanc, MD, MSc, FRCPC
Associate Professor
Pediatrics, Psychiatry, Community Health and Epidemiology
Dalhousie University
IWK Health Centre Work phone: (902) 470-8930
5850 University Avenue Work fax: (902) 470-6913
Halifax, Nova Scotia Email: [email protected]
B3K 6R8 CANADA Pager: (902) 470-8888
________________________________________________________________
*
* 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/
