Comparing models using AIC/BIC

[Joe Tuckles]

I am comparing clogit models using AIC/BIC. I have an outcome variable (outcome), three independent variables (inde1, inde2, inde3) two variables which I may stratify to assess for effect modification (strat1, strat2). I then have two additional independent variables which I do not know whether to include or not (query1, query2).
How should I run the

 etstat ic

test?

Should model 1 include:

clogit outcome inde1 inde2 inde3 strat1 strat2 , group(set) or
estimates store m1

Do I include the strat variables here?

Model 2:

clogit outcome inde1 inde2 inde3 strat1 strat2 query1 , group(set) or
estimates store m2
estat ic

Then to run again presumably model 1 remains unchanged but this time model 2 includes query2? Or do I include query1 and query2 in model 2?
What if there is barely any difference between the AIC/BIC values?

[Richard Williams]

Without knowing anything about what the variables are or even what the general topic is it is hard to say. If they all seem theoretically sound why not include them all? If you want to test whether subsets of variables belong in the model you can either use test commands or set it up as nested models and use lrtest commands or BIC/AIC values to make your decision.

You seem to think you should use query1 or query2 but not both. Why? Are they thought to be two different ways of measuring the same thing? I suppose you could use whichever one, empirically, seems to work better, especially if that makes good substantive sense. But it may be best just to use both, or perhaps add them together to make a single measure.