LR test for Interaction in Fine and Gray test

[Mich Prov]

Dear statalist,

we know that it is possibile to check the effect modification between two categorical variables using lrtest; for example
stset fu_1, fail(morte=1) id(id)
stcox i.agegroup##i.sex
est store a
stcox i.agegroup i.sex
est store b
lrtest a b

and an interaction term can be used if p<0.05 (null hypothesis is no interaction between age and sex)

how ti perform the same test afferma stcrreg with the fine and gray model?

[Clyde Schechter]

Well, -stcrreg- doesn't estimate by maximum likelihood, so I don't think you can do a likelihood ratio test here. But you can do a Wald test of the interaction term using -test-.

That said, I wouldn't necessarily base my decision about including an interaction term in the model based on a statistical significance test. If the interaction term coefficient(s) are fairly large, suggesting that the effects may well differ appreciably, then I'd be inclined to retain the interaction, assuming that there was a good reason to include it in the model in the first place.

[Mich Prov]

perfect! I have computed test 2.sex#1.agegroup = 0 and the output is chi2 = 0.49 (p=0.484). Suppose that this coefficient is far from 0, so if they are all non significant i can maintain interaction ?

[Clyde Schechter]

Model selection is always difficult, and when done based on test statistics it is usually done poorly. The key considerations are model fit, and substantive considerations. If you would like to post the actual output, I will tell you what I might do in your circumstances.

[Mich Prov]

Thank you Professr Schechter, I understood the meaning of this test a computed a cumulative effect of interaction coefficients... the result is a chi2 with correct D.F. but a p value of 0.07... Like you said we can however speak of interaction effect...depending on the single coefficient and in what measure the coefficients are larger than standard errore... is it true??

Michele

[Clyde Schechter]

So, don't make a fetish out of p = 0.05. 0.07 is close, and interaction terms, by their very nature, have lower power associated with their tests, so they should not be held to the same stringency as "main" effects. In any case, as I have said before, I don't think reliance on any statistical test to decide what terms to include in a model is a bad practice.

Similarly, it's not how the interaction coefficients compare to their standard errors that matters here. It's how they affect the model predictions that really matters here. What I would do is look at the output of -margins agegroup#sex- to see if the predicted outcomes differ much by agegroup within sex, or by sex within agegroups. If they do, then the interaction terms are important and should be retained. The issue is not whether these differences are statistically significant, it is whether they are large enough to matter from a practical perspective. I have often retained non-statistically-significant interaction effects, and I think I have equally often jettisoned statistically-significant ones. The important issue is whether they affect model outcomes enough to matter for practical purposes.