Interaction terms

[Laura Myles]

Hi Listers,

I am studying alcohol use (yes/no) in young people. I am interested in assessing whether there is an interaction between age (1 15-19 vs 20-24 vs 25-29) and sex (1 male vs. 2 female) .

I find the interaction term is significant

logistic alcohol_use i.sex##i.age
testparm i.sex#i.age

I would then like to estimate the ORs for females (vs. men) at each age group so I set the code as:

logistic alcohol_use i.sex#i.age i.age

The output makes sense; however, it does not produce the exact same output as when I run 3 separate logistic regressions for each age group:

logistic alcohol_use sex if age==1
logistic alcohol_use sex if age==2
logistic alcohol_use sex if age==3

I am assuming the difference is due to having more terms in: logistic alcohol_use i.sex#i.age i.age - is that correct?

I am just now unsure which one is the preferred approach.

Thanks in advance!

[Clyde Schechter]

So, I'll assume your age variable is coded 1/2/3 for the three age-ranges you mention. Then to get the female:male odds ratios in each age group you run the original logistic regression and then do:

forvalues i = 1/3 {
    lincom 2.female + 2.female#`i'.age, or
}

The model -logistic alcohol_use i.sex#i.age i.age- is a mis-specified model because it contains sex#age interaction but fails to include a "main" effect for sex. So it's just invalid and you can't use any results it gives you, even if they appear to make sense.

Using

logistic alcohol_use sex if age==1
logistic alcohol_use sex if age==2
logistic alcohol_use sex if age==3

would be valid, and, in this simple model at least, would give you the same results as the approach I show above.

[Laura Myles]

Thanks Clyde!