odds ratios with interactions

[mcap]

Using unconditional logistic regression - I have a predictor - job
stress (yes/no). I am adjusting for age, gender and work hours (all
dichotomous). The outcome is work injury.

There is an interaction between job stress and gender that is both
significant and theoretically plausible.

The question is.....is there a relatively easy way in SPSS for me to
get the ORs for job stress for men and women separately? In Stata
this is fairly straight forward (lincom or adjust commands).

Thanks!
Marc

[Bruce Weaver]

Hi Marc. In your model, the odds ratio for job stress is for the
reference category of your sex variable. So run the model twice,
once with variable MALE (i.e., 1=M, 0=F), and again with variable
FEMALE (1=F, 0=M). With MALE in the model, you'll get the job
stress odds ratio for females (the reference category for MALE),
and with FEMALE, you'll get the job stress odds ratio for males
(the reference category for FEMALE). You'll have to use different
product terms in the two models, obviously.

HTH.

--
Bruce Weaver
bwe...@lakeheadu.ca
www.angelfire.com/wv/bwhomedir
"When all else fails, RTFM."

[mcap]

> bwea...@lakeheadu.cawww.angelfire.com/wv/bwhomedir

> "When all else fails, RTFM."

Thanks!!!!!!

[mcap]

Hi Bruce:

One last thing. I typically leave my dichotmous predictors as 0,1
for this kind of thing. I think we have discussed this previously but
why does that work? You end up with an interaction term with levels
0, 0, 0 and 1. Instead of having both predictors 1 and 2 where your
interaction is 1, 2, and 4.

All the texts I have seen leave it as 0,1. Just curious.

Marc

[Bruce Weaver]

I don't have a good answer to that one, Marc. I have a feeling
someone (Ray Koopman, maybe?) has addressed it in the past.

In the case of SPSS, I think that even if you code the variables as
1-2, the logistic regression procedure recodes them to 0-1. IIRC,
there's a table near the beginning of the output that shows the coding
scheme used.

Cheers,
Bruce

--
Bruce Weaver
bwe...@lakeheadu.ca
www.angelfire.com/wv/bwhomedir

[mcap]

Thanks Bruce. I think he did explain it to me once (or someone
else). The amount of content that I am capable of forgetting is
simply amazing. Thanks for getting back to me. Reversing the coding
did the trick perfectly. I finally got my new Stata installed also.
Lincom is helpful for these things.

Perhaps with interaction terms that run from 0-4 instead of 0 to 1 the
results are similar because the OR of the larger spread (0-4)
represents the change in the OR (a ratio of the odds ratios in this
case) associated with a one level change in the interaction term -
which would be the same as a jump from 0 to 1???? Just a guess.

Marc