Breaking Down Analysis into subgroups

[Ellen Kiley]

This is probably an Epi question, and a very basic one at that, but I am learning...and so throwing this out there. I have a pregnancy dataset "All Deliveries", and a composite outcome Severe Maternal Morbidity (SMM). I want to compare prevalence of SMM between those with Pregnancy Related Hypertension (PHYP) and those without PHYP, so far no problem. I have other variables to consider as well, and again no problem. Then I am hoping to break it down and compare the outcomes for RACE between PHYP and NOPHYP with SMM. Is it just a matter of a logistic regression equation: "logistic SMM PHYP i.RACE" ? Because for the heck of it I divided the dataset into two, those with PHYP and those without PHYP, and looked at the association of race with the outcome in each dataset, but there is there a way to compare those OR's? Does that even make sense? I think probably not - I just have to do the logistic...but curious nevertheless. THANK YOU to anyone who read this far!

[Clyde Schechter]

Then I am hoping to break it down and compare the outcomes for RACE between PHYP and NOPHYP with SMM. Is it just a matter of a logistic regression equation: "logistic SMM PHYP i.RACE" ?

No, it's not quite that simple.

There are two ways you can compare the ORs.

One of them is to run -logistic SMM i.PHYP##i.RACE-. You don't say how many categories your RACE variable has. For simplicity, I'll assume here that it is dichotomous and coded 0/1. Then the coefficient you get for 1.PHYP#1.RACE is the ratio of odds ratios (a multiplicative analog of difference in differences) between the effect of PHYP on SMM in race 0 and the effect of PHYP on SMM in race 1. If RACE has multiple categories, then, for all values j of that variable, the coefficient of 1.PHYP#j.RACE is the ratio of odds ratios between the effect of PHYP on SMM in race j and the effect of PHYP on SMM in race 0.

I recommend you read the excellent Richard Williams' https://www3.nd.edu/~rwilliam/stats2/l53.pdf for a lucid explanation of interactions. The examples in that document use ordinary linear regression, not logistic, but it works exactly the same way with logistic regression except for the results being odds ratios and ratios of odds ratios.

Another way you can do this is with the -suest- command:

logistic SMM PHYP if RACE == 0
estimates store race0

logistic SMM PHYP if RACE == 1
estimates store race1

suest race0 race1
suest, coefl

The output of the -suest, coefl- command will show you the names of the coefficients for the joint estimation of these models. Then use -lincom- with the appropriate coefficient names to get the difference between the coefficient of PHYP between the two race groups. Actually, add the -or- option to -lincom- so you get the ratio of odds ratios, rather than the difference in coefficients.

[Ellen Kiley]

Thank you, Clyde. This gives me a lot to work with.