Analysis of propensity-matched data.

[Fraser Rubens]

I have been working on a project in which I am comparing outcomes in patients who have undergone one of two surgical approaches (treatment, "t").

I identified and matched 441 pairs of patients using 22 covariates.

I have a clinical outcome of interest "y".

I compared immediate outcomes using paired analysis techniques (mcnemar) - this makes sense and the results were fine. I also did a stratified cox proportional hazards analysis that stratified on the matched pairs. all is good.

now here's my problem

I want to test the effect of sex on the outcome "y" in each of the 2 groups. Turns out there are around 80 females in each of the 2 groups, but these females are not matched necessarily to other females.
Graphically I can see that there is a dramatic effect of the treatment by sex (males appear to benefit whereas females don't).

The data is no longer paired, so I cannot use mcnemar. How do I analyze this data?

[Clyde Schechter]

I don't understand what the problem is. Unless you have removed the variable that gives the pairing information, your data are still paired. It is true that the sex variable may differ between the members of the pair, but it's still paired data nonetheless. In fact, if the pairs were always in agreement on sex, you would, in principle, be unable to estimate any sex effect from the data!

Now, it is probably not a good idea to simply due a crude (paired) comparison of outcomes in males and females since you also have a treatment effect in your data, and you are, it appears, thinking that the treatment effect differs between the two sexes. So this is just a conditional logistic regression problem:

xtset pair_id
xtlogit y i.t##i.sex
margins sex, dydx(t)
margins sex, dydx(t) pwcompare(effects)

where pair_id is the variable that identifies which pair each observation belongs to.

[Fraser Rubens]

That makes sense. Thank you Clyde

Fraser