Overlapping groups, statistical dilemma

[Jesper Eriksson]

Hi,

Currently there is a debate in my medical field, recently new definitions of a feared syndrome, sepsis, came out. The debate is partly about which definition has the best specificity for mortality and other adverse outcomes.

I am currently analysing data from 700 patients. I want to compare the two definitions of sepsis, let's say definition A & B. When I am applying definition A, 286 patients are diagnosed with sepsis. When I am applying definition B, 149 patients are diagnosed with sepsis.
Now I would like to compare sepsis-A (n=286) with Sepsis-B (n=149) against different outcomes, for example 30-day mortality. Which definition has the strongest association with 30-day mortality, and is there a significant difference between the two definitions in terms of 30-day mortality.
The problem is that the two groups are not independent, patients can belong to A only (n=139), Sepsis-B only (n=2) or both Sepsis-A and Sepsis-B (n=147). My data are not assumed to be normally distributed.

Using standard tests such as chi2 or Mann-whitney does not make sense. Any suggestions how to approach this problem?

All the best,

Jesper Eriksson

[Carlo Lazzaro]

Jesper:
you can create a three-level factor variable (A; B; A_B) and use it as a predictor (I suppose you have other independent variables) in, say a Cox regression (-help stcox-) against mortality or a composed outcome (it mainly depends on what makes sense/is the standard in your research field).
Obviously, is expected that Group B will be the weakest part of the model; but it is also expected that new classifications do not turn the older ones completely upside down (that's why the A_B stratum collects a pretty relevant share of your sample)..

[Jesper Eriksson]

Jesper:
you can create a three-level factor variable (A; B; A_B) and use it as a predictor (I suppose you have other independent variables) in, say a Cox regression (-help stcox-) against mortality or a composed outcome (it mainly depends on what makes sense/is the standard in your research field)

Thanks! Yes, that is a solution. The drawback with that approach is that the group "B" would only contain 2 patients. Which I guess would be too few too draw any conclusions on. Basically i would compare predictors A vs A_B. Or am I missing your point?

Kind regards,
Jesper Eriksson

Edit: Sorry I just now saw your edit.

[Carlo Lazzaro]

Jesper:
I have gone too straight to Cox regression and lost, at a first sight, that the B stratum would be composed of 2 patients only!