Case-case comparisons

[MarieK]

Dear all,

I have a population of 2000 patients - all of whom have had a stroke. I have calculated the ORs of the stroke subtypes (there are 7 subtypes) according the presence or absence of kidney disease. However, some of the results don't really make sense and kidney disease appears to be protective against certain stroke subtypes (which isn't true) and increases the risk of others. I think that these unexpected results are likely as a result of using case-case comparisons (i.e. stroke subtype 1 vs all other subtypes) as opposed to case-control comparisons but unfortunately we don't have a suitable control population. Would anyone have any suggestions for a way for modelling or correcting for this?

BW,
Marie

 

[hlsmith]

I would guess two things may be going on here.

1.) You don't have a complete causal model. So you are missing confounding variables.
2.) Perhaps you don't have a large enough sample.

Bonus, 3.) What model are you running (e.g., multinomial logistic, ordinal logistic, simple logistic, multiple logistic? This, as well as your selection for a reference group may impact your estimates

 

[MarieK]

I would guess two things may be going on here.

1.) You don't have a complete causal model. So you are missing confounding variables.
2.) Perhaps you don't have a large enough sample.

Bonus, 3.) What model are you running (e.g., multinomial logistic, ordinal logistic, simple logistic, multiple logistic? This, as well as your selection for a reference group may impact your estimates

Thank you for these suggestions! I don't think we are missing any confounding variables but you might be right about sample size. I have used a simple logistic model for crude ORs and multiple logistic for age-adjusted ORs.

 

[hlsmith]

Is there a proven dose response relationship with renal disease and stroke? Meaning the lowest subtype may have a 10% prevalence of renal disease and the highest group should have say 80% prevalence of renal disease? Also, is it appropriate to convert renal disease to just a binary variable? I am imagining it isn't just diseased and not diseased. I would picture a gradient in disease severity that should be quantified/parameterized, there are things like % kidney function or dialysis.