log-rank test for paired samples?

#1
Is there a way to compare survival of paired samples (paired samples variant of the logrank test/Cox regression)?

I would like to assess difference in survival of the "same" cohort classified by two different classifications of disease. Most of patients with the disease would be the same in compared groups, but some will be included or excluded due to the difference in classification criteria. Would standard log-rank test be appropriate? Is there a method that would account for shared patients?
 
Last edited:

hlsmith

Omega Contributor
#2
I did not quite following what you are trying to do. Is it comparing survival curves of the exact same model on two slightly dependent datasets (partial overlap in subjects)?

What is the purpose of trying to do what you asking about?
 
#3
i would like to assess whether there is "spurious" improvement in survival due to different inclusion/exclusion of alternative-diagnoses-with-adverse-outcomes when one or other classification of disease is used. Most patients would be the same, but some will differ due to different criteria
 

hlsmith

Omega Contributor
#4
I am not sure off the top of my head. I will note that you will have different sample sizes, thus different precision solely based on n-values.

So the models are the exact same but the inclusion/exclusion criteria are different, with a majority overlap in the two samples, correct? I would just run them both and call it a sensitivity analysis, but my gut feeling is that you cannot make direct comparisons between them due to incomplete dependency issues. But I am not an expert. I would imagine your question is a general question, that could be generalized to any type of regression, not just survival.

P.S., @GretaGarbo - are you familiar with a way to compare two identical models with a partial (majority) overlap in the subjects in each sample?
 
Last edited:
#5
@Markica85 - It sounds like a mixed-effects Cox model might be appropriate. The model would have a random effect for subjects and a fixed effect for diagnostic criterion. Googling "mixed effects Cox regression" should turn up useful info. There's an R package, coxme, for this, though I've never used it.