7
$\begingroup$

I'm wondering about this question:

short version: how to adequately compare the effect of a reclassification of the same subjects on survival

long version: I have one cancer cohort that was sorted into TNM classes long time ago. The definition of the classes has now been updated and I like to know if, say, class 1 of the old classification has a better/worse survival than class 1 of the updated classification. My problem is that the cohort stays the same, only the sorting into classes changes.

Mini example: patients 1,2 are in class I_old and patients 3,4 in class II_old, patient 1 stays in class I_new but patient 2 goes to class II_new, patient 3 stays in II_new, patient 4 changes to I_new. If I want to compare the effect of the reclassification in class I, I would have to compare the survival of patients 1 and 2 versus the the survival of patients 1 and 4. So patient 1 is in both of the groups.

I doubt that the standard procedures like logrank test are adequate here. But what can be done? I have a very vague idea about proportional hazards regression with a time varying covariate (namely indicator for "before" and "after" update) or a frailty model but cannot tell at all if I'm going in the right direction.

What do you think?

psj

$\endgroup$
2
  • $\begingroup$ Just as a note from what I can gather time varying in this context does not make sense, unless the criteria used to classify individuals in either before or after is time varying itself (in which case I would say you want those characteristics themselves as the covariates not the classification). $\endgroup$
    – Andy W
    Oct 15, 2010 at 12:50
  • $\begingroup$ @ Andy W: yes, you're right. Maybe this was why I was uncomfortable with the idea of time varying covariates. $\endgroup$
    – psj
    Oct 15, 2010 at 14:42

1 Answer 1

4
$\begingroup$

I'll concentrate on your example question: does class 1 of the old classification scheme have a better or worse survival than class 1 of the updated classification scheme?

We can form four mutually exclusive groups of patients:

(a) Patients who weren't in class 1 under either scheme. Clearly, they don't help us answer the question.

(b) Patients who were in class 1 under both schemes. Clearly, they don't help us answer the question either.

(c) Patients who were in class 1 under the old scheme, but aren't in class 1 under the new scheme.

(d) Patients who weren't in class 1 under the old scheme, but are in class 1 under the new scheme.

Compare survival in groups (c) and (d). If survival if better in (c), then class 1 of the old scheme has better survival than class 1 of the new scheme. If survival if better in (d), then class 1 of the new scheme has better survival than class 1 of the old scheme.

$\endgroup$
1
  • $\begingroup$ onestop, that's a good idea! Definitively valid with a clear conclusion to be drawn from the results. I don't have the final data until now, but hopefully I will have enough fluctuation between classes. If not, I will have a very small sample size in groups (c) and (d) and inference will be difficult. But that's the best strategy so far... :) $\endgroup$
    – psj
    Oct 15, 2010 at 14:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.