0
$\begingroup$

While trying to interpret some of the results I got from the Cox model, I've read that the centered survival curves where you calculate the mean value of each variable is pretty useless but wouldn't it be a somewhat decent indicator of how well the model explains the gap between 2 groups seen in Kaplan Meier estimator?

In my data when I calculate survival of males and females using Kaplan-Meier and then Cox with mean values for all other variables and 0-1 for sex, both graphs are pretty similar and with a similar gap (women have better survival than men).

Then if I do the same with race (white, non-white) the gap using centered Cox survival is a lot smaller than with Kaplan-Meier.

Maybe it's an obvious question but does that mean that other variables in the model explain the race gap seen on Kaplan-Meier estimator but not the sex gap?

$\endgroup$

1 Answer 1

0
$\begingroup$

Your plot for race differences from the Cox model with the "centered" predictors might not be useless, but it doesn't represent any realistic situation. In your case, it's using some average between male and female predictions, for example.

It can be better to specify a particular realistic set of covariate values to illustrate differences in plots like that. That's the approach used for example by the R rms package, which first evaluates the distributions of predictor values with a datadist() function and then (with a corresponding option setting) uses as the default for predictions a set of representative rather than "average" covariate values. You can do the same by specifying representative newdata to the R survfit() function before plotting the predicted curves. Then anyone looking at your results will be reassured that what you're showing has some connection to a potentially observable physical reality.

That said, unless race was involved in an interaction with some of the other predictors in the model, the "gap" between curves distinguished by race will be the same regardless of whether you use representative, centered, or other covariate values as the background condition. Once you specify a proportional hazards model, the race-associated hazard (absent interactions) is necessarily the same at any single set of other covariate values used for the predicted Cox survival curves.

does that mean that other variables in the model explain the race gap seen on Kaplan-Meier estimator but not the sex gap?

Probably. You can address that better by looking into your data in more detail. For a start, are there race-associated differences in other outcome-associated covariates in the model? This type of analysis can be a bit trickier in Cox modeling than in some other types of regression models, as hazard ratios are "non-collapsible." An outcome-associated covariate therefore doesn't even need to be correlated with race to affect the estimate for the race coefficient. My guess, however, is that race-associated covariate differences will play a major role.

$\endgroup$

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.