Clustered log-rank test for equality of survivor functions

[Rob Wood]

Hi All,

I'm using the built in Stata command -sts test- to perform a log-rank test for the equality of my survivor functions in a Kaplan Meier plot. However, is there a way to adjust this test to account for clustered data?

In a Cox-proportional hazards model, for example, I can simply specify -vce(cluster clustervar)- as an option in the -stcox- command, where clustervar is my cluster indicator variable.

I have looked at the -strata()- option within the -stcox- command, but I'm not sure this is appropriate as I have a lot of clusters, approx. 100.

Thanks,

Rob.

[Brooke Jarrett]

Hi Rob (from 8 years in the future!) — I was curious whether you ever found an answer to your post here. I've looked through "An Introduction to Survival Analysis Using Stata" by Cleves and Gould, but I've come up empty handed. Thanks in advance and happy new year!

[Leonardo Guizzetti]

Assuming you are referring to cluster randomised designs (in which whole clusters are allocated to one group), an example is presented in the PDF manual for -help power logrank-, specifically the section beginning "Compare two survivor functions with clustered data".

You can fit a Cox regression model with the sole covariate being the factor variable for group, and using the cluster robust estimator. The test then is whether the HR for group (assuming 2 groups) is equal to 1. This is not a log-rank test per se, which is non-parametric, but it is semiparametric and a reasonable approach.