Similar to this question, I would like a good metric for the distance between multiple ecdfs. However, in this case I have survival functions that are right-censored.


Is there a metric that is commonly used with survival functions?

  • If not, is there any metric that weights divergence at earlier values of $t$ more than later values? Early failures are much less desirable than later ones, so it seems natural to devise a metric with such a property.

    • If not, does it make sense to devise one, and does anyone have any pointers for doing this? I have some intuition that cdfs might naturally penalize early failures, but I don't know enough about the subject to get from there to a metric.
  • 1
    $\begingroup$ Have you considered having a look at the log-rank test that compares two Kaplan-Meier curves? $\endgroup$ – ocram Mar 24 '13 at 15:11

I've discovered that the Peto and Peto generalization of the Wilcoxon test fits this purpose. It weights divergence at earlier values of t more than later values. You can use the test statistic as the metric, or the p-value if you need to compare pairs of curves with different sample sizes.

This metric has the added bonus of being readily available in R as the survdiff function in the survival package. Some citations (and the R implemenation) point at the Harrington and Fleming $G^\rho$ family of tests, where $\rho = 1$ gives the Peto-Peto test described above.

[1] Peto, R., & Peto, J. (1972). Asymptotically efficient rank invariant test procedures. Journal of the Royal Statistical Society. Series A (General), 185-207.

[2] Harrington, D. P., & Fleming, T. R. (1982). A class of rank test procedures for censored survival data. Biometrika, 69(3), 553-566.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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