Search code examples
rsurvival-analysis

Concordance index (c-index) for censored survival data returned NA

I am using R package 'survcomp' to calculate the c-index of predicted survival vs. known test set survival. But it returned NA as result. Data are as following:

Test set known survival:

     OS.Time    OS.event
1   1686    0
2    517    0
3    7    0
4    1965    0
5    5748    0
6    495    0
7    10    0
8    365    1
9    1733    0

Predicted survival (may not be a good estimate, but should not affect c-index calculation anyhow):

0.11946905 0.27748710 0.88212024 0.19557143 0.09421667 0.27926806 1.53901710 1.03344646 0.73955159

The R code used to calculate c-index is: concordance.index(pred, test_surv$OS.Time, test_surv$OS.event, method="noether")$c.index. As said, NA returned.

Is this because there is only one event in the test set? But I checked the c-index calculation formular. It seems only when the denominator is 0, the c-index can be NA, but it is not the case above.

Solution

  • If you look at the code you wil see this was intended behavior:

    if (sum(cc.ix) < 3) {
    if (msurv) {
        data <- list(x = x, surv.time = surv.time, surv.event = surv.event)
    }
    else {
        data <- list(x = x, cl = cl)
    }
    return(list(c.index = NA, se = NA, lower = NA, upper = NA, 
        p.value = NA, n = sum(cc.ix), data = data, comppairs = NA))

    When the number of events were 2 or less the authors clearly intend to return NA for pretty much everything. (It's not that difficult to construct an experiment that replicates this conclusion.)