In therneau/survival: Survival Analysis

The data set was sent by Mika Nurmikolu. It is fairly small, 484 observations and 69 events, with only 1 tied event time. The key issue is a covariate which is almost all zero.

ndata <- read.csv("ndata.csv")
with(ndata, table(status=status, snp=snp))
subset(ndata, snp==1)

coxph(Surv(time1, time2, status) ~ snp, ndata, iter=1)

At the first iteration the coefficient is 46.7 and the LPL is NaN. What drives this example? The first death in the data set is at time 60, so the first observation with snp=1 doesn't even count. The only window of time that matters is (254,298] where the data has variation in the covariate.

fit0 <- coxph(Surv(time1, time2, status) ~snp, ndata, iter=0)
dt0  <- coxph.detail(fit0)
index <- which(dt0$time > 254 & dt0$time <= 298)
index
dt0$score[index]
range(dt0$score[-index])
range(dt0$imat[-index])
dt0$time[index]
beta1 <- sum(dt0$score)/sum(dt0$imat)

The relevant computation only involves 4 time points. Why does the loglik blow up?

The loglik calculation is

n <- dt0$nrisk[index]
denom <- (n-1) + exp(beta1)
numer <- c(0,0,0, beta1)
numer - log(denom)

The four loglik terms from the last line above are all reasonable. The issue comes up for the next event at time 302, and exp(beta1) is subtracted from the denominator. The new value should be 187, but becomes less than zero.

therneau/survival documentation built on April 26, 2024, 8:23 a.m.