Dear All, I have a question regarding the output of survfit() when I supply a Cox model. Lets say for example: library(survival) fit <- coxph(Surv(time, status == 2) ~ factor(spiders), data = pbc) fit # HR for spiders is significant newdata <- data.frame(spiders = factor(0:1)) sf <- survfit(fit, newdata = newdata) sum.sf <- summary(sfit, times = c(2000, 2500, 3000)) # survival estimates for the yes/no spiders # and the 3 follow up times sum.sf$surv # corresponding lower limits of the # 95% CI sum.sf$low # corresponding upper limits of the # 95% CI sum.sf$up we observe that the 95% CIs overlap!! How is this possible since the HR for spiders is significant. Regards, Mura
On Thu, 22 Oct 2009, Mura Tamakou wrote:> Dear All, > > I have a question regarding the output of survfit() when I supply a Cox model. Lets say for example: ><snipped: code that doesn't quite run>> > we observe that the 95% CIs overlap!! How is this possible since the HR for spiders is significant. >It's perfectly natural. To a good approximation, the p-value for the comparison will be significant when the point estimate for each group is outside the confidence interval for the other group. Suppose you had two point estimates with standard error equal to 1.0. The standard error of the difference would be 1.414, so the p-value would be less than 0.05 if the two point estimates differ by more than 1.96*1.414. The confidence intervals will overlap if the point estimates differ by less than 1.96*2. -thomas Thomas Lumley Assoc. Professor, Biostatistics tlumley at u.washington.edu University of Washington, Seattle