Hi Does anyone know if there is a function like survdiff which can also handle left-truncated and right-censored data? When I use it on left-truncated and right-censored data I get an error message saying Right censored data only. Many thanks Rajen [[alternative HTML version deleted]]

> Does anyone know if there is a function like survdiff which can also handle > left-truncated and right-censored data? When I use it on left-truncated and > right-censored data I get an error message saying Right censored data only.coxph(Surv(time1, time2, status) ~ factor(group), data=mydata) The ''score'' test from a Cox model is identical to the logrank test. (Well, almost identical - if there are two deaths on the same day the LR calculation uses an n-1 at one point where the Cox uses an n. Neither is right/wrong, just the choice of the authors of the two different papers. The difference is never of any consequence, usually several digits out in the test statistic: just enough to force addendums like this one.) To recreate the observed -expected columns fit0 <- coxph(Surv(time1, time2, status) ~ factor(group), data=mydata, iter=0, na.action=na.exclude) o.minus.e <- tapply(resid(fit0), mydata$group, sum) obs <- tapply(mydata$status, mdata$group, sum) cbind(observed=obs, expected= obs- o.minus.e, "o-e"=o.minus.e) Terry Therneau

Thank you very much that''s perfect. Rajen 2009/8/4 Terry Therneau <therneau@mayo.edu>> > Does anyone know if there is a function like survdiff which can also > handle > > left-truncated and right-censored data? When I use it on left-truncated > and > > right-censored data I get an error message saying Right censored data > only. > > coxph(Surv(time1, time2, status) ~ factor(group), data=mydata) > > The ''score'' test from a Cox model is identical to the logrank test. > > (Well, almost identical - if there are two deaths on the same day the LR > calculation uses an n-1 at one point where the Cox uses an n. Neither is > right/wrong, just the choice of the authors of the two different papers. > The > difference is never of any consequence, usually several digits out in the > test > statistic: just enough to force addendums like this one.) > > To recreate the observed -expected columns > > fit0 <- coxph(Surv(time1, time2, status) ~ factor(group), data=mydata, > iter=0, na.action=na.exclude) > o.minus.e <- tapply(resid(fit0), mydata$group, sum) > obs <- tapply(mydata$status, mdata$group, sum) > cbind(observed=obs, expected= obs- o.minus.e, "o-e"=o.minus.e) > > > Terry Therneau > >[[alternative HTML version deleted]]