I have a stratified Cox model with covariates ("big") which violate proportional hazard assumptions:
model0 <- coxph(Surv(t_start, t_stop, status) ~ big + strata(prov), data=labor)
cox.zph output is this:
chisq df p
big 10.8 1 0.001
GLOBAL 10.8 1 0.001
I performed step function to address this:
labor_cut=survSplit(Surv(t_stop,status) ~ ., data=labor, cut = c(410, 1461, 2191), episode='t_group')
model0.strat <- coxph(Surv(t_start, t_stop, status) ~ big:strata(t_group), data=labor_cut)
I now have two questions. 1) Can I still stratify on clusters (e.g. strata(prov)) while also using step function to stratify by time cut-offs?
And 2) how is using survSplit() to cut the dataset at various timepoints different from simply splitting the dataset by hand and re-running each Cox regression individually? I am getting different coefficients from each method.
In addition, splitting the dataset at the same timepoints by hand yields results that are now proportional. However, using survSplit and interacting my non-proportional variable with strata(t_group) still yields results that are non-proportional. Why is this?
