stratified estimator

[benkeser]

Adding in a way to compute stratified Kaplan-Meier-style estimator would also be a nice addition. The code currently can be hacked to do this as follows:

set.seed(1234)
n <- 100
t_0 <- 6
trt <- rbinom(n, 1, 0.5)
# e.g., study site
adjustVars <- data.frame(site = (rbinom(n,1,0.5) + 1))
ftime <- round(1 + runif(n, 1, 350) - trt + adjustVars$site)
ftype <- round(runif(n, 0, 1))

#' make a formula that will implement an empirical hazard
#' estimator when called in estimateHazard
get.ftimeForm <- function(trt, site){
	form <- "-1"
	for(i in unique(trt)){
		for(s in unique(site)){
			form <- c(form, 
			  paste0("I(trt==",i,"& site == ",s," & t==",
			         unique(ftime[ftype>0 & trt==i & site == s]),")",
			         collapse="+"))
		}
	}
	return(paste(form,collapse="+"))
}


#' make a formula that will implement an empirical hazard
#' estimator when called in estimateCensoring
get.ctimeForm <- function(trt, site){
	form <- "-1"
	for(i in unique(trt)){
		for(s in unique(site)){
			form <- c(form, 
			  paste0("I(trt==",i,"& site == ",s," & t==",
			         unique(ftime[ftype==0 & trt==i & site == s]),")",
			         collapse="+"))
		}
	}
	return(paste(form,collapse="+"))
}

form.ftime <- get.ftimeForm(trt = trt, site = adjustVars$site)
form.ctime <- get.ctimeForm(trt = trt, site = adjustVars$site)

fit <- survtmle(ftime = ftime, 
                ftype = ftype,
                trt = trt,
                adjustVars = adjustVars,
                glm.ftime = form.ftime,
                glm.ctime = form.ctime,
                glm.trt = "1", t0 = 300)

This example runs fine for n <- 100 but not for n <- 1000. Changing the call to glm within estimateCensoring and estimateHazards to family = gaussian() allows larger sample sizes to run, albeit still somewhat slowly (e.g., ~ 1 minute for n <- 1000). speedglm helps get this down to around 10 seconds.

[nhejazi]

We have the speedglm functionality working now, but did we maybe want to write a wrapper for the main survtmle function to allow computation of the stratified estimator be more straightforward?

[benkeser]

Why don't we not worry about it and include an example in the vignette of how to do it with direct calls to hazard_tmle (or mean_tmle, which I suspect may be faster).

[nhejazi]

Sure, sounds good to me. I'll just re-open to remind us to keep it on the radar in terms of adding this to the vignette soon enough.