158 lines
5.7 KiB
R
Raw Normal View History

2025-01-12 00:52:51 +08:00
library(survival)
options(na.action=na.exclude) # preserve missings
options(contrasts=c('contr.treatment', 'contr.poly')) #ensure constrast type
aeq <- function(x,y,...) all.equal(as.vector(x), as.vector(y),...)
#
# Test out the results for competing risks. Survfit does this directly as just
# one case of the Aalen-Johansen, but also known as 'cumulative incidence'.
#
# First trivial test
tdata <- data.frame(time=c(1,2,2,3,3,3,5,6),
status = c(0,1,0,1,0,1,0,1),
event = c(1,1,2,2,1,2,3,2),
grp = c(1,2,1,2,1,2,1,2),
id = 1:8)
old <- survfit(Surv(time, status*event, type="mstate") ~1, tdata) #old style
fit <- survfit(Surv(time, factor(status*event)) ~1, tdata)
# test that the old (should be depricated) form gives the same answer
indx <- match("call", names(fit))
all.equal(unclass(old)[-indx], unclass(fit)[-indx])
byhand <- function() {
#everyone starts in state 0
p1 <- c(1,0,0)
p2 <- c(6/7, 1/7, 0) # 0-1 transition at time 2
u2 <- matrix(rep(c(1/49, -1/49, 0), each=8), ncol=3) #leverage matrix at time 2
u2[1,] <- 0 #subject 1 is not present
u2[2,1:2] <- u2[2, 1:2] + c(-1/7, 1/7)
p3 <- c((6/7)*(3/5), 1/7, 12/35) # 0-2 transition at time 3, 5 at risk
h3 <- matrix(c(3/5, 0, 2/5, 0,1,0, 0,1,0), byrow=T, ncol=3) #hazard mat
u3 <- u2 %*% h3
u3[4:8,1] <- u3[4:8,1] + p2[1]*2/25
u3[4:8,3] <- u3[4:8,3] -p2[1]*2/25
u3[4,] <- u3[4,] + c(-p2[1]/5, 0, p2[1]/5)
u3[6,] <- u3[4,]
p6 <- c(0, 1/7, 6/7) # 0-2 at time 6, 1 at risk
h6 <- matrix(c(-1,0,1,0,1,0,0,1,0), byrow=T, ncol=3)
u6 <- cbind(0, u3[,2], -u3[,2])
V <- rbind(0, colSums(u2^2),
colSums(u3^2),
colSums(u3^2),
colSums(u6^2))
list(P=rbind(p1, p2, p3, p3, p6), u2=u2, u3=u3, u6=u6, V=V)
}
bfit <- byhand()
aeq(fit$pstate, bfit$P)
aeq(fit$n.risk[,1], c(8,7,5,2,1))
aeq(fit$n.event[,2:3], c(0,1,0,0,0, 0,0 ,2,0,1))
aeq(fit$std.err, sqrt(bfit$V))
# Check the influence directly, per row
eps <- 1e-6
deltaU <- array(0, dim= c(nrow(tdata), dim(fit$pstate)))
deltaC <- array(0, dim= c(nrow(tdata), dim(fit$cumhaz)))
deltaA <- deltaU
auc <- function(fit) {
nr <- length(fit$time)
rbind(fit$p0*fit$time[1],
apply(diff(fit$time) * fit$pstate[-nr,], 2, cumsum))
}
for (i in 1:nrow(tdata)) {
twt <- rep(1, nrow(tdata))
twt[i] <- twt[i] + eps
tfit <- survfit(Surv(time, factor(status*event)) ~1, tdata, id=id,
weights= twt)
deltaU[i,,] <- (tfit$pstate - fit$pstate)/eps # approx derivative
deltaC[i,,] <- (tfit$cumhaz - fit$cumhaz)/eps
deltaA[i,,] <- (auc(tfit) - auc(fit))/eps
}
aeq(bfit$u2, deltaU[,2,], tol=eps)
aeq(bfit$u3, deltaU[,3,], tol=eps)
aeq(bfit$u6, deltaU[,5,], tol=eps)
sqmean <- function(x) sqrt(sum(x^2))
aeq(fit$std.chaz, apply(deltaC, 2:3, sqmean), tol=eps)
aeq(fit$std.err, apply(deltaU, 2:3, sqmean), tol=eps)
aeq(fit$std.auc, apply(deltaA, 2:3, sqmean), tol=eps)
# Times purposely has values that are before the start, exact, intermediate
# and after the end of the observed times
sfit <- summary(fit, times=c(0, 1, 3.5, 6, 7), extend=TRUE)
aeq(sfit$pstate, rbind(c(1,0,0), bfit$P[c(1,3,5,5),]))
aeq(sfit$n.risk[,1], c(8,8, 2, 1, 0))
aeq(sfit$n.event, matrix(c(0,0,0,0,0, 0,0,1,0,0, 0,0,2,1,0), ncol=3))
#
# For this we need the competing risks MGUS data set, first
# event
#
tdata <- mgus1[mgus1$enum==1,]
# Ensure the old-style call using "etype" works (backwards compatability)
fit1 <- survfit(Surv(stop, status) ~ 1, etype=event, tdata)
fit1b <-survfit(Surv(stop, event) ~1, tdata)
indx <- match("call", names(fit1))
all.equal(unclass(fit1)[-indx], unclass(fit1b)[-indx])
# Now get the overall survival, and the hazard for progression
fit2 <- survfit(Surv(stop, status) ~1, tdata) #overall to "first bad thing"
fit3 <- survfit(Surv(stop, status*(event=='pcm')) ~1, tdata,
type='fleming')
fit4 <- survfit(Surv(stop, status*(event=='death')) ~1, tdata,
type='fleming')
aeq(fit1$n.risk[,1], fit2$n.risk)
aeq(rowSums(fit1$n.event), fit2$n.event)
# Classic CI formula
# integral [hazard(t) S(t-0) dt], where S= "survival to first event"
haz1 <- diff(c(0, -log(fit3$surv))) #Aalen hazard estimate for progression
haz2 <- diff(c(0, -log(fit4$surv))) #Aalen estimate for death
tsurv <- c(1, fit2$surv[-length(fit2$surv)]) #lagged survival
ci1 <- cumsum(haz1 *tsurv)
ci2 <- cumsum(haz2 *tsurv)
aeq(cbind(ci1, ci2), fit1$pstate[,2:3])
#
# Now, make sure that it works for subgroups
#
fit1 <- survfit(Surv(stop, event) ~ sex, tdata)
fit2 <- survfit(Surv(stop, event) ~ 1, tdata,
subset=(sex=='female'))
fit3 <- survfit(Surv(stop, event) ~ 1, tdata,
subset=(sex=='male'))
aeq(fit2$pstate, fit1$pstate[1:fit1$strata[1],])
aeq(fit2$std, fit1$std[1:fit1$strata[1],])
aeq(fit3$pstate, fit1$pstate[-(1:fit1$strata[1]),])
# A second test of cumulative incidence
# compare results to Bob Gray's functions
# The file gray1 is the result of
# library(cmprsk)
# tstat <- ifelse(tdata$status==0, 0, 1+ (tdata$event=='death'))
# gray1 <- cuminc(tdata$stop, tstat)
load("gray1.rda")
fit2 <- survfit(Surv(stop, event) ~ 1, tdata)
if (FALSE) {
# lines of the two graphs should overlay
plot(gray1[[1]]$time, gray1[[1]]$est, type='l',
ylim=range(c(gray1[[1]]$est, gray1[[2]]$est)),
xlab="Time")
lines(gray1[[2]]$time, gray1[[2]]$est, lty=2)
matlines(fit2$time, fit2$pstate, col=2, lty=1:2, type='s')
}
# To formally match these is a bit of a nuisance.
# The cuminc function returns a full step function, and survfit only
# the bottoms of the steps.
temp1 <- tapply(gray1[[1]]$est, gray1[[1]]$time, max)[-1] #toss time 0
indx1 <- match(names(temp1), fit2$time)
aeq(temp1, fit2$pstate[indx1,2])