library(survival) aeq <- function(x, y) all.equal(as.vector(x), as.vector(y)) # # Compute the hazard functions for a multi-state Cox model # coxhaz <- function(y, id, risk, wt, expm=TRUE) { # y should be a multi-state survival if (!inherits(y, "Surv") || attr(y, "type") != "mcounting") stop("invalid response") n <- nrow(y) if (missing(id) || length(id) !=n) stop("invalid id") if (missing(wt)) wt <- rep(1.0, n) else if (length(wt) !=n || any(wt <=0)) stop("invalid wt") # get the current state, and the list of transtions # transitions to censor don't count mcheck <- survcheck(y~1, id= id) states <- mcheck$states nstate <- length(states) istate <- mcheck$istate event <- y[,3] > 0 temp <- attr(y, 'states')[y[event,3]] tmat <- table(y[event,2], from=istate[event], to=temp) tmat2 <- tapply(wt[event], list(y[event,2], from=mcheck$istate[event], to=temp), sum) tmat2 <- ifelse(is.na(tmat2), 0, tmat2) # Hazards can be done one at a time. For each of them the risk # weight vector for the subjects can be different. # First organize the material as a 2 dim matrix temp <- apply(tmat, 2:3, sum) # count of transtions keep <- which(temp>0) from <- states[row(temp)[keep]] hlab <- outer(match(rownames(temp), states), match(colnames(temp), states), paste, sep=':')[keep] nhaz <- length(keep) nevent <- matrix(tmat2, nrow(tmat2))[,keep] dtime <- sort(unique(y[event,2])) ntime <- length(dtime) dimnames(nevent) <- list(NULL, hlab) aindex <- cbind(as.numeric(substring(hlab,1,1)), as.numeric(substring(hlab,3,3))) if (missing(risk)) risk <- matrix(1, nrow=n, ncol=nhaz) if (!is.matrix(risk) || nrow(risk) != n || ncol(risk) != nhaz) stop("invalid risk matrix") risk <- risk * wt # get the weighted at risk at each time wtrisk <- matrix(, length(dtime), nhaz) statematch <- outer(istate, from, function(x, y) x==y) risk <- ifelse(statematch, risk, 0) for (i in 1:ntime) { atrisk <- (y[,1]< dtime[i] & y[,2] >= dtime[i]) wtrisk[i,] <- colSums(risk[atrisk,, drop=FALSE]) } haz <- nevent/ifelse(wtrisk==0, 1, wtrisk) # avoid 0/0 chaz<- apply(haz, 2, cumsum) # compute the probability in state, with p(0)= 1,0, .. pstate <- matrix(0, ntime+1, nstate) pstate[1,1] <- 1 for (i in 1:ntime) { tmat <- matrix(0, nstate, nstate) tmat[aindex] <- haz[i,] if (expm) { diag(tmat) <- -rowSums(tmat) pstate[i+1,] <- pstate[i,] %*% as.matrix(Matrix::expm(tmat)) } else { diag(tmat) <- 1-rowSums(tmat) pstate[i+1,] <- pstate[i,] %*% tmat } } list(time=dtime, nrisk=wtrisk, nevent=nevent, haz=haz, cumhaz=chaz, states=states, pstate= pstate[-1,]) } mtest <- data.frame(id= c(1, 1, 1, 2, 3, 4, 4, 4, 5, 5), t1= c(0, 4, 9, 0, 2, 0, 2, 8, 1, 3), t2= c(4, 9, 10, 5, 9, 2, 8, 9, 3, 11), state= c(1, 2, 1, 2, 3, 1, 3, 0, 2, 0), x = c(0, 0, 0, 1, 1, 0, 0, 0, 2, 2)) mtest$state <- factor(mtest$state, 0:3, c("censor", "a", "b", "c")) # True results # #time at risk events # entry a b c # #2 1245 4 -> a #3 1235 4 5 -> b #4 123 4 5 1 -> a #5 23 14 5 2 -> b, exits #8 3 14 5 4 -> c #9 3 1 5 4 1->b, 3->c & exit, 4 censored #10 15 1->a, exit #11 5 censor # with all coefficients =0 check1 <- with(mtest, coxhaz(Surv(t1, t2, state), id)) fit1 <- survfit(Surv(t1, t2, state) ~1, mtest, id=id) aeq(check1$cumhaz, fit1$cumhaz[match(check1$time, fit1$time),]) dummy <- data.frame(x=1:2) cox0 <- coxph(Surv(t1, t2, state) ~x, iter=0, mtest, id=id) cfit0 <- survfit(cox0, newdata=dummy) indx <- match(check1$time, cfit0$time) aeq(check1$cumhaz, cfit0$cumhaz[indx,1,]) aeq(check1$cumhaz, cfit0$cumhaz[indx,2,]) aeq(check1$pstate, cfit0$pstate[indx,1,]) # a fixed coefficient mfit <- coxph(Surv(t1, t2, state) ~x, iter=0, mtest, id=id, init= log(1:6)) msurv <- survfit(mfit, newdata=list(x=0:1)) mrisk <- exp(outer(mtest$x, log(1:6), '*')) # hazards for each transition check2 <- with(mtest, coxhaz(Surv(t1, t2, state), id=id, risk=mrisk)) aeq(check2$cumhaz, msurv$cumhaz[indx,1,]) aeq(check2$pstate, msurv$pstate[indx,1,]) # a different predicted x multiplies the risk weights # now use exp(x - target) as the risk score mrisk2 <- mrisk %*% diag(1/(1:6)) check2b <- with(mtest, coxhaz(Surv(t1, t2, state), id=id, risk=mrisk2)) aeq(check2b$cumhaz, msurv$cumhaz[indx,2,]) aeq(check2b$pstate, msurv$pstate[indx,2,]) # since pstate depends only on the hazards and p(0), if the hazards are # right I don't have to check pstate for every subcase if (FALSE) { # this graph is very useful temp <- survcheck(Surv(t1, t2, state) ~1, mtest, id=id) plot(c(0,11), c(1,5.1), type='n', xlab="Time", ylab= "Subject") with(mtest, segments(t1+.1, id, t2, id, col=as.numeric(temp$istate))) event <- subset(mtest, state!='censor') text(event$t2, event$id+.2, as.character(event$state)) } # slight change, add a few censored subjects # all the events happen on even numbered days test2 <- data.frame(id= c(1, 1, 1, 2, 3, 4, 4, 4, 5, 5, 6, 7, 8, 9), t1= c(0, 8, 18, 0, 4, 0, 4, 16, 2, 6, 0, 0, 7, 8), t2= c(8, 18, 20, 10, 18, 4, 16, 18, 6, 22, 5, 10, 10, 15), state= c(1, 2, 1, 2, 3, 1, 3, 0, 2, 0,0,0,0,0), x = c(0, 0, 0, 1, 1, 0, 0, 0, 2, 2, 1, 1, 2, 0)) test2$state <- factor(test2$state, 0:3, c("censor", "a", "b", "c")) if (FALSE) { # this graph is very useful when debugging temp <- survcheck(Surv(t1, t2, state) ~1, test2, id=id) plot(c(0,22), c(1,9.1), type='n', xlab="Time", ylab= "Subject") with(test2, segments(t1+.1, id, t2, id, col=as.numeric(temp$istate))) event <- subset(test2, state!='censor') text(event$t2, event$id+.2, as.character(event$state)) } # s0 to a, cumhaz of 1/6 (t=4) + 1/5 (t=8) # b to a, cumhaz of 1/2 at 20 # s0 to b, cumhaz of 1/5 at 6, +1/5 at 10 # a to b, cumhaz of 1/1 at 18 # s0 to c, cumhaz of 1/1 at 18 # a to c, cumhaz of 1/2 at 16 time2 <- c(4,5,6,8,10,15,16,18,20, 22) chaz2 <- matrix(0, nrow=10, ncol=6, dimnames=list(time2, c("1:2", "1:3", "1:3", "2:3", "1:4", "2:4"))) chaz2['4',1] <- 1/6; chaz2['8',1] <- 1/5 chaz2['20',2] <- 1/2 chaz2['6', 3] <- 1/5; chaz2['10', 3] <- 1/5 chaz2['18',4:5] <- 1 chaz2['16', 6] <- 1/2 chaz2 <- apply(chaz2, 2, cumsum) cox3 <- coxph(Surv(t1, t2, state) ~x, id=id, test2, iter=0) # no weights csurv3 <- survfit(cox3, newdata=data.frame(x=0:1)) aeq(csurv3$time, time2) aeq(csurv3$cumhaz[,1,], chaz2) aeq(csurv3$cumhaz[,2,], chaz2) check3 <- with(test2, coxhaz(Surv(t1, t2, state), id=id)) indx3 <- match(check3$time, csurv3$time) aeq(check3$cumhaz, chaz2[indx3,]) # a check on the coxhaz function above aeq(check3$pstate, csurv3$pstate[indx3,1,]) cox4 <- coxph(Surv(t1,t2, state) ~ x, id=id, test2, init=log(1:6), iter=0) csurv4 <- survfit(cox4, newdata=data.frame(x=0:1)) mrisk4 <- exp(outer(test2$x, log(1:6), '*')) # hazards for each transition check4 <- with(test2, coxhaz(Surv(t1, t2, state), id=id, risk=mrisk4)) aeq(check4$cumhaz, csurv4$cumhaz[indx3,1,]) aeq(check4$pstate, csurv4$pstate[indx3,1,]) aeq(csurv4$cumhaz[,2,], csurv4$cumhaz[,1,] %*% diag(1:6)) # Check the stype=1 option csurv4b <- survfit(cox4, newdata= data.frame(x=0:1), stype=1) check4b <- with(test2, coxhaz(Surv(t1, t2, state), id=id, risk=mrisk4, expm=FALSE)) aeq(check4b$cumhaz, csurv4b$cumhaz[indx3,1,]) aeq(check4b$pstate, csurv4b$pstate[indx3,1,])