library(survival) aeq <- function(x, y, ...) all.equal(as.vector(x), as.vector(y), ...) # Test the survival curve for a fit with shared hazards. # Use the pbcseq data set, and turn bilirubin into a time-dependent state with # 4 levels, and a shared baseline hazard for the 4 transitions to death. # The subtlety is that coefficients for a shared (proportional) baseline hazard # are attached to a state, not to an observation. # (A bilirubin value of <1 is normal.) pbc1 <- pbcseq pbc1$bili4 <- cut(pbc1$bili, c(0,1, 2,4, 100), c("normal", "1-2", "2-4", ">4")) ptemp <- subset(pbc1, !duplicated(id)) # first row of each pbc2 <- tmerge(ptemp[, c("id", "age", "sex")], ptemp, id, death= event(futime, status==2)) pbc2 <- tmerge(pbc2, pbc1, id=id, bili = tdc(day, bili), bili4 = tdc(day, bili4), bstat = event(day, as.numeric(bili4))) btemp <- with(pbc2, ifelse(death, 5, bstat)) # a row with the same starting and ending bili4 level is not an event b2 <- ifelse(((as.numeric(pbc2$bili4)) == btemp), 0, btemp) pbc2$bstat <- factor(b2, 0:5, c("censor", "normal", "1-2", "2-4", ">4", "death")) check1 <- survcheck(Surv(tstart, tstop, bstat) ~ 1, istate= bili4, id = id, data=pbc2) check1$transitions fit2 <- coxph(list(Surv(tstart, tstop, bstat) ~ 1, c(1:4):5 ~ age / common + shared), id= id, istate=bili4, data=pbc2) # Before we tackle fit2, start small with just 9 subjects, coefs fixed to # simple values to make hand computation easier. There are no transitions # from state 3 to death in this subset, so there is one age coefficient and # 2 PH coefs. pbc3 <- subset(pbc2, id < 10) pbc3$age <- round(pbc3$age) # easier to do "by hand" sums fit3 <- coxph(list(Surv(tstart, tstop, bstat) ~ 1, c(1:4):5 ~ age / common + shared), x=TRUE, id= id, istate=bili4, data=pbc3, init= c(.05, .6, 1.1), iter=0) # a mixed p0 gives a stronger test than our usual (1, 0,0,0,0) surv3 <- survfit(fit3, newdata=list(age=50), p0=c(.4, .3, .2, .1, 0)) etime <- sort(unique(pbc3$tstop[pbc3$bstat != "censor"])) # At event time 1 (182), all 9 are at risk, (3,3,2,1) in initial states 1-4 atrisk <- pbc3$tstart < etime[1] & pbc3$tstop >= etime[1] # all 9 at risk table(pbc3$bili4[atrisk]) # One event occurs at 182, a 2:1 transition (1-2 to normal) # Risk scores for the non-death transitions are all exp(0) =1, # so the hazard matrix H will have second row of (1/3, -1/3, 0,0,0) and all # other rows are 0. with(subset(pbc3, tstop== 182), table(istate= bili4, state=bstat)) # The next four events are from 3:4, 3:2, 2:3, and 1:2, so also have # simple transtions, i.e., no covariates so all risk scores are exp(0) =1 # hmat <- array(0, dim=c(5,5,6)) # first 6 hazard matrices, start with 3,3,2,1 hmat[2,1,1] <- 1/3; hmat[2,2,1] <- -1/3 # new count= 4,2,2,1 hmat[3,4,2] <- 1/2; hmat[3,3,2] <- -1/2 # new count= 4,2,1,2 hmat[3,2,3] <- 1 ; hmat[3,3,3] <- -1 # new count= 4,3,0,2 hmat[2,3,4] <- 1/3; hmat[2,2,4] <- -1/3 # new count= 4,2,1,2 hmat[1,2,5] <- 1/4; hmat[1,1,5] <- -1/4 # new count= 3,3,1,2 # Event 6 is a transition from state 4 to death, at day 400 # For the shared hazard, the denominator is all those in states 1,2, or 4. atrisk <- with(pbc3, tstart < etime[6] & tstop >= etime[6]) table(pbc3$bili4[atrisk]) # current states just before time 6 # The subject in state 2-4 is not considered to be at risk for a death. # The coxph routine assumes that the set of transitions that CAN happen = the # set that did happen at least once. adata <- subset(pbc3, atrisk & bili4 != '2-4') eta <- with(adata, .05*(age-50) + .6*(bili4=="1-2") + 1.1*(bili4 == ">4")) cbind(adata[,c('id', 'age', 'tstop', 'bili4', 'bstat')], eta, risk=exp(eta)) basehaz <- 1/sum(exp(eta)) hmat[1,5,6] <- basehaz; hmat[1,1,6] <- -basehaz hmat[2,5,6] <- basehaz * exp(.6); hmat[2,2,6] <- -basehaz*exp(.6) hmat[4,5,6] <- basehaz * exp(1.1); hmat[4,4,6] <- -basehaz*exp(1.1) # double check: sum of per-subject hazards at this time point = number of # events at this time point sum(basehaz * exp(eta)) ==1 tmat <- array(0., dim= dim(hmat)) # transition matrices pstate <- matrix((4:0)/10, nrow=1) for (i in 1:6) { tmat[,,i] <- as.matrix(Matrix::expm(hmat[,,i])) pstate <- rbind(pstate, pstate[i,]%*% tmat[,,i]) } dtime <- c(1, which(surv3$time %in% etime)) # skip censored rows aeq(surv3$pstate[dtime[1:7],1,], pstate) # # A function to do the above "by hand" calculations, over all time points # It is verified for the particular fit we did, but written for # more generality. # fit: a multi-state fit, with shared baselines # istate: the inital state for each row of data # p0: starting dist for compuation # x0: curve for this set of covariates # mysurv <- function(fit, istate, p0, x0, debug=0) { if (!inherits(fit, 'coxphms')) stop("invalid fit") smap <- fit$smap from <- as.numeric(sub(":.*$", "", colnames(smap))) to <- as.numeric(sub("^.*:", "", colnames(smap))) shared <- duplicated(smap[1,]) nshare <- sum(shared) bcoef <- rep(1, ncol(smap)) # coefficients for shared baseline beta <- coef(fit, matrix=TRUE) if (nshare >0) { # coefficients for shared baseline will be the last nshare of them i <- seq(length=nshare, to=length(fit$coefficients)) bcoef[shared] <- exp(fit$coefficients[i]) # remove shared coef rows from beta phrow <- apply(fit$cmap, 1, function(x) any(x %in% i)) beta <- beta[!phrow,, drop=FALSE] } # Make the values for istate and state match the 1:2, etc of the fit, # i.e., the order of fit$states # istate and state are used in tables, using factors makes sure the result # is always the right size nstate <- length(fit$states) state <- factor(fit$y[,3], 1:nstate) # endpoint of a transition if (length(istate) != nrow(fit$y)) stop ("mismatched istate") istate <- factor(as.character(istate), fit$states) # set up output ntran <- ncol(smap) # number of transitions utime <- sort(unique(fit$y[!is.na(state), 2])) # unique event times ntime <- length(utime) tmat <- matrix(0, nstate, nstate) # transtion matrix at this time point pmat <- diag(nstate) # product of transitions nrisk <- matrix(0., ntime, nstate) #number at risk wtrisk<- matrix(0., ntime, ntran) # weighted number per transtion nevent <- matrix(0L, ntime, nstate) # number of events of each type pstate <- matrix(0L, ntime, nstate) # probability in state hmat <- matrix(0., nstate, nstate) # working matrix of hazards # eta is a matrix of (x for subject - x0) %*% coef, one row per subject, # one column per transition eta <- (fit$x - rep(x0, each= nrow(fit$y))) %*% beta rwt <- exp(eta) # the risk weight for each obs t1 <- fit$y[,1] t2 <- fit$y[,2] for (i in 1:ntime) { atrisk <- (t1 < utime[i] & utime[i] <= t2) # risk set at this time event <- which(utime[i] == t2) # potential events, at this time nrisk[i,] <- c(table(istate[atrisk])) # number at risk in each state nevent[i,] <- c(table(state[event])) # The linear predictor and hence the number at risk is different for # every transition. Also, some will not be at risk for the transition. # for (k in 1:ntran) { atrisk2 <- (atrisk & (as.numeric(istate) == from[k])) wtrisk[i,k] <- sum(rwt[atrisk2,k]) } dtemp <- table(istate[event], state[event]) #censors don't count # fill in hmat, one hazard at a time hmat <- 0*hmat for (j in unique(smap)) { # for each baseline hazard k <- which(smap == j) # transitons that share this hazard deaths <- sum(dtemp[cbind(from[k], to[k])]) # total events if (deaths==0) hmat[cbind(from[k], to[k])] <- 0 # avoid 0/0 else { hazard <- deaths/ sum(wtrisk[i, k] * bcoef[k]) #shared baseline hmat[cbind(from[k], to[k])] <- hazard * bcoef[k] # PH } } diag(hmat) <- diag(hmat) - rowSums(hmat) # rows sum to zero tmat <- as.matrix(Matrix::expm(hmat)) # transtion matrix # if (i >= debug) browser() pmat <- pmat %*% tmat pstate[i,] <- drop(p0 %*% pmat) } list(time=utime, nrisk=nrisk, nevent=nevent, pstate=pstate, wtrisk= wtrisk, P=pmat) } test3 <- mysurv(fit3, pbc3$bili4, p0= 4:0/10, x0 =50) aeq(test3$pstate, surv3$pstate[match(test3$time, surv3$time),1,]) # Now with the full data set fit2 <- coxph(list(Surv(tstart, tstop, bstat) ~ 1, c(1:4):5 ~ age / common + shared), id= id, istate=bili4, data=pbc2, ties='breslow', x=TRUE) surv2 <- survfit(fit2, newdata=list(age=50), p0=c(.4, .3, .2, .1, 0)) test2 <- mysurv(fit2, pbc2$bili4, p0= 4:0/10, fit2, x0 =50) aeq(test2$pstate, surv2$pstate[match(test2$time, surv2$time),1,]) if (FALSE){ # for testing, make a plot xfun <- function(i) { j <- match(test2$time[i], surv2$time) all.equal(test2$pstate[i,], surv2$pstate[j,1,]) } plot(surv2, col=1:5, lwd=2) matpoints(test2$time, test2$pstate, col=1:5, pch='o') }