https://github.com/cran/rstpm2
Tip revision: d415ac4eba2a0f4dd170f58ced8f09afe73047ea authored by Mark Clements on 28 February 2023, 15:02:29 UTC
version 1.6.1
version 1.6.1
Tip revision: d415ac4
working_code.R
## package.skeleton(name="rstpm2", path="c:/usr/src/R", force=TRUE, namespace=TRUE, code_files="pm2-3.R")
## Local Windows setup:
## Rtools.bat
## R CMD INSTALL --html "c:/usr/src/R/rstpm2/pkg"
## R CMD build "c:/usr/src/R/rstpm2/pkg"
## R CMD INSTALL --build "c:/usr/src/R/rstpm2/pkg"
## R CMD CHECK "c:/usr/src/R/rstpm2/pkg"
##
## Local Ubuntu setup:
## R CMD INSTALL --html ~/src/R/rstpm2/pkg --library=~/R/x86_64-pc-linux-gnu-library/2.12
## R CMD build ~/src/R/rstpm2/pkg
## R CMD build --binary ~/src/R/rstpm2/pkg
##
## testPackage <- TRUE
## if (testPackage) {
## require(splines)
## require(survival)
## require(bbmle)
## }
## Email from Paul S
library(rstpm2)
fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,df=3,tvc=list(hormon=2))
## Plot marginal HR
## Straight from plot function
plot(fit, type = "meanhr", newdata = transform(brcancer, hormon = 0), exposed = function(data) transform(data, hormon = 1))
## Using ggplot (need to manually extract predictions and plot)
library(ggplot2)
## debugonce(rstpm2:::predict.stpm2.base)
system.time(pred <- predict(fit,
newdata = transform(brcancer, hormon = 0),
grid = TRUE, full = TRUE, se.fit = TRUE, type = "meanhr", recent=TRUE,
exposed = function(data) transform(data, hormon = 1)))
system.time(pred2 <- predict(fit,
newdata = transform(brcancer, hormon = 0),
grid = TRUE, full = TRUE, se.fit = TRUE, type = "meanhr",
exposed = function(data) transform(data, hormon = 1)))
## Plot
ggplot(pred, aes(x = rectime, y = Estimate)) +
geom_line() +
geom_ribbon(aes(ymin = lower, ymax = upper), alpha = 0.3) +
ylim(0,2) +
xlab('Time (days)') + ylab('HR') +
theme_bw(base_size = 15)
## Email for episig
## install.packages("extRemes")
library(extRemes)
library(survival)
set.seed(12345)
n<-500000
x1<-rnorm(n)
x2<-rnorm(n)
error<-revd(n, loc = 0, scale = 1)
b0<- 1
b1<-0.5
b2<--1.2
c<-1
time<-exp(b0+b1*x1+b2*x2-c*error)
status<-rep(1,n)
survreg(Surv(time, status==1) ~ x1+x2,dist="exponential")
library(rstpm2)
object <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,df=3,tvc=list(hormon=2))
newdata = data.frame(hormon=0)
simulate(object, newdata=newdata) # currently returns just the times -- this will later return a data-frame
simulate(object, newdata=newdata, start=1500) # left-truncation
plot(object, newdata=newdata)
grep_call = function(name,x) {
local_function = function(x)
if(length(x)==1) x==name else any(sapply(x, local_function))
which(sapply(x, local_function))
}
gsm_design = function(object, newdata, inflate=100) {
stopifnot(inherits(object, "stpm2"),
is.list(newdata),
is.numeric(inflate),
length(inflate) == 1)
## Assumed patterns:
## timeEffect := (ns|nsx)(log(timeVar),knots,Boundary.knots,centre=FALSE,derivs=(c(2,2)|c(2,1)))
## effect := timeEffect | otherEffect:timeEffect | timeEffect:otherEffect
terms = attr(object@model.frame, "terms")
factors = attr(terms, "factors")[-1,,drop=FALSE]
variables = attr(terms, "variables")[-(1:2)]
predvars = attr(terms, "predvars")[-(1:2)]
indices = grep_call(object@timeVar, variables)
if(length(indices)==0) stop("No timeVar in the formula -- unexpected error")
index_time_variables = grep_call(object@timeVar, variables) # time variables
index_time_effects = grep(object@timeVar,colnames(factors)) # components in the rhs with time variables
## I want to know how wide is each term
nms = names(coef(object))
term.labels = attr(terms, "term.labels")
coef_index <-
sapply(strsplit(nms, ":"), function(c) {
if (length(c)>2)
stop("current implementation only allows for main effects and two-way interactions")
pmatchp = function(x, table)
!is.na(pmatch(x, table))
if(length(c)==1) {
for (i in seq_along(term.labels)) {
t = term.labels[i]
## browser()
if (pmatchp(t,c))
return(i)
}
if (c == "(Intercept)")
return(0)
return(-1)
}
## => length(c) == 2
term.labels.split = strsplit(term.labels, ":")
for (i in seq_along(term.labels)) {
t = term.labels.split[[i]]
if (all(pmatchp(t,c)))
return(i)
}
return(-1)
})
## how to map from term.labels to factors to predvars?
## Can I replace each term with its predvars?
parse_ns = function(mycall,x,index_time_effect) {
df = length(c(mycall$knots, mycall$Boundary.knots)) - 1
stopifnot(mycall[[1]] == quote(nsx) || mycall[[1]] == quote(ns),
length(mycall[[2]])>1,
mycall[[2]][[1]] == quote(log), # assumes log
mycall[[2]][[2]] == object@timeVar, # what about a scalar product or divisor?
is.null(mycall$deriv) || (mycall$derivs[1] == 2 && mycall$derivs[2] %in% 1:2),
mycall$centre == FALSE) # doesn't allow for centering
cure = !is.null(mycall$derives) && all(mycall$derivs == c(2,1))
time = object@args$time
q_const = attr(nsx(log(mean(time)), knots=mycall$knots,
Boundary.knots=mycall$Boundary.knots,
intercept=mycall$intercept),
"q.const")
## browser()
list(call = mycall,
knots=mycall$knots,
Boundary_knots=mycall$Boundary.knots,
intercept=as.integer(mycall$intercept),
gamma=coef(object)[which(coef_index %in% index_time_effect)],
q_const = q_const,
cure = as.integer(cure),
x=x)
}
time = object@args$time
newdata[[object@timeVar]] = mean(time) # NB: time not used
Xp = predict(object, newdata=newdata, type="lpmatrix")
index2 = which(coef_index %in% index_time_effects)
etap = drop(Xp[, index2, drop=FALSE] %*% coef(object)[index2])
list(type="gsm",
link_name=object@args$link,
tmin = min(time), # not currently used?
tmax = max(time),
inflate=as.double(inflate),
etap=etap,
log_time=TRUE,
terms =
lapply(index_time_effects,
function(i) {
j = which(factors[,i] != 0)
if (length(j)==1)
return(parse_ns(predvars[[j]], rep(1, nrow(newdata)), i))
else {
if(length(j)>3)
stop("Current implementation only allows for two-way interaction terms")
if (j[1] %in% index_time_variables)
return(parse_ns(predvars[[j[1]]], eval(predvars[[j[2]]], newdata), i))
else return(parse_ns(predvars[[j[2]]], eval(predvars[[j[1]]], newdata), i))
}
})
)
}
library(rstpm2)
library(microsimulation)
object <- gsm(Surv(rectime,censrec==1)~hormon,data=brcancer,df=3,tvc=list(hormon=2))
newdata = data.frame(hormon=0:1)
inflate = 100
(design <- gsm_design(object,newdata))
replicate(10,.Call("test_read_gsm", design, PACKAGE="microsimulation"))
set.seed(1002)
simulate(fit1, nsim=10, newdata=nd)
simulate(fit1, newdata=data.frame(hormon=0:1))
set.seed(1002)
replicate(10,.Call("test_read_gsm", design, PACKAGE="rstpm2"))
set.seed(1002)
range(replicate(10,.Call("test_read_gsm", gsm_design(fit1,nd), PACKAGE="rstpm2")) -
simulate(fit1, nsim=10, seed=1002, newdata=nd))
library(microsimulation)
set.seed(1002)
replicate(10,.Call("test_read_gsm", gsm_design(fit1,nd), PACKAGE="microsimulation"))
simulate.stpm2 <- function(object, nsim=1, seed=NULL,
newdata=NULL, lower=1e-6, upper=1e5, ...) {
stopifnot(!is.null(newdata))
if (!is.null(seed)) set.seed(seed)
## assumes nsim replicates per row in newdata
n = nsim * nrow(newdata)
newdata = newdata[rep(1:nrow(newdata), each=nsim), , drop=FALSE]
e <- rexp(n)
objective <- function(time) {
newdata[[object@timeVar]] <- time
predict(object, type="cumhaz", newdata=newdata) - e
}
vuniroot(objective, lower=rep(lower,length=n), upper=rep(upper,length=n))$root
}
setMethod("simulate", "stpm2", simulate.stpm2)
setMethod("simulate", "pstpm2", simulate.stpm2)
## Can we estimate the hessian externally?
library(rstpm2)
library(numDeriv)
fit = stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,df=4)
summary(fit)
fit@vcov = solve(numDeriv::hessian(fit@minuslogl, coef(fit)))
summary(fit)
## Can we also use the gradient?
(solve(numDeriv::jacobian(fit@args$gradnegll, coef(fit))) - solve(numDeriv::hessian(fit@minuslogl, coef(fit)))) |> range()
(solve(numDeriv::jacobian(fit@args$gradnegll, coef(fit))) - vcov(fit)) |> range()
## Can we improve on accuracy using nsxD? This would also include predictions using C++ code.
##
library(rstpm2)
m <- rstpm2::stpm2(Surv(time, status) ~ sex, data = lung)
predict(m, newdata = data.frame(sex = c(1,1), time = c(100,200)), type = "rmst", se.fit=TRUE) # ok
plot(m, newdata = data.frame(sex = 1), type="rmst")
plot(m, newdata = data.frame(sex = 1), var="sex", type="rmstdiff")
## Predictions for differences
## Can we return the predictions, gradients and covariance matrix and use those for differences?
## preddiff = pred1 - pred0
## graddiff = grad1 - grad0
## vardiff = graddiff^T Sigma graddiff
## Note that we can silently return attributes
d = structure(data.frame(a=1), attr1=1:100, class=c("Test","data.frame"))
print(d) # print.Test() does not exist, so this uses print.data.frame()
## What about gradients on transformed scales,
## with a transformation g and an inverse transformation G?
## var(eta1) = grad_eta1^T Sigma grad_eta1
## and pred1 = G(eta1)
## => var(pred1) = var(G(eta1)) = grad(G(eta1))^T Sigma grad(G(eta1))
## where grad1 = grad(G(eta1)) = G'(eta1) grad_eta1?? (chain rule)
## preddiff = pred1 - pred0
## graddiff = grad1 - grad0
## vardiff = graddiff^T Sigma graddiff
## add cure for the AFT models
library(rstpm2)
fit0 = aft(Surv(rectime,censrec==1)~hormon,data=brcancer,df=4)
fit = aft(Surv(rectime,censrec==1)~hormon,data=brcancer,df=4,cure=TRUE)
par(mfrow=1:2)
plot(fit0,newdata=data.frame(hormon=0),main="Without cure")
plot(fit,newdata=data.frame(hormon=0),main="With cure")
cov2cor(vcov(fit0))
## cure models paper
library(cuRe)
colonDC <- subset(cuRe::colonDC, stage %in% c("Regional","Distant"))
colonDC <- transform(colonDC, stage=factor(stage),
stageDistant=as.numeric(stage == "Distant"),
bhaz = general.haz(time = "FU",
rmap = list(age="agedays", sex = "sex", year = "dx"),
data = colonDC, ratetable = survexp.dk))
fit.lat <- stpm2(Surv(FUyear, status) ~ stageDistant + bhazard(bhaz),
data = colonDC, df = 4, cure = TRUE)
fit.lat.timevar <- stpm2(Surv(FUyear, status) ~ stageDistant + bhazard(bhaz),
data = colonDC, df = 4, cure = TRUE,
tvc.formula = ~nsx(log(FUyear), df=4, cure=TRUE):stageDistant)
fit.lat.timevar2 <- stpm2(Surv(FUyear, status) ~ stageDistant + bhazard(bhaz),
data = colonDC, df = 4, cure = TRUE, tvc = list(stageDistant = 4))
predict(fit.lat, newdata = data.frame(FUyear = 0, stageDistant = 0), type = "probcure", se.fit = TRUE)
predict(fit.lat, newdata = data.frame(FUyear = 2, stageDistant = 0), type = "uncured", se.fit = TRUE)
predict(fit.lat.timevar, newdata = data.frame(FUyear = 2, stageDistant = 0), type = "probcure", se.fit = TRUE)
plot(fit.lat, newdata = data.frame(stageDistant = 0), type = "probcure")
plot(fit.lat, newdata = data.frame(stageDistant = 0), type = "surv")
plot(fit.lat, newdata = data.frame(stageDistant = 0), type = "uncured")
## solve A x = b
set.seed(12345)
A = matrix(rnorm(18),3,byrow=TRUE)
b = c(0,0,1)
x = qr.solve(A,b) # under-determined system
drop(A %*% x - b) # this is a solution -- noice!
qmat = qr.Q(qr(t(A)), complete=TRUE)[, -(1:3), drop=FALSE]
A %*% qmat %*% rnorm(3) # should be close to zero...
A %*% qmat
library(rstpm2)
library(splines)
print(ns1 <- nsx(1:10,df=4,intercept=TRUE))
## check how to calculate q.const
Bstar <- splineDesign(sort(c(rep(1,4),attr(ns1,"knots"),rep(10,4))),
x=c(1,10),derivs=c(2,2))
qr.Q(qr(t(Bstar)), complete=TRUE)[,-(1:2), drop=FALSE] - attr(ns1, "q.const") # OK
## Now, solve: Bstar x = c(0,0,1)
Bstar <- splineDesign(sort(c(rep(1,4),attr(ns1,"knots"),rep(10,4))),
x=c(1,10,10),derivs=c(2,2,1))
b = c(0,0,1)
x = qr.solve(Bstar,b) # magic sauce
drop(Bstar %*% x) - b # OK
qmat = qr.Q(qr(t(Bstar)), complete=TRUE)[,-(1:3), drop=FALSE]
theta = rnorm(3)
drop(Bstar %*% (drop(qmat %*% theta) + x)) - b # OK
Bstar %*% qmat %*% theta + Bstar %*% x - b # OK
## transformation: theta -> B %*% (drop(qmat %*% theta) + x) = B %*% qmat %*% theta + B %*% x
## nsx(..., derivs, target=rep(0,length(derivs))) for target=b
## returns an attribute for offset=x
## For fixed design matrix, we get N = B %*% qmat and C = B %*% x, so that theta -> N %*% theta + C
## How will this affect the linear predictor?
## Rather than N(t)^T theta, we will have N(t)^T theta + C
## For the derivative, we get N' = B' %*% qmat and C' = B' %*% x,
## so that we have N'(t)^T theta + C'
xs = seq(1,10,l=301)
B <- splineDesign(sort(c(rep(1,4),attr(ns1,"knots"),rep(10,4))),
x=xs)
N = B %*% qmat
C = B %*% x
theta = rnorm(3)
plot(xs, N %*% theta + C, type="l")
## Plot of derivatives...
B <- splineDesign(sort(c(rep(1,4),attr(ns1,"knots"),rep(10,4))),
x=xs, derivs=1)
N = B %*% qmat
C = B %*% x
theta = rnorm(3)
plot(xs, N %*% theta + C, type="l")
## How to predict outside of the boundaries?
## test for smoothpwc
library(rstpm2)
library(dplyr)
relative_survival <- function(model1, smoothpwc, ...) {
transmat = matrix(c(NA,1,2,
NA,NA,NA,
NA,NA,NA),3,3,byrow=TRUE)
rownames(transmat) <- colnames(transmat) <- c("Initial","Cause-specific death","Other causes of death")
rstpm2::markov_msm(list(model1,smoothpwc), ..., trans = transmat)
}
## use popmort
colon2 = inner_join(survival::colon |> filter(etype==2),
mutate(popmort,sex=2-sex,rate) |> filter(year==2000), by=c("age","sex")) |>
mutate(t=time/365.25)
excess = gsm(Surv(t,status)~factor(rx)+bhazard(rate), data=colon2, df=3)
smoothpwc1 = with(filter(popmort,sex==1 & year==2000),smoothpwc(age+0.5-70,rate)) # example is for men aged 70 years
rs = relative_survival(excess, smoothpwc1, newdata=data.frame(rx="Obs"), t = seq(0,7, length=301))
plot(rs,ggplot=TRUE)
## Spline interpolation
library(schumaker)
d <- data.frame(x=(0:9)+0.5,y=((1:10)-5.5)^2)
xs=seq(0,20,length=301)
plot(xs,Schumaker(d$x,d$y,Extrapolation="Constant")$Spline(xs),type="l")
points(y~x,data=d)
library(splines)
plot(xs,splinefun(d$x,d$y,method="natural")(xs),type="l")
points(y~x,data=d)
d2 <- rbind(d,transform(tail(d,1)[rep(1,2),],x=c(9.5+1e-7,9.5+2e-7)))
xs=seq(9.5-1e-3,9.5+1e-3,length=301)
plot(xs,splinefun(d2$x,d2$y,method="natural")(xs),type="l")
points(y~x,data=d2)
abline(h=20.25,lty=2)
splinefun(d2$x,d2$y,method="natural")(15)-20.25
splinefunx <- function(x, y, method="natural", constant.left=FALSE, constant.right=FALSE, ...) {
xstar <- x
ystar <- y
if (constant.right) {
xstar <- c(xstar,tail(x,1)+(1:10)*1e-7)
ystar <- c(ystar,rep(tail(y,1),10))
}
if (constant.left) {
xstar <- c(xstar,x[1]-(1:10)*1e-7)
ystar <- c(ystar,rep(y[1],10))
}
splinefun(xstar, ystar, ..., method=method)
}
splinefunx(d$x,d$y,constant.right=TRUE)(15)-20.25
splinefunx(d$x,d$y,constant.right=TRUE)(15000)-20.25
splinefunx(d$x,d$y)(9.5)-20.25
##
library(numDeriv)
xs=seq(0,20,length=301)
plot(xs, numDeriv::grad(splinefunx(d$x,d$y,constant.right=TRUE), xs), type="l")
rug(d$x)
xs=seq(9.5-1e-3,9.5+1e-3,length=301)
plot(xs, numDeriv::grad(splinefunx(d$x,d$y,constant.right=TRUE), xs), type="l")
## Rather than using mid-points, we could use intervals and then assume a monotone smoother...
numdiff <- function(f,x,eps=1e-5) (f(x+eps)-f(x-eps))/2/eps
d <- transform(data.frame(x0=0:9,x1=(0:9)+1,delta=1),
xmid=(x0+x1)/2,
yint=((x1-5)^3-(x0-5)^3)/3)
set.seed(12345)
d <- transform(d,
yint_ = yint*rgamma(NROW(d),10,10))
xs=seq(0,20,length=301)
## cumd <- transform(d, H=cumsum(c(0,diff(x0)*head(y0,-1))))
cumd <- with(d, data.frame(x=c(x0,tail(x1,1)), H=cumsum(c(0,diff(c(x0,tail(x1,1)))*yint))))
sf.natural <- splinefun(cumd$x,cumd$H, method="natural")
sf.mono <- splinefun(cumd$x,cumd$H, method="hyman")
plot(xs, numdiff(sf.natural,xs), type="l", ylim=c(0,25)) # constant from rhs of last interval
lines(xs, splinefunx(d$xmid,d$yint,constant.right=TRUE)(xs), col="blue") # constant from mid-point
curve((x-5)^2, col="green", add=TRUE)
lines(xs, numdiff(sf.mono,xs), type="l", col="orange") # extrapolates nicely if quadratic
## now with variation in the outcome
cumd <- with(d, data.frame(x=c(x0,tail(x1,1)+1), H=cumsum(c(0,diff(c(x0,tail(x1,1)+1))*yint_))))
sf <- splinefun(cumd$x,cumd$H, method="hyman")
plot(xs, numdiff(sf,xs), type="l", ylim=c(0,35))
lines(xs, splinefunx(d$xmid,d$yint_,constant.right=TRUE)(xs), col="blue")
curve((x-5)^2, col="green", add=TRUE)
## gradient for penalty in the AFT model
library(Rcpp)
library(RcppArmadillo)
fdiff = \(f,x,eps=1e-5)
sapply(1:length(x),
\(i) (f("[<-"(x,i,x[i]+eps)) - f("[<-"(x,i,x[i]-eps)))/2/eps)
src = "#include <RcppArmadillo.h>
using namespace arma;
// [[Rcpp::depends(RcppArmadillo)]]
// [[Rcpp::export]]
mat differenceMatrix(int n) {
return join_rows(zeros(n-1,1),eye(n-1,n-1)) - join_rows(eye(n-1,n-1),zeros(n-1,1));
}
// [[Rcpp::export]]
vec quadraticPenalty(mat Q, vec beta) {
vec delta = differenceMatrix(Q.n_rows) * Q * beta;
return delta % delta % (delta<0);
}
// [[Rcpp::export]]
vec gradientPenalty(mat Q, vec beta) { // Q: (nbeta+2) x nbeta
size_t n = Q.n_rows;
mat D = join_rows(zeros(n-1,1),eye(n-1,n-1)) - join_rows(eye(n-1,n-1),zeros(n-1,1)); // (nbeta+1) x (nbeta+2)
vec delta = D * Q * beta; // nbeta+1
mat M = Q.t() * D.row(0).t() * D.row(0) * Q * (delta(0)<0.0); // nbeta x nbeta
for(size_t j=1; j<delta.size(); j++) {
if (delta(j)<0.0)
M += Q.t() * D.row(j).t() * D.row(j) * Q;
}
return 2*M*beta;
}
// [[Rcpp::export]]
vec gradientPenalty2(mat Q, vec beta) { // Q: (nbeta+2) x nbeta; beta: nbeta
vec betaStar = Q * beta; // nbeta+2
size_t n = Q.n_rows; // nbeta+2
mat D(n-1,n); // nbeta+1 x nbeta+2
D.zeros();
for(size_t i=0; i<n-1; i++)
if (betaStar(i)>betaStar(i+1)) {
D(i,i) = -1.0;
D(i,i+1) = 1.0;
}
mat M = Q.t() * D.t() * D * Q; // nbeta x nbeta
return 2*M*beta;
}
"
sourceCpp(code=src)
##
library(rstpm2)
Qmat = attr(nsx(1:10,df=3,intercept = TRUE),"q.const") # 5x3
differenceMatrix(nrow(Qmat)) # 4x5
quadraticPenalty(Qmat, c(1,-1,1)) # 4x1
as.vector(gradientPenalty(Qmat, c(1,-1,1)))
as.vector(gradientPenalty2(Qmat, c(1,-1,1)))
fdiff(\(x) sum(quadraticPenalty(Qmat, x)), c(1,-1,1), 1e-5)
as.vector(gradientPenalty(Qmat, c(1,-1,1))) -
fdiff(\(x) sum(quadraticPenalty(Qmat, x)), c(1,-1,1), 1e-3)
quadraticPenalty(Qmat, c(1,-1,2)) # 4x1
gradientPenalty(Qmat, c(1,-1,2))
## Error report by Joshua (now fixed)
## Load package
library(rstpm2)
## Define start time
brcancer2 <- transform(brcancer,
startTime = ifelse(hormon == 0, rectime * 0.5, 0))
## Stpm2 call without explicit specification of the type argument
stpm2(Surv(startTime, rectime, censrec == 1) ~ hormon,
data = brcancer2,
df = 3)
## Stpm2 call with explicit specification of the type argument
## Do not run, this will cause your R-session to terminate
## debug(gsm)
stpm2(Surv(startTime, rectime, censrec == 1, type = "counting") ~ hormon,
data = brcancer2,
df = 3)
## Are 2D thin-plate spline approximations in mgcv dense matrices?
library(mgcv)
df=expand.grid(u=seq(0,1,length=100),v=seq(0,1,length=100))
set.seed(12345)
df=transform(df,y=rnorm(nrow(df),sin(u*2*pi)*cos(v*2*pi),0.1))
gam1=gam(y~s(u,v,bs="ts"),data=df)
## Email from Grace
library(rstpm2)
fit = gsm(Surv(rectime,censrec==1)~hormon,data=brcancer,df=3)
fit@lm
attr(fit@lm$terms, "predvars")
nd = data.frame(rectime=1000, hormon=0)
rstpm2:::lpmatrix.lm(fit@lm, nd) %*% coef(fit) # or, more simply
predict(fit@lm, nd)
##
eps=1e-4
(predict(fit@lm, data.frame(rectime=1000+eps, hormon=0)) -
predict(fit@lm, data.frame(rectime=1000-eps, hormon=0)))/2/eps
## or
(predict(fit@lm, data.frame(rectime=1000*exp(eps), hormon=0)) -
predict(fit@lm, data.frame(rectime=1000*exp(-eps), hormon=0)))/2/eps/1000
##
exp(predict(fit@lm, nd))*
(predict(fit@lm, data.frame(rectime=1000*exp(eps), hormon=0)) -
predict(fit@lm, data.frame(rectime=1000*exp(-eps), hormon=0)))/2/eps/1000
predict(fit, newdata=nd, type="haz")
plot(fit,type="haz", newdata=data.frame(hormon=0))
## biostat3: coxphHaz and coxphHazList
as.data.frame.coxphHaz = function(x, row.names=NULL, optional = FALSE, ...) {
newdata = attr(x,"newdata")
## To avoid "row names were found from a short variable and have been discarded":
rownames(newdata) = NULL
data.frame(newdata, x=x$x, y=x$y)
}
as.data.frame.coxphHazList = function(x, row.names=NULL, optional = FALSE, ...) {
do.call(rbind, lapply(x, as.data.frame))
}
fit <- coxph(Surv(surv_mm/12,status=="Dead: cancer")~agegrp, data=colon)
newdata <- data.frame(agegrp=levels(colon$agegrp))
haz <- suppressWarnings(coxphHaz(fit,newdata))
library(ggplot2)
head(as.data.frame(haz))
ggplot(as.data.frame(haz), aes(x=x,y=y,ymin=y.lower,ymax=y.upper,col=agegrp,fill=agegrp)) +
geom_ribbon(alpha=0.5) + geom_line(col="black") # + facet_grid(~agegrp)
## Constrained MLE
## relative survival and LEL
library(rstpm2)
library(biostat3)
head(melanoma)
head(popmort)
merged = merge(transform(melanoma, ye = floor(yexit), se=ifelse(sex=="Male",1,2), event=status != "Alive"),
popmort, by.x=c("age","ye","se"), by.y=c("_age","_year","sex"), all.x=TRUE)
merged = subset(merged, sex="Female")
## sum(is.na(merged$rate))
model1 = stpm2(Surv(surv_mm,event)~bhazard(rate),data=merged,df=3) # only intercept - was buggy
## model2 = stpm2(Surv(surv_mm,event)~age+bhazard(rate)+cluster(age),data=merged,df=3)
## model1 = stpm2(Surv(surv_mm,event)~1,data=merged,df=3,bhazard=merged$rate)
## mean(subset(merged,event)$yexit)
## We need event counts and person-time in the popmort data-frame
library(dplyr)
temp = cbind(filter(popmort,sex==1) %>% mutate(males=rate,rate=NULL,sex=NULL),
data.frame(females=filter(popmort,sex==2)$rate))
plot(males~females,temp,log="xy")
abline(0,1,col=2)
##
d = data.frame(y=c(1000,2000),sex=0:1,pt=1e5) # some data
fit = glm(y~sex+offset(log(pt)),data=d, family=poisson) # Poisson regression
pred = predict(fit,newdata=transform(d,pt=1),type="link",se.fit=TRUE)
with(pred,
exp(data.frame(Estimate=fit,Lower=fit-1.96*se.fit,Upper=fit+1-96*se.fit)))
library(rstpm2)
exp(confint(predictnl(fit,predict,newdata=transform(d,pt=1),type="response")))
set.seed(12345)
n = 1000
x = factor(rbinom(n,1,0.5))
y = rnorm(n, x==1)
fit = lm(y~x)
predict(fit, newdata=data.frame(x=0)) # fails
predict(fit, newdata=data.frame(x=1)) # fails
predict(fit, newdata=data.frame(x=factor(0:1))) # okay
predict(fit, newdata=data.frame(x=factor(0,0:1))) # okay
##
library(expm)
##
K = 2
Z = c(1,0,0)
Y = c(1000,2000); N = 1e4; Sigma=diag(Y/N^2)
lambda=Y/N
##
alpha=cbind(c(-sum(lambda),lambda),matrix(0,K+1,K))
expm(alpha*10) # ok
##
Z = c(1,rep(0,(1+2)*3-1))
alpha=cbind(c(-sum(lambda),lambda,-1,-1,1,0,0,1),matrix(0,length(Z),length(Z)-1))
expm(alpha)
m=matrix(expm(alpha)[4:9,1],2,3)
t(m) %*% Sigma %*% m # includes covariance terms
expm(alpha*10)
m=matrix(expm(alpha*10)[4:9,1],2,3)
sqrt(diag(t(m) %*% Sigma %*% m)) # includes covariance terms
## This approach works!
## We can add further differential equations for the covariance matrix for P (for a time-dependent covariance matrix for the parameters)
library(deSolve)
func <- function(t, y, parms, ...) {
## browser()
index = findInterval(t,parms$t)
lambda = parms$lambda[index,] # length=K
K = length(lambda)
q = rbind(c(-sum(lambda),lambda),matrix(0,K,K+1)) # length=K+1
Qm = lapply(1:K,function(k) {
m = matrix(0,K+1,K+1)
m[1,1] = -1
m[1,k+1] = 1
m
})
Sigma = parms$Sigma[[index]] # dim=K^2
P = y[1:(K+1)] # K+1 rowvec
Pm = matrix(y[(2+K):(1+K+K*(K+1))],K+1,K) # (K+1)*K
dPdt = P %*% q
dPmdt = sapply(1:K, function(k) Pm[,k] %*% q + P %*% Qm[[k]]) # (K+1)*K
dphidt = dPmdt %*% (Sigma %*% t(Pm)) + Pm %*% (Sigma %*% t(dPmdt)) # (K+1)^2
list(c(dPdt, dPmdt, dphidt)) # length=1+K+(K+1)*K+(K+1)^2
}
Y = c(1000,2000); N = 1e4; Sigma=diag(Y/N^2); K=2
lambda = Y/N
y = c(1,rep(0,K+K*(K+1)+(K+1)^2))
parms = list(t=c(0,Inf), lambda=rbind(lambda), Sigma=list(Sigma))
ode1 = ode(y,seq(0,10),func,parms)
ode1[11,2:4]
sqrt(diag(matrix(ode1[11,11:19],3,3)))
ode1
## Using markov_msm (with an *identity* link for comparability)
fit1=glm(rate~1, data=data.frame(rate=0.1,N=1e4), family=poisson(link="identity"), weight=N)
fit2=glm(rate~1, data=data.frame(rate=0.2,N=1e4), family=poisson(link="identity"), weight=N)
test=markov_msm(list(fit1,fit2),
trans=matrix(c(NA,1,2,NA,NA,NA,NA,NA,NA),3,3,TRUE), t=0:10, newdata=data.frame(N=1),
tmvar="t")
print(test,se=TRUE)
test$Pu
## fit1=glm(rate~factor(findInterval(t,0:2))-1, data=data.frame(t=0:2,rate=c(0.1,0.2,0.3),N=1e4), family=poisson(link="identity"), weight=N)
fit1=glm(Y~offset(log(N)), data=data.frame(Y=1000,N=1e4), family=poisson)
fit2=glm(Y~offset(log(N)), data=data.frame(Y=2000,N=1e4), family=poisson)
test=markov_msm(list(fit1,fit2),
trans=matrix(c(NA,1,2,NA,NA,NA,NA,NA,NA),3,3,TRUE), t=0:10, newdata=data.frame(N=1),
tmvar="t")
print(test,se=TRUE)
## Inquiry
library(survival)
library(rstpm2)
m <- rstpm2::stpm2(Surv(time, status) ~ sex, data = lung)
plot(m, newdata=data.frame(sex=2))
predict(m, newdata = data.frame(sex = 2, time = 100), type = "rmst", se.fit=TRUE)
## plot(m, newdata=data.frame(sex=2), type="rmst") # fails
## predict(m, newdata = data.frame(sex = 1:2, time = 100), type = "rmst", se.fit=TRUE) # fails
## Inquiry
## Cross-validation with bias correction (Davison and Hinkley, 1997, p. 295)
crossValidation = function(object,data,prediction,K = 10,correction=TRUE) {
N = nrow(data)
ndex = 1:N
part = as.factor(sample(1:K, N, replace = TRUE))
folds = split(ndex, part)
p = sapply(folds, length)/N
sum(sapply(1:K, function(k) {
training = data[unlist(folds[-k]), ]
validation = data[folds[[k]], ]
objectStar = update(object,data=training)
prediction(objectStar,validation) - p[k]*prediction(objectStar,data)*correction
})) + prediction(object,data)*correction
}
library(rstpm2)
setMethod("update", "stpm2", function(object, ...) { # I have added this to rstpm2 on GitHub
object@call = object@Call
update.default(object, ...)
})
fit = gsm(Surv(rectime,censrec==1)~hormon+x1+x2+x3,data=brcancer,df=3)
set.seed(420)
## Currently, there is no simpler way to predict out-of-sample -2*log-likelihood for an stpm2 model
predictDeviance = function(eventVar) function(object,newdata) {
events = newdata[[eventVar]]>0
-2*(sum(predict(object,newdata=newdata[events,,drop=FALSE],
type="loghazard"))-
sum(predict(object,newdata=newdata,type="cumhaz")))
}
crossValidation(fit, brcancer,predictDeviance("censrec"))
set.seed(420)
crossValidation(fit, brcancer,predictDeviance("censrec"),correction=FALSE)
AIC(fit)
BIC(fit)
library(rstpm2)
library(survival)
library(timereg)
library(ggplot2)
library(lattice)
## Two states: Initial -> Final
## Note: this shows how to use markov_msm to estimate survival and risk probabilities based on
## smooth hazard models.
two_states <- function(model, ...) {
transmat = matrix(c(NA,1,NA,NA),2,2,byrow=TRUE)
rownames(transmat) <- colnames(transmat) <- c("Initial","Final")
markov_sde(list(model), ..., trans = transmat)
}
## Note: the first argument is the hazard model. The other arguments are arguments to the
## markov_msm function, except for the transition matrix, which is defined by the new function.
colon2 <- transform(survival::colon, Obs=(rx=="Obs")+0, Lev=(rx=="Lev")+0,Lev_5FU=(rx=="Lev+5FU")+0)
death = aalen(Surv(time,status)~Lev+Lev_5FU, data=subset(colon2,etype==2))
cr = two_states(death, newdata=data.frame(Lev=0,Lev_5FU=0))
plot(cr,ggplot=TRUE,stacked=FALSE,which="L") # ok
plot(cr,ggplot=TRUE,which="L") # ok - is this sensible?
plot(cr) # ok -- no legend
plot(cr,ggplot=TRUE) # ok
plot(cr,lattice=TRUE) # ok -- no legend
plot(cr,ggplot=TRUE,stacked=FALSE) # ok
##
competing_risks <- function(models, ...) {
transmat = matrix(c(NA,1,2,
NA,NA,NA,
NA,NA,NA),3,3,byrow=TRUE)
rownames(transmat) <- colnames(transmat) <- c("Initial","Event 1","Event 2")
markov_sde(models, ..., trans = transmat)
}
## Note: the first argument is the hazard model. The other arguments are arguments to the
## markov_msm function, except for the transition matrix, which is defined by the new function.
colon2 <- transform(survival::colon, Obs=(rx=="Obs")+0, Lev=(rx=="Lev")+0,Lev_5FU=(rx=="Lev+5FU")+0)
progression = aalen(Surv(time,status)~Lev+Lev_5FU, data=subset(colon2,etype==1))
death = aalen(Surv(time,status)~Lev+Lev_5FU, data=subset(colon2,etype==2))
cr = competing_risks(list(progression,death), newdata=data.frame(Lev=0,Lev_5FU=0))
plot(cr)
##
illness_death <- function(models, ...) {
transmat = matrix(c(NA,1,2,
NA,NA,3,
NA,NA,NA),3,3,byrow=TRUE)
rownames(transmat) <- colnames(transmat) <- c("Initial","Illness","Death")
markov_sde(models, ..., trans = transmat)
}
## Note: the first argument is the hazard model. The other arguments are arguments to the
## markov_msm function, except for the transition matrix, which is defined by the new function.
colon2 <- transform(survival::colon, Obs=(rx=="Obs"), Lev=(rx=="Lev"),Lev_5FU=(rx=="Lev+5FU"))
d1 = subset(colon2,etype==1)
d2 = subset(colon2,etype==2)
index = d1$time < d2$time & d1$status==1 # index for initial -> recurrence
d12 = d1 # initial -> recurrence
d23 = transform(d2[index,], enter=d1$time[index]) # recurrence -> death
d13 = transform(d2, # initial -> death
time=pmin(d1$time,d2$time),
status=ifelse(d1$time==d2$time,d2$status,0))
progression = aalen(Surv(time,status)~Lev+Lev_5FU, data=d12)
death1 = aalen(Surv(time,status)~Lev+Lev_5FU, data=d13)
death2 = aalen(Surv(enter,time,status)~Lev+Lev_5FU, data=d23)
cr = illness_death(list(progression,death1,death2), newdata=data.frame(Lev=FALSE,Lev_5FU=FALSE)) # fails
plot(cr)
##
illness_death = function(models, ...) {
transmat = matrix(c(NA,1,2,NA,
NA,NA,NA,3,
NA,NA,NA,NA,
NA,NA,NA,NA),4,4,byrow=TRUE)
rownames(transmat) <- colnames(transmat) <- c("Initial","Illness","DirectDeath","DeathAfterRecurrence")
markov_sde(models, ..., trans = transmat)
}
## Note: the first argument is the hazard model. The other arguments are arguments to the
## markov_msm function, except for the transition matrix, which is defined by the new function.
colon2 = transform(survival::colon, Obs=(rx=="Obs")+0, Lev=(rx=="Lev")+0, Lev_5FU=(rx=="Lev+5FU")+0)
d1 = subset(colon2,etype==1)
d2 = subset(colon2,etype==2)
index = d1$time < d2$time & d1$status==1 # index for initial -> recurrence
d12 = d1 # initial -> recurrence
d23 = transform(d2[index,], enter=d1$time[index]) # recurrence -> death
d13 = transform(d2, # initial -> death
time=pmin(d1$time,d2$time),
status=ifelse(d1$time==d2$time,d2$status,0))
progression = aalen(Surv(time,status)~Lev+Lev_5FU, data=d12)
death1 = aalen(Surv(time,status)~Lev+Lev_5FU, data=d13)
death2 = aalen(Surv(enter,time,status)~Lev+Lev_5FU, data=d23)
cr = illness_death(list(progression,death1,death2), newdata=data.frame(Lev=0,Lev_5FU=0:1),
weights=c(-1,1), los=FALSE)
plot(cr,ggplot=TRUE) + facet_grid(~Lev_5FU) # ok
plot(cr,ggplot=TRUE,stacked=FALSE) + facet_grid(state~Lev_5FU) # ok
plot(cr,ggplot=TRUE,stacked=FALSE,which="L") + facet_grid(state~Lev_5FU) # ok
plot(standardise(cr),stacked=FALSE,ggplot=TRUE) # CIs are *very* wide -- error?
## Non-parametric baseline: SDE approach due to Ryalen and colleagues
markov_sde <- function(models, trans, newdata, init=NULL, nLebesgue=1e4+1, los=FALSE, nOut=300,
weights=1) {
transfun <- function(tmat) {
indices <- sort(as.vector(tmat)); indices <- setdiff(indices,NA)
nStates <- nrow(tmat)
out <- do.call(rbind,
lapply(indices, function(i) {
index2 <- which(tmat == i)
from <- (index2-1) %% nStates +1
to <- (index2-1) %/% nStates + 1
data.frame(from=from,to=to)
}))
matrix(as.integer(as.matrix(out)),nrow(out))-1L
}
## TODO check parameters
nStates <- nrow(trans)
nTrans <- sum(!is.na(trans))
if (is.null(init)) {
init <- c(1,rep(0,nStates-1))
}
stopifnot(length(init)==nStates)
## init <- rep(init,nrow(newdata))
n <- sum(sapply(models,attr,"orig.max.clust"))
cumHazList <- lapply(models, function(object)
-t(log(predict(object, newdata=newdata, se=FALSE)$S0)))
timesList <- lapply(models, function(model) model$cum[,1])
eventTimes <- times <- sort(unique(unlist(timesList)))
hazList <- lapply(1:nrow(newdata), function(i) {
hazMatrix <- matrix(0,nrow=nTrans,ncol=length(times))
for (j in 1:nTrans) {
hazMatrix[j,match(timesList[[j]],times)] <- diff(c(0,cumHazList[[j]][,i]))
}
hazMatrix
})
hazMatrix <- do.call(rbind,hazList)
if (length(weights)==1 && weights != 0)
weights <- rep(weights,nrow(newdata))/(weights*nrow(newdata))
vcov <- matrix(1,1,1)
if (!los) {
out <- .Call("plugin_P_by", n, nrow(newdata), hazMatrix, init, transfun(trans), weights,
vcov,
PACKAGE="rstpm2")
out$P <- out$X
out$P.se <- sqrt(out$variance)
} else {
out <- .Call("plugin_P_L_by",
n, nrow(newdata), hazMatrix, init, transfun(trans), times, weights, nOut,
vcov, nLebesgue,
PACKAGE="rstpm2")
PIndex <- 1:(nrow(out$X)/2)
tr <- function(x) array(as.vector(x), dim=c(nStates,nrow(newdata),ncol(x)))
out$P <- tr(out$X[PIndex,])
out$L <- tr(out$X[-PIndex,])
out$P.se <- sqrt(tr(out$variance[PIndex,]))
out$L.se <- sqrt(tr(out$variance[-PIndex,]))
}
out$n <- n
out$times <- if (los) out$time else times
out$newdata <- newdata
out$trans <- trans
out$los <- los
out$init <- init
out$weights <- weights
class(out) <- "markov_sde"
if (!all(weights==0)) {
stand <- out # Warning: copy! This may be a bad idea...
if (!los) {
stand$P <- stand$Y
stand$P.se <- sqrt(stand$varY)
} else {
PIndex <- 1:(nrow(out$Y)/2)
stand$P <- stand$Y[PIndex,]
stand$L <- stand$Y[-PIndex,]
stand$P.se <- sqrt(stand$varY[PIndex,])
stand$L.se <- sqrt(stand$varY[-PIndex,])
}
## tidy up
stand$X <- stand$Y
stand$variance <- stand$varY
## out$X <- out$variance <- out$Y <- out$varY <- stand$Y <- stand$varY <- NULL
stand$newdata <- stand$newdata[1,,drop=FALSE]
class(stand) <- "markov_sde"
out$stand <- stand
}
out
}
standardise.markov_sde <- function(object) {
structure(object$stand, class="markov_sde")
}
cr = two_states(death, newdata=data.frame(rx=levels(survival::colon$rx)))
cr = two_states(death, newdata=data.frame(Obs=1))
plot.markov_sde <- function(x, y, stacked=TRUE, which=c("P","L"),
xlab="Time", ylab=NULL, col=2:6, border=col,
ggplot2=FALSE, lattice=FALSE, alpha=0.2,
strata=NULL,
...) {
stopifnot(inherits(x,"markov_sde"))
which <- match.arg(which)
if (!missing(y)) warning("y argument is ignored")
## ylab defaults
if (is.null(ylab))
ylab <- if(which=='P') "Probability" else "Length of stay"
if (ggplot2)
rstpm2:::ggplot.markov_msm(x, which=which, stacked=stacked, xlab=xlab, ylab=ylab,
alpha=alpha, ...)
else if (lattice)
rstpm2:::xyplot.markov_msm(x, which=which, stacked=stacked, xlab=xlab, ylab=ylab,
col=col, border=border, strata=strata, ...)
else {
if (is.null(index) && nrow(x$newdata)>1) {
warning("More than one set of covariates; defaults to weighted estimator")
x <- x$stand # Warning: replacement
index <- 1
}
browser()
if (is.null(index)) index <- 1
df <- merge(x$newdata[index,,drop=FALSE], as.data.frame(x))
states <- unique(df$state)
if (stacked) {
out <- graphics::plot(range(x$times, na.rm=TRUE),0:1, type="n", xlab=xlab, ylab=ylab, ...)
lower <- 0
for (i in length(states):1) { # put the last state at the bottom
df2 <- df[df$state==states[i],]
if (length(lower)==1) lower <- rep(0,nrow(df2))
upper <- lower+df2[[which]]
graphics::polygon(c(df2$time,rev(df2$time)), c(lower,rev(upper)),
border=border[i], col=col[i])
lower <- upper
}
graphics::box()
invisible(out)
}
else stop('Unstacked plot not implemented in base graphics; use ggplot2=TRUE or lattice=TRUE')
}
}
plot(standardise(cr)) # ok (no legend)
plot(standardise(cr), ggplot=TRUE) # ok (includes legend)
plot(standardise(cr), lattice=TRUE) # ok (no legend)
plot(cr, ggplot=TRUE) + facet_grid(~ rx) # ok
plot(cr, index=1) # incorrect
plot(cr, ggplot=TRUE, index=1) # incorrect
plot(cr, lattice=TRUE, index=1) # incorrect
plot(standardise(cr), ggplot=TRUE, stacked=FALSE) # incorrect: wrong CIs for standardised estimates
plot(cr, ggplot=TRUE, stacked=FALSE) + facet_grid(~ rx) # incorrect
as.data.frame.markov_sde <- function(x, row.names=NULL, ci=TRUE,
P.conf.type="logit", L.conf.type="log",
P.range=c(0,1), L.range=c(0,Inf),
...) {
if (any(x$weights<0))
P.conf.type <- L.conf.type <- "plain"
.id. <- 1:nrow(x$newdata)
nStates <- nrow(x$trans)
state.names <- rownames(x$trans)
stateNames <- if (!is.null(rownames(x$trans))) rownames(x$trans) else 1:nrow(x$trans)
out <- expand.grid(state=stateNames, .id.=.id., time=x$times)
out <- cbind(x$newdata[out$.id.,],out)
names(out)[1:ncol(x$newdata)] <- colnames(x$newdata)
out$P <- as.vector(x$P)
out$P.se <- as.vector(x$P.se)
out <- out[order(out$.id.,out$state,out$time),]
out$.id. <- NULL
if (ci) {
tmp <- rstpm2:::surv.confint(out$P,out$P.se, conf.type=P.conf.type, min.value=P.range[1], max.value=P.range[2])
out$P.lower <- tmp$lower
out$P.upper <- tmp$upper
}
if (x$los) {
out$L <- as.vector(x$L)
out$L.se <- as.vector(x$L.se)
if (ci) {
tmp <- rstpm2:::surv.confint(out$L,out$L.se, conf.type=L.conf.type, min.value=L.range[1], max.value=L.range[2])
out$L.lower <- tmp$lower
out$L.upper <- tmp$upper
}
}
if(!is.null(row.names)) rownames(out) <- row.names
out
}
temp = as.data.frame(cr)
head(temp)
ggplot(temp,aes(x=time,y=P,fill=rx,ymin=P.lower,ymax=P.upper)) + geom_line() +
geom_ribbon(alpha=0.5) + facet_grid(state~rx) # ok
rstpm2:::ggplot.markov_msm(temp) + facet_grid(~rx) # ok
## Competing risks
## Note: this shows how to adapt the markov_msm model for competing risks.
competing_risks <- function(listOfModels, ...) {
nRisks = length(listOfModels)
transmat = matrix(NA,nRisks+1,nRisks+1)
transmat[1,1+(1:nRisks)] = 1:nRisks
rownames(transmat) <- colnames(transmat) <- c("Initial",names(listOfModels))
rstpm2::markov_msm(listOfModels, ..., trans = transmat)
}
## Note: The first argument for competing_risks is a list of models. Names from that list are
## used for labelling the states. The other arguments are as per the markov_msm function,
## except for the transition matrix, which is defined by the competing_risks function.
recurrence = gsm(Surv(time,status)~factor(rx), data=survival::colon, subset=(etype==1), df=3)
death = gsm(Surv(time,status)~factor(rx), data=survival::colon, subset=(etype==2), df=3)
cr = competing_risks(list(Recurrence=recurrence,Death=death),
newdata=data.frame(rx=levels(survival::colon$rx)),
t = seq(0,2500, length=301))
## Plot the probabilities for each state for three different treatment arms
plot(cr, ggplot=TRUE) + facet_grid(~ rx)
## And: differences in probabilities
cr_diff = diff(subset(cr,rx=="Lev+5FU"),subset(cr,rx=="Obs"))
plot(cr_diff, ggplot=TRUE, stacked=FALSE)
## Example using Crowther and Lambert (2018)
library(readstata13)
library(transform.hazards)
library(timereg)
library(rstpm2) # standardise
mex.1 <- read.dta13("https://fmwww.bc.edu/repec/bocode/m/multistate_example.dta")
transmat <- rbind("Post-surgery"=c(NA,1,2),
"Relapsed"=c(NA,NA,3),
"Died"=c(NA,NA,NA))
colnames(transmat) <- rownames(transmat)
mex.2 <- transform(mex.1,osi=(osi=="deceased")+0)
levels(mex.2$size)[2] <- ">20-50 mm" # fix typo
mex <- mstate::msprep(time=c(NA,"rf","os"),status=c(NA,"rfi","osi"),
data=mex.2,trans=transmat,id="pid",
keep=c("age","size","nodes","pr_1","hormon"))
mex <- transform(mex,
size2=(unclass(size)==2)+0, # avoids issues with TRUE/FALSE
size3=(unclass(size)==3)+0,
hormon=(hormon=="yes")+0,
Tstart=Tstart/12,
Tstop=Tstop/12)
## Slow fitting...
c.ar <- aalen(Surv(Tstart,Tstop,status) ~ const(age) + size2 + size3 + const(nodes) + pr_1 + const(hormon),
data = subset(mex, trans==1))
c.ad <- aalen(Surv(Tstart, Tstop, status) ~ const(age) + const(size) + const(nodes) + const(pr_1) + const(hormon),
data = subset(mex, trans==2))
c.rd <- aalen( Surv(Tstart,Tstop,status) ~ const(age) + const(size) + const(nodes) + pr_1 + const(hormon),
data=subset(mex, trans==3))
##
nd <- expand.grid(nodes=seq(0,20,10), size=levels(mex$size))
nd <- transform(nd, age=54, pr_1=3, hormon=0, size2=(unclass(size)==2)+0, size3=(unclass(size)==3)+0)
system.time(fit1 <- markov_sde(list(c.ar,c.ad,c.rd), trans=transmat, newdata=nd[c(1,2),], los=TRUE))
system.time(fit0 <- markov_sde(list(c.ar,c.ad,c.rd), trans=transmat, newdata=nd[c(1,2),]))
plot(fit1)
plot(fit1, ggplot=TRUE)
plot(fit1, lattice=TRUE, stacked=FALSE, which="L")
## plot(fit1,which="L",xlim=NULL) # does this make sense?
df1 <- as.data.frame(fit1)
fit2 <- standardise(fit1)
plot(fit2,ggplot=TRUE)
df2 <- as.data.frame(fit2)
diff1 <- diff(fit1)
plot(diff1)
## type="af"
library(rstpm2)
fit = gsm(Surv(rectime,censrec==1)~hormon,data=brcancer,df=3)
plot(fit,type="af",newdata=brcancer,
exposed=function(data) transform(data,hormon=1))
pred2 <- predict(fit,type="af",newdata=brcancer, grid=TRUE,
exposed=function(data) transform(data,hormon=1),
se.fit=TRUE,full=TRUE)
with(pred2, matplot(rectime,cbind(Estimate,lower,upper),type="l",lty=c(1,2,2), col=1,
xlab="Time since treatment (days)", ylab="PAF", ylim=c(0,0.5)))
## Bug report from Joshua
# Loading rstpm2 package
library(rstpm2)
# Fit stmp2 model using breastcancer data set
fpm_model <- stpm2(Surv(rectime, censrec) ~ hormon,
data = brcancer,
df = 3,
tvc = list("hormon" = 3))
# Predict hazard difference comparing hormon users and non-users
# using full=TRUE option for obtaining a full data set for ggplot()
predict(
fpm_model,
type = "hdiff",
newdata = data.frame(hormon = 0),
var = "hormon",
grid = TRUE,
se.fit = TRUE,
full = TRUE)
## test plots for PO models with random effects
library(rstpm2)
set.seed(12345)
logit <- binomial()$linkfun
expit <- binomial()$linkinv
Spo <- function(eta) expit(eta)
eta <- function(t) {}
rPOgamma <- function(n,eta,theta,eps=1e-16) {
## U <- pmin(runif(n),1-eps)
U <- runif(n)
## solve_t(U=expit(-eta(t))^theta), where eta(0)=-Inf and eta(Inf)=Inf
V <- -logit(U^(1/theta))
if (any(is.na(V))) browser()
c(list(U=U),vuniroot(function(t) eta(t)-V, lower=rep(1e-50,n), upper=rep(1e100,n)))
}
t0 <- rPOgamma(1e5, function(t) log(t), 1)
with(t0,plot(U,root,log="y"))
t1 <- rPOgamma(n=1e5, function(t) log(t), rgamma(10,1))
range(t0$root)
range(t1$root)
plot(density(log(t0$root)))
## Plots with three levels
library(rstpm2)
table(brcancer$x4) # cancer stage
## define indicators for the "exposed" levels (*R* needs these to be numeric)
d <- transform(brcancer, x4.2=(x4==2)+0, x4.3=(x4==3)+0)
## fit the model with the indicators as main effects and with tvc
fit <- stpm2(Surv(rectime,censrec==1)~x4.2+x4.3,data=d,df=3,tvc=list(x4.2=2,x4.3=2))
## predict for each exposure level
pred2 <- predict(fit,newdata=data.frame(x4.2=0,x4.3=0),type="hr",var="x4.2", grid=TRUE, full=TRUE,
se.fit=TRUE)
pred3 <- predict(fit,newdata=data.frame(x4.2=0,x4.3=0),type="hr",var="x4.3", grid=TRUE, full=TRUE,
se.fit=TRUE)
pred <- transform(rbind(pred2,pred3), x4=factor(ifelse(x4.2,2,3)))
library(ggplot2)
ggplot(pred, aes(x=rectime, y=Estimate, ymin=lower, ymax=upper, fill=x4, col=x4)) +
geom_line() + coord_cartesian(ylim=c(0,10)) + geom_ribbon(alpha=0.3, col=NA)
## how to extract the design information
library(rstpm2)
fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,df=3)
names(fit@model.frame)
attributes(fit@model.frame[[3]])
fit@lm$terms
fit@x <- matrix()
ls(environment(fit@model.frame))
##
withEnvs <- sapply(slotNames(fit),function(nm) !is.null(environment(slot(fit,nm))))
slotNames(fit)[withEnvs]
lapply(slotNames(fit)[withEnvs],function(nm) ls(environment(slot(fit,nm))))
##
withEnvs <- sapply(fit@args,function(obj) !is.null(environment(obj)))
names(fit@args)[withEnvs]
lapply(fit@args[withEnvs],function(obj) ls(environment(obj)))
## bug in predict for meansurv
library(rstpm2)
library(ggplot2)
fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,df=3)
## Easy to use plot and lines functions
plot(fit, newdata=transform(brcancer, hormon=0), type="meansurv")
lines(fit, newdata=transform(brcancer, hormon=1), type="meansurv",lty=2)
## More tedious to do for different covariate patterns with ggplot2
pred0 <- predict(fit, newdata=transform(brcancer, hormon=0), type="meansurv", full=TRUE, se.fit=TRUE,
grid=TRUE)
pred1 <- predict(fit, newdata=transform(brcancer, hormon=1), type="meansurv", full=TRUE, se.fit=TRUE,
grid=TRUE)
## bug with values returned AsIs - I'll try to fix this
unAsIs <- function(object)
if (inherits(object,"AsIs")) "class<-"(object, setdiff(class(object), "AsIs")) else object
pred <- rbind(transform(unAsIs(pred1),hormon=1),transform(unAsIs(pred0),hormon=0))
pred <- transform(pred, Hormone=ifelse(hormon==1,"Yes","No"))
ggplot(pred, aes(x=rectime,y=Estimate,ymin=lower,ymax=upper,fill=Hormone)) +
xlab("Time since diagnosis (years)") +
ylab("Standardised survival") +
geom_ribbon(alpha=0.6) +
geom_line()
## bug in predict for meansurv
library(devtools)
install_github("mclements/rstpm2", ref="develop")
library(rstpm2)
fit.tvc <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,df=3, tvc=list(hormon=3))
out <- predict(fit.tvc, newdata=transform(brcancer,hormon=1),type="meansurv",grid=TRUE, se.fit=TRUE,
full=TRUE)
## bug in plot(..., xlab="Something")
library(rstpm2)
fit <- stpm2(Surv(rectime, censrec)~hormon, data=brcancer,df=3)
plot(fit, newdata=data.frame(hormon=1), type="hr", xlab="Time since diagnosis (years)", var="hormon",
ylab="Hazard ratio", main="Lung cancer-BMI")
## predict linear predictor
library(rstpm2)
fit <- stpm2(Surv(rectime, censrec)~hormon, data=brcancer,df=3)
predict(fit, newdata=data.frame(hormon=0:1, rectime=1000), type="link")
predict(fit, newdata=data.frame(hormon=1, rectime=1000), type="link") -
predict(fit, newdata=data.frame(hormon=0, rectime=1000), type="link")
##
library(rstpm2)
library(Hmisc)
d <- local({
set.seed(12345)
x <- rep(0:1,length=200)
y <- rexp(length(x), exp(-3+x))
data.frame(x,y,e=TRUE)
})
fit0 <- stpm2(Surv(y, e)~1, data=d,df=3)
fit <- stpm2(Surv(y, e)~x, data=d,df=3)
rcorr.cens(predict(fit0,type="link")-predict(fit0,newdata=transform(d,x=0),type="link"),
with(d,Surv(y,e)))
rcorr.cens(-predict(fit,type="link")+
predict(fit,newdata=transform(d,x=0),type="link"),
with(d,Surv(y,e)))
summary(fit)
## testing - Bug requires loading devtools *before* rstpm2
setwd("~/src/R/rstpm2")
library(devtools)
devtools::test()
## offset
library(rstpm2)
brcancer2 <- transform(brcancer,off=0.1)
fit <- stpm2(Surv(rectime,censrec==1)~hormon, df = 2, data=brcancer2)
fit2 <- stpm2(Surv(rectime,censrec==1)~hormon+offset(off), df = 2, data=brcancer2)
fit
fit2
zeroModel(fit)
## p-value for survival differences
library(rstpm2)
fit <- stpm2(Surv(rectime,censrec==1)~hormon, df = 2, data=brcancer,tvc=list(hormon=2))
test <- predict(fit, type="sdiff", var="hormon", newdata=data.frame(hormon=0,rectime=500),se.fit=TRUE)
z <- test[1,1]/((test[1,3]-test[1,2])/2/qnorm(0.975))
2*pnorm(-abs(z))
##
## test the difference in survival rates
test <- predictnl(fit, function(object,newdata=NULL) {
lp1 <- predict(object, newdata=data.frame(hormon=1,rectime=500), type="surv")
lp2 <- predict(object, newdata=data.frame(hormon=0,rectime=500), type="surv")
lp1-lp2
})
with(test, c(fit=fit,
se.fit=se.fit,
statistic=fit/se.fit,
p=2*pnorm(-abs(fit/se.fit))))
## same p-value as predict(..., type="sdiff")
##
s18 <- summary(survfit(Surv(rectime,censrec==1)~hormon,data=brcancer),time=500)
z2<-diff(s18$surv)/sqrt(sum(s18$std.err^2))
2*pnorm(-abs(z2))
##
library(bpcp)
with(brcancer,
fixtdiff(rectime,censrec,hormon,500,doall = TRUE))
## bug #3 on GitHub
library(rstpm2)
# Function with model_formula paramter **ERRORS**
f_bad <- function(model_formula)
stpm2(formula = model_formula, df = 2, data=brcancer, link.type = "PH")
# Function with formula paramter to match the name pf stpm2's formula parameter **PASSES**
f_good <- function(formula)
stpm2(formula = formula, df = 2, data=brcancer, link.type = "PH")
f_bad (model_formula = Surv(rectime,censrec==1)~hormon)
f_good (formula = Surv(rectime,censrec==1)~hormon)
## random draws
library(rstpm2)
predict.cumhaz <-
function(object, newdata=NULL, ...)
{
stopifnot(inherits(object,"stpm2") || inherits(object,"pstpm2"))
args <- object@args
beta <- coef(object)
if (is.null(newdata))
X <- args$X
else if (inherits(object, "stpm2")) {
X <- object@args$transX(lpmatrix.lm(object@lm, newdata), newdata)
}
else if (inherits(object, "pstpm2")) {
X <- object@args$transX(predict(object@gam, newdata, type="lpmatrix"), newdata)
}
link <- object@link # cf. link for transformation of the predictions
eta <- as.vector(X %*% beta)
link$H(eta)
}
simulate <- function(object, nsim=nrow(as.data.frame(newdata)),
newdata=as.data.frame(object@data), lower=1e-6, upper=1e5, ...) {
e <- rexp(nsim)
objective <- function(time) {
newdata[[object@timeVar]] <- time
predict.cumhaz(object, type="cumhaz", newdata=newdata) - e
}
vuniroot(objective, lower=rep(lower,length=nsim), upper=rep(upper,length=nsim))$root
}
fit1 <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer)
set.seed(12345)
d <- do.call(rbind,lapply(1:100,function(i) brcancer))
system.time(r <- simulate(fit1, newdata=d))
length(r)
library(rstpm2)
fit1 <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer)
negll2 <- function(beta,svalues,missingIndicator,k=2)
sapply(svalues, function(s)
fit1@args$logli2(coef(fit1),ifelse(missingIndicator, s*coef(fit1)[k], 0)))
head(negll2(coef(fit1),c(0,1),(1:686) %% 2))
## values for the tests
library(rstpm2)
brcancer2 <- transform(rstpm2::brcancer, w=1)
brcancer2$w[1] <- NA
fit1 <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2,weights=w)
nd <- data.frame(hormon=1,rectime=1000)
predict(fit1, newdata=nd, type="surv")
predict(fit1, newdata=nd, type="fail")
predict(fit1, newdata=nd, type="haz")
predict(fit1, newdata=nd, type="hr",var="hormon")
predict(fit1, newdata=nd, type="hdiff", var="hormon")
## predictions for relative survival (email from Anke Richters)
set.seed(12345)
d <- with(list(t0=rexp(10000), # constant hazard
t1=rweibull(10000, 1.5), # rising hazard
bg=rexp(2*10000)), # constant background hazard (rate=1)
data.frame(t=pmin(bg,c(t0,t1)), x=rep(0:1,each=10000), e=TRUE, bhazard=1))
library(rstpm2)
uniroot(function(x) pexp(x)-pweibull(x,shape=1.5), c(1e-6,10)) # survival overlaps at 1
uniroot(function(x) 1-dweibull(x,shape=1.5)/pweibull(x,shape=1.5,lower=FALSE),
c(1e-6,10)) # hazards overlap at 0.444
fit <- stpm2(Surv(t,e)~x,data=d,tvc=list(x=3),bhazard=d$bhazard)
## fit <- pstpm2(Surv(t,e)~1,data=d,smooth.formula=~s(log(t))+s(log(t),by=x),bhazard=d$bhazard)
ts <- seq(0,4,length=301)[-1]
## hazard plot - ok
plot(fit,newdata=data.frame(x=0),type="hazard",ylim=c(0,4))
lines(fit,newdata=data.frame(x=1),type="hazard",lty=2)
abline(h=1,col="blue") # theorical
lines(ts,dweibull(ts,shape=1.5)/pweibull(ts,shape=1.5,lower.tail=FALSE),lty=2,col="blue") # theorical
## survival plot - ok
plot(fit,newdata=data.frame(x=0),type="surv")
lines(fit,newdata=data.frame(x=1),type="surv",lty=2)
lines(ts, pexp(ts,lower.tail=FALSE), col="blue") # theoretical
lines(ts, pweibull(ts,shape=1.5,lower.tail=FALSE), lty=2, col="blue") # theoretical
## hr - now fixed
plot(fit,newdata=data.frame(x=0),type="hr",var="x")
lines(ts, dweibull(ts,shape=1.5)/pweibull(ts,shape=1.5,lower.tail=FALSE), lty=2, col="blue") # theoretical
abline(h=1,lty=2)
## plot(fit,newdata=data.frame(x=0),type="hr",exposed=function(data) transform(data,x=1)) # same mistake
plot(fit,newdata=data.frame(x=0),type="sdiff",var="x",ylim=c(-1,1))
S0 <- predict(fit,newdata=data.frame(x=0),grid=TRUE,full=TRUE)
S1 <- predict(fit,newdata=data.frame(x=1),grid=TRUE,full=TRUE)
lines(S0$t,S1$Estimate-S0$Estimate,col="blue")
abline(h=0)
abline(v=1)
library(rstpm2)
fit1 <- coxph(Surv(rectime,censrec==1)~ns(x1,df=2),data=brcancer)
## Missing values in predictions
library(rstpm2)
brcancer2 <- brcancer
brcancer$rectime[1] <- NA
fit1 <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2)
pred1 <- predict(fit1,newdata=brcancer2)
head(pred1)
fit1
summary(fit1)
brcancer2 <- brcancer
brcancer$hormon[1] <- NA
fit1 <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2)
pred1 <- predict(fit1,newdata=brcancer2)
head(pred1)
fit1
summary(fit1)
##
library(rstpm2)
brcancer2 <- brcancer
brcancer$rectime[1] <- NA
fit1 <- pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2)
pred1 <- predict(fit1,newdata=brcancer2)
head(pred1)
fit1
summary(fit1)
brcancer2 <- brcancer
brcancer$hormon[1] <- NA
fit1 <- pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2)
pred1 <- predict(fit1,newdata=brcancer2)
head(pred1)
fit1
summary(fit1)
## Examples using predictnl for Alessandro
library(rstpm2)
brcancer2 <- transform(brcancer, x4.23=x4 %in% 2:3)
fit1 <- stpm2(Surv(rectime,censrec==1)~hormon*x4.23,data=brcancer2,df=3)
summary(fit1)
newd <- data.frame(hormon=0,x4.23=FALSE)
plot(fit1, newdata=newd)
RERI <- function(object, newdata,
var1, val1=1,
var2, val2=1) {
exp1 <- function(data) {data[[var1]] <- val1; data}
exp2 <- function(data) {data[[var2]] <- val2; data}
s00 <- predict(object, newdata, type="surv")
s10 <- predict(object, newdata=exp1(newdata), type="surv")
s01 <- predict(object, newdata=exp2(newdata), type="surv")
s11 <- predict(object, newdata=exp1(exp2(newdata)), type="surv")
-(s11-s10-s01+s00)/(1-s00)
}
times <- seq(0,2500,length=301)[-1]
reri <- RERI(fit1,newdata=transform(newd,rectime=times),var1="hormon",var2="x4.23",val2=TRUE)
plot(times,reri,type="l")
reri2 <- predictnl(fit1,fun=RERI,newdata=transform(newd,rectime=times),var1="hormon",var2="x4.23",val2=TRUE)
with(reri2, matplot(times,fit+cbind(0,-1.96*se.fit,+1.96*se.fit),type="l",lty=c(1,2,2),col=1,
xlab="Time since diagnosis", ylab="RERI"))
abline(h=0,lty=3)
RERI.hr <- function(object, newdata,
var1, val1=1,
var2, val2=1) {
exp1 <- function(data) {data[[var1]] <- data[[var1]]+val1; data}
exp2 <- function(data) {data[[var2]] <- data[[var2]]+val2; data}
h00 <- predict(object, newdata, type="haz")
h10 <- predict(object, newdata=exp1(newdata), type="haz")
h01 <- predict(object, newdata=exp2(newdata), type="haz")
h11 <- predict(object, newdata=exp1(exp2(newdata)), type="haz")
(h11-h10-h01+h00)/h00
}
RERI.hr(fit1,newdata=transform(newd,rectime=1000),var1="hormon",var2="x4.23",val2=TRUE)
predictnl(fit1,fun=RERI.hr,newdata=transform(newd,rectime=1000),var1="hormon",var2="x4.23",val2=TRUE)
## testing of relative survival
library(rstpm2)
ayear <- 365.24
brcancer2 <- transform(brcancer, age=80*ayear, sex="male", year=as.Date("1980-01-01"), time=1, recyear=rectime/ayear)
rate0 <- survexp(time~1,data=brcancer2,method="individual.h",scale=ayear)
(fit1 <- stpm2(Surv(recyear,censrec==1)~hormon,data=brcancer2,df=2,cure=T,bhazard=rate0))
head(predict(fit1,type.relsurv="excess"))
head(predict(fit1,type.relsurv="total"))
head(brcancer2)
ayear <- 365.24
timeVar <- substitute(times)
scale <- ayear
rmap <- substitute(list())
newdata <- data.frame(sex=c("male",rep("male",5)),age=ayear*60,year=2002,times=c(1,1:5))
survexp1 <- do.call(survexp, list(substitute(I(timeVar*scale)~1,list(timeVar=timeVar)),
ratetable=survexp.us,
scale=scale,
rmap=rmap,
cohort=FALSE,
data=newdata))
plot(fit1, newdata=data.frame(hormon=1,age=80,sex="male",year=1980))
## lines(fit1, newdata=data.frame(hormon=1,age=80,sex="male",year=1980))
## Bug report from Alessandro for 1.4.0
library(rstpm2)
data(kidney)
fitg = stpm2(Surv(time, status) ~ age + sex, cluster = kidney$id, data = kidney,
RandDist = "Gamma")
head(predict(fitg))
fitln = stpm2(Surv(time, status) ~ age + sex, cluster = kidney$id, data = kidney,
RandDist = "LogN")
head(predict(fitln))
fitln = stpm2(Surv(time, status) ~ age + sex, cluster = kidney$id, data = kidney, Z=~age-1,
RandDist = "LogN")
head(predict(fitln))
## test meanhr
library(rstpm2)
fit <- stpm2(Surv(rectime, censrec==1) ~ x4+x5, data = brcancer, df=3)
fit <- stpm2(Surv(rectime, censrec==1) ~ x4+x5, data = brcancer, df=3)
summary(fit)
eform(fit)
plot(fit, newdata=data.frame(hormon=0,x4=0,x5=0))
plot(fit, newdata=data.frame(hormon=0,x4=0,x5=0),type="hazard")
plot(fit, newdata=data.frame(hormon=0,x4=0,x5=0), type="hr", exposed=function(data) transform(data, x4=1))
plot(fit, newdata=transform(brcancer,x4=1), type="meanhr", exposed=function(data) transform(data, x4=2))
plot(fit, newdata=transform(brcancer,x4=1), type="meanhaz")
## test rmst
library(rstpm2)
fit <- stpm2(Surv(rectime, censrec==1) ~ hormon, data = brcancer, df=3)
plot(fit, newdata=data.frame(hormon=1))
predict(fit, newdata=data.frame(hormon=1,rectime=1000), type="rmst", se.fit=TRUE)
predict(fit, newdata=data.frame(hormon=0,rectime=1000), type="rmst", se.fit=TRUE)
library(devtools)
install.packages("bbmle")
devtools::install_github("mclements/rstpm2",ref="develop")
## 2017-06-21
## Verify: the choice of basis dimension (default: k=10) for penalized regression splines is not sensitive to estimates
## Adjusted by a constant coefficient (e.g. alpha=2) to correct potential overfitting by GCV for lambda
## alpha = 1.4 suggested by Kim and Gu (2004)
library(rstpm2)
## k = 7
pfit7 <- pstpm2(Surv(rectime, censrec==1) ~ hormon, data = brcancer, smooth.formula = ~ s(log(rectime), k=7), alpha=2)
plot(pfit7, newdata = data.frame(hormon=0), type="hazard")
## k = 27
pfit27 <- pstpm2(Surv(rectime, censrec==1) ~ hormon, data = brcancer, smooth.formula = ~ s(log(rectime), k=27), alpha=2)
plot(pfit27, newdata = data.frame(hormon=0), type="hazard")
## Estimated effective degree of freedom (EDF)
pfit7@edf ## 5.36
pfit27@edf ## 5.96
require(coxme)
## Fix error in code for gradli
library(rstpm2)
data(brcancer)
fit <- stpm2(Surv(rectime,censrec) ~ hormon,data=transform(brcancer,censrec=1))
fit <- stpm2(Surv(rectime,censrec==1) ~ hormon,data=brcancer,cure=TRUE)
fit <- stpm2(Surv(rectime,censrec==1) ~ hormon,data=brcancer)
plot(fit,newdata=data.frame(hormon=1),type="uncured",exposed=function(data) transform(data,rectime=2500))
X <- fit@args$X
XD <- fit@args$XD
args <- fit@args
beta.est <- coef(fit)
eta <- as.vector(X %*% beta.est)
etaD <- as.vector(XD %*% beta.est)
link <- switch(fit@args$link,PH=rstpm2:::link.PH,PO=rstpm2:::link.PO)
h <- link$h(eta,etaD) # - as.vector(predict(fit, type="haz")) ## Ok!
H <- link$H(eta) #- as.vector(predict(fit, type="cumhaz")) ## Ok!
gradh <- as.matrix(link$gradh(eta,etaD, args))
gradH <- as.matrix(link$gradH(eta, args))
gradli <- residuals(fit, type="gradli") ## n*npar
dim(gradli)
gradli2 <- gradH - ifelse(fit@args$event,1/h,0)*gradh
head(gradli + gradli2)
## Gamma frailty
require(rstpm2)
brcancer2 <- transform(brcancer, id=rep(1:(nrow(brcancer)/2),each=2))
brcancer2$hormon[1] <- NA
fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2, cluster=brcancer2$id)
summary(fit)
plot(fit,newdata=data.frame(hormon=1),type="margsurv")
fit2 <- pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2, cluster=brcancer2$id)
summary(fit2)
plot(fit2,newdata=data.frame(hormon=1),type="margsurv")
# Aranda-Ordaz link
refresh
require(rstpm2)
## PH
summary(fit <- stpm2(Surv(rectime,censrec==1)~1,data=brcancer,link="PH", df=3))
summary(fit <- stpm2(Surv(rectime,censrec==1)~1,data=brcancer,link="AO", df=3)) # Same: OK
summary(fit <- stpm2(Surv(rectime,censrec==1)~1,data=brcancer,link="PO", df=3))
summary(fit <- stpm2(Surv(rectime,censrec==1)~1,data=brcancer,link="AO", theta.AO=1, df=3)) # Same: OK
summary(fit <- pstpm2(Surv(rectime,censrec==1)~1,data=brcancer,link="AO", theta.AO=0.5))
refresh
require(rstpm2)
## PH
summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,link="PH", df=3))
predict(fit, newdata=transform(brcancer,rectime=1000),type="meansurv",keep.attributes=FALSE,se.fit=TRUE,use.gr=F)
predict(fit, newdata=transform(brcancer,rectime=1000),type="meansurv",keep.attributes=FALSE,se.fit=TRUE,use.gr=T)
plot(fit,newdata=transform(brcancer,hormon=1),type="meansurv",times=seq(10,1500,by=10))
plot(fit,newdata=transform(brcancer,hormon=2),type="meansurv",times=seq(10,1500,by=10),lty=2,add=TRUE)
newd <- merge(transform(brcancer,rectime=NULL), data.frame(rectime=c(500,1000)))
unlist(predict(fit,newdata=newd,type="af",exposed=function(data) transform(data,hormon=1),keep.attributes=FALSE,se.fit=TRUE) -
predict(fit,newdata=newd,type="af",exposed=function(data) transform(data,hormon=1),keep.attributes=FALSE,se.fit=TRUE,use.gr=FALSE))
system.time(plot(fit,type="af",exposed=function(data) transform(data,hormon=1),recent=TRUE))
system.time(plot(fit,type="af",exposed=function(data) transform(data,hormon=1),recent=FALSE))
plot(fit,newdata=NULL,type="meansurv",ci=F)
plot(fit,newdata=NULL,type="meansurvdiff",exposed=function(data) transform(data,hormon=1))
plot(fit,newdata=data.frame(hormon=1),type="surv",ci=F)
plot(fit,newdata=data.frame(hormon=1),type="fail",ci=T)
unlist(predict(fit,newdata=newd,type="meansurvdiff",exposed=function(data) transform(data,hormon=1),keep.attributes=FALSE,se.fit=TRUE) -
predict(fit,newdata=newd,type="meansurvdiff",exposed=function(data) transform(data,hormon=1),keep.attributes=FALSE,se.fit=TRUE,use.gr=FALSE))
unlist(predict(fit,newdata=newd,type="meansurv",keep.attributes=FALSE,se.fit=TRUE)-
predict(fit,newdata=newd,type="meansurv",keep.attributes=FALSE,se.fit=TRUE,use.gr=FALSE))
## comparison with AF
## Example 1: clustered data with frailty U
require(AF)
set.seed(12345)
expit <- function(x) 1 / (1 + exp( - x))
n <- 100
m <- 2
alpha <- 1.5
eta <- 1
phi <- 0.5
beta <- 1
id <- rep(1:n,each=m)
U <- rep(rgamma(n, shape = 1 / phi, scale = phi), each = m)
Z <- rnorm(n * m)
X <- rbinom(n * m, size = 1, prob = expit(Z))
## Reparametrize scale as in rweibull function
weibull.scale <- alpha / (U * exp(beta * X)) ^ (1 / eta)
t <- rweibull(n * m, shape = eta, scale = weibull.scale)
## Right censoring
cen <- runif(n * m, 0, 10)
delta <- as.numeric(t < cen)
t <- pmin(t, cen)
d <- data.frame(t, delta, X, Z, id)
require(rstpm2)
fit2 <- stpm2(formula = Surv(t, delta) ~ X + Z + X * Z, data = d, df=1, cluster=d$id, smooth.formula=~log(t))
predict(fit2, type="af", newdata=transform(d,t=1),exposed=function(data) transform(data, X=0), se.fit=TRUE)
plot(fit2, type="af", exposed=function(data) transform(data, X=0))
plot(fit2, type="meansurvdiff", exposed=function(data) transform(data, X=0))
plot(fit2, type="meansurv")
fit3 <- pstpm2(formula = Surv(t, delta) ~ X + Z + X * Z, data = d, df=1, cluster=d$id)
predict(fit3, type="af", newdata=transform(d,t=1),exposed=function(data) transform(data, X=0), se.fit=TRUE)
plot(fit3, type="meansurv")
predict(fit2, newdata=transform(d,t=1), type="meansurv")
## check analytical gradients for margsurv and marghaz
predict(fit2, newdata=data.frame(t=1,X=1,Z=1), type="margsurv", use.gr=TRUE, se.fit=TRUE)-predict(fit2, newdata=data.frame(t=1,X=1,Z=1), type="margsurv", use.gr=FALSE, se.fit=TRUE)
predict(fit2, newdata=data.frame(t=1,X=1,Z=1), type="marghaz", use.gr=TRUE, se.fit=TRUE)-predict(fit2, newdata=data.frame(t=1,X=1,Z=1), type="marghaz", use.gr=FALSE, se.fit=TRUE)
predict(fit2, newdata=data.frame(t=1,X=1,Z=1), type="hazard", use.gr=TRUE, se.fit=TRUE)-predict(fit2, newdata=data.frame(t=1,X=1,Z=1), type="hazard", use.gr=FALSE, se.fit=TRUE)
require(boot)
meansurv <- function(data,index) predict(fit2, newdata=transform(data[index,,drop=FALSE],t=1), type="meansurv")
meansurv(d,TRUE)
boot1 <- boot(d, meansurv, R=1000)
boot.ci(boot1)
require(rstpm2)
fit <- stpm2(formula = Surv(t, delta) ~ X + Z + X * Z, data = d, df=1)
diag(vcov(fit))
fit <- stpm2(formula = Surv(t, delta) ~ X + Z + X * Z, data = d, frailty=FALSE, cluster=d$id, df=1)
diag(vcov(fit))
fit <- stpm2(formula = Surv(t, delta) ~ X + Z + X * Z, data = d, cluster=d$id, df=1)
diag(vcov(fit))
##
fit <- stpm2(formula = Surv(t, delta) ~ X + Z + X * Z, data = d, cluster = d$id, df=1)
predict(fit,type="af",newdata=transform(d,t=1),exposed=function(data) transform(data,X=0),keep.attributes=FALSE,se.fit=TRUE)
fit <- stpm2(formula = Surv(t, delta) ~ X + Z + X * Z, data = d, cluster = d$id, df=4)
predict(fit,type="af",newdata=transform(d,t=1),exposed=function(data) transform(data,X=0),keep.attributes=FALSE,se.fit=TRUE)
fit <- stpm2(formula = Surv(t, delta) ~ X + Z + X * Z, data = d, cluster=d$id, df=1)
predict(fit,type="af",newdata=transform(d,t=1),exposed=function(data) transform(data,X=0),keep.attributes=FALSE,se.fit=TRUE)
## Fit a frailty object
library(stdReg)
fit <- stdReg::parfrailty(formula = Surv(t, delta) ~ X + Z + X * Z, data = d, clusterid = "id")
summary(fit)
## Estimate the attributable fraction from the fitted frailty model
time <- c(seq(from = 0.2, to = 1, by = 0.2))
time <- 1
## debug(AFfrailty)
library(AF)
AFfrailty_est <- AFparfrailty(object = fit, data = d, exposure = "X", times = time, clusterid = "id")
AFfrailty_est
##AF:::summary.AF(AFfrailty_est)
## tvc for Maarten Coemans
require(rstpm2)
brcancer <- transform(brcancer, x1c=x1-mean(x1))
summary(fit.tvc <- stpm2(Surv(rectime,censrec==1)~ hormon+x1c, data=brcancer, df=3, tvc=list(hormon=2,x1c=2)))
plot(fit.tvc,newdata=data.frame(hormon=0,x1c=-10),type="hr", var="hormon")
plot(fit.tvc,newdata=data.frame(hormon=0,x1c=+10),type="hr", var="hormon", add=TRUE,ci=FALSE,line.col=2)
## same model
summary(fit.tvc <- stpm2(Surv(rectime,censrec==1)~ hormon+x1c, data=brcancer,
smooth.formula = ~ns(log(rectime),df=3)+hormon:ns(log(rectime),df=2)+x1c:ns(log(rectime),df=2)))
## and again...
summary(fit.tvc <- stpm2(Surv(rectime,censrec==1)~ x1c, data=brcancer,
smooth.formula = ~ns(log(rectime),df=3)+hormon:ns(log(rectime),df=3,intercept=TRUE)+x1c:ns(log(rectime),df=2)))
## new model with time different time transformation for the TVCs
summary(fit.tvc <- stpm2(Surv(rectime,censrec==1)~ hormon+x1c, data=brcancer,
smooth.formula = ~ns(log(rectime),df=3)+hormon:ns(rectime,df=2)+x1c:ns(rectime,df=2)))
plot(fit.tvc,newdata=data.frame(hormon=0,x1c=-10),type="hr", var="hormon")
plot(fit.tvc,newdata=data.frame(hormon=0,x1c=+10),type="hr", var="hormon", add=TRUE,ci=FALSE,line.col=2)
##
## not including the main effect and no intercept is not the same
summary(fit.tvc <- stpm2(Surv(rectime,censrec==1)~ x1c, data=brcancer,
smooth.formula = ~ns(log(rectime),df=3)+hormon:ns(log(rectime),df=2)+x1c:ns(log(rectime),df=2)))
## Standardised survival
require(rstpm2)
plot.meansurv <- function(x, y=NULL, times=NULL, newdata=NULL, add=FALSE, ci=!add, rug=!add, recent=FALSE,
xlab=NULL, ylab="Mean survival", lty=1, line.col=1, ci.col="grey", ...) {
if (is.null(times)) stop("plot.meansurv: times argument should be specified")
if (is.null(newdata)) newdata <- x@data
times <- times[times !=0]
if (recent) {
newdata <- do.call("rbind",
lapply(times,
function(time) {
newdata[[x@timeVar]] <- newdata[[x@timeVar]]*0+time
newdata
}))
pred <- predict(x, newdata=newdata, type="meansurv", se.fit=ci) # requires recent version
pred <- if (ci) rbind(c(Estimate=1,lower=1,upper=1),pred) else c(1,pred)
} else {
pred <- lapply(times,
function(time) {
newdata[[x@timeVar]] <- newdata[[x@timeVar]]*0+time
predict(x, newdata=newdata, type="meansurv", se.fit=ci)
})
pred <- if (ci) rbind(c(Estimate=1,lower=1,upper=1),do.call("rbind", pred)) else c(1,unlist(pred))
}
times <- c(0,times)
if (is.null(xlab)) xlab <- deparse(x@timeExpr)
if (!add) matplot(times, pred, type="n", xlab=xlab, ylab=ylab, ...)
if (ci) {
polygon(c(times,rev(times)),c(pred$lower,rev(pred$upper)),col=ci.col,border=ci.col)
lines(times,pred$Estimate,col=line.col,lty=lty)
} else {
lines(times,pred,col=line.col,lty=lty)
}
if (rug) {
Y <- x@y
eventTimes <- Y[Y[,ncol(Y)]==1,ncol(Y)-1]
rug(eventTimes,col=line.col)
}
return(invisible(y))
}
brcancer <- transform(brcancer, x1c=x1-mean(x1))
summary(fit.tvc <- stpm2(Surv(rectime,censrec==1)~ hormon+x1c, data=brcancer, df=3, tvc=list(hormon=2,x1c=2)))
times <- seq(0,3000,by=100)
plot.meansurv(fit.tvc, newdata=transform(brcancer, hormon=1), times=times,
ylim=c(0.2,1))
plot.meansurv(fit.tvc, times=times, newdata=transform(brcancer, hormon=0), line.col=2, add=TRUE)
## Examples using ns() for covariates - this was buggy.
refresh
require(rstpm2)
summary(fit <- stpm2(Surv(rectime,censrec==1)~1,
smooth.formula=~ns(log(rectime),df=3)+ns(x1,df=3),
data=brcancer,link="PH"))
summary(fit <- stpm2(Surv(rectime,censrec==1)~ns(x1,df=3), df=3,data=brcancer,link="PH"))
summary(fit <- pstpm2(Surv(rectime,censrec==1)~ns(x1,df=3), data=brcancer,link="PH"))
grad <- function(f,x,eps=1e-5)
sapply(1:length(x), function(i) {
lower <- upper <- x
upper[i] <- x[i]+eps
lower[i] <- x[i]-eps
(f(upper)-f(lower))/2/eps
})
link <- function(S,theta=0.5) log((S^(-theta)-1)/theta)
S <- ilink <- function(eta,theta=0.5) exp(-log(theta*exp(eta)+1)/theta)
H <- function(eta,theta=0.5) -log(S(eta,theta))
h <- function(eta,etaD,theta=0.5) exp(eta)*etaD/(theta*exp(eta)+1)
gradH <- function(eta,X,theta=0.5) exp(eta)*X/(1+theta*exp(eta))
gradh <- function(eta,etaD,X,XD,theta=0.5) {
eta <- as.vector(eta)
etaD <- as.vector(etaD)
((theta*exp(2*eta)+exp(eta))*XD+exp(eta)*etaD*X) /
(theta*exp(eta)+1)^2
}
X <- cbind(1,1:2,1) # (constant, t, x)
XD <- cbind(0,1:2,0)
beta <- c(0.1, 0.2, 0.3)
eta <- as.vector(X %*% beta)
etaD <- as.vector(XD %*% beta)
S(eta)
H(eta)
h(eta,etaD) - grad(function(t) H(cbind(1,t,1) %*% beta), 1) # OK
gradH(eta,X) - grad(function(beta) H(X %*% beta), beta) # OK
gradh(eta,etaD,X,XD)
grad(function(beta) h(X %*% beta, XD %*% beta), beta)
ilink(link(.1))
link(ilink(.1))
require(abind)
X <- matrix(seq(0,1,length=5*10),nrow=10)
beta <- seq(0,1,length=5)
H <- exp(as.vector(X %*% beta))
dHdbeta <- X * H # row=indiv, col=beta
d2Hdbeta2 <- aperm(abind(lapply(1:ncol(X), function(k) X[,k] * X * H),along=3),c(2,3,1))
abind(lapply(1:nrow(X), function(i) (X[i,] %*% t(X[i,])) * H[i]),along=3) -
aperm(abind(lapply(1:ncol(X), function(k) X[,k] * X * H),along=3),c(2,3,1))
numder <- function(f,x,eps=1e-8) (f(x+eps)-f(x-eps))/2/eps
expit <- function(x) 1/(1+exp(-x))
numder(expit,2)
expit(2)*expit(-2)
numder(dnorm,2)
-dnorm(2)*2
require(mgcv)
d <- data.frame(x = seq(0,1,length=100), x2=rnorm(100), y = rnorm(100))
fit <- gam(y~s(x)+s(x2,by=x), data=d)
X <- predict(fit,d,type="lpmatrix")
X0 <- predict(fit,transform(d,x=0),type="lpmatrix")
Xstar <- X-X0
index0 <- rstpm2:::which.dim(Xstar)
lapply(fit$smooth, function(s) {
which((1:ncol(X) %in% index0)[s$first.para:s$last.para]) # index for S'
})
lapply(fit$smooth, function(s) {
range(which((1:ncol(X) %in% s$first.para:s$last.para)[index0]))
})
lapply(fit$smooth,"[[","S")
## outline: given a full index=1:n, a reduced index set index0 and a smoother with first.para, last.para and a square matrix S, return a revised first.para', last.para' and matrix S'
## For S':
refresh
require(rstpm2)
## additive
fit <- stpm2(Surv(rectime,censrec==1)~1,data=brcancer,link="AH",
smooth.formula=~ns(rectime,df=4)+hormon:ns(rectime,df=3), optimiser="NelderMead")
summary(fit)
fit2 <- stpm2(Surv(rectime,censrec==1)~1,data=brcancer,link="AH",
smooth.formula=~ns(rectime,df=4)+hormon:ns(rectime,df=3))
summary(fit2)
plot(fit2,newdata=data.frame(hormon=0),type="haz")
plot(fit2,newdata=data.frame(hormon=1),add=TRUE,lty=2,type="haz")
fit <- pstpm2(Surv(rectime,censrec==1)~1,data=brcancer,link="AH",
smooth.formula=~s(rectime)+s(rectime,by=hormon))
plot(fit,newdata=data.frame(hormon=0),type="haz")
plot(fit,newdata=data.frame(hormon=1),add=TRUE,lty=2,type="haz")
## test robust estimators from penalized models
## most spline coefficients become statistically significant
require(rstpm2)
summary(pstpm2(Surv(rectime/365,censrec==1)~hormon,data=brcancer,robust=FALSE))
summary(pstpm2(Surv(rectime/365,censrec==1)~hormon,data=brcancer,robust=TRUE))
## robust standard errors for clustered data
refresh
require(rstpm2)
brcancer2 <- transform(brcancer,
id=rep(1:(nrow(brcancer)/2),each=2))
fit <- stpm2(Surv(rectime,censrec==1)~1,data=brcancer)
summary(fit)
fit <- stpm2(Surv(rectime,censrec==1)~1,data=brcancer, cluster=brcancer2$id, robust=TRUE)
summary(fit)
##
require(rstpm2)
brcancer2 <- transform(brcancer, id=rep(1:(nrow(brcancer)/2),each=2))
fit <- stpm2(Surv(rectime,censrec==1)~1,data=brcancer, cluster=brcancer2$id)
summary(fit)
predict(fit,type="gradli")
## Stata estimated coef for hormon
## PH: -.3614357
## PO: -.474102
## Probit: -.2823338
system.time(print( stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer, stata=TRUE)))
system.time(print(pfit <- pstpm2(Surv(rectime,censrec==1)~hormon,smooth.formula=~s(log(rectime))+s(x1),data=brcancer)))
##
system.time(print( stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,type="PO")))
system.time(print(pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,type="PO")))
##
system.time(print( stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,type="probit")))
system.time(print(pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,type="probit"))) # slow
summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,smooth.formula=~nsx(log(rectime), df=4, stata.stpm2.compatible = TRUE)))
if (FALSE) {
debug(pstpm2)
pfit <- pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,sp=1)
## towards the end of the pstpm2 function...
sum(diag(solve(optimHess(coef(mle2),negllsp,sp=1)) %*% optimHess(coef(mle2),negll0sp,sp=1)))
sum(diag(solve(optimHess(coef(mle2),negllsp,sp=fit$sp)) %*% optimHess(coef(mle2),negll0sp,sp=fit$sp)))
negllsp(coef(mle2),sp=1)
negll0sp(coef(mle2),sp=1)
}
update.list <- function(list,...) {
args <- list(...)
for (name in names(args))
list[[name]] <- args[[name]]
list
}
## right censored
## Stata estimated coef for hormon (PH): -.3614357
refresh
require(rstpm2)
summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,
smooth.formula=~nsx(log(rectime),df=3,stata=TRUE),trace=0))
summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,
smooth.formula=~nsx(log(rectime),df=3,stata=TRUE),trace=0,optimiser="NelderMead"))
summary(fit2 <- pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer))
summary(fit2 <- pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,optimiser="NelderMead"))
## delayed entry
## Stata estimated coef for hormon (PH): -1.162504
library(rstpm2)
brcancer2 <- transform(brcancer,startTime=ifelse(hormon==0,rectime/2,0))
## brcancer2 <- transform(brcancer,startTime=0.1)
##debug(rstpm2:::meat.stpm2)
summary(fit <- stpm2(Surv(startTime,rectime,censrec==1)~hormon,data=brcancer2,
smooth.formula=~nsx(log(rectime),df=3,stata=TRUE)))
summary(fit2 <- pstpm2(Surv(startTime,rectime,censrec==1)~hormon,data=brcancer2,control=list(optimiser="NelderMead"))) # OK
summary(fit2 <- pstpm2(Surv(startTime,rectime,censrec==1)~hormon,data=brcancer2))
plot(fit,newdata=data.frame(hormon=1))
lines(fit2,newdata=data.frame(hormon=1),lty=2)
head(predict(fit)) # OK
head(predict(fit,se.fit=TRUE))
## delayed entry and tvc (problems?)
summary(fit <- stpm2(Surv(startTime,rectime,censrec==1)~hormon,data=brcancer2,
smooth.formula=~nsx(rectime,df=3)+hormon:nsx(rectime,df=3,stata=TRUE)))
head(predict(fit,se.fit=TRUE))
pstpm2(Surv(startTime,rectime,censrec==1)~hormon,data=brcancer2)
## left truncated with clusters
library(rstpm2)
brcancer2 <- transform(brcancer,
startTime=ifelse(hormon==0,rectime/2,0),
id=rep(1:(nrow(brcancer)/2),each=2))
##debug(stpm2)
summary(fit <- stpm2(Surv(startTime,rectime,censrec==1)~hormon,data=brcancer2,
cluster=brcancer2$id, control=list(optimiser="NelderMead"),recurrent=TRUE,
smooth.formula=~nsx(log(rectime),df=3,stata=TRUE)))
summary(fit0 <- stpm2(Surv(startTime,rectime,censrec==1)~hormon,data=brcancer2,
control=list(optimiser="NelderMead"),
smooth.formula=~nsx(log(rectime),df=3,stata=TRUE)))
require(foreign)
require(rstpm2)
stmixed <- read.dta("http://fmwww.bc.edu/repec/bocode/s/stmixed_example2.dta")
stmixed2 <- transform(stmixed, start = ifelse(treat,stime/2,0))
summary(r2 <- stpm2(Surv(stime,event)~treat,data=stmixed2,cluster=stmixed$trial,RandDist="LogN",df=3,Z=~treat-1,adaptive=TRUE,optimiser="NelderMead"))
summary(r2 <- stpm2(Surv(stime,event)~treat,data=stmixed2,cluster=stmixed$trial,RandDist="LogN",df=3,Z=~treat-1,adaptive=TRUE))
##
library(foreign)
library(rstpm2)
stmixed <- read.dta("http://fmwww.bc.edu/repec/bocode/s/stmixed_example2.dta")
summary(r <- stpm2(Surv(stime,event)~treat,data=stmixed,cluster=stmixed$trial,RandDist="LogN",df=3,Z=~treat-1))
## non-adaptive
system.time(print(summary(r2 <- stpm2(Surv(stime,event)~treat,data=stmixed2,cluster=stmixed$trial,RandDist="LogN",df=3,Z=~treat,adaptive=FALSE,nodes=20,optimiser="NelderMead")))) # slow and gradients not close to zero
system.time(print(summary(r2 <- stpm2(Surv(stime,event)~treat,data=stmixed2,cluster=stmixed$trial,RandDist="LogN",df=3,Z=~treat,nodes=20,adaptive=FALSE)))) # gradients close to zero
## random intercept and random slope with 20 nodes
summary(r2 <- stpm2(Surv(stime,event)~treat,data=stmixed2,cluster=stmixed$trial,RandDist="LogN",df=3,Z=~treat,nodes=20,adaptive=FALSE))
summary(r2 <- stpm2(Surv(stime,event)~treat,data=stmixed2,cluster=stmixed$trial,RandDist="LogN",df=3,Z=~treat,adaptive=FALSE,nodes=20)) # gradients close to zero
## Simple examples with no random effects and with a random intercept (check: deviances)
summary(r2 <- stpm2(Surv(stime,event)~treat,data=stmixed,df=3))
summary(r2 <- stpm2(Surv(stime,event)~treat,data=stmixed,cluster=stmixed$trial,RandDist="LogN",df=3,Z=~treat-1))
## check modes and sqrttau
args <- r2@args
args$return_type <- "modes"
.Call("model_output", args, package="rstpm2")
args$return_type <- "variances" # fudge
.Call("model_output", args, package="rstpm2")
## check gradients
args <- r2@args
args$return_type <- "gradient"
.Call("model_output", args, package="rstpm2")
fdgrad <- function(obj,eps=1e-6) {
args <- obj@args
args$return_type <- "objective"
sapply(1:length(args$init), function(i) {
largs <- args
largs$init[i] <- args$init[i]+eps
f1 <- .Call("model_output", largs, package="rstpm2")
largs$init[i] <- args$init[i]-eps
f2 <- .Call("model_output", largs, package="rstpm2")
data.frame(f1,f2,gradient=(f1-f2)/2.0/eps)
})
}
fdgrad(r2,1e-3)
## random intercept and random slope
summary(r2 <- stpm2(Surv(stime,event)~treat,data=stmixed2,cluster=stmixed$trial,RandDist="LogN",df=3,Z=~treat))
summary(stpm2(Surv(start,stime,event)~treat,data=stmixed2))
summary(r2 <- stpm2(Surv(stime,event)~treat,data=stmixed2,cluster=stmixed$trial,RandDist="LogN"))
##
summary(r2 <- pstpm2(Surv(stime,event)~treat,data=stmixed2,cluster=stmixed$trial,RandDist="LogN"))
##
summary(r2 <- stpm2(Surv(stime,event)~treat+factor(trial),data=stmixed2,cluster=stmixed$trial,RandDist="LogN",Z=~treat-1))
##
summary(r2 <- pstpm2(Surv(stime,event)~treat+factor(trial),data=stmixed2,cluster=stmixed$trial,RandDist="LogN",Z=~treat-1))
## gradients for RE parameters
require(expm)
link.PH <- list(link=function(S) log(-log(S)),
ilink=function(eta) exp(-exp(eta)),
h=function(eta,etaD) etaD*exp(eta),
H=function(eta) exp(eta),
gradh=function(eta,etaD,obj) obj$XD*exp(eta)+obj$X*etaD*exp(eta),
gradH=function(eta,obj) obj$X*exp(eta))
link.AH <- list(link=function(S) -log(S),
ilink=function(eta) exp(-eta),
h=function(eta,etaD) etaD,
H=function(eta) eta,
gradh=function(eta,etaD,obj) obj$XD,
gradH=function(eta,obj) obj$X)
corrtrans <- function(x) (1-exp(-x)) / (1+exp(-x))
l <- function(gamma=c(log(.3),log(.4),0.5)) {
## initially assume an additive model
H <- function(eta) eta
h <- function(eta,etaD) etaD
delta <- 1
eta <- 1
etaD <- 2
Z <- c(1,2)
d <- c(1,1)
mode <- c(1,2)
## define sqrtmat
var <- exp(gamma[c(1,2)])
rho <- corrtrans(gamma[3])
cov <- rho*sqrt(var[1]*var[2])
Sigma <- matrix(c(var[1],cov,cov,var[2]),2,2)
## SqrtSigma <- chol(Sigma)
SqrtSigma <- sqrtm(as(Sigma,"symmetricMatrix"))
b <- mode + SqrtSigma %*% d
Zb <- sum(Z*b)
l <- delta*log(h(eta+Zb,etaD))-H(eta+Zb)
l
}
g <- c(log(.3),log(.4),0.5)
## gradient wrt g using finite differences
sapply(1:3, function(i,eps=1e-6) {
g[i] <- g[i]+eps
val <- l(g)
g[i] <- g[i]-2*eps
(val - l(g))/2/eps
})
bf <- function(gamma=c(log(.3),log(.4),0.5)) {
Z <- c(1,2)
d <- c(1,1)
mode <- c(1,2)
## define sqrtmat
var <- exp(gamma[c(1,2)])
rho <- corrtrans(gamma[3])
cov <- rho*sqrt(var[1]*var[2])
Sigma <- matrix(c(var[1],cov,cov,var[2]),2,2)
## SqrtSigma <- chol(Sigma)
SqrtSigma <- sqrtm(as(Sigma,"symmetricMatrix"))
b <- mode + SqrtSigma %*% d
b
}
## gradient for b wrt g using finite differences
(gradbwrtg <- sapply(1:3, function(i,eps=1e-6) {
g[i] <- g[i]+eps
val <- bf(g)
g[i] <- g[i]-2*eps
(val - bf(g))/2/eps
}))
var <- exp(g[c(1,2)])
rho <- corrtrans(g[3])
cov <- rho*sqrt(var[1]*var[2])
## wrt g[1]
matrix(c(g[1]*var[1],cov*g[1]/2,cov*g[1]/2,0),2,2)
##
A=1; B=0.5; D=2
M=matrix(c(A,B,B,D),2,2)
tau <- A+D
delta <- A*D-B*B
s <- sqrt(delta)
t <- sqrt(tau+2*s)
sqrtM <- matrix(c(A+s,B,B,D+s),2,2)/t
sqrtM %*% sqrtM
lb <- function(b) {
## initially assume an additive model
H <- function(eta) eta
h <- function(eta,etaD) etaD
delta <- 1
eta <- 1
etaD <- 2
Z <- c(1,2)
d <- c(1,1)
mode <- c(1,2)
Zb <- sum(Z*b)
l <- delta*log(h(eta+Zb,etaD))-H(eta+Zb)
l
}
## gradient wrt b using finite differences
gradlwrtb <- sapply(1:2, function(i,eps=1e-6) {
b[i] <- b[i]+eps
val <- lb(b)
b[i] <- b[i]-2*eps
(val - lb(b))/2/eps
})
t(gradbwrtg) %*% gradlwrtb
## gradient for a multivariate normal
require(mvtnorm)
fdgrad <- function(f,x, ..., eps=1.0e-6)
sapply(1:length(x), function(i) {
e <- rep(0,length(x))
e[i] <- 1
(f(x+e*eps, ...)-f(x-e*eps, ...))/2/eps
})
## hess(i,i) = (-f2 +16.0*f1 - 30.0*f0 + 16.0*fm1 -fm2)/(12.0*hi*hi);
fdhessian <- function(f,x, ..., eps=1.0e-5)
sapply(1:length(x), function(i) {
ei <- rep(0,length(x))
ei[i] <- 1
sapply(1:length(x), function(j) {
ej <- rep(0,length(x))
ej[j] <- 1
if (i==j) (-f(x+2*ei*eps, ...)+16*f(x+ei*eps, ...)-30*f(x,...)+16*f(x-ei*eps, ...)-f(x-2*ei*eps, ...))/12/eps/eps else (f(x+eps*ei+eps*ej)-f(x+eps*ei-eps*ej)-f(x-eps*ei+eps*ej)+f(x-eps*ei-eps*ej))/4/eps/eps
})
})
-fdgrad(mvtnorm::dmvnorm, c(1,1), c(0,0), Sigma <- matrix(c(1,.5,.5,1),2), log=TRUE)
## fdhessian(mvtnorm::dmvnorm, c(1,1), c(0,0), Sigma <- matrix(c(1,.5,.5,1),2))
as.vector(solve(Sigma) %*% c(1,1))
summary(fit <- stpm2(Surv(start,stime,event)~treat,data=stmixed2,cluster=stmixed$trial,optimiser="NelderMead",recurrent=TRUE))
summary(fit <- stpm2(Surv(start,stime,event)~treat,data=stmixed2,cluster=stmixed$trial,recurrent=TRUE))
summary(fit <- stpm2(Surv(start,stime,event)~treat,data=stmixed2,cluster=stmixed$trial,RandDist="LogN",
optimiser="NelderMead", recurrent=TRUE))
summary(r2 <- stpm2(Surv(start,stime,event)~treat,data=stmixed2,cluster=stmixed$trial,RandDist="LogN",
recurrent=TRUE))
summary(fit <- stpm2(Surv(start,stime,event)~treat,data=stmixed2,cluster=stmixed$trial,optimiser="NelderMead"))
summary(fit <- stpm2(Surv(start,stime,event)~treat,data=stmixed2,cluster=stmixed$trial))
summary(fit <- stpm2(Surv(start,stime,event)~treat,data=stmixed2,cluster=stmixed$trial,RandDist="LogN",
optimiser="NelderMead"))
summary(r2 <- stpm2(Surv(start,stime,event)~treat,data=stmixed2,cluster=stmixed$trial,RandDist="LogN"))
summary(r <- stpm2(Surv(start,stime,event)~treat,data=stmixed2))
## check gradients
args <- fit@args
args$return_type <- "gradient"
.Call("model_output", args, package="rstpm2")
fdgrad <- function(obj,eps=1e-6) {
args <- obj@args
args$return_type <- "objective"
sapply(1:length(args$init), function(i) {
largs <- args
largs$init[i] <- args$init[i]+eps
f1 <- .Call("model_output", largs, package="rstpm2")
largs$init[i] <- args$init[i]-eps
f2 <- .Call("model_output", largs, package="rstpm2")
data.frame(f1,f2,gradient=(f1-f2)/2.0/eps)
})
}
fdgrad(fit,1e-3)
require(rstpm2)
require(mgcv)
x=seq(0,1,length=5001)
set.seed(12345)
y=rnorm(length(x),sin(2*pi*x))
i <- x>0.65
d=data.frame(x=x[i],y=y[i])
fit <- gam(y~s(x),data=d)
## plot(fit)
plot(x,predict(fit,newdata=data.frame(x=x)),type="l")
plot(x,y)
## weighted estimates
refresh
require(rstpm2)
## unequal weights
brcancer2 <- transform(brcancer,w=ifelse(hormon==0,10,1))
## unweighted
summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2,
smooth.formula=~nsx(log(rectime),df=3,stata=TRUE)))
## weighted estimates
## stpm2
summary(stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2,weights=w,
smooth.formula=~nsx(log(rectime),df=3,stata=TRUE)))
## stpm2 robust
summary(stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2,weights=w,robust=TRUE,
smooth.formula=~nsx(log(rectime),df=3,stata=TRUE)))
summary(pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2))
## pstpm2
summary(pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2,weights=w))
summary(pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2,weights=w,robust=TRUE))
##
## equal weights
brcancer2 <- transform(brcancer,w=4)
## unweighted
summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2,
smooth.formula=~nsx(log(rectime),df=3,stata=TRUE)))
## weighted estimates
## stpm2
summary(stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2,weights=w,
smooth.formula=~nsx(log(rectime),df=3,stata=TRUE)))
summary(stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2,weights=w,robust=TRUE,
smooth.formula=~nsx(log(rectime),df=3,stata=TRUE)))
## pstpm2
summary(pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2))
summary(pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2,weights=w))
summary(pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2,weights=w,robust=TRUE))
refresh
require(rstpm2)
brcancer2 <- transform(brcancer,w=ifelse(hormon==0,10,1))
##debug(rstpm2:::meat.stpm2)
summary(fit <- pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2,weights=w,robust=TRUE,
smooth.formula=~nsx(log(rectime),df=3,stata=TRUE)))
summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer2,
logH.formula=~nsx(log(rectime),df=3,stata=TRUE)))
## code for the SAS PROC ICPHREG examples
read.textConnection <- function(text, ...) {
conn <- textConnection(text)
on.exit(close(conn))
read.table(conn, ...)
}
hiv <- read.textConnection("0 16 0 0 0 1
15 26 0 0 0 1
12 26 0 0 0 1
17 26 0 0 0 1
13 26 0 0 0 1
0 24 0 0 1 0
6 26 0 1 1 0
0 15 0 1 1 0
14 26 0 1 1 0
12 26 0 1 1 0
13 26 0 1 0 1
12 26 0 1 1 0
12 26 0 1 1 0
0 18 0 1 0 1
0 14 0 1 0 1
0 17 0 1 1 0
0 15 0 1 1 0
3 26 1 0 0 1
4 26 1 0 0 1
1 11 1 0 0 1
13 19 1 0 0 1
0 6 1 0 0 1
0 11 1 1 0 0
6 26 1 1 0 0
0 6 1 1 0 0
2 12 1 1 0 0
1 17 1 1 1 0
0 14 1 1 0 0
0 25 1 1 0 1
2 11 1 1 0 0
0 14 1 1 0 0")
names(hiv) <- c("Left","Right","Stage","Dose","CdLow","CdHigh")
##hiv <- transform(hiv, Left=pmax(1e-5,Left))
hiv <- transform(hiv,Event = ifelse(Left==0,2,
ifelse(Right>=26,0,
3)))
hiv2 <- transform(hiv,
Left = ifelse(Event==2,Right,
ifelse(Event==0,Left,
Left)),
Right = ifelse(Event==2,NA,
ifelse(Event==0,NA,
Right)))
library(rstpm2)
summary(stpm2(Surv(Left,Right,Event,type="interval")~Stage, data=hiv2, df=2))[2]
survreg(Surv(Left, Right, Event, type = "interval")~Stage, data=hiv2,dist="exponential")
library(rms)
psm(Surv(Left, Right, Event, type = "interval")~Stage, data=hiv2,dist="exponential")
##
library(rstpm2)
my.brcancer = brcancer
my.brcancer$left = my.brcancer$rectime
my.brcancer$right = ifelse(my.brcancer$censrec==1, my.brcancer$rectime, Inf)
test.stpm2.C = rstpm2::stpm2(Surv(left, right, censrec, type = "interval")~hormon,
data=my.brcancer,df=3)
## additive model
summary(fit <- stpm2(Surv(startTime,rectime,censrec==1)~hormon,data=brcancer2,
logH.formula=~nsx(rectime,df=3),
tvc.formula=~hormon:nsx(rectime,df=3,stata=TRUE)))
require(foreign)
require(rstpm2)
stmixed <- read.dta("http://fmwww.bc.edu/repec/bocode/s/stmixed_example2.dta")
system.time(r <- stpm2(Surv(stime,event)~treat,data=stmixed,cluster=stmixed$trial))
system.time(r <- stpm2(Surv(stime,event)~treat,data=stmixed,cluster=stmixed$trial,RandDist="LogN",
nodes=20))
summary(r)
require(mexhaz)
system.time(mix <-
mexhaz(formula=Surv(stime,event)~treat, data=stmixed, base="exp.bs",degree=3,
random="trial", verbose=0))
## Frailty model
require(rstpm2)
require(frailtypack)
data(dataAdditive)
##debug(pstpm2)
system.time(mod2n <- pstpm2(Surv(t1,t2,event)~var1,
data=dataAdditive,
RandDist="LogN",
##optimiser="NelderMead",
smooth.formula=~s(log(t2)),
sp.init=0.07723242,
adaptive=TRUE,
cluster=dataAdditive$group, nodes=10, trace=0))
summary(mod2n)
localargs <- mod2n@args
localargs$init <- mod2n@args$init*1.1
localargs$adaptive=TRUE
localargs$return_type <- "gradient"
.Call("model_output", localargs, package="rstpm2")
fdgrad <- function(obj,eps=1e-6) {
args <- obj@args
args$init <- args$init*1.1
sapply(1:length(args$init), function(i) {
args$return_type <- "objective"
args$init[i] <- args$init[i]+eps
f1 <- .Call("model_output", args, package="rstpm2")
args$init[i] <- args$init[i]-2*eps
f2 <- .Call("model_output", args, package="rstpm2")
(f1-f2)/2/eps
})
}
fdgrad(mod2n)
## OK for adaptive=FALSE
localargs <- mod2n@args
localargs$return_type <- "variances"
.Call("model_output", localargs, package="rstpm2")
localargs$return_type <- "modes"
.Call("model_output", localargs, package="rstpm2")
system.time(mod2nb <- stpm2(Surv(t1,t2,event)~var1,
data=dataAdditive,
RandDist="LogN",
logH.formula=~ns(log(t2),df=7),
cluster=dataAdditive$group, nodes=20))
system.time(mod2g <- pstpm2(Surv(t1,t2,event)~var1,
data=dataAdditive,
RandDist="Gamma",
smooth.formula=~s(log(t2)),
cluster=dataAdditive$group))
mod1 <- frailtyPenal(Surv(t1,t2,event)~cluster(group)+var1,data=dataAdditive,
n.knots=8,kappa=0.1,cross.validation=TRUE)
mod1n <- frailtyPenal(Surv(t1,t2,event)~cluster(group)+var1,data=dataAdditive,
n.knots=8,kappa=0.1,cross.validation=TRUE, RandDist="LogN")
system.time(mod2 <- stpm2(Surv(t1,t2,event)~var1, # Gamma
data=dataAdditive,
logH.formula=~ns(t2,df=7),
cluster=dataAdditive$group))
system.time(coxph1 <- coxph(Surv(t1,t2,event)~var1+frailty(group,distribution="gaussian"),
data=dataAdditive))
summary(coxph1)
system.time(mod2n <- stpm2(Surv(t1,t2,event)~var1,
data=dataAdditive,
RandDist="LogN",
optimiser="NelderMead",
logH.formula=~ns(log(t2),df=7),
cluster=dataAdditive$group, nodes=20))
system.time(mod2nb <- stpm2(Surv(t1,t2,event)~var1,
data=dataAdditive,
RandDist="LogN",
logH.formula=~ns(log(t2),df=7),
cluster=dataAdditive$group, nodes=20))
system.time(mod3 <- pstpm2(Surv(t1,t2,event)~var1,
data=dataAdditive,
RandDist="LogN",
criterion="BIC",
smooth.formula=~s(log(t2)),
cluster=dataAdditive$group, nodes=20))
system.time(mod3 <- coxph(Surv(t1,t2,event)~frailty(group,distribution="gamma")+var1,data=dataAdditive))
summary(mod2)
coef2 <- coef(summary(mod2))
theta <- exp(coef2[nrow(coef2),1])
se.logtheta <- coef2[nrow(coef2),2]
se.theta <- theta*se.logtheta
test.statistic <- 1/se.logtheta
pchisq(test.statistic,df=1,lower.tail=FALSE)/2
library(rstpm2)
library(ICE)
data(ICHemophiliac)
ICHemophiliac2 <- transform(as.data.frame(ICHemophiliac),event=3)
## fit1 <- stpm2(Surv(left,right,event,type="interval")~1,data=ICHemophiliac2,df=3)
fit1 <- pstpm2(Surv(left,right,event,type="interval")~1,data=ICHemophiliac2,
smooth.formula=~s(left,k=7))
estimate <- ickde(ICHemophiliac, m=200, h=0.9)
plot(estimate, type="l", ylim=c(0,0.20))
tt <- seq(0,20,length=301)[-1]
lines(tt,predict(fit1,newdata=data.frame(left=tt),type="density"),col="blue")
## reg1 <- survreg(Surv(left,right,event,type="interval")~1,data=ICHemophiliac2)
## weibullShape <- 1/reg1$scale
## ## weibullScale <- exp(predict(reg1,type="lp"))
## weibullScale <- predict(reg1);
## tt <- seq(0,20,length=301)
## estimate <- ickde(ICHemophiliac, m=200, h=0.9)
## plot(estimate, type="l", ylim=c(0,0.15))
## lines(tt,dweibull(tt,weibullShape,weibullScale),lty=2)
library(rstpm2)
library(survival)
data(veteran)
## Re-define variables
veteran <- dplyr::mutate(veteran,
squamous = ifelse(celltype=="squamous",1,0),
smallcell = ifelse(celltype=="smallcell",1,0),
adeno = ifelse(celltype=="adeno",1,0),
large = ifelse(celltype=="large",1,0),
prior.ty = ifelse(prior==0,0,1),
trt = ifelse(trt==2,1,0),
high = ifelse(karno > 50,1,0))
lung<-subset(veteran, prior==0) ## patients with no prior therapy
## Why no optimal smoothing parameters?? divergence with version 1.3.3
pfit <-pstpm2(Surv(time,status==1) ~ adeno + smallcell + squamous,
smooth.formula = ~s(log(time)) + s(karno), data=lung, link.type="PO", trace = 1)
## two-dimensional smoothers
x1 <- x2 <- seq(0,1,length=11)
dat <- expand.grid(x1=x1,x2=x2)
dat$y <- rnorm(nrow(dat))
require(mgcv)
fit <- gam(y~s(x1,x2),data=dat)
fit$smooth
system.time(print(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,type="probit")))
system.time(print(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,type="probit",use.rcpp=FALSE)))
system.time(print(fit2 <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer, type="PO")))
system.time(print(fit2 <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer, type="PO",use.rcpp=FALSE)))
system.time(print(stpm2Gen(Surv(rectime,censrec==1)~hormon,data=brcancer)))
system.time(print(stpm2Gen(Surv(rectime,censrec==1)~hormon,data=brcancer, use.rcpp=FALSE)))
head(predict(fit,se.fit=TRUE))
head(predict(fit,type="haz",se.fit=TRUE))
brcancer <- brcancer[rep(1:nrow(brcancer),each=500),]
system.time(print(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer))) # faster than Stata!
system.time(print(pfit <- pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer)))
plot(pfit,newdata=data.frame(hormon=0))
refresh
require(rstpm2)
data(brcancer)
system.time(print(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,
tvc=list(hormon=3))))
system.time(print(pfit <- pstpm2(Surv(rectime,censrec==1)~1,data=brcancer,sp.init=c(0.0001,0.0001),
tvc.formula=~s(log(rectime),by=hormon))))
print(pstpm2(Surv(rectime,censrec==1)~1,data=brcancer,init=coef(pfit)*100,
tvc.formula=~s(log(rectime),by=hormon)))
summary(pfit)
plot(pfit,newdata=data.frame(hormon=0))
plot(pfit,newdata=data.frame(hormon=1),add=TRUE)
plot(pfit,newdata=data.frame(hormon=0),type="haz")
plot(pfit,newdata=data.frame(hormon=1),type="haz",add=TRUE)
pfit <- pstpm2(Surv(rectime/365,censrec==1)~1,data=brcancer) # OK
plot(pfit,newdata=data.frame(hormon=0))
system.time(print(pfit <- pstpm2(Surv(rectime/365,censrec==1)~1,data=brcancer,
tvc.formula=~s(log(rectime/365),by=hormon))))
plot(pfit,newdata=data.frame(hormon=0)) # OK
times <- seq(500,2000,by=500)
meansurv1 <- t(sapply(times,function(time) predict(pfit,transform(brcancer,rectime=time,hormon=1),type="meansurv",se.fit=TRUE)))
meansurv0 <- t(sapply(times,function(time) predict(pfit,transform(brcancer,rectime=time,hormon=0),type="meansurv",se.fit=TRUE)))
matplot(times,meansurv1,type="l",lty=c(1,2,2),col=1)
matlines(times,meansurv0,type="l",lty=c(1,2,2),col=2)
meansurvdiff1 <- t(sapply(times,function(time) predict(pfit,transform(brcancer,rectime=time,hormon=0),type="meansurvdiff",var="hormon",se.fit=TRUE)))
matplot(times,meansurvdiff1,type="l",lty=c(1,2,2),col=1)
system.time(print(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,control=list(parscale=100,reltol=1e-10),use.rcpp=FALSE)))
summary(fit)
system.time(print(fit2 <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,control=list(parscale=10000.0),reltol=1e-10,init=0.0001*coef(fit))))
summary(fit2)
plot(fit2,newdata=data.frame(hormon=1))
brcancerN <- brcancer[rep(1:nrow(brcancer),each=100),]
system.time(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancerN,use.rcpp=FALSE,
control=list(parscale=0.1,reltol=1e-10)))
summary(fit)
system.time(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancerN,use.rcpp=TRUE))
summary(fit)
###### penalised likelihood
## environment(pstpm2) <- environment(rstpm2::pstpm2)
## require(rstpm2)
try(detach("package:rstpm2",unload=TRUE))
## source("/home/MEB/marcle/src/R/rstpm2/R/pm2-3.R")
refresh
require(rstpm2)
data(brcancer)
brcancer$recyear <- brcancer$rectime/365
system.time(fit0 <- stpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,df=5))
system.time(pfit0 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,sp.init=1))
system.time(pfit0.1 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
smooth.formula=~s(log(recyear),k=15),sp.init=10,alpha=2))
system.time(pfit1.1 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
smooth.formula=~s(log(recyear)),sp.init=10,criterion="BIC"))
system.time(pfit2 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
smooth.formula=~s(recyear),sp.init=10))
plot(pfit0,newdata=data.frame(hormon=1),line.col="red",type="hazard")
plot(pfit0.1,newdata=data.frame(hormon=1),line.col="blue",add=TRUE,type="hazard")
plot(pfit1.1,newdata=data.frame(hormon=1),line.col="orange",add=TRUE,type="hazard")
plot(fit0,newdata=data.frame(hormon=1),line.col="green",type="hazard",add=TRUE)
plot(pfit2,newdata=data.frame(hormon=1),line.col="black",type="hazard",add=TRUE)
plot(pfit0,newdata=data.frame(hormon=1),line.col="red")
plot(pfit0.1,newdata=data.frame(hormon=1),line.col="blue",add=TRUE)
plot(pfit1.1,newdata=data.frame(hormon=1),line.col="pink",add=TRUE)
plot(fit0,newdata=data.frame(hormon=1),line.col="green",add=TRUE)
plot(pfit2,newdata=data.frame(hormon=1),line.col="black",add=TRUE)
system.time(pfit0.check <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer, sp=pfit0@sp, use.rcpp=FALSE))
system.time(pfit0.check2 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer, sp=pfit0@sp))
summary(pfit0)
summary(pfit0.check)
summary(pfit0.check2)
system.time(pfit0 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(log(recyear),k=30),sp.init=1))
system.time(pfit0.1 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(log(recyear),k=30),sp.init=1,criterion="BIC"))
plot(pfit0,newdata=data.frame(hormon=1),line.col="red",type="hazard")
plot(pfit0.1,newdata=data.frame(hormon=1),line.col="blue",add=TRUE,type="hazard")
system.time(pfit1 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(log(recyear),k=30),sp=10,pen="h",
smoother.parameters=list("log(recyear)"=list(var="recyear",
inverse=exp,
transform=log))))
plot(pfit1,newdata=data.frame(hormon=1),line.col="green",add=TRUE,type="hazard")
system.time(pfit1 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(log(recyear),k=30),sp=10,criterion="GCV"))
system.time(pfit2 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(log(recyear),k=20),sp=10,criterion="BIC"))
plot(pfit1,newdata=data.frame(hormon=1),type="hazard",ylim=c(0,0.25))
plot(pfit2,newdata=data.frame(hormon=1),add=TRUE,line.col="blue",type="hazard")
system.time(pfit1 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(log(recyear),k=30),sp=1,criterion="GCV"))
system.time(pfit2 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(recyear,k=20),sp=1,use.rcpp=F,penalty="h"))
system.time(pfit2 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(recyear,k=20),sp=1,penalty="h"))
plot(pfit2,newdata=data.frame(hormon=1),type="hazard")
system.time(pfit2 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(recyear,k=30),sp=1,use.rcp=FALSE))
plot(pfit1,newdata=data.frame(hormon=1),type="hazard")
plot(pfit2,newdata=data.frame(hormon=1),type="hazard")
system.time(pfit2.0 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(recyear,k=30),sp=0.055,penalty="h",cr="GCV"))
system.time(pfit2 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(recyear,k=30),sp=0.055,use.rcp=FALSE,penalty="h"))
rstpm2:::gcv(pfit2)
plot(pfit2,newdata=data.frame(hormon=1),line.col="red",add=TRUE,type="hazard")
plot(pfit2.0,newdata=data.frame(hormon=1),line.col="green",add=TRUE,type="hazard")
require(frailtypack)
fpack1 <- frailtyPenal(Surv(recyear,censrec==1)~hormon, data=brcancer, cross.validation=TRUE, n.knots=10, kappa1=0.1)
plot(fpack1)
system.time(pfit1 <- pstpm2(Surv(recyear,censrec==1)~hormon+x3,data=brcancer,
logH.formula=~s(log(recyear),k=20)+s(x3),sp=c(0.1,0.1)))
system.time(pfit1 <- pstpm2(Surv(recyear,censrec==1)~hormon+x3,data=brcancer,
logH.formula=~s(log(recyear),k=20)+s(x3)))
system.time(pfit1 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(log(recyear),k=20)+s(x3)))
plot(pfit1,newdata=data.frame(hormon=1,x3=20))
plot(pfit1,newdata=data.frame(hormon=0,x3=20),type="hazard")
plot(pfit1,newdata=data.frame(hormon=1,x3=20),type="hazard",add=TRUE,line.col="blue",lty=1)
summary(pfit1)
brcancerN <- brcancer[rep(1:nrow(brcancer),each=100),]
system.time(pfit1 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancerN,
logH.formula=~s(log(recyear),k=20)))
plot(pfit1,newdata=data.frame(hormon=1))
pfit1@gam$sp
par(mfrow=c(2,2))
plot(pfit1,newdata=data.frame(hormon=1))
summary(pfit1@gam)$edf
rstpm2:::gcv(pfit1)
rstpm2:::gcvc(pfit1,nn)
sps <- as.list(10^(seq(-4,2,by=0.5)))
system.time(pfit2 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(log(recyear),k=20), sp=sps))
gcvs <- lapply(pfit2,rstpm2:::gcv)
plot(sps,unlist(gcvs),type="l",log="x")
plot(sapply(gcvs,attr,"negll"),sapply(gcvs,attr,"trace"),type="l",asp=1)
plot(sapply(gcvs,attr,"trace"),sapply(gcvs,attr,"negll"),type="l",asp=1)
plot(sps,sapply(pfit2,rstpm2:::aicc,nn=nn),type="l",log="x")
plot(sps,sapply(pfit2,rstpm2:::bic,nn=nn),type="l",log="x")
##gcvc
brcancer$recyear <- brcancer$rectime/365
sps <- 10^(seq(-4,2,by=0.5))
gcvcs <- sapply(sps, function(sp) {
gcvc(pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(recyear,k=30), sp=sp),length(brcancer$recyear))
})
plot(sps,gcvcs,type="l",log="x")
###bic
brcancer$recyear <- brcancer$rectime/365
sps <- 10^(seq(-4,2,by=0.5))
gcvcs <- sapply(sps, function(sp) {
bic(pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(recyear,k=30), sp=sp),length(brcancer$recyear))
})
plot(sps,gcvcs,type="l",log="x")
###aicc
brcancer$recyear <- brcancer$rectime/365
sps <- 10^(seq(-4,2,by=0.5))
gcvcs <- sapply(sps, function(sp) {
aicc(pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(recyear,k=30), sp=sp),length(brcancer$recyear))
})
plot(sps,gcvcs,type="l",log="x")
#########################
### penalty functions
require(mgcv)
require(gaussquad)
## Outline:
## get w, lambda, X0, X1, X2, X3
## calculate s0, s1, s2, s3
## calculate h2 and pfun=integrate(h2^2,t)
## calculate dh2sq.dbeta and dpfun=integrate(dh2sq.dbeta,t)
##
## calculate w, lambda, X0, X1, X2, X3
derivativeDesign <-
function (functn, lower = -1, upper = 1, rule = NULL,
...)
{
pred <- if (length(list(...)) && length(formals(functn)) >
1)
function(x) functn(x, ...)
else functn
if (is.null(rule))
rule <- ## gaussquad::legendre.quadrature.rules(20)[[20]]
data.frame(x = c(0.993128599185095, 0.963971927277914, 0.912234428251326,
0.839116971822219, 0.746331906460151, 0.636053680726515, 0.510867001950827,
0.37370608871542, 0.227785851141646, 0.0765265211334977, -0.0765265211334974,
-0.227785851141645, -0.373706088715418, -0.510867001950827, -0.636053680726516,
-0.746331906460151, -0.839116971822219, -0.912234428251326, -0.963971927277913,
-0.993128599185094),
w = c(0.0176140071391522, 0.040601429800387,
0.0626720483341092, 0.0832767415767053, 0.101930119817241, 0.11819453196152,
0.131688638449176, 0.14209610931838, 0.149172986472603, 0.152753387130726,
0.152753387130726, 0.149172986472603, 0.142096109318381, 0.131688638449175,
0.11819453196152, 0.10193011981724, 0.0832767415767068, 0.0626720483341075,
0.0406014298003876, 0.0176140071391522))
lambda <- (upper - lower)/(2)
mu <- (lower + upper)/(2)
x <- lambda * rule$x + mu
w <- rule$w
eps <- .Machine$double.eps^(1/8)
X0 <- pred(x)
X1 <- (-pred(x+2*eps)+8*pred(x+eps)-8*pred(x-eps)+pred(x-2*eps))/12/eps
X2 <- (-pred(x+2*eps)/12+4/3*pred(x+eps)-5/2*pred(x)+4/3*pred(x-eps)-pred(x-2*eps)/12)/eps/eps
X3 <- (-pred(x+3*eps)/8+pred(x+2*eps)-13/8*pred(x+eps)+
13/8*pred(x-eps)-pred(x-2*eps)+pred(x-3*eps)/8)/eps/eps/eps
return(list(x=x,w=w,lambda=lambda,X0=X0,X1=X1,X2=X2,X3=X3))
}
hpfun <- function(beta,design) {
lapply(design,function(obj) {
s0 <- as.vector(obj$X0 %*% beta)
s1 <- as.vector(obj$X1 %*% beta)
s2 <- as.vector(obj$X2 %*% beta)
s3 <- as.vector(obj$X3 %*% beta)
h2 <- (s3+3*s1*s2+s1^3)*exp(s0)
obj$lambda*sum(obj$w*h2^2)
})
}
hdpfun <- function(beta,design) {
lapply(design, function(obj) {
s0 <- as.vector(obj$X0 %*% beta)
s1 <- as.vector(obj$X1 %*% beta)
s2 <- as.vector(obj$X2 %*% beta)
s3 <- as.vector(obj$X3 %*% beta)
h2 <- (s3+3*s1*s2+s1^3)*exp(s0)
dh2sq.dbeta <- 2*h2*(exp(s0)*(obj$X3+3*(obj$X1*s2+obj$X2*s1)+3*s1^2*obj$X1)+h2*obj$X0)
obj$lambda*colSums(obj$w*dh2sq.dbeta)
})
}
smootherDesign <- function(gamobj,data) {
d <- data[1,,drop=FALSE] ## how to get mean prediction values, particularly for factors?
makepred <- function(var) {
function(value) {
d <- d[rep(1,length(value)),]
d[[var]] <- value
predict(gamobj,newdata=d,type="lpmatrix")
}
}
lapply(gamobj$smooth, function(smoother) {
var <- smoother$term
pred <- makepred(var)
derivativeDesign(pred,lower=min(data[[var]]),upper=max(data[[var]]))
})
}
## example data
d <- within(data.frame(x=seq(0,1,length=301)), {
mu <- exp(x)
y <- rnorm(301,mu,0.01)
})
fit <- gam(y~s(x),data=d,family=gaussian(link="log"))
beta <- coef(fit)
design <- smootherDesign(fit,d)
hpfun(beta,design)
hdpfun(beta,design)
## Testing...
require(mgcv)
d <- within(data.frame(x=seq(0,2*pi,length=301)), {
mu <- sin(x)
dmu <- cos(x)
y <- rnorm(301,mu,0.001)
})
fit <- gam(y~s(x),data=d)
mat <- predict(fit,type="lpmatrix")
with(d,plot(x,y))
with(d,lines(x,mu,lwd=2))
with(d,lines(x,predict(fit),col="blue",lwd=2))
par(mfrow=c(3,2))
pred <- function(eps,obj=fit,data=d,var="x") {
nd <- d
nd[[var]] <- nd[[var]]+eps
predict(obj,newdata=nd,type="lpmatrix")
}
## First derivative
eps <- .Machine$double.eps^(1/8)
matD <- (pred(eps) - pred(-eps)) / 2 / eps
with(d,plot(x,dmu,lwd=2,type="l"))
with(d,lines(x,matD %*% coef(fit),col="blue",lwd=2))
##
## 1/12 −2/3 0 2/3 −1/12
eps <- .Machine$double.eps^(1/8)
matD <- (-pred(2*eps)+8*pred(eps)-8*pred(-eps)+pred(-2*eps))/12/eps
with(d,plot(x,dmu,lwd=2,type="l"))
with(d,lines(x,matD %*% coef(fit),col="blue",lwd=2))
##
## Second derivative
eps <- .Machine$double.eps^(1/8)
matD2 <- (pred(eps)-2*pred(0)+pred(-eps))/eps/eps
with(d,plot(x,-mu,lwd=2,type="l"))
with(d,lines(x,matD2 %*% coef(fit),col="blue",lwd=2))
##
## −1/12 4/3 −5/2 4/3 −1/12
eps <- .Machine$double.eps^(1/8)
matD2 <- (-pred(2*eps)/12+4/3*pred(eps)-5/2*pred(0)+4/3*pred(-eps)-pred(-2*eps)/12)/eps/eps
with(d,plot(x,-mu,lwd=2,type="l"))
with(d,lines(x,matD2 %*% coef(fit),col="blue",lwd=2))
##
## Third derivatives
eps <- .Machine$double.eps^(1/8)
matD3 <- (pred(2*eps)-
2*pred(eps)+
2*pred(-eps)-
pred(-2*eps))/2/eps/eps/eps
with(d,plot(x,-dmu,lwd=2,type="l"))
with(d,lines(x,matD3 %*% coef(fit),col="blue",lwd=2))
##
## 1/8 −1 13/8 0 −13/8 1 −1/8
eps <- .Machine$double.eps^(1/8)
matD3 <- (-pred(3*eps)/8+pred(2*eps)-13/8*pred(eps)+
13/8*pred(-eps)-pred(-2*eps)+pred(-3*eps)/8)/eps/eps/eps
with(d,plot(x,-dmu,lwd=2,type="l"))
with(d,lines(x,matD3 %*% coef(fit),col="blue",lwd=2))
## (browse-url "http://en.wikipedia.org/wiki/Finite_difference_coefficients")
require(mgcv)
data <- data.frame(x=1:10,y=1:10)
fit <- gam(y~s(x,k=5,bs="ps"),data=data)
round(cbind(1,(spline.des(knots=fit$smooth[[1]]$knots,x=data$x)$design %*%
qr.Q(attr(fit$smooth[[1]],"qrc"),complete=TRUE))[,-1]) -
predict(fit,type="lpmatrix"),1e-10)
round(cbind(1,(spline.des(knots=fit$smooth[[1]]$knots,x=5:6)$design %*%
qr.Q(attr(fit$smooth[[1]],"qrc"),complete=TRUE))[,-1]) -
predict(fit,newdata=data.frame(x=5:6),type="lpmatrix"),1e-10)
cbind(0,(spline.des(knots=fit$smooth[[1]]$knots,x=data$x,derivs=rep(1,nrow(data)))$design %*%
qr.Q(attr(fit$smooth[[1]],"qrc"),complete=TRUE))[,-1]) -
(predict(fit,newdata=transform(data,x=x+1e-5),type="lpmatrix")-
(predict(fit,newdata=transform(data,x=x-1e-5),type="lpmatrix")))/2e-5
#######Optimal fitting#######
###GCV,AICC,BIC or GCVC to choose smoothing parameters###
opt.val<-function(pstpm2.fit,nn){
like<-pstpm2.fit@like
Hl<-numDeriv::hessian(like,coef(pstpm2.fit))
Hinv<-vcov(pstpm2.fit)
trace<-sum(diag(Hinv%*%Hl))
loglike<-(like(coef(pstpm2.fit)))/nn
gcv<-(trace-loglike)/nn
aicc<-(-2*loglike+2*trace*nn/(nn-trace-1))/nn
bic<-(-2*loglike+trace*log(nn))/nn
gcvc<-(-2*loglike-2*nn*log(1-trace/nn))/nn
out<-c(loglike,gcv,aicc,bic,gcvc)
return(out)
}
###############################
###############################
# setClass("opt.fit", representation(
# num.ind = "numeric",
# cr = "numeric",
# tops = "data.frame",
# sp.opt = "numeric",
# fun.min = "numeric"
# ),
# contains="pstpm2")
# #########################
# opt.fit<-function(formula,data,logH.formula,sp.low,sp.upp,num.sp,timeVar = NULL){
# ###number of individual
# num.ind <- nrow(data)
# #####Censoring rate####
# ## set up the data
# ## ensure that data is a data frame
# data <- get_all_vars(formula, data)
# # ## parse the function call
# # Call <- match.call()
# # mf <- match.call(expand.dots = FALSE)
# # m <- match(c("formula", "data", "subset", "contrasts", "weights"),
# # names(mf), 0L)
# # mf <- mf[c(1L, m)]
# stopifnot(length(lhs(formula))>=2)
# eventExpr <- lhs(formula)[[length(lhs(formula))]]
# delayed <- length(lhs(formula))==4
# timeExpr <- lhs(formula)[[if (delayed) 3 else 2]] # expression
# if (is.null(timeVar))
# timeVar <- all.vars(timeExpr)
# time <- eval(timeExpr, data)
# if (delayed) {
# time0Expr <- lhs(formula)[[2]]
# time0 <- eval(time0Expr, data)
# }
# event <- eval(eventExpr,data)
# cr <- sum(event > min(event))/num.ind
# #
# # cr=table(lhs(formula)[[if (delayed) 4 else 3]][2])/nn
# ##nn<-length(brcancer$recyear)
# # system.time(pfit1 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
# # logH.formula=~s(recyear,k=30), sp=1e-1))
# # plot(pfit1,newdata=data.frame(hormon=1))
#
# #sps <- 10^(seq(-4,4,by=0.5))
# # sp.low=10^-4
# # sp.upp=4000
# # num.sp=30
# sps <- 10^(seq(log10(sp.low),log10(sp.upp),length=num.sp))
# optvals <- sapply(sps, function(sp) {
# opt.val(pstpm2(formula,data,logH.formula=NULL, sp=sp),num.ind)
# })
# tops<-t(optvals)
# colnames(tops) <- c("loglike","gcv","aicc","bic","gcvc")
# rownames(tops) <- rownames(tops, do.NULL = FALSE, prefix = "Obs.")
# # tops<-as.data.frame(tops)
# tops<-as.data.frame(tops)
# ####Plot#########
# #par(mfrow=c(1,2))
# ###to choose optimal smoothing parameter ###
# ind.min <- sapply(2:5,function(x) order(tops[,x])[1])
# sp.opt <- sps[ind.min]
# obj<-pstpm2(formula,data,logH.formula=NULL, sp=sp.opt[1])
# fun.min <- sapply(2:5,function(x) min(tops[,x]))
# # if(ind.min[1]==1)
# # stop("Hit left boundary, make sp.low smaller.")
# # if(ind.min[1]==num.sp)
# # stop("Hit right boundary, make sp.upp bigger.")
# # with(tops,matplot(sps,tops[,-1],type="l",col=1:4,lty=1:4,xlab="x",ylab="y"))
# # points(sp.opt,fun.min,pch=4,lwd=2,cex=1.2)
# # lines(sp.opt,fun.min,err=-1,col=1:4,lty=1:4)
#
# ###Estimate final model with optimal value of sp###
#
# #
# # summary(pfit.obj)
# #########################################
# out <- as(obj,"opt.fit")
# out <- new("opt.fit",
# coef = pstpm2@coef,
# fullcoef = pstpm2@fullcoef,
# vcov = pstpm2@vcov,
# min = pstpm2@min,
# details = pstpm2@details,
# minuslogl = pstpm2@minuslogl,
# method = pstpm2@method,
# data = data,
# formula = pstpm2@formula,
# optimizer = "optim",
# xlevels = .getXlevels(pstpm2@terms,pstpm2@model.frame),
# ##contrasts = attr(X, "contrasts"),
# contrasts = NULL, # wrong!
# logli = pstpm2@logli,
# ##weights = weights,
# Call = pstpm2@Call,
# terms = pstpm2@terms,
# model.frame = pstpm2@model.frame,
# gam = pstpm2@gam,
# timeVar = pstpm2@timeVar,
# timeExpr = pstpm2@timeExpr,
# like = pstpm2@like,
# negll<-pstpm2@negll,
# call.formula = pstpm2@call.formula,
# x = pstpm2@x,
# xd = pstpm2@xd,
# termsd = pstpm2@termsd, # wrong!
# y = pstpm2@y,
# num.ind = num.ind,
# cr = cr,
# tops = tops,
# sp.opt = sp.opt,
# fun.min = fun.min)
#
# return(out)
# }
#####load data####
load("brcancer.rda")
data(brcancer)
brcancer$recyear <- brcancer$rectime/365
####model fit###
opt.fit(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(recyear), sp.low=10^-4,sp.upp=4000,
num.sp=30,timeVar = NULL)
# ###methods for Plot ###
# setMethod(
# f= "plot",
# signature(x="opt.fit", y="missing"),
# definition=function (x,y,...){
# matplot(x@sps,x@tops[,-1],type="l",col=1:4,lty=1:4,xlab="",ylab="")
# points(x@sp.opt,x@fun.min,pch=4,lwd=2,cex=1.2)
# lines(x@sp.opt,x@fun.min,err=-1,col=1:4,lty=1:4)
# }
# )
# ####methods for print####
# setMethod ("print",signature(x="opt.fit", y="missing"),
# function(x,...){
# cat("*** Class opt.fit, method Print *** \n")
# cat("* Optimal SP ="); print (x@sp.opt)
# cat("* GCV = \n"); print (x@fun.min[1])
# cat("******* End Print (opt.fit) ******* \n")
# }
# )
##########################
aplot <-
function (x, y, ...)
{
.local <- function (x, y, newdata, type = "surv", xlab = NULL,
ylab = NULL, line.col = 1, ci.col = "grey", lty = par("lty"),
add = FALSE, ci = !add, rug = !add, var = NULL, ...)
{
browser()
y <- predict(x, newdata, type = type, var = var, grid = TRUE,
se.fit = TRUE)
if (is.null(xlab))
xlab <- deparse(x@timeExpr)
if (is.null(ylab))
ylab <- switch(type, hr = "Hazard ratio", hazard = "Hazard",
surv = "Survival", sdiff = "Survival difference",
hdiff = "Hazard difference", cumhaz = "Cumulative hazard")
xx <- attr(y, "newdata")
xx <- eval(x@timeExpr, xx)
if (!add)
matplot(xx, y, type = "n", xlab = xlab, ylab = ylab,
...)
if (ci)
polygon(c(xx, rev(xx)), c(y[, 2], rev(y[, 3])), col = ci.col,
border = ci.col)
lines(xx, y[, 1], col = line.col, lty = lty)
if (rug) {
Y <- x@y
eventTimes <- Y[Y[, ncol(Y)] == 1, ncol(Y) - 1]
rug(eventTimes, col = line.col)
}
return(invisible(y))
}
.local(x, y, ...)
}
aplot(fit,newdata=data.frame(hormon=1))
apredict <- function (object, ...)
{
.local <- function (object, newdata = NULL, type = c("surv",
"cumhaz", "hazard", "hr", "sdiff", "hdiff", "loghazard",
"link"), grid = FALSE, seqLength = 300, se.fit = FALSE,
link = NULL, exposed = incrVar(var), var, ...)
{
local <- function(object, newdata = NULL, type = "surv",
exposed) {
## browser()
tt <- object@terms
if (is.null(newdata)) {
X <- object@x
XD <- object@xd
y <- object@y
time <- as.vector(y[, ncol(y) - 1])
}
else {
lpfunc <- function(delta, fit, data, var) {
data[[var]] <- data[[var]] + delta
lpmatrix.lm(fit, data)
}
X <- lpmatrix.lm(object@lm, newdata)
XD <- grad(lpfunc, 0, object@lm, newdata, object@timeVar)
XD <- matrix(XD, nrow = nrow(X))
if (type %in% c("hazard", "hr", "sdiff", "hdiff",
"loghazard")) {
time <- eval(object@timeExpr, newdata)
}
if (object@delayed) {
newdata0 <- newdata
newdata0[[object@timeVar]] <- newdata[[object@time0Var]]
X0 <- lpmatrix.lm(object@lm, newdata0)
}
if (type %in% c("hr", "sdiff", "hdiff")) {
if (missing(exposed))
stop("exposed needs to be specified for type in ('hr','sdiff','hdiff')")
newdata2 <- exposed(newdata)
X2 <- lpmatrix.lm(object@lm, newdata2)
XD2 <- grad(lpfunc, 0, object@lm, newdata2,
object@timeVar)
XD2 <- matrix(XD, nrow = nrow(X))
}
}
beta <- coef(object)
cumHaz = as.vector(exp(X %*% beta))
Sigma = vcov(object)
if (type == "link") {
return(as.vector(X %*% beta))
}
if (type == "cumhaz") {
if (object@delayed)
return(cumHaz - as.vector(X0 %*% beta))
else return(cumHaz)
}
if (type == "surv") {
return(exp(-cumHaz))
}
if (type == "sdiff")
return(as.vector(exp(-exp(X2 %*% beta))) - exp(-cumHaz))
if (type == "hazard") {
return(as.vector(XD %*% beta) * cumHaz)
}
if (type == "loghazard") {
return(as.vector(log(XD %*% beta)) + log(cumHaz))
}
if (type == "hdiff") {
return(as.vector((XD2 %*% beta) * exp(X2 %*% beta) - (XD %*%
beta)/time * cumHaz))
}
if (type == "hr") {
cumHazRatio = exp((X2 - X) %*% beta)
return(as.vector((XD2 %*% beta)/(XD %*% beta) * cumHazRatio))
}
}
type <- match.arg(type)
if (is.null(newdata) && type %in% c("hr", "sdiff", "hdiff"))
stop("Prediction using type in ('hr','sdiff','hdiff') requires newdata to be specified.")
if (grid) {
Y <- object@y
event <- Y[, ncol(Y)] == 1
time <- object@data[[object@timeVar]]
eventTimes <- time[event]
X <- seq(min(eventTimes), max(eventTimes), length = seqLength)[-1]
data.x <- data.frame(X)
names(data.x) <- object@timeVar
newdata <- merge(newdata, data.x)
}
pred <- if (!se.fit) {
local(object, newdata, type = type, exposed = exposed,
...)
}
else {
if (is.null(link))
link <- switch(type, surv = "cloglog", cumhaz = "log",
hazard = "log", hr = "log", sdiff = "I", hdiff = "I",
loghazard = "I", link = "I")
predictnl(object, local, link = link, newdata = newdata,
type = type, exposed = exposed, ...)
}
attr(pred, "newdata") <- newdata
return(pred)
}
.local(object, ...)
}
environment(apredict) <- environment(stpm2)
dim(apredict(fit,newdata=data.frame(hormon=1),grid=T)) # n=300 or 299??
apredict(fit,newdata=data.frame(hormon=1),grid=T)
dim(apredict(fit,newdata=data.frame(hormon=1),grid=T,se.fit=T)) # n=300 or 299??
apredict(fit,newdata=data.frame(hormon=1),grid=T,se.fit=T)
debug(rstpm2:::numDeltaMethod)
try(suppressWarnings(detach("package:rstpm2",unload=TRUE)),silent=TRUE)
require(rstpm2)
data(brcancer)
system.time(fit2 <- stpm2(Surv(rectime/365,censrec==1)~hormon,data=brcancer,df=5))
system.time(fit3 <- pstpm2(Surv(rectime/365,censrec==1)~hormon,data=brcancer,use.gr=F))
plot(fit3,newdata=data.frame(hormon=0),type="hazard")
plot(fit2,newdata=data.frame(hormon=0),type="hazard",add=TRUE,ci=FALSE,rug=FALSE,
line.col=2)
## penalised likelihood
brcancer$recyear <- brcancer$rectime/365
system.time(pfit1 <- pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(log(recyear),k=30), sp=1e-1))
system.time(fit1 <- stpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~ns(log(recyear),df=4)))
plot(pfit1,newdata=data.frame(hormon=1))
plot(fit1,newdata=data.frame(hormon=1),lty=2,add=TRUE,ci=F)
rstpm2:::gcv(pfit1)
sps <- 10^(seq(-4,2,by=0.5))
gcvs <- sapply(sps, function(sp) {
rstpm2:::gcv(pstpm2(Surv(recyear,censrec==1)~hormon,data=brcancer,
logH.formula=~s(recyear,k=30), sp=sp))
})
plot(sps,gcvs,type="l",log="x")
##
system.time(fit <- rstpm2:::stpm2Old(Surv(rectime,censrec==1)~hormon,df=5,data=brcancer))
system.time(fit2 <- stpm2(Surv(rectime,censrec==1)~hormon,df=5,data=brcancer))
system.time(fit3 <- pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer))
##
plot(fit3,newdata=data.frame(hormon=0),type="hazard")
plot(fit2,newdata=data.frame(hormon=0),type="hazard",add=TRUE,line.col=2,ci=FALSE)
##
system.time(fit <- stpm2(Surv(rectime/365,censrec==1)~hormon,df=5,data=brcancer))
system.time(fit2 <- rstpm2:::stpm2Old(Surv(rectime/365,censrec==1)~hormon,df=5,data=brcancer))
##
system.time(fit3 <- pstpm2(Surv(rectime/365,censrec==1)~hormon,data=brcancer))
plot(fit3,newdata=data.frame(hormon=0),type="hazard")
##
plot(fit2,newdata=data.frame(hormon=0),type="hazard",add=TRUE,line.col=2,ci=FALSE)
##
plot(fit.tvc,newdata=data.frame(hormon=1),type="hr",var="hormon")
##
summary(fit.tvc <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,df=3,
tvc=list(hormon=3)))
anova(fit,fit.tvc) # compare with and without tvc
summary(fit.tvc <- stpm2Old(Surv(rectime,censrec==1)~hormon,data=brcancer,df=3,
tvc=list(hormon=3)))
anova(fit,fit.tvc) # compare with and without tvc
##
plot(fit.tvc,newdata=data.frame(hormon=0),type="hr",var="hormon")
# no lines method: use add=TRUE
plot(fit.tvc,newdata=data.frame(hormon=1),type="hr",var="hormon",
add=TRUE,ci=FALSE,line.col=2)
##
## plain: identical results (good)
stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer)
stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,
logH.formula=~ns(log(rectime),3))
rstpm2:::stpm2Old(Surv(rectime,censrec==1)~hormon,data=brcancer)
## cure: identical (requires bhazard to be sensible)
rate0 <- 10^(-5+brcancer$x1/100)
(fit1 <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,df=2,cure=T,bhazard=rate0))
(fit2 <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,
logH.formula=~nsx(log(rectime),df=2,cure=T,log=T),bhazard=rate0))
(fit3 <- rstpm2:::stpm2Old(Surv(rectime,censrec==1)~hormon,data=brcancer,cure=T,df=2,bhazard=rate0))
(fit4 <- rstpm2:::stpm2Old(Surv(rectime,censrec==1)~hormon,data=brcancer,bhazard=rate0,
logH.formula=~nsx(log(rectime),2,cure=T)))
##### examples #####
require(foreign)
if (FALSE) { # testing in open code
install.packages("bbmle", repos="http://R-Forge.R-project.org")
require(bbmle)
brcancer=read.dta("brcancer.dta")
brcancer=transform(brcancer,rate0=10^(-5+x1/100))
}
try(suppressWarnings(detach("package:bbmle",unload=TRUE)),silent=TRUE)
try(suppressWarnings(detach("package:rstpm2",unload=TRUE)),silent=TRUE)
## require(rstpm2)
data(brcancer)
summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,
logH.formula=~nsx(log(rectime),df=3,stata=TRUE)))
brcancer <- transform(brcancer,w=10)
summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,
weights=w,robust=TRUE,
logH.formula=~nsx(log(rectime),df=3,stata=TRUE)))
## sandwich variance estimator (from the sandwich package)
coeftest.stpm2 <-
function (x, vcov. = NULL, df = NULL, ...)
{
est <- coef(x)
if (is.null(vcov.))
se <- vcov(x)
else {
if (is.function(vcov.))
se <- vcov.(x)
else se <- vcov.
}
se <- sqrt(diag(se))
if (!is.null(names(est)) && !is.null(names(se))) {
anames <- names(est)[names(est) %in% names(se)]
est <- est[anames]
se <- se[anames]
}
tval <- as.vector(est)/se
pval <- 2 * pnorm(abs(tval), lower.tail = FALSE)
cnames <- c("Estimate", "Std. Error", "z value", "Pr(>|z|)")
mthd <- "z"
rval <- cbind(est, se, tval, pval)
colnames(rval) <- cnames
class(rval) <- "coeftest"
attr(rval, "method") <- paste(mthd, "test of coefficients")
return(rval)
}
## weights.stpm2 <-
## function (object, ...)
## {
## wts <- object@weights
## if (is.null(wts))
## wts
## else napredict(object@na.action, wts)
## }
require(sandwich)
coxph1 <- coxph(Surv(rectime,censrec==1)~hormon,data=brcancer)
update(coxph1,robust=TRUE)
sandwich(coxph1)
sandwich.stpm2(fit) # hurrah!
## require(lmtest)
## coeftest(coxph1)
## coeftest(coxph1,vcov.=sandwich(coxph1))
## coeftest(fit,sandwich(fit))
sandwich(fit)
sandwich(fit,bread.=bread.stpm2,meat.=meat.stpm2)
## some predictions
head(predict(fit,se.fit=TRUE,type="surv"))
head(predict(fit,se.fit=TRUE,type="hazard"))
## some plots
plot(fit,newdata=data.frame(hormon=0),type="hazard")
## time-varying coefficient
summary(fit.tvc <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,df=3,
tvc=list(hormon=3)))
anova(fit,fit.tvc) # compare with and without tvc
plot(fit.tvc,newdata=data.frame(hormon=0),type="hr",var="hormon")
# no lines method: use add=TRUE
plot(fit.tvc,newdata=data.frame(hormon=1),type="hr",var="hormon",
add=TRUE,ci=FALSE,line.col=2)
plot(fit.tvc,newdata=data.frame(hormon=0),type="sdiff",var="hormon")
plot(fit.tvc,newdata=data.frame(hormon=0),type="hdiff",var="hormon")
plot(fit.tvc,newdata=data.frame(hormon=0),type="hazard")
plot(fit.tvc,newdata=data.frame(hormon=1),type="hazard",line.col=2,ci=FALSE,add=TRUE)
## trace("predict", browser, exit=browser, signature = "stpm2")
set.seed(10101)
brcancer <- transform(brcancer, x=rlnorm(nrow(brcancer)))
summary(fit.tvc <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,df=3,
tvc.formula=~hormon:nsx(log(rectime),df=3)))
## cure model
## cf. http://www.pauldickman.com/survival/solutions/q37.do
### Data setup
require(foreign)
colon <- read.dta("http://www.pauldickman.com/survival/colon.dta")
popmort <- read.dta("http://www.pauldickman.com/survival/popmort.dta")
brcancer <- read.dta("http://www.stata-press.com/data/r11/brcancer.dta")
popmort <- transform(popmort, age=`_age`, year=`_year`, `_age`=NULL, `_year`=NULL)
save(colon,file="c:/usr/src/R/rstpm2/pkg/data/colon.rda")
save(popmort,file="c:/usr/src/R/rstpm2/pkg/data/popmort.rda")
save(brcancer,file="c:/usr/src/R/rstpm2/pkg/data/brcancer.rda")
## require(rstpm2)
popmort2 <- transform(popmort,exitage=age,exityear=year,age=NULL,year=NULL)
colon2 <- within(colon, {
status <- ifelse(surv_mm>120.5,1,status)
tm <- pmin(surv_mm,120.5)/12
exit <- dx+tm*365.25
sex <- as.numeric(sex)
exitage <- pmin(floor(age+tm),99)
exityear <- floor(yydx+tm)
})
colon2 <- merge(colon2,popmort2)
## compare relative survival without and with cure
summary(fit0 <- stpm2(Surv(tm,status %in% 2:3)~I(year8594=="Diagnosed 85-94"),
data=colon2,
bhazard=colon2$rate, df=5)) ## CHECKED: same year8594 estimate as Stata
head(predict(fit0))
## estimate of failure at the end of follow-up
1-predict(fit0,data.frame(year8594 = unique(colon2$year8594),tm=max(colon2$tm)),type="surv",se.fit=TRUE)
plot(fit0,newdata=data.frame(year8594 = "Diagnosed 85-94"),ylim=0:1)
plot(fit0,newdata=data.frame(year8594 = "Diagnosed 75-84"),add=TRUE,line.col="red",rug=FALSE)
##
summary(fit <- stpm2(Surv(tm,status %in% 2:3)~I(year8594=="Diagnosed 85-94"),
data=colon2,
bhazard=colon2$rate,
df=5,cure=TRUE))
head(predict(fit))
## cure fractions (I need to add this to the predict function)
1-predict(fit,data.frame(year8594 = unique(colon2$year8594),tm=max(colon2$tm)),type="surv",se.fit=TRUE)
newdata1 <- data.frame(year8594 = "Diagnosed 85-94")
plot(fit,newdata=newdata1,add=TRUE,ci=FALSE,lty=2,rug=FALSE)
plot(fit,newdata=data.frame(year8594="Diagnosed 75-84"),add=TRUE,rug=FALSE,line.col="red",ci=FALSE,lty=2)
plot(fit,newdata=newdata1,type="hazard")
plot(fit,newdata=newdata1,type="cumhaz")
## http://www.pauldickman.com/survival/r/melanoma.relsurv.r
library(foreign)
library(survival)
library(relsurv)
# Download rates files from http://www.mortality.org/
# # 6. Life Tables By year of death (period) 1x1
# Save tables by gender in text files
# The transrate.hmd command translate these to R ratetables
Finlandpop <- transrate.hmd("c:/usr/tmp/mltper_1x1.txt","c:/usr/tmp/fltper_1x1.txt")
## The relsurv package requires time in days (exit and dx are dates of exit and diagnosis)
colon3 <- transform(colon2,tm.dd=as.numeric(exit-dx))
colon3$sex <- ifelse(colon2$sex==1,"male","female")
as.date <- function(x)
if (inherits(x,"Date")) as.date(as.numeric(x)+3653) else date::as.date(x)
model1 <- rs.surv(Surv(tm.dd,status %in% 2:3)~year8594+ratetable(age=(X_age+0.5)*365.25,sex=sex,year=as.date(exit)),colon3,ratetable=Finlandpop)
plot(model1,lty=1:2)
oldx <- 0:100
oldy <- (oldx-50)^2
oldy[c(20,30)] <- 0
old <- data.frame(x=oldx,y=oldy)
predict(lm(y~nsx(x,knots=c(25,50,75,95)),old)) # as per Stata
newx <- seq(min(oldx)/1.05,max(oldx)*1.05,length=101)
new <- data.frame(x=newx)
plot(oldx,oldy)
predict(lm(y~nsx(x,df=5,cure=TRUE),old))
sum(oldy)
terms(lm(y~nsx(x,df=5,cure=TRUE),old))
lm(y~nsx(x,df=5),old)
lines(newx,
predict(lm(y~nsx(x,df=4,cure=FALSE),old),newdata=new),
type="l") # oops
lines(newx,
predict(lm(y~nsx(x,df=3),old),newdata=new),
lty=2)
summary(fit <- stpm2(Surv(tm,status %in% 2:3)~I(year8594=="Diagnosed 85-94"),
data=colon2,
bhazard=colon2$rate,
logH.formula=~nsx(log(tm),df=6,stata=TRUE))) # okay
summary(fit <- stpm2(Surv(tm,status %in% 2:3)~I(year8594=="Diagnosed 85-94"),
data=colon2,
logH.formula=~nsx(log(tm),df=6,stata=TRUE))) # okay
## Stata
## stata.knots=c(4.276666164398193, 6.214608192443848, 6.7833251953125, 7.806289196014404)
stataKnots <- function(x,df) {
intKnots <- round((1:(df-1))/df,2) # yes, Paul implicitly rounded to 2 dp
logx <- log(x)
c(min(logx),quantile(logx,intKnots,type=2),max(logx))
}
stata.knots <- stataKnots(subset(brcancer,censrec==1)$rectime,3)
## sapply(1:9,function(type) log(quantile(subset(brcancer,censrec==1)$rectime,c(0.33,0.67),type=type)))
summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,
logH.args=list(knots=stata.knots[2:3],
Boundary.knots=stata.knots[c(1,4)])))
## formula specification for logH
summary(stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,
logH.formula=~ns(log(rectime),df=3)))
pred <- predict(fit.tvc,newdata=data.frame(hormon=0:3),grid=TRUE,se.fit=TRUE,type="cumhaz")
pred.all <- cbind(pred,attr(pred,"newdata"))
## require(lattice)
## xyplot(Estimate ~ rectime, data=pred.all, group=hormon,type="l",xlab="Time")
## relative survival
brcancer <- transform(brcancer,rate0=10^(-5+x1/100))
summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer,bhazard=brcancer$rate0,df=3))
head(predict(fit,se.fit=TRUE))
## delayed entry
brcancer2 <- transform(brcancer,startTime=ifelse(hormon==0,rectime*0.5,0))
## debug(stpm2)
summary(fit <- stpm2(Surv(startTime,rectime,censrec==1)~hormon,data=brcancer2,
logH.formula=~nsx(log(rectime),df=3,stata=TRUE)))
head(predict(fit,se.fit=TRUE))
## delayed entry and tvc
summary(fit <- stpm2(Surv(startTime,rectime,censrec==1)~hormon,data=brcancer2,
tvc.formula=~hormon:nsx(log(rectime),df=3,stata=TRUE)))
head(predict(fit,se.fit=TRUE))
## multiple time scales
brcancer <- transform(brcancer,recyr=rectime/365.25)
## predictions from a simple model
summary(fit <- stpm2(Surv(recyr,censrec==1)~hormon+x1,data=brcancer,
logH.formula=~nsx(log(recyr),df=3,centre=log(50))))
head(predict(fit))
grid.x1 <- with(brcancer, seq(40,70,length=300))
newdata0 <- with(brcancer, data.frame(recyr=5,x1=grid.x1,hormon=0))
matplot(grid.x1,
predict(fit,type="hr",newdata=newdata0,var="hormon",se.fit=TRUE), type="l")
## predictions with multiple time scales
summary(fit <- stpm2(Surv(recyr,censrec==1)~hormon,data=brcancer,
logH.formula=~nsx(log(recyr),df=3,centre=log(50)),
tvc.formula=~hormon:nsx(log(recyr+x1),df=2)))
matplot(grid.x1,
predict(fit,type="hr",newdata=newdata0,var="hormon",se.fit=TRUE), type="l")
brcancer <- transform(brcancer,recyr=rectime/365.25,entry=recyr/2)
summary(fit <- stpm2(Surv(entry,recyr,censrec==1)~hormon,data=brcancer,
logH.formula=~nsx(log(recyr),df=3,centre=log(50)),
tvc.formula=~hormon:nsx(log(recyr+x1),df=2)))
summary(fit <- stpm2(Surv(recyr,censrec==1)~hormon+x1,data=brcancer,
logH.formula=~nsx(log(recyr),df=3,centre=log(50))))
plot(grid.x1,
predict(fit,type="hr",newdata=newdata0,var="hormon",se.fit=TRUE)$fit, type="l")
plot(fit,newdata=data.frame(hormon=0,x1=50),var="hormon",type="hr")
head(predict(fit,type="hazard",newdata=newdata0))
head(predict(fit,type="hazard",newdata=transform(newdata0,hormon=1)))
newdata0 <- with(brcancer, data.frame(recyr=5+1,x1=grid.x1-1,hormon=0))
predict(fit,type="hr",newdata=newdata0,var="hormon")
summary(fit <- stpm2(Surv(recyr,censrec==1)~hormon+x1,data=brcancer,
logH.formula=~nsx(log(recyr),df=3,centre=log(50)),tvc=list(hormon=3)))
brcancer <- transform(brcancer, startAge=x1, endAge=x1+rectime/365)
summary(fit <- stpm2(Surv(startAge,endAge,censrec==1)~hormon,data=brcancer,
logH.formula=~nsx(log(endAge),df=3,centre=log(50)),tvc=list(hormon=3)))
## some simulated data: H_{weibull}(t)=(t/b)^a
n <- 1000
sim1 <- data.frame(age=seq(20,70,length=n),x=rep(0:1,each=n/2))
y <- rweibull(1000,shape=1,scale=1)
with(brcancer, plot(density(x1[censrec==1])))
summary(fit <- stpm2(Surv(recyr,censrec==1)~hormon,data=brcancer,logH.formula=~nsx(log(recyr),df=3,stata=TRUE)))
brcancer <- transform(brcancer,ageStart=rnorm(length(rectime),50,5))
brcancer <- transform(brcancer,ageStop=ageStart+rectime)
summary(fit <- stpm2(Surv(ageStart,ageStop,censrec==1)~hormon,data=brcancer,df=3))
brcancer3 <- transform(brcancer,startTime=ifelse(censrec==1,0,10))
summary(fit <- stpm2(Surv(startTime,rectime,censrec==1)~hormon,data=subset(brcancer,rectime>10),df=3))
summary(fit <- stpm2(Surv(startTime,rectime,censrec==1)~hormon,data=subset(brcancer3,rectime>10),df=3))
## check the performance time
refresh
require(rstpm2)
data(brcancer)
brcancer10 = do.call("rbind",lapply(1:10,function(i) brcancer))
system.time(summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,df=3,data=brcancer10)))
system.time(summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,df=3,data=brcancer10)))
system.time(summary(fit <- stpm2(Surv(rectime,censrec==1)~hormon,data=brcancer10, logH.formula=~ns(log(rectime),df=4)+hormon:ns(log(rectime),df=3))))
system.time(summary(fit <- pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer10)))
system.time(summary(fit <- pstpm2(Surv(rectime,censrec==1)~1,data=brcancer10, smooth.formula=~s(log(rectime))+s(log(rectime),by=hormon))))
fit <- pstpm2(Surv(rectime,censrec==1)~hormon,data=brcancer10,trace=1)
fit <- pstpm2(Surv(rectime,censrec==1)~1,data=brcancer10, smooth.formula=~s(log(rectime))+s(log(rectime),by=hormon),trace=1,sp.init=c(1,1), reltol=list(outer=1e-5,search=1e-10,final=1e-10))
system.time(summary(fit <- pstpm2(Surv(rectime,censrec==1)~1,data=brcancer10,
smooth.formula=~s(log(rectime))+s(log(rectime),by=hormon),
sp=c(0.006,0.0031),trace=1,outer_optim=2,criterion="GCV",
reltol=list(outer=1e-5,search=1e-10,final=1e-10))))
## > fit@sp
## [1] 0.06104312 0.31430954
system.time(fit <- pstpm2(Surv(rectime,censrec==1)~1,data=brcancer10, smooth.formula=~s(log(rectime))+s(log(rectime),by=hormon),sp=c(1,1)))
nsx(1:10,df=3) - ns(1:10,df=3)
nsx(1:10,df=3,centre=3)
nsx(1:10,df=3,centre=3,Boundary.knots=c(2,8),derivs=c(1,1))
nsx(1:10,df=3,cure=TRUE)
nsxDeriv(1:10,df=3) - nsDeriv(1:10,df=3)
nsxDeriv(1:10,df=3,centre=5,derivs=c(1,1))
nsxDeriv(1:10,df=3,centre=5,cure=TRUE)
nsDeriv(1:10,df=3) - nsDeriv2(1:10,df=3)
## bug with calling mle2
require(bbmle)
mle2a <- function(...)
mle2(...)
## some data
x <- 0:10
y <- c(26, 17, 13, 12, 20, 5, 9, 8, 5, 4, 8)
d <- data.frame(x,y)
## some fits
(fit0 <- mle2(y~dpois(lambda=ymean),start=list(ymean=mean(y)),data=d)) # okay
(fit0.2 <- mle2(y~dpois(lambda=ymean),start=list(ymean=mean(y)),data=d,
control=list(parscale=2))) # okay
(fit1 <- mle2a(y~dpois(lambda=ymean),start=list(ymean=mean(y)),data=d)) # okay
(fit1.2 <- mle2a(y~dpois(lambda=ymean),start=list(ymean=mean(y)),data=d,
control=list(parscale=2))) # FAILS
## stdReg::parfrailty documentation
library(stdReg)
library(survival)
## simulate data
n <- 1000
m <- 3
alpha <- 1.5
eta <- 1
phi <- 0.5
beta <- 1
id <- rep(1:n, each=m)
U <- rep(rgamma(n, shape=1/phi,scale=phi), each=m)
X <- rnorm(n*m)
## reparametrize scale as in rweibull function
weibull.scale <- alpha/(U*exp(beta*X))^(1/eta)
T <- rweibull(n*m, shape=eta, scale=weibull.scale)
## right censoring
C <- runif(n*m, 0,10)
D <- as.numeric(T<C)
T <- pmin(T, C)
## strong left-truncation
L <- runif(n*m, 0, 2)
incl <- T>L
incl <- ave(x=incl, id, FUN=sum)==m
dd <- data.frame(L, T, D, X, id)
dd <- dd[incl, ]
##
fit <- parfrailty(formula=Surv(L, T, D)~X, data=dd, clusterid="id")
summary(fit)
##
library(rstpm2)
fit2 <- stpm2(formula=Surv(L, T, D)~X, data=dd, cluster=dd$id, smooth.formula=~log(T))
summary(fit2)
## ignore left truncation
fit <- parfrailty(formula=Surv(T, D)~X, data=dd, clusterid="id")
summary(fit)
fit2 <- stpm2(Surv(T, D)~X, data=dd, cluster=dd$id, smooth.formula=~log(T))
summary(fit2)
## normal random effect
fit2 <- stpm2(formula=Surv(T, D)~X, data=dd, cluster=dd$id, smooth.formula=~log(T), RandDist="LogN")
summary(fit2)
## end of examples ##
## ## * stata
## cd c:\Users\marcle\Documents\Home\
## clear
## webuse brcancer
## use brcancer
## stset rectime, f(censrec==1)
## cap program drop dopredictions
## program define dopredictions
## preserve
## predict hr, hrnumerator(hormon 1) ci
## predict haz, hazard ci
## predict surv, surv ci
## predict sdiff, sdiff1(hormon 1) ci
## list hr* in 1/5
## list haz* surv* in 1/5
## list sdiff* in 1/5
## restore
## end
## * basic model
## stpm2 hormon, df(3) scale(h)
## dopredictions
## * cure
## gen rate0=10^(-5+x1/100)
## stpm2 hormon, df(3) scale(h) cure bhazard(rate0)
## dopredictions
## * tvc
## stpm2 hormon, df(3) tvc(hormon) dftvc(3) scale(h)
## dopredictions
## * delayed entry
## preserve
## replace _t0 = rectime*0.5 if hormon==0
## stpm2 hormon, df(3) scale(h)
## dopredictions
## restore
## * relative survival
## preserve
## gen rate0=10^(-5+x1/100)
## stpm2 hormon, df(3) scale(h) bhazard(rate0)
## dopredictions
## restore
## * test speed
## clear all
## set mem 100m
## use brcancer
## stset rectime, f(censrec==1)
## expand 100
## timer clear
## timer on 1
## stpm2 hormon, df(3) scale(h)
## timer off 1
## timer list
## hazard.pm = function(obj,tm,X,XD) # obj$par
## {
## Xlocal=predict(X,newx=log(tm))
## XDlocal=predict(XD,newx=log(tm))
## with(obj,
## c((XDlocal %*% par)/tm*exp(Xlocal %*% par)))
## }
## with(list(df=df,x=seq(0,3,length=100)[-1]),
## {
## plot(x,hazard.pm(fit,x,X,XD),type="l",ylim=c(0,2))
## lines(x,dweibull(x,shape=1)/pweibull(x,shape=1,lower=FALSE),lty=2)
## })
## ##
## require(deSolve)
## temp <- as.data.frame(ode(y=0,times=seq(0,10,length=100)[-1],
## func=function(t,state,parameters=NULL) list(exp(sin(2*pi*log(t))))))
## plot(temp,type="l")
## temp <- transform(temp, cum=`1`,logcum=log(`1`))
## with(temp,plot(log(time),logcum))
## temp1 <- temp[-1,]
## fit <- glm(log(cum)~log(time)+sin(2*pi*log(time))+cos(2*pi*log(time)),data=temp1)
## lines(log(temp1$time),predict(fit))
## ## In summary:
## ## we can model using sine and cosine terms for the log-cumulative hazard - for log(time).