Smooth modeling of right-censored survival data with {icpack}"
In icpack: Survival Analysis of Interval-Censored Data

knitr::opts_chunk$set(fig.width = 6, fig.height = 4,
                      echo = TRUE, warning = FALSE, message = FALSE)

Introduction

The purpose of this vignette is to illustrate the use of {icpack}, using the right-censored Ova data, and to contrast the smooth modeling using {icpack} with the more standard non- and semi-parametric approaches of the survival package.

Ovarian cancer data

The ovarian cancer data are available after loading the package.

library(icpack)

Preliminaries

There are three covariates (see the help file). Here is a summary of their frequencies and distribution.

table(Ova$Diameter)
table(Ova$FIGO)
table(Ova$Karnofsky)

Diameter is the diameter of the tumor, FIGO is the FIGO classification, where in fact FIGO=0 corresponds to a FIGO III and FIGO=1 to a FIGO IV classification, and Karnofsky is the often used Karnofsky score divided by 10, so for instance Karnofsky=9 corresponds to a Karnofksy score of 90.

Using icfit

The time-to-event data are right-censored. Both Diameter and Karnofsky are ordinal, so we could use them both as categorical and as continuous variables. Let's first use them both as categorical, and see whether it makes sense to use them as continuous variables by inspecting their (log) hazard ratios.

ic_fit <- icfit(Surv(time, d) ~ factor(Diameter) + FIGO + factor(Karnofsky),
                data = Ova)
ic_fit

It looks like both for Diameter and Karnofsky the increments of the log hazard ratios are reasonably equal for neighboring factor levels, which could argue for using both as continuous covariates in the future. Before we do that, however, let's compare the result with that from coxph in the survival package.

cox_fit <- coxph(Surv(time, d) ~ factor(Diameter) + FIGO + factor(Karnofsky), data = Ova)
cox_fit

The results, in terms of the hazard ratios and their standard errors, are very similar indeed.

Let's refit the smooth and semi-parametric model, now using Diameter and Karnofsky as continuous variables.

ic_fit <- icfit(Surv(time, d) ~ Diameter + FIGO + Karnofsky, data = Ova)
cox_fit <- coxph(Surv(time, d) ~ Diameter + FIGO + Karnofsky, data = Ova)
ic_fit
cox_fit

Without claiming in any way that using Diameter and Karnofsky as continuous variables is better, we continue with these last two models, mainly for simplicity.

Plotting

The main difference between the smooth and semi-parametric fits of course resides in the estimates of the baseline hazard. We illustrate this by extracting the baseline hazards and plotting the corresponding baseline survival curves.

The icpack package uses ggplot2 for plotting. The default plot is the (baseline) hazard; by setting type="survival" we get the (baseline) survival function. Default option is to add shaded pointwise confidence intervals; they have been switched off here.

plot(ic_fit, type="survival", conf.int = FALSE,
     title = "Baseline survival function from icfit")

Here is the corresponding baseline survival function from coxph, with the smooth baseline survival curve from icfit added to it in blue.

bh <- basehaz(cox_fit, centered=FALSE)
bh$surv <- exp(-bh$hazard)
plot(c(0, bh$time), c(1, bh$surv), type="s",
     xlab="Time", ylab="Survival", main="Baseline survival functions from coxph and icfit")
nd_baseline <- data.frame(Diameter = 0, FIGO = 0, Karnofsky = 0)
pred_baseline <- predict(ic_fit, newdata = nd_baseline)[[1]]
lines(pred_baseline$time, pred_baseline$Surv, col = "blue")
legend("topright", c("coxph", "icfit"), lwd=1, col=c("black", "blue"), bty="n")

The plot does not show pointwise confidence intervals; it is possible to obtain those using predict, which is illustrated in the next section. Obviously, the estimated baseline survival curve form coxph is less smooth; their values however are very similar.

The estimate of the baseline hazard from the Cox model fit is not smooth; plotting it does not produce anything remotely useful.

plot(bh$time, diff(c(0, bh$hazard)),
     xlab="Time", ylab="Hazard", main="Baseline hazard function from coxph")

The baseline hazard obtained from icfit is smooth and does give a useful illustration of the hazard, along with its pointwise confidence intervals.

plot(ic_fit, type = 'hazard', title = 'Baseline hazard function from icfit')

The cumulative hazard can also be shown using type="cumhazard".

Predicting

Actually the baseline hazards and survival function are not really that interesting, in the sense that they correspond to a subject with Diameter=0, FIGO=0 and Karnofsky=0. Such persons (at least Karnofsky=0) are not in the data and are even biologically impossible.

Modeled after the survival package, the function predict with a newdata object can be used to obtain hazards and survival functions for subjects with given characteristics. Let us illustrate this with subject A, who has Diameter=0, FIGO=0 and Karnofsky=10, the most optimistic possibility, and subject B, who has Diameter=4, FIGO=0 and Karnofsky=6, the most pessimistic possibility, apart from FIGO. This is simply entered into a data frame with the same names as the original data.

newd <- data.frame(subject = c("A", "B"),
                   Diameter = c(0, 4),
                   FIGO = c(0, 0),
                   Karnofsky = c(10, 6))
newd

picf <- predict(ic_fit, newdata=newd)

The result is an object of class predict.icfit, which is a list with 2 items (one for each subject), each containing a data frame with time points and hazards, cumulative hazards, survival functions at these time points with associated standard errors and confidence bounds. The summary method can be used to extract the hazards etc at specified time points.

summary(picf, times=(0:7) * 365.25)

Finally, the hazards or cumulative hazards or survival functions can be plotted.

library(gridExtra)
plot(picf, type="surv", ylim=c(0, 1), xlab="Days since randomisation",
     title = c("Subject A", "Subject B"), nrow=1)

The results from coxph for these patients are shown in the plots below as black lines, in blue the results from icpack.

sfAB <- survfit(cox_fit, newdata=newd)
sfA <- sfAB[1]
sfB <- sfAB[2]
# Retrieve separate plots from icfit for comparison
picf1 <- picf[[1]]
class(picf1) <- c("predict.icfit", "data.frame")
plt1 <- plot(picf1, type='surv', ylim=c(0, 1),
             xlab = "Days since randomisation",
             title = "Subject A")
picf2 <- picf[[2]]
class(picf2) <- c("predict.icfit", "data.frame")
plt2 <- plot(picf2, type='surv', ylim=c(0, 1),
             xlab = "Days since randomisation",
             title = "Subject B")

oldpar <- par(mfrow=c(1, 2))
plot(sfA, xlab="Days since randomisation", ylab="Survival", main="Subject A")
lines(plt1$data$time, plt1$data$Surv, col = "blue")
lines(plt1$data$time, plt1$data$lowerSurv, col = "blue")
lines(plt1$data$time, plt1$data$upperSurv, col = "blue")
legend("topright", c("coxph", "icfit"), lwd=1, col=c("black", "blue"), bty="n")
plot(sfB, xlab="Days since randomisation", ylab="Survival", main="Subject B")
lines(plt2$data$time, plt2$data$Surv, col = "blue")
lines(plt2$data$time, plt2$data$lowerSurv, col = "blue")
lines(plt2$data$time, plt2$data$upperSurv, col = "blue")
legend("topright", c("coxph", "icfit"), lwd=1, col=c("black", "blue"), bty="n")

par(oldpar)