Site icon R-bloggers

An enhanced Kaplan-Meier plot

[This article was first published on Stat Bandit » R, and kindly contributed to R-bloggers]. (You can report issue about the content on this page here)
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

We often see, in publications, a Kaplan-Meier survival plot, with a table of the number of subjects at risk at different time points aligned below the figure. I needed this type of plot (or really, matrices of such plots) for an upcoming publication. Of course, my preferred toolbox was R and the ggplot2 package.

There were other attempts to do this type of plot in ggplot2, mainly by Gary Collins and an anonymous author as seen on the ggplot2 mailing list. There was also an invaluable blog on Learn-R on plotting a table using geom_text.

I decided to use Gary’s code as my starting point. I wanted to add a log-rank test p-value for stratified KM curves on the plot, and I wanted to code it so that the text in the table remained aligned to the x-axis even when my strata names changed with different data sets. The former was relatively easy, but the latter was challenging (and seemed to be unsolved by Gary, who had posed it as a question). Both were instructive, and took me back a layer to the grid graphics engine (on which both ggplot2 and lattice are based) and the gridExtra package.

The placement of the p-value in the plot is a case of a single character string being placed at a certain point on the graph, which I had done with the text command in base R. However, the geom_text layer seems to be intended for plotting text at all points in the data and not for inserting a single text string. However, I learned of the grid.text function from the grid package, which would enable me to do this for grid-based graphics.

The second part is to plot a table aligned to the x-axis of the plot. The plotting was achieved using Gary’s code using the method outlined on the Learn-R website. The problem was alignment. A few experiments showed that alignment depended on the size of the strata labels used in the plot. I started playing with the plot.margin option in ggplot2 and then figured out that I could maintain the alignment (more or less) if the left margin was computed based on the length of the labels. Some experimentation got me to a reasonable formula for this computation.

Now I had to put the 3 parts together. Both the plot and the table were ggplot objects, so I knew I could stitch them together with grid.arrange from the gridExtra package. However, the p-value was being written using a separate grid.text command. I had to learn more about grid. I decided to use basic grid commands to convert the KM plot to a grid object using ggplotGrob and add the p-value text string to it using addGrob and textGrob. I could now stitch this together with the plotted table using grid.arrange, using the same layout specs.

However, this still wasn’t ideal since the plots were overlapping a bit. I decided to put a buffer blank image in between. In base graphics, I could achieve this with plot(..., type="n"). In ggplot, I learned about geom_blank() and theme_blank() to create a blank image. This I then put in a thin row between the plot and the table.

I incorporated this process into the function ggkm which is given below. An example plot generated by this code:

data(colon)
fit <- survfit(Surv(time,status)~rx, data=colon)
ggkm(fit, timeby=500)

is shown here:
rstudio-V79384

The ggkm function is given below:
Update:The following code is modified thanks to Baptiste Auguie’s comments, which were very helpful. I also make the p-value text optional for the plot, with the default settings omitting the p-values.

Update (3/26/2011) I’ve further updated the code based on comments which allow it to work on a naive install (albeit with the required packages), as well as replacing “df” with “dframe” to get around the fact that “df” is a R function.

Update (9/1/2011) Both Gil Thomas and Mark Cowley have made improvements to the code that have made it more stable and universally run-able (is that a word?). The following reflects Mark’s last edits to the code.

#’ Create a Kaplan-Meier plot using ggplot2
#’
#’ @param sfit a \code{\link[survival]{survfit}} object
#’ @param table logical: Create a table graphic below the K-M plot, indicating at-risk numbers?
#’ @param returns logical: if \code{TRUE}, return an arrangeGrob object
#’ @param xlabs x-axis label
#’ @param ylabs y-axis label
#’ @param ystratalabs The strata labels. \code{Default = levels(summary(sfit)$strata)}
#’ @param ystrataname The legend name. Default = “Strata”
#’ @param timeby numeric: control the granularity along the time-axis
#’ @param main plot title
#’ @param pval logical: add the pvalue to the plot?
#’ @return a ggplot is made. if return=TRUE, then an arrangeGlob object
#’ is returned
#’ @author Abhijit Dasgupta with contributions by Gil Tomas
#’ \url{http://statbandit.wordpress.com/2011/03/08/an-enhanced-kaplan-meier-plot/}
#’ @export
#’ @examples
#’ \dontrun{
#’ data(colon)
#’	fit <- survfit(Surv(time,status)~rx, data=colon)
#'	ggkm(fit, timeby=500)
#' }
ggkm <- function(sfit, table = TRUE, returns = FALSE,
xlabs = "Time", ylabs = "survival probability",
ystratalabs = NULL, ystrataname = NULL,
timeby = 100, main = "Kaplan-Meier Plot",
pval = TRUE, ...) {
require(ggplot2)
require(survival)
require(gridExtra)
if(is.null(ystratalabs)) {
   ystratalabs <- as.character(levels(summary(sfit)$strata))
}
m <- max(nchar(ystratalabs))
if(is.null(ystrataname)) ystrataname <- "Strata"
times <- seq(0, max(sfit$time), by = timeby)
.df <- data.frame(time = sfit$time, n.risk = sfit$n.risk,
    n.event = sfit$n.event, surv = sfit$surv, strata = summary(sfit, censored = T)$strata,
    upper = sfit$upper, lower = sfit$lower)
levels(.df$strata) <- ystratalabs
zeros <- data.frame(time = 0, surv = 1, strata = factor(ystratalabs, levels=levels(.df$strata)),
    upper = 1, lower = 1)
.df <- rbind.fill(zeros, .df)
d <- length(levels(.df$strata))
p <- ggplot(.df, aes(time, surv, groups = strata)) +
    geom_step(aes(linetype = strata), size = 0.7) +
    theme_bw() +
    opts(axis.title.x = theme_text(vjust = 0.5)) +
    scale_x_continuous(xlabs, breaks = times, limits = c(0, max(sfit$time))) +
    scale_y_continuous(ylabs, limits = c(0, 1)) +
    opts(panel.grid.minor = theme_blank()) +
    opts(legend.position = c(ifelse(m < 10, .28, .35), ifelse(d < 4, .25, .35))) +
    opts(legend.key = theme_rect(colour = NA)) +
    labs(linetype = ystrataname) +
    opts(plot.margin = unit(c(0, 1, .5, ifelse(m < 10, 1.5, 2.5)), "lines")) +
    opts(title = main)

## Create a blank plot for place-holding
## .df <- data.frame()
blank.pic <- ggplot(.df, aes(time, surv)) +
    geom_blank() +
    theme_bw() +
    opts(axis.text.x = theme_blank(), axis.text.y = theme_blank(),
        axis.title.x = theme_blank(), axis.title.y = theme_blank(),
        axis.ticks = theme_blank(), panel.grid.major = theme_blank(),
        panel.border = theme_blank())
if(pval) {
    sdiff <- survdiff(eval(sfit$call$formula), data = eval(sfit$call$data))
    pval <- pchisq(sdiff$chisq, length(sdiff$n) – 1, lower.tail = FALSE)
    pvaltxt <- ifelse(pval < 0.0001, "p < 0.0001", paste("p =", signif(pval, 3)))
    p <- p + annotate("text", x = 0.6 * max(sfit$time), y = 0.1, label = pvaltxt)
}
if(table) {
    ## Create table graphic to include at-risk numbers
    risk.data <- data.frame(strata = summary(sfit, times = times, extend = TRUE)$strata,
        time = summary(sfit, times = times, extend = TRUE)$time,
        n.risk = summary(sfit, times = times, extend = TRUE)$n.risk)
    data.table <- ggplot(risk.data, aes(x = time, y = strata, label = format(n.risk, nsmall = 0))) +
        #, color = strata)) +
        geom_text(size = 3.5) +
        theme_bw() +
        scale_y_discrete(breaks = as.character(levels(risk.data$strata)), labels = ystratalabs) +
        # scale_y_discrete(#format1ter = abbreviate,
        # breaks = 1:3,
        # labels = ystratalabs) +
        scale_x_continuous("Numbers at risk", limits = c(0, max(sfit$time))) +
        opts(axis.title.x = theme_text(size = 10, vjust = 1), panel.grid.major = theme_blank(),
        panel.grid.minor = theme_blank(), panel.border = theme_blank(),
        axis.text.x = theme_blank(), axis.ticks = theme_blank(),
        axis.text.y = theme_text(face = "bold", hjust = 1))
    data.table <- data.table + opts(legend.position = "none") +
        xlab(NULL) + ylab(NULL)
    data.table <- data.table +
        opts(plot.margin = unit(c(-1.5, 1, 0.1, ifelse(m < 10, 2.5, 3.5) – 0.28 * m), "lines"))
## Plotting the graphs
## p <- ggplotGrob(p)
## p <- addGrob(p, textGrob(x = unit(.8, "npc"), y = unit(.25, "npc"), label = pvaltxt,
    ## gp = gpar(size = 12)))
    grid.arrange(p, blank.pic, data.table,
        clip = FALSE, nrow = 3, ncol = 1,
        heights = unit(c(2, .1, .25),c("null", "null", "null")))
    if(returns) {
        a <- arrangeGrob(p, blank.pic, data.table, clip = FALSE,
            nrow = 3, ncol = 1, heights = unit(c(2, .1, .25),c("null", "null", "null")))
        return(a)
    }
}
else {
    ## p <- ggplotGrob(p)
    ## p <- addGrob(p, textGrob(x = unit(0.5, "npc"), y = unit(0.23, "npc"),
    ## label = pvaltxt, gp = gpar(size = 12)))
    print(p)
    if(returns) return(p)
    }
}

To leave a comment for the author, please follow the link and comment on their blog: Stat Bandit » R.

R-bloggers.com offers daily e-mail updates about R news and tutorials about learning R and many other topics. Click here if you're looking to post or find an R/data-science job.
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.