In R ggplot2, include stat_ecdf() endpoints (0,0) and (1,1)

Question

I'm trying to use `stat_ecdf()` to plot cumulative successes as a function of a rank score created by a predictive model. ``` #libraries require(ggplot2) require(scales) # fake data for reproducibility set.seed(123) n <- 200 df <- data.frame(model_score= rexp(n=n,rate=1:n), obs_set= sample(c("training","validation"),n,replace=TRUE)) df$model_rank <- rank(df$model_score)/n df$target_outcome <- rbinom(n,1,1-df$model_rank) # Plot Gain Chart using stat_ecdf() ggplot(subset(df,target_outcome==1),aes(x = model_rank)) + stat_ecdf(aes(colour = obs_set), size=1) + scale_x_continuous(limits=c(0,1), labels=percent,breaks=seq(0,1,.1)) + xlab("Model Percentile") + ylab("Percent of Target Outcome") + scale_y_continuous(limits=c(0,1), labels=percent) + geom_segment(aes(x=0,y=0,xend=1,yend=1), colour = "gray", linetype="longdash", size=1) + ggtitle("Gain Chart") ``` All I want to do is force the ECDF to start at (0,0) and end at (1,1) so that there are no gaps at the beginning or end of the curve. If…

Accepted Answer

Unfortunately, the definition of `stat_ecdf` gives no wiggle room here; it determines the endpoints internally.

There is a somewhat advanced solution. With the latest version of ggplot2 (`devtools::install_github("hadley/ggplot2")`), the extensibility is improved, to the point where it is possible to override this behavior, but not without some boilerplate.

```
stat_ecdf2 <- function(mapping = NULL, data = NULL, geom = "step",
                      position = "identity", n = NULL, show.legend = NA,
                      inherit.aes = TRUE, minval=NULL, maxval=NULL,...) {
  layer(
    data = data,
    mapping = mapping,
    stat = StatEcdf2,
    geom = geom,
    position = position,
    show.legend = show.legend,
    inherit.aes = inherit.aes,
    stat_params = list(n = n, minval=minval,maxval=maxval),
    params = list(...)
  )
}

StatEcdf2 <- ggproto("StatEcdf2", StatEcdf,
  calculate = function(data, scales, n = NULL, minval=NULL, maxval=NULL, ...) {
    df <- StatEcdf$calculate(data, scales, n, ...)
    if (!is.null(minval)) { df$x[1] <- minval }
    if (!is.null(maxval)) { df$x[length(df$x)] <- maxval }
    df
  }
)
```

Now, `stat_ecdf2` will behave the same as `stat_ecdf`, but with an optional `minval` and `maxval` parameter. So this will do the trick:

```
ggplot(subset(df,target_outcome==1),aes(x = model_rank)) +
  stat_ecdf2(aes(colour = obs_set), size=1, minval=0, maxval=1) +
  scale_x_continuous(limits=c(0,1), labels=percent,breaks=seq(0,1,.1)) +
  xlab("Model Percentile") + ylab("Percent of Target Outcome") +
  scale_y_continuous(limits=c(0,1), labels=percent) +
  geom_segment(aes(x=0,y=0,xend=1,yend=1),
               colour = "gray", linetype="longdash", size=1) +
  ggtitle("Gain Chart")
```

The big caveat here is that I don't know if the current extensibility model will be supported in the future; it has changed several times in the past, and the change to use "ggproto" is recent -- like July 15th 2015 recent.

As a plus, this gave me a chance to really dig into ggplot's internals, which is something that I've been meaning to do for a while.

In R ggplot2, include stat_ecdf() endpoints (0,0) and (1,1)

Solution

Problem

Related problems