How to make a generic stat in ggplot2
For a while now I’ve been thinking that, yes, ggplot2
is awesome and offers a lot of geoms
and stats
, but it would be great if it could be extended with new user-generated geoms
and stats
. Then I learnt that ggplot2
actually has a pretty great extension system so I could create my own geoms I needed for my work or just for fun. But still, creating a geom from scratch is an involved process that doesn’t lend itself to simple transformations.
Finally, I thought of a possible solution: create a generic stat
–a tabula rasa, if you will– that can work on the data with any function. Natively ggplot2
offers stat_summary()
, but it’s only meant to be used with, well, summary statistics. What I wanted was something completely generic and this is my first try.
Below is the code for stat_rasa()
(better name pending). It works just like any other stat
except that it works with any function that takes a data.frame and returns a transformed data.frame that can be interpreted by the chosen geom
.
# ggproto object
StatRasa <- ggplot2::ggproto("StatRasa", ggplot2::Stat,
compute_group = function(data, scales, fun, fun.args) {
# Change default arguments of the function to the
# values in fun.args
args <- formals(fun)
for (i in seq_along(fun.args)) {
if (names(fun.args[i]) %in% names(fun.args)) {
args[[names(fun.args[i])]] <- fun.args[[i]]
}
}
formals(fun) <- args
# Apply function to data
fun(data)
})
# stat function used in ggplot
stat_rasa <- function(mapping = NULL, data = NULL,
geom = "point",
position = "identity",
fun = NULL,
...,
show.legend = NA,
inherit.aes = TRUE) {
# Check arguments
if (!is.function(fun)) stop("fun must be a function")
# Pass dotted arguments to a list
fun.args <- match.call(expand.dots = FALSE)$`...`
ggplot2::layer(
data = data,
mapping = mapping,
stat = StatRasa,
geom = geom,
position = position,
show.legend = show.legend,
inherit.aes = inherit.aes,
check.aes = FALSE,
check.param = FALSE,
params = list(
fun = fun,
fun.args = fun.args,
na.rm = FALSE,
...
)
)
}
For example, let’s say we want to quickly glance at detrended data. We then create a very simple function
Detrend <- function(data, method = "lm", span = 0.2) {
if (method == "lm") {
data$y <- resid(lm(y ~ x, data = data))
} else {
data$y <- resid(loess(y ~ x, span = span, data = data))
}
as.data.frame(data)
}
and pass it to stat_rasa()
library(ggplot2)
set.seed(42)
x <- seq(-1, 3, length.out = 30)
y <- x^2 + rnorm(30)*0.5
df <- data.frame(x = x, y = y)
ggplot(df, aes(x, y)) +
geom_line() +
stat_rasa(geom = "line", fun = Detrend, method = "smooth",
color = "steelblue")
We can get better legibility and less typing by creating a wrapper function with a more descriptive name.
stat_detrend <- function(...) {
stat_rasa(fun = Detrend, ...)
}
ggplot(df, aes(x, y)) +
geom_line() +
stat_detrend(method = "lm", color = "blue", geom = "line")
Another case could be calculating contours from an irregular grid. Since ggplot2::stat_contour()
uses grDevices::contourLines()
, it needs values defined in a regular grid, but there’s a package called contoureR
that can compute contours from irregularly spaced observations. With stat_rasa()
we can integrate it with ggplot2
effortlessly by creating a small function and using geom = "path"
.
IrregularContour <- function(data, breaks = scales::fullseq,
binwidth = NULL,
bins = 10) {
if (is.function(breaks)) {
# If no parameters set, use pretty bins to calculate binwidth
if (is.null(binwidth)) {
binwidth <- diff(range(data$z)) / bins
}
breaks <- breaks(range(data$z), binwidth)
}
cl <- contoureR::getContourLines(x = data$x, y = data$y, z = data$z,
levels = breaks)
if (length(cl) == 0) {
warning("Not possible to generate contour data", call. = FALSE)
return(data.frame())
}
cl <- cl[, 3:7]
colnames(cl) <- c("piece", "group", "x", "y", "level")
return(cl)
}
stat_contour_irregular <- function(...) {
stat_rasa(fun = IrregularContour, geom = "path", ...)
}
set.seed(42)
df <- data.frame(x = rnorm(500),
y = rnorm(500))
df$z <- with(df, -x*y*exp(-x^2 - y^2))
ggplot(df, aes(x, y)) +
geom_point(aes(color = z)) +
stat_contour_irregular(aes(z = z, color = ..level..), bins = 15) +
scale_color_viridis_c()
And voilà.
There’s always things to improve. For example, the possibility of using a custom function to compute parameters that depend on the data, but I believe that as it stands covers 80% of simple applications. I should also use a better name, but naming things is hard work.