How to apply separate coord_cartesian() to "zoom in" into individual panels of a facet_grid()?

Question

Inspired by the Q Finding the elbow/knee in a curve I started to play around with smooth.spline().

In particular, I want to visualize how the parameter df (degree of freedom) influences the approximation and the first and second derivative. Note that this Q is not about approximation but about a specific problem (or edge case) in visualisation with ggplot2.

First attempt: simple facet_grid()

library(ggplot2)
ggplot(ap, aes(x, y)) +
  geom_point(data = dp, alpha = 0.2) +
  geom_line() + 
  facet_grid(deriv ~ df, scales = "free_y", labeller = label_both) + 
  theme_bw()

dp is a data.table containing the data points for which an approximation is sought and ap is a data.table with the approximated data plus the derivatives (data are given below).

For each row, facet_grid() with scales = "free_y" has choosen a scale which displays all data. Unfortunately, one panel has kind of "outliers" which make it difficult to see details in the other panels. So, I want to "zoom in".

"Zoom in" using coord_cartesian()

ggplot(ap, aes(x, y)) +
  geom_point(data = dp, alpha = 0.2) +
  geom_line() + 
  facet_grid(deriv ~ df, scales = "free_y", labeller = label_both) + 
  theme_bw() +
  coord_cartesian(ylim = c(-200, 50))

With the manually selected range, more details in the panels of row 3 have been made visible. But, the limit has been applied to all panels of the grid. So, in row 1 details hardly can been distinguished.

What I'm looking for is a way to apply coord_cartesian() with specific parameters separately to each individual panel (or group of panels, e.g., rowwise) of the grid. For instance, is it possible to manipulate the ggplot object afterwards?

Workaround: Combine individual plots with cowplot

As a workaround, we can create three separate plots and combine them afterwards using the cowplot package:

g0 <- ggplot(ap[deriv == 0], aes(x, y)) +
  geom_point(data = dp, alpha = 0.2) +
  geom_line() + 
  facet_grid(deriv ~ df, scales = "free_y", labeller = label_both) + 
  theme_bw()

g1 <- ggplot(ap[deriv == 1], aes(x, y)) +
  geom_line() + 
  facet_grid(deriv ~ df, scales = "free_y", labeller = label_both) + 
  theme_bw() +
  coord_cartesian(ylim = c(-50, 50))

g2 <- ggplot(ap[deriv == 2], aes(x, y)) +
  geom_line() + 
  facet_grid(deriv ~ df, scales = "free_y", labeller = label_both) + 
  theme_bw() +
  coord_cartesian(ylim = c(-200, 100))

cowplot::plot_grid(g0, g1, g2, ncol = 1, align = "v")

Unfortunately, this solution

• requires to write code to create three separate plots,
• duplicates strips and axes and adds whitespace which isn't available for display of the data.

Is facet_wrap() an alternative?

We can use facet_wrap() instead of facet_grid():

ggplot(ap, aes(x, y)) +
  # geom_point(data = dp, alpha = 0.2) + # this line causes error message
  geom_line() + 
  facet_wrap(~ deriv + df, scales = "free_y", labeller = label_both, nrow = 3) + 
  theme_bw()

Now, the y-axes of every panel are scaled individually exhibiting details of some of the panels. Unfortunately, we still can't "zoom in" into the bottom right panel because using coord_cartesian() would affect all panels.

In addition, the line

geom_point(data = dp, alpha = 0.2)

strangely causes

Error in gList(list(x = 0.5, y = 0.5, width = 1, height = 1, just = "centre", : only 'grobs' allowed in "gList"

I had to comment this line out, so the the data points which are to be approximated are not displayed.

Data

library(data.table)
# data points
dp <- data.table(
  x = c(6.6260, 6.6234, 6.6206, 6.6008, 6.5568, 6.4953, 6.4441, 6.2186,
        6.0942, 5.8833, 5.7020, 5.4361, 5.0501, 4.7440, 4.1598, 3.9318,
        3.4479, 3.3462, 3.1080, 2.8468, 2.3365, 2.1574, 1.8990, 1.5644,
        1.3072, 1.1579, 0.95783, 0.82376, 0.67734, 0.34578, 0.27116, 0.058285),
  y = 1:32,
  deriv = 0)
# approximated data points and derivatives
ap <- rbindlist(
  lapply(seq(2, length(dp$x), length.out = 4),
         function(df) {
           rbindlist(
             lapply(0:2, 
                    function(deriv) {
                      result <- as.data.table(
                        predict(smooth.spline(dp$x, dp$y, df = df), deriv = deriv))
                      result[, c("df", "deriv") := list(df, deriv)]
                    })
           )
         })
)  

Answer 1

Late answer, but the following hack just occurred to me. Would it work for your use case?

Step 1. Create an alternative version of the intended plot, limiting the range of y values such that scales = "free_y" gives a desired scale range for each facet row. Also create the intended facet plot with the full data range:

library(ggplot2)
library(dplyr)

# alternate plot version with truncated data range
p.alt <- ap %>%
  group_by(deriv) %>%
  mutate(upper = quantile(y, 0.75),
         lower = quantile(y, 0.25),
         IQR.multiplier = (upper - lower) * 10) %>%
  ungroup() %>%
  mutate(is.outlier = y < lower - IQR.multiplier | y > upper + IQR.multiplier) %>%
  mutate(y = ifelse(is.outlier, NA, y)) %>%

  ggplot(aes(x, y)) +
  geom_point(data = dp, alpha = 0.2) +
  geom_line() + 
  facet_grid(deriv ~ df, scales = "free_y", labeller = label_both) + 
  theme_bw()

# intended plot version with full data range
p <- p.alt %+% ap

Step 2. Use ggplot_build() to generate plot data for both ggplot objects. Apply the panel parameters of the alt version onto the intended version:

p <- ggplot_build(p)
p.alt <- ggplot_build(p.alt)

p$layout$panel_params <- p.alt$layout$panel_params
rm(p.alt)

Step 3. Build the intended plot from the modified plot data, & plot the result:

p <- ggplot_gtable(p)

grid::grid.draw(p)

Note: in this example, I truncated the data range by setting all values more than 10*IQR away from the upper / lower quartile in each facet row as NA. This can be replaced by any other logic for defining outliers.