Number of bins off by one in geom_contour()

[clauswilke]

If we explicitly provide a number of bins, we get one too few. This is particularly striking if we set bins = 2, since that removes all contouring. This example of an empty plot is currently on the ggplot2 website (scroll down to 4th example):
https://ggplot2.tidyverse.org/reference/geom_contour.html#examples

I assume this is a regression introduced in 3.3.0.

library(ggplot2)

p <- ggplot(faithfuld, aes(waiting, eruptions, z = density))

p + geom_contour(bins = 3)

p + geom_contour_filled(bins = 3, aes(fill = after_stat(ordered(piece)))) + 
  scale_fill_viridis_d()

p + geom_contour(bins = 6)

p + geom_contour_filled(bins = 6, aes(fill = after_stat(ordered(piece)))) + 
  scale_fill_viridis_d()

^{Created on 2020-03-09 by the reprex package (v0.3.0)}

[clauswilke]

Ok, this appears to be dataset dependent, so it's likely a bug in scales::fullseq().

library(tidyverse)

volcano %>%
  as_tibble() %>%
  rowid_to_column(var = "x") %>%
  gather(key = "y", value = "z", -1) %>%
  mutate(y = as.numeric(gsub("V", "", y))) %>%
  ggplot(aes(x, y, z = z)) + 
  geom_contour_filled(bins = 5, aes(fill = after_stat(ordered(piece)))) + 
  scale_fill_viridis_d()

^{Created on 2020-03-10 by the reprex package (v0.3.0)}

[clauswilke]

At a minimum, though, the example with bins = 2 needs to be changed.

[thomasp85]

have you dug deeper into this? should an issue in scales be opened?

[clauswilke]

should an issue in scales be opened?

I think so. I gave up because I didn't immediately understand what data input causes it to work or not, but scales::fullseq() is definitely not behaving quite right.

[thomasp85]

Can I get you to do that and close this issue?

[clauswilke]

Actually, looking over everything one more time, it's not clear to me that we can just blame scales::fullseq(). We may not be using it correctly. It does what it promises to do. Because it takes a range and a binwidth as input, it doesn't guarantee a specific number of bins.

In the following reprex, the function make_breaks() encapsulates the logic used by ggplot2. We see that scales::fullseq() does what it is asked to do, but the results aren't necessarily what we want, depending on how the boundaries of the range line up with the binwidth.

library(tidyverse)

volcano_tbl <- volcano %>%
  as_tibble() %>%
  rowid_to_column(var = "x") %>%
  gather(key = "y", value = "z", -1) %>%
  mutate(y = as.numeric(gsub("V", "", y))) 

make_breaks <- function(data, bins) {
  z_range <- range(data, na.rm = TRUE, finite = TRUE)
  binwidth <- diff(z_range) / (bins - 1)
  breaks <- scales::fullseq(z_range, binwidth)
  cat("Range:", z_range, "\n")
  cat("Binwidth", binwidth, "\n")
  cat("Breaks:", breaks, "\n")
  invisible(breaks)
}

# one bin too few
make_breaks(faithfuld$density, 3)  
#> Range: 1.25925e-24 0.0369878 
#> Binwidth 0.0184939 
#> Breaks: 0 0.0184939 0.0369878
make_breaks(faithfuld$density, 5)  
#> Range: 1.25925e-24 0.0369878 
#> Binwidth 0.00924695 
#> Breaks: 0 0.00924695 0.0184939 0.02774085 0.0369878

# shifting the data creates the desired number of bins
make_breaks(faithfuld$density + .01, 3)  
#> Range: 0.01 0.0469878 
#> Binwidth 0.0184939 
#> Breaks: 0 0.0184939 0.0369878 0.0554817
make_breaks(faithfuld$density + .01, 5)  
#> Range: 0.01 0.0469878 
#> Binwidth 0.00924695 
#> Breaks: 0.00924695 0.0184939 0.02774085 0.0369878 0.04623475 0.0554817

# also correct number of bins
make_breaks(volcano_tbl$z, 3)
#> Range: 94 195 
#> Binwidth 50.5 
#> Breaks: 50.5 101 151.5 202
make_breaks(volcano_tbl$z, 5)
#> Range: 94 195 
#> Binwidth 25.25 
#> Breaks: 75.75 101 126.25 151.5 176.75 202

^{Created on 2020-04-30 by the reprex package (v0.3.0)}

[thomasp85]

👍 we’ll fix it in ggplot2 then

[clauswilke]

The question is fundamentally whether we should divide by bins - 1 or bins in this line:

ggplot2/R/stat-contour.r

Line 158 in 53815f8

binwidth <- diff(z_range) / (bins - 1)

A lazy fix would be to divide by bins - 1, calculate the breaks, check them out, and if it's one too few rerun the calculation but divide by bins.

[thomasp85]

I remember fiddling with that in the bin scale as well. I think your suggestion is pragmatic and good

[clauswilke]

Ok. I'll prepare a PR.