How do I use facetting correctly in ggplot geom_tile, while keeping the aspect ratio intact?

Question

I am trying to create a 'likeliness plot' intended to quickly show an items likeliness vs other items in a table.

A quick example:

'property_data.csv' file to use:

"","Country","Town","Property","Property_value"
"1","UK","London","Road_quality","Bad"
"2","UK","London","Air_quality","Very bad"
"3","UK","London","House_quality","Average"
"4","UK","London","Library_quality","Good"
"5","UK","London","Pool_quality","Average"
"6","UK","London","Park_quality","Bad"
"7","UK","London","River_quality","Very good"
"8","UK","London","Water_quality","Decent"
"9","UK","London","School_quality","Bad"
"10","UK","Liverpool","Road_quality","Bad"
"11","UK","Liverpool","Air_quality","Very bad"
"12","UK","Liverpool","House_quality","Average"
"13","UK","Liverpool","Library_quality","Good"
"14","UK","Liverpool","Pool_quality","Average"
"15","UK","Liverpool","Park_quality","Bad"
"16","UK","Liverpool","River_quality","Very good"
"17","UK","Liverpool","Water_quality","Decent"
"18","UK","Liverpool","School_quality","Bad"
"19","USA","New York","Road_quality","Bad"
"20","USA","New York","Air_quality","Very bad"
"21","USA","New York","House_quality","Average"
"22","USA","New York","Library_quality","Good"
"23","USA","New York","Pool_quality","Average"
"24","USA","New York","Park_quality","Bad"
"25","USA","New York","River_quality","Very good"
"26","USA","New York","Water_quality","Decent"
"27","USA","New York","School_quality","Bad"

Code:

prop <- read.csv('property_data.csv')

Property_col_vector <- c("NA" = "#e6194b",
                "Very bad" = "#e6194B",
                "Bad" = "#ffe119",
                "Average" = "#bfef45",
                "Decent" = "#3cb44b",
                "Good" = "#42d4f4",
                "Very good" = "#4363d8")

plot_likeliness <- function(town_property_table){
    g <- ggplot(town_property_table, aes(Property, Town)) +
      geom_tile(aes(fill = Property_value, width=.9, height=.9)) +
      theme_classic() +
      theme(axis.text.x = element_text(angle = 90, hjust = 1, vjust=0.5),
            strip.text.y = element_text(angle = 0)) +
      scale_fill_manual(values = Property_col_vector) +
      coord_fixed()
    return(g)
}

summary_town_plot <- plot_likeliness(prop)

Output:

This is looking great! Now I've created a plot that looks nice because I used the coord_fixed() function, but now I want to create the same plot, facetted by Country.

To do this I created the following function:

plot_likeliness_facetted <- function(town_property_table){
  g <- ggplot(town_property_table, aes(Property, Town)) +
    geom_tile(aes(fill = Property_value, width=.9, height=.9)) +
    theme_classic() +
    theme(axis.text.x = element_text(angle = 90, hjust = 1, vjust=0.5),
          strip.text.y = element_text(angle = 0)) +
    scale_fill_manual(values = Property_col_vector) +
    facet_grid(Country ~ .,
               scale = 'free_y')
  return(g)
}

facetted_town_plot <- plot_likeliness_facetted(prop)
facetted_town_plot

Result:

However, now my tiles are stretched and if i try to use '+ coords_fixed()' I get the error:

Error: coord_fixed doesn't support free scales

How can I get the plot to facet, but maintain the aspect ratio ? Please note that I'm plotting these in a series, so hardcoding the heights of the plot with manual values is not a solution I'm after, I need something that dynamically scales with the amount of values in the table.

Many thanks for any help!

Edit: Although the same question was asked in slightly different context elsewhere, it had multiple answers with none marked as solving the question.

Sophistication

try adding space = "free_y" to facet_grid – Whiten

I did, this does not fix the issue for me. – Sophistication

Basic problem is that most coord_* do not play well with facet_*. See github.com/tidyverse/ggplot2/issues/1492 – Whiten

Well, I know what the issue is, or at least I spotted that it was worked on before: github.com/tidyverse/ggplot2/issues/2651 but this code is marked as merged into ggplot 3.1.1 and marked as resolved, yet it is still not working as intended for me, so i'm looking for a way (even some solution using the height derived from the ggplot object somehow) to set the height dynamically based on the number of elements per facet – Sophistication

If you specify both free scales and free space, then at least the UK and USA tile heights will be the same. + facet_grid(Country ~ ., scales="free_y", space="free_y"). Still not compatible with coord_fixed() but it's a minor improvement in my opinion. – Montague

You said it's not feasible to hardcode the plot dimensions, but maybe a sort of semi-automated compromise would work well enough? Along with the free space and free scales arguments, one could experiment with different additive and multiplicative plot height units, based on the number of rows in the plot. ggsave("test.png", plot=facetted_town_plot, units="in", width=6.5, height=0.5 * length(unique(prop$Town)) + 1.2) – Montague

Well, this is a duplicate: #16258595 – Vibraculum

@M-M Although true, the answer was not accepted and I didn't find it initially when searching for geom_tile. thanks for the pointer though. – Sophistication

Possible duplicate of Facet with free scales but keep aspect ratio fixed – Vibraculum

Answer 1

theme(aspect.ratio = 1) and space = 'free' seems to work.

plot_likeliness_facetted <- function(town_property_table){
  g <- ggplot(town_property_table, aes(Property, Town)) +
    geom_tile(aes(fill = Property_value, width=.9, height=.9)) +
    theme_classic() +
    theme(axis.text.x = element_text(angle = 90, hjust = 1, vjust=0.5),
          strip.text.y = element_text(angle = 0), aspect.ratio = 1) +
    scale_fill_manual(values = Property_col_vector) +
    facet_grid(Country ~ .,
               scale = 'free_y', space = 'free')
  return(g)
}

Marrilee

Answer 2

This might not be a perfect answer, but I'm going to give it a spin anyway. Basically, it is going to be difficult to do this with base ggplot because -as you mentioned- coord_fixed() or theme(aspect.ratio = ...) don't play nice with facets.

The first solution I'll propose, is to use gtables to programatically set the width of panels to match the number of variables on your x-axis:

plot_likeliness_gtable <- function(town_property_table){
  g <- ggplot(town_property_table, aes(Property, Town)) +
    geom_tile(aes(fill = Property_value, width=.9, height=.9)) +
    theme_classic() +
    theme(axis.text.x = element_text(angle = 90, hjust = 1, vjust=0.5),
          strip.text.y = element_text(angle = 0)) +
    scale_fill_manual(values = Property_col_vector) +
    facet_grid(Country ~ .,
               scale = 'free_y', space = "free_y")
  # Here be the gtable bits
  gt <- ggplotGrob(g)
  # Find out where the panel is stored in the x-direction
  panel_x <- unique(gt$layout$l[grepl("panel", gt$layout$name)])[1]
  # Set that width based on the number of x-axis variables, plus 0.2 because
  # of the expand arguments in the scales
  gt$widths[panel_x] <- unit(nlevels(droplevels(town_property_table$Property)) + 0.2, "null")
  # Respect needs to be true to have 'null' units match in x- and y-direction
  gt$respect <- TRUE
  return(gt)
}

Which would work in the following way:

library(grid)
x <- plot_likeliness_gtable(prop)
grid.newpage(); grid.draw(x)

And gives this plot:

This all works reasonably well but at this point, it would probably be good to discuss some of the drawbacks of having gtables instead of ggplot objects. First, you can't edit it anymore with ggplot, so you can't add another + geom_myfavouriteshape() or anything of the sort. You could still edit parts of the plot in gtable/grid though. Second, it has the quirky grid.newpage(); grid.draw() syntax, which needs the grid library. Third, we're kind of relying on the ggplot facetting to set the y-direction panel heights correctly (2.2 and 1.2 null-units in your example) while this might not be appropriate in all cases. On the upside, you're still defining dimensions in flexible null-units, so it'll scale pretty well with whatever plotting device you're using.

The second solution I'll propose could be a bit hacky for many a taste, but it'll take away the first two drawbacks of using gtables. Some time ago, I had similar issues with the weird panel size behaviour when facetting, so I wrote these functions to set panel sizes. The essence of what is does is to copy the panel drawing function from whatever plot you're making and wrap it inside a new function that sets the panel sizes to some pre-defined numbers. It has to be called after any facetting function though. It would work like this:

plot_likeliness_forcedsizes <- function(town_property_table){
  g <- ggplot(town_property_table, aes(Property, Town)) +
    geom_tile(aes(fill = Property_value, width=.9, height=.9)) +
    theme_classic() +
    theme(axis.text.x = element_text(angle = 90, hjust = 1, vjust=0.5),
          strip.text.y = element_text(angle = 0)) +
    scale_fill_manual(values = Property_col_vector) +
    facet_grid(Country ~ .,
               scale = 'free_y', space = "free_y") +
    force_panelsizes(cols = nlevels(droplevels(town_property_table$Property)) + 0.2,
                     respect = TRUE)
  return(g)
}

myplot <- plot_likeliness_forcedsizes(prop)
myplot

It still relies on ggplot setting the y-direction heights correctly though, but you could override these within force_panelsizes() if things go awry.

Hope this helped, good luck!

Telespectroscope

Teun, the link to your magic functions is dead :( – Shipmaster

Thanks for letting me know tjebo! I've re-linked it to a portion of a vignette in ggh4x. At the time of writing that wasn't on CRAN, but now it should be no trouble to install it. – Telespectroscope

ah, oops, I only now realised this is even an exported function in ggh4x ... should have checked first :D – Shipmaster