Jitter text/labels with position_stack

Question

Consider the following data.frame and chart:

library(ggplot2)
library(scales)
df <- data.frame(L=rep(LETTERS[1:2],each=4),
                 l=rep(letters[1:4],2),
                 val=c(96.5,1,2,0.5,48,0.7,0.3,51))
#   L l  val
# 1 A a 96.5
# 2 A b  1.0
# 3 A c  2.0
# 4 A d  0.5
# 5 B a 48.0
# 6 B b  0.7
# 7 B c  0.3
# 8 B d 51.0

ggplot(df,aes(x=L,y=val,fill=l)) +
  geom_bar(stat="identity") +
  geom_text(aes(label=percent(val/100)),position=position_stack(vjust =0.5))

Some labels are hard to read due to small values. I'd like to jitter those vertically. I'm aware of position_jitter but it doesn't seem compatible with a stacked bar chart.

Answer 1

We can create a new Position, position_jitter_stack().

 position_jitter_stack <- function(vjust = 1, reverse = FALSE, 
                                  jitter.width = 1, jitter.height = 1,
                                  jitter.seed = NULL, offset = NULL) {
  ggproto(NULL, PositionJitterStack, vjust = vjust, reverse = reverse, 
          jitter.width = jitter.width, jitter.height = jitter.height,
          jitter.seed = jitter.seed, offset = offset)
}

PositionJitterStack <- ggproto("PositionJitterStack", PositionStack,
  type = NULL,
  vjust = 1,
  fill = FALSE,
  reverse = FALSE,
  jitter.height = 1,
  jitter.width = 1,
  jitter.seed = NULL,
  offset = 1,

  setup_params = function(self, data) {
    list(
      var = self$var %||% ggplot2:::stack_var(data),
      fill = self$fill,
      vjust = self$vjust,
      reverse = self$reverse,
      jitter.height = self$jitter.height,
      jitter.width = self$jitter.width,
      jitter.seed = self$jitter.seed,
      offset = self$offset
    )
  },

  setup_data = function(self, data, params) {
    data <- PositionStack$setup_data(data, params)
    if (!is.null(params$offset)) {
      data$to_jitter <- sapply(seq(nrow(data)), function(i) {
        any(abs(data$y[-i] - data$y[i]) <= params$offset)
      })
    } else {
      data$to_jitter <- TRUE
      }
    data
  },

  compute_panel = function(data, params, scales) {
    data <- PositionStack$compute_panel(data, params, scales)

    jitter_df <- data.frame(width = params$jitter.width,
                            height = params$jitter.height)

    if (!is.null(params$jitter.seed)) jitter_df$seed = params$jitter.seed
    jitter_positions <- PositionJitter$compute_layer(
      data[data$to_jitter, c("x", "y")],
      jitter_df
    )

    data$x[data$to_jitter] <- jitter_positions$x
    data$y[data$to_jitter] <- jitter_positions$y

    data
  }
)

And plot it ...

ggplot(df,aes(x=L,y=val,fill=l)) +
  geom_bar(stat="identity") +
  geom_text(aes(label=percent(val/100)),
            position = position_jitter_stack(vjust =0.5,
             jitter.height = 0.1,
             jitter.width =  0.3, offset = 1))

Alternatively, we could write a very simple repel function.

library(rlang)

position_stack_repel <- function(vjust = 1, reverse = FALSE, 
                                 offset = 1) {
  ggproto(NULL, PositionStackRepel, vjust = vjust, reverse = reverse,
          offset = offset)
}

PositionStackRepel <- ggproto("PositionStackRepel", PositionStack,
  type = NULL,
  vjust = 1,
  fill = FALSE,
  reverse = FALSE,
  offset = 1,

  setup_params = function(self, data) {
    list(
      var = self$var %||% ggplot2:::stack_var(data),
      fill = self$fill,
      vjust = self$vjust,
      reverse = self$reverse,
      offset = self$offset
    )
  },

  setup_data = function(self, data, params) {
    data <- PositionStack$setup_data(data, params)
    data <- data[order(data$x), ]
    data$to_repel <- unlist(by(data, data$x, function(x) {
      sapply(seq(nrow(x)), function(i) {
        (x$y[i]) / sum(x$y) < 0.1 & (
          (if (i != 1) (x$y[i-1] / sum(x$y)) < 0.1 else FALSE) | (
            if (i != nrow(x)) (x$y[i+1] / sum(x$y)) < 0.1 else FALSE))
      })
    }))
    data
  },

  compute_panel = function(data, params, scales) {
    data <- PositionStack$compute_panel(data, params, scales)
    data[data$to_repel, "x"] <- unlist(
      by(data[data$to_repel, ], data[data$to_repel, ]$x, 
         function(x) seq(x$x[1] - 0.3, x$x[1] + 0.3, length.out = nrow(x))))
    data
  }
)

Plot it:

ggplot(df,aes(x=L,y=val,fill=l)) +
  geom_bar(stat="identity") +
  geom_text(aes(label=percent(val/100)),
            position = position_stack_repel(vjust =0.5))

Answer 2

I found 2 solutions that involve computing the base position of labels beforehand, one using position_jitter and one using ggrepel (suggested by user @gfgm in deleted answer)

create positions:

Note that I need to put NAs first here so I used: How to have NA's displayed first using arrange()

library(dplyr)
df <- df %>%
  group_by(L) %>%
  arrange(!is.na(l), desc(l)) %>% 
  mutate(pos = cumsum(val) - val/2)) # the -val/2 is to center the text

position_jitter solution

set.seed(2)
ggplot(df,aes(x=L,y=val,fill=l)) +
  geom_bar(stat="identity") +
  geom_text(aes(y=pos,label=percent(val/100)),position = position_jitter(width = 0,height=4))

ggrepel solution

library(ggrepel)
ggplot(df,aes(x=L,y=val,fill=l)) +
  geom_bar(stat="identity") +
  geom_text_repel(aes(y=pos,label=percent(val/100)),direction="y",box.padding=0)

comparison of both

ggrepel solution doesn't require manual calibration, the output isn't perfect but it's consistent, it also has great flexibility though and would be the solution of choice for most variants of my issue. Note that geom_text_repel has a seed parameter, but in my case it doesn't affect the results.

position_jitter doesn't give consistent result, positions are randomized, and for most cases it's a less good solution as text overlays (I think it's jittering as if we were dealing with points). For a given chart though it can give a better solution than ggrepel using set.seed beforehand, so maybe better for some reporting, worse the rest of the time.

If geom_text_repel supported position_stack I wouldn't have to go through the pain of the first step, but it doesn't unfortunately.

Both solutions have the slightly annoying effect of jittering isolated labels that shouldn't be jittered at all (this issue is handled by @erocoar's solution).