The goal of gghat is to add functionality to ggplot, such that hat
graphs are included in its repertoire. To this end, gghat contains
geom_hat() as well as a compatible position function to be used with
geom_text(). You can install the development version of gghat from
GitHub. Obviously, gghat also
requires that ggplot is installed. In fact, I would recommend that the
user installs the entire tidyverse too, as it contains a host of
useful packages.
# install.packages("devtools")
# library(devtools)
# devtools::install_github("eliashyde/gghat")library(gghat)
library(patchwork)
library(randomizr)
library(estimatr)
library(tidyverse)
library(vayr)Comparative bar graphs — in which bars are paired or otherwise grouped — are useful because they give a strong visual impression. As always, though, we must be conscientious of how we are curating this impression. Starting the y-axis at zero has long since been the golden rule with these types of plots, lest the relative lengths of truncated bars mislead the viewer with a lie factor greater than one (Tufte, 2001). That said, a y-axis that begins at zero may minimize statistically significant treatment effects in comparative bar graphs, rendering them visually insignificant. In a recent paper, Jessica Witt suggests that we derive our y-ranges from standard deviation when possible to increase sensitivity, reducing the discrepancy between actual and perceived effect size (Witt, 2019a).
How, then, might we mitigate the negative effects of an abbreviated y-axis? We could plot only the differences between groups, allowing us to start our y-axis at zero while also limiting it to a reasonable range. In doing so, though, we would lose contextualizing information — namely, the raw values. Or instead of bars, we could plot points, which visually retain only relative position and thus do not suffer from non-zero baselines. But Witt argues that points do not cater to the principles of grouping, failing to “facilitate the perceptual grouping of pairs” (Witt, 2019b). Of course, we could connect the grouped points with lines, but this isn’t ideal either because lines imply continuous variables; they are less suited for discrete data (Zacks & Tversky, 1999).
Witt explains that comparative bar graphs need not fall by the wayside, at least not entirely; they need only be redesigned. Rather than equivocate, negotiating the aforementioned trade-offs, she presents the hat graph, a new visualization of bar-graph descent that allows for “proper interpretation of discrete categories” and the use of “unrestricted settings for the y-axis” (Witt, 2019b). From a pair of bars in a bar graph, a hat graph would retain both heights as horizontal lines; it would also retain a bit of the body of the second bar to represent the difference between the two heights, resulting in a hat shape. The steps are illustrated below. The “brim” of each hat marks one value (e.g., either pre-treatment or without treatment) and the “crown” marks another (e.g., either post-treatment or with treatment), while the profile reflects the difference therebetween (Witt, 2019b). As Tufte would recommend, this gets rid of redundancy. Moreover, as a graphic for difference, the profile has a lie factor of one, while the brim and crown are retained as annotations so as not to lose sight of the actual data values Hat graphs are efficient, requiring very little work on behalf of the viewer — although they are still novel, so it is to be expected that they occasion brief pause.
It is difficult to create hat graphs with ggplot. In essence, the user
must “hack” a stacked and dodged bar graph — and to be clear, a stacked
and dodged bar graph is itself no easy feat within current constraints.
But this modest package, gghat, adds the necessary functionality to
allow the user to quickly and easily make hat graphs. In other words, it
adds hat graphs to ggplot’s repertoire. First and foremost, gghat
introduces geom_hat(), which takes the same parameters as a
pre-summarized comparative bar graph: a discrete x variable, a
quantitative y variable compatible with stat = identity, and a
group variable. Do note that hat graphs generally posit a
pre-treatment or control group on the left, so with respect to the
grouping variable users must take care to order factor levels
appropriately. Then the typical aesthetic parameters apply: fill,
color, alpha, width, linewidth, and so on. The linewidth, 0.75
by default, is of particular importance because it controls the weight
of the hat’s brim.
geom_hat() is based on a position function, dodgediff(), which is
written within the gghat package. That is to say, much of the work
that distinguishes geom_hat() from, say, the geom_bar() or
geom_col() is actually outsourced to this position function. The
benefit of this sort of organization is that variations of the same
position function can be applied to other geoms to make them compatible
with hat graphs. In particular the dodgedifftext() position function
can be used with geom_text(). A nudge argument can be passed to
dodgedifftext(), moving text vertically away from the hats.
my_data <- tibble( # Create summarized data by treatment group, before and after
treatment = c("A", "B", "C", "D"),
pre_treatment_mean = c(50, 49, 48, 50),
post_treatment_mean = c(51, 53, 55, 52)
) |>
pivot_longer(!treatment, names_to = "condition", values_to = "mean") |> # Format as long data
mutate(condition = fct_relevel(condition, "pre_treatment_mean",
"post_treatment_mean")) # Order properly
ggplot(my_data, aes(x = treatment, y = mean, group = condition)) +
geom_hat(width = 0.75, linewidth = 0.75, # Add hat geoms
fill = "black", color = "black", alpha = 0.25) +
geom_text(aes(label = mean), family = "Courier", # And direct labels
position = position_dodgedifftext(width = 0.75, nudge = 0.6)) +
scale_y_continuous(expand = c(0, 0)) +
coord_cartesian(ylim = c(40, 60)) +
theme_minimal() +
theme(
text = element_text(family = "Courier"),
legend.position = "none",
panel.grid = element_blank(),
axis.line = element_line(linewidth = 0.5),
axis.ticks = element_line(linewidth = 0.5),
plot.title = element_text(hjust = 0.5)
) +
xlab("Treatment") +
ylab("Measurement") +
ggtitle("Measurements PRE- and POST-treatment") # An example of just one use for hat graphs
The backwards parameter, which takes a boolean argument, is unique to
gghat. It determines the behavior of geom_hat() given a negative
treatment effect. By default, backwards is equal to FALSE, such that
the leftmost group is the brim of the hat, representing the baseline.
Thus, a negative treatment effect would yield an upside-down hat. This
is recommended, as it’s easier to intuit the underlying experimental
process. But if the user prefers, they can instead set backwards equal
to TRUE, which makes the group with the minimum value the brim, such
that a negative treatment effect would be reflected by a backwards hat.
In this case, if the user is applying the dodgedifftext() position
function to geom_text(), they must also explicitly set backwards
equal to TRUE within these functions to ensure proper behavior. For
instance, text will sit below negative treatment effects when
backwards is FALSE but above all groups when backwards is TRUE. Note
that the nudge parameter for dodgedifftext() is particular in its
ability to handle either case, as it pushes text above the hats upwards
and text below the hats downwards.
my_data <- tibble( # New data wherein treatment B has a NEGATIVE effect
treatment = c("A", "B", "C", "D"),
pre_treatment_mean = c(50, 53, 48, 50),
post_treatment_mean = c(51, 49, 55, 52)
) |>
pivot_longer(!treatment, names_to = "condition", values_to = "mean") |>
mutate(condition = fct_relevel(condition, "pre_treatment_mean",
"post_treatment_mean"),
hue = if_else(treatment == "B", "YES", "NO")) # To highlight backwards hats
forwards <- ggplot(my_data, aes(x = treatment, y = mean, group = condition)) +
geom_hat(aes(fill = hue, color = hue), width = 0.75, backwards = FALSE) + # The default value
geom_text(aes(label = mean), family = "Courier",
position = position_dodgedifftext(width = 0.75, nudge = 0.6)) +
scale_y_continuous(expand = c(0, 0)) +
scale_color_manual(values = c("grey80", "black")) +
scale_fill_manual(values = c("grey80", "black")) +
coord_cartesian(ylim = c(40, 60)) +
theme_minimal() +
theme(
text = element_text(family = "Courier"),
legend.position = "none",
panel.grid = element_blank(),
axis.line = element_line(linewidth = 0.5),
axis.ticks = element_line(linewidth = 0.5),
plot.title = element_text(hjust = 0.5)
) +
xlab("Treatment") +
ylab("Measurement") +
ggtitle("Backwards = FALSE")
backwards <- ggplot(my_data, aes(x = treatment, y = mean, group = condition)) +
geom_hat(aes(fill = hue, color = hue), width = 0.75, backwards = TRUE) + # Backwards here
geom_text(aes(label = mean), family = "Courier", # Backwards parameter used again below
position = position_dodgedifftext(width = 0.75, nudge = 0.6, backwards = TRUE)) +
scale_y_continuous(expand = c(0, 0)) +
scale_color_manual(values = c("grey80", "black")) +
scale_fill_manual(values = c("grey80", "black")) +
coord_cartesian(ylim = c(40, 60)) +
theme_minimal() +
theme(
text = element_text(family = "Courier"),
legend.position = "none",
panel.grid = element_blank(),
axis.line = element_line(linewidth = 0.5),
axis.ticks = element_line(linewidth = 0.5),
plot.title = element_text(hjust = 0.5)
) +
xlab("Treatment") +
ylab("Measurement") +
ggtitle("Backwards = TRUE")
forwards + backwards
Although hat graphs are designed with paired groups in mind,
geom_hat() can tolerate more than two groups without any issue.
backwards can still be FALSE or TRUE per the user’s preference, and
the supplementary position function dodgedifftext() will behave
properly.
my_data <- tibble(
treatment = c("A", "B", "C", "D"),
pre_treatment_mean = c(50, 53, 51, 50),
post_treatment_mean_round1 = c(51, 50, 49, 52),
post_treatment_mean_round2 = c(49, 48, 52, 53) # Add a round of treatment
) |>
pivot_longer(!treatment, names_to = "condition", values_to = "mean") |>
mutate(condition = fct_relevel(condition, "pre_treatment_mean",
"post_treatment_mean_round1",
"post_treatment_mean_round2")) # Always order thoughtfully
ggplot(my_data, aes(x = treatment, y = mean, group = condition)) +
geom_hat(width = 0.75, fill = "black", color = "black", alpha = 0.25) +
geom_text(aes(label = mean), family = "Courier",
position = position_dodgedifftext(width = 0.75, nudge = 0.6)) +
scale_y_continuous(expand = c(0, 0)) +
coord_cartesian(ylim = c(40, 60)) +
theme_minimal() +
theme(
text = element_text(family = "Courier"),
legend.position = "none",
panel.grid = element_blank(),
axis.line = element_line(linewidth = 0.5),
axis.ticks = element_line(linewidth = 0.5),
plot.title = element_text(hjust = 0.5)
) +
xlab("Treatment") +
ylab("Measurement") +
ggtitle("Measurements after MULTIPLE rounds of treatment")
Tufte, E. R. (2001). The Visual Display of Quantitative Information (2nd ed.). Graphics Press.
Witt, J. K. (2019a). “Graph Construction: An Empirical Investigation on Setting the Range of the Y-Axis.” Meta-Psychology, 3(1). https://doi.org/10.31234/osf.io/jmqkx.
—. (2019b). “Introducing Hat Graphs.” Cognitive Research, 4(31). https://doi.org/10.1186/s41235-019-0182-3.
Zacks, J., & Tversky, B. (1999). “Bars and Lines: A Study of Graphic Communication.” Memory & Cognition, 27(6), 1073–1079. https://doi.org/10.3758/BF03201236.