--- title: "Visualize as you randomize" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{Visualize as you randomize} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>", out.width = "100%" ) ``` The goal of `vayr` is to provide `ggplot2` extensions that foster "visualize as you randomize" principles. These principles are outlined in detail in "Visualize As You Randomize: Design-based Statistical Graphs for Randomized Experiments," a chapter in *Advances in Experimental Political Science* ([PDF](https://alexandercoppock.com/coppock_2021.pdf), [DOI](https://doi.org/10.1017/9781108777919.022)). The package includes position adjustments that avoid over-plotting, which helps organize "data-space" to better contextualize statistical models. ## Installation The release version of `vayr` can be installed from CRAN, and the development version can be installed from [GitHub](https://github.com/) using a package like `remotes`, `devtools`, or `pak`. `vayr` relies on `ggplot2`, `packcircles`, and `withr`, so these must be installed as well. ``` r # From CRAN install.packages("vayr") # From GitHub # install.packages("pak") pak::pak("acoppock/vayr") ``` ## Contents `vayr` contains a handful of `ggplot2` functions that apply as position adjustments to "point-like" geoms such as `geom_point` or `geom_text`: + `position_jitter_ellipse()` and `position_jitterdodge_ellipse()` + `position_bluenoise()` and `position_bluenoisedodge()` + `position_sunflower()` and `position_sunflowerdodge()` + `position_honeycomb()` and `position_honeycombdodge()` + `position_circlepack()` and `position_circlepackdodge()` These functions avoid over-plotting, so they are especially useful when plotting discrete rather than continuous data. To demonstrate, we use them below to visualize synthetic data, over-plotted at the origin. ```{r, echo = FALSE} suppressMessages(library(dplyr)) ``` ```{r setup} library(dplyr) library(estimatr) library(ggplot2) library(patchwork) library(vayr) set.seed(1) dat <- data.frame( x = c(rep(0, 200)), y = c(rep(0, 200)), group = (rep(c("A", "B", "B", "B"), 50)), size = runif(200, 0, 1) ) ``` If position is the product of discrete variables alone, then over-plotting is of particular concern. `position_jitter()` can mitigate it. It introduces variation by randomly sampling points on a rectangle. This approach is effective but can be unattractive. The position adjustments in `vayr` aim to do better. ```{r contents_0, dpi = 96, fig.width = 12, fig.height = 6, fig.alt = 'perfect over-plotting and position_jitter()'} # perfectly over-plotted points over_plot <- ggplot(dat, aes(x = x, y = y)) + geom_point() + coord_equal(xlim = c(-1.1, 1.1), ylim = c(-1.1, 1.1)) + theme_bw() + theme(axis.title = element_blank(), plot.title = element_text(hjust = 0.5, face = "bold")) + ggtitle('"perfect over-plotting"') # position_jitter() jitter_plot <- ggplot(dat, aes(x = x, y = y)) + geom_point(position = position_jitter(width = 0.5, height = 0.5)) + coord_equal(xlim = c(-1.1, 1.1), ylim = c(-1.1, 1.1)) + theme_bw() + theme(axis.title = element_blank(), plot.title = element_text(hjust = 0.5, face = "bold")) + ggtitle("position_jitter()") over_plot + jitter_plot ``` ### Position Jitter Ellipse `position_jitter_ellipse()` adds elliptical random noise to perfectly over-plotted points, offering a pleasing way to visualize many points that represent the same position. The benefit of sampling on an ellipse of a given `height` and `width` rather than on a rectangle is that the resulting dispersion retains the impression of a single point. The size of the ellipses stays constant, while their density varies depending on the amount of data. ```{r contents_1, dpi = 96, fig.width = 12, fig.height = 6, fig.alt = 'position_jitter_ellipse() and position_jitterdodge_ellipse()'} # position_jitter_ellipse() jitter_ellipse_plot <- ggplot(dat, aes(x = x, y = y)) + geom_point(position = position_jitter_ellipse(width = 0.5, height = 0.5)) + coord_equal(xlim = c(-1.1, 1.1), ylim = c(-1.1, 1.1)) + theme_bw() + theme(axis.title = element_blank(), plot.title = element_text(hjust = 0.5, face = "bold")) + ggtitle("position_jitter_ellipse()") # position_jitterdodge_ellipse() jitterdodge_ellipse_plot <- ggplot(dat, aes(x = x, y = y, color = group)) + geom_point(position = position_jitterdodge_ellipse(dodge.width = 2, jitter.width = 0.5, jitter.height = 0.5)) + coord_equal(xlim = c(-1.1, 1.1), ylim = c(-1.1, 1.1)) + theme_bw() + theme(legend.position = "none", axis.title = element_blank(), plot.title = element_text(hjust = 0.5, face = "bold")) + ggtitle("position_jitterdodge_ellipse()") jitter_ellipse_plot + jitterdodge_ellipse_plot ``` ### Position Blue Noise `position_bluenoise()` fills the same elliptical field as `position_jitter_ellipse()`, but places the points so that no two land much closer together than the rest. Sampling uniformly at random, which is what jittering does, leaves visible knots and voids: in a draw of 250 points the closest pair typically sits about a fifteenth of the median spacing apart, and a reader cannot tell those knots from real structure. The arrangement here has the even spacing of a sunflower while still looking unstructured, so nobody mistakes a spiral arm for a finding. It is the pattern the eye's own photoreceptors are laid out in. ```{r contents_1B, dpi = 96, fig.width = 12, fig.height = 6, fig.alt = 'position_bluenoise() and position_bluenoisedodge()'} # position_bluenoise() bluenoise_plot <- ggplot(dat, aes(x = x, y = y)) + geom_point(position = position_bluenoise(width = 0.5, height = 0.5)) + coord_equal(xlim = c(-1.1, 1.1), ylim = c(-1.1, 1.1)) + theme_bw() + theme(axis.title = element_blank(), plot.title = element_text(hjust = 0.5, face = "bold")) + ggtitle("position_bluenoise()") # position_bluenoisedodge() bluenoisedodge_plot <- ggplot(dat, aes(x = x, y = y, color = group)) + geom_point(position = position_bluenoisedodge(dodge.width = 2, scatter.width = 0.5, scatter.height = 0.5)) + coord_equal(xlim = c(-1.1, 1.1), ylim = c(-1.1, 1.1)) + theme_bw() + theme(legend.position = "none", axis.title = element_blank(), plot.title = element_text(hjust = 0.5, face = "bold")) + ggtitle("position_bluenoisedodge()") bluenoise_plot + bluenoisedodge_plot ``` ### Position Sunflower `position_sunflower()` arranges perfectly over-plotted points using a sunflower algorithm, which produces a pattern that resembles the seeds of a sunflower, working from the inside out in the order of the data. The parameters for this position adjustment are `density` and `aspect_ratio`. The size of the flowers varies depending on the amount of over-plotting, but the density of the pattern remains constant. A point with nothing over-plotting it stays where it is. We generally recommend pairing the position adjustment with `coord_equal()`, in which case the default aspect ratio of 1 yields perfectly circular flowers, but the aspect ratio of the flowers can be adjusted if need be. ```{r contents_2A, dpi = 96, fig.width = 12, fig.height = 6, fig.alt = 'position_sunflower() and position_sunflowerdodge()'} # position_sunflower() sunflower_plot <- ggplot(dat, aes(x = x, y = y)) + geom_point(position = position_sunflower(density = 1, aspect_ratio = 1)) + coord_equal(xlim = c(-2.1, 2.1), ylim = c(-2.1, 2.1)) + theme_bw() + theme(axis.title = element_blank(), plot.title = element_text(hjust = 0.5, face = "bold")) + ggtitle("position_sunflower()") # position_sunflowerdodge() sunflowerdodge_plot <- ggplot(dat, aes(x = x, y = y, color = group)) + geom_point(position = position_sunflowerdodge(width = 4, density = 1, aspect_ratio = 1)) + coord_equal(xlim = c(-2.1, 2.1), ylim = c(-2.1, 2.1)) + theme_bw() + theme(legend.position = "none", axis.title = element_blank(), plot.title = element_text(hjust = 0.5, face = "bold")) + ggtitle("position_sunflowerdodge()") sunflower_plot + sunflowerdodge_plot ``` The `density` parameter controls the density of the pattern. A density of 1 is normalized to 100 points in a unit circle; a density of 2, 200 points; and a density of 0.5, 50 points. Because density is normalized relative to Cartesian units, its visual effect depends on the ranges of the axes and the dimensions of the saved image. Smaller ranges or larger dimensions require a greater density to produce the same visual effect. Point size matters too. ``` {r contents_2B, dpi = 96, fig.width = 12, fig.height = 12, fig.alt = 'density', echo = FALSE} densities <- rep(c(0.5, 1, 2), each = 3) ns <- rep(c(50, 100, 200), 3) density_plots <- list() for (i in 1:9) { density_dat <- data.frame(x = c(rep(0, ns[i])), y = c(rep(0, ns[i]))) density_plots[[i]] <- ggplot(density_dat, aes(x, y)) + geom_point(position = position_sunflower(density = densities[i])) + coord_equal(xlim = c(-2, 2), ylim = c(-2, 2)) + theme_bw() + labs(title = paste0("n = ", ns[i], ", density = ", densities[i])) + theme(axis.title = element_blank(), panel.grid.minor = element_blank(), plot.title = element_text(face = "bold")) } wrap_plots(density_plots, ncol = 3) ``` The `aspect_ratio` parameter changes the aspect ratio of the flowers, which is their width divided by their height. The parameter earns its keep when the position adjustment is used without `coord_equal()`. The flowers can be made wider or taller to compensate for the aspect ratio of the axes or the image. Set `aspect_ratio` to the reciprocal of the distortion you are correcting: flowers that render twice as wide as they are tall need an `aspect_ratio` of 0.5. For instance, consider a plot with an x axis that ranges from 0 to 1, and a y axis that ranges from 0 to 2. Saving this plot as a square image would squish the y axis, resulting in flowers twice as wide as they are tall. An `aspect_ratio` of 0.5 offsets that distortion. Under `coord_fixed()` the arithmetic is simpler, because `ratio` already expresses the distortion you need to undo: set `aspect_ratio` to the same value as `ratio` and the flowers come out circular. The two parameters agree in value while being defined in opposite directions, since `aspect_ratio` is width to height and `ratio` is height to width. The grid below crosses the two, and the circular flowers appear where the values match, running from the top right to the bottom left. ``` {r contents_2C, dpi = 96, fig.width = 12, fig.height = 12, fig.alt = 'aspect_ratio', echo = FALSE} flower_ratios <- rep(c(0.5, 1, 2), each = 3) axis_ratios <- rep(c(2, 1, 0.5), 3) aspect_ratio_dat <- data.frame(x = c(rep(0, 100)), y = c(rep(0, 100))) aspect_ratio_plots <- list() for (i in 1:9) { aspect_ratio_plots[[i]] <- ggplot(aspect_ratio_dat, aes(x, y)) + geom_point(position = position_sunflower(aspect_ratio = flower_ratios[i])) + coord_fixed(xlim = c(-2, 2), ylim = c(-2, 2), ratio = axis_ratios[i]) + theme_bw() + labs(title = paste0("aspect_ratio = ", flower_ratios[i], "\ncoord_fixed(ratio = ", axis_ratios[i], ")")) + theme(axis.title = element_blank(), panel.grid.minor = element_blank(), plot.title = element_text(face = "bold", size = 10)) } wrap_plots(aspect_ratio_plots, ncol = 3) ``` ### Position Honeycomb `position_honeycomb()` arranges the same points on a hexagonal lattice, the densest packing of equal circles in the plane. It takes the same `density` and `aspect_ratio` as `position_sunflower()` and covers the same footprint at the same density, so the two are interchangeable and the choice between them is about looks: the lattice reads as countable and orderly, the spiral as organic and without a preferred direction. `position_beeswarm()` in the 'ggbeeswarm' package also offers a hexagonal method, and does a different job. A beeswarm spreads points along one axis to show the shape of a distribution, so perfectly over-plotted points come out as a line rather than a cluster, and its hexagonal and square methods move points off their true value on the data axis. Reach for a beeswarm to show a distribution, and for this to show a count. ```{r contents_2D, dpi = 96, fig.width = 12, fig.height = 6, fig.alt = 'position_honeycomb() and position_honeycombdodge()'} # position_honeycomb() honeycomb_plot <- ggplot(dat, aes(x = x, y = y)) + geom_point(position = position_honeycomb(density = 1, aspect_ratio = 1)) + coord_equal(xlim = c(-2.1, 2.1), ylim = c(-2.1, 2.1)) + theme_bw() + theme(axis.title = element_blank(), plot.title = element_text(hjust = 0.5, face = "bold")) + ggtitle("position_honeycomb()") # position_honeycombdodge() honeycombdodge_plot <- ggplot(dat, aes(x = x, y = y, color = group)) + geom_point(position = position_honeycombdodge(width = 4, density = 1, aspect_ratio = 1)) + coord_equal(xlim = c(-2.1, 2.1), ylim = c(-2.1, 2.1)) + theme_bw() + theme(legend.position = "none", axis.title = element_blank(), plot.title = element_text(hjust = 0.5, face = "bold")) + ggtitle("position_honeycombdodge()") honeycomb_plot + honeycombdodge_plot ``` ### Position Circle Pack `position_circlepack()` uses a circle packing algorithm from `packcircles` to arrange perfectly over-plotted points of varying sizes into an elliptical area. It also takes `density` and `aspect_ratio` as parameters. Do not confuse it with `geom_circlepack()` from `ggcirclepack`, which can be found on [GitHub](https://github.com/EvaMaeRey/ggcirclepack). ```{r contents_3A, dpi = 96, fig.width = 12, fig.height = 6, fig.alt = 'position_circlepack() and position_circlepackdodge()'} # position_circlepack() circlepack_plot <- ggplot(dat, aes(x = x, y = y, size = size)) + geom_point(alpha = 0.25, position = position_circlepack(density = 0.25, aspect_ratio = 1)) + coord_equal(xlim = c(-1, 1), ylim = c(-1.1, 1.1)) + theme_bw() + theme(legend.position = "none", axis.title = element_blank(), plot.title = element_text(hjust = 0.5, face = "bold")) + ggtitle("position_circlepack()") # position_circlepackdodge() circlepackdodge_plot <- ggplot(dat, aes(x = x, y = y, color = group, size = size)) + geom_point(alpha = 0.25, position = position_circlepackdodge(width = 2, density = 0.25, aspect_ratio = 1)) + coord_equal(xlim = c(-1, 1), ylim = c(-1.1, 1.1)) + theme_bw() + theme(legend.position = "none", axis.title = element_blank(), plot.title = element_text(hjust = 0.5, face = "bold")) + ggtitle("position_circlepackdodge()") circlepack_plot + circlepackdodge_plot ``` Like `position_sunflower()`, `position_circlepack()` works from the inside out in the order of the data. Arranging the data by size therefore organizes the points accordingly. ```{r contents_3B, dpi = 96, fig.width = 12, fig.height = 4, fig.alt = 'random, ascending, descending'} # random size, base plot random <- ggplot(dat, aes(x = x, y = y, size = size)) + geom_point(alpha = 0.25, position = position_circlepack(density = 0.075, aspect_ratio = 1)) + coord_equal(xlim = c(-1, 1), ylim = c(-1.1, 1.1)) + theme_bw() + theme(legend.position = "none", axis.title = element_blank(), plot.title = element_text(hjust = 0.5, face = "bold")) + ggtitle("random") # ascending size ascending <- random %+% arrange(dat, size) + ggtitle("ascending") # descending size descending <- random %+% arrange(dat, desc(size)) + ggtitle("descending") random + ascending + descending ``` ## Example `vayr` also includes data from the Patriot Act experiment described in [*Persuasion in Parallel*](https://alexandercoppock.com/coppock_2022.html). The Patriot Act was an anti-terrorism law, and the `patriot_act` dataset comes from an experiment that measured support for this law after randomly exposing participants to statements that cast the legislation in either a negative or positive light. The experiment was conducted in 2009 with a nationwide sample, and it was replicated in 2015 with a sample of MTurkers. In both instances, the treatments had a similar effect on Democrats and Republicans. The data hold four variables: + `sample_label`, the study to which the participant belonged + `pid_3`, the partisanship of the participant + `T1_content`, the statements to which the participant was exposed + `PA_support`, the participant's post-treatment support for the Patriot Act The figure below visualizes the data using `position_sunflowerdodge()` from `vayr`. It adjusts both `density` and `aspect_ratio`: a high `density` compensates for the small point size, and a tall `aspect_ratio` compensates for the wide plot. ```{r patriot_act_visualization, dpi = 96, fig.width = 12, fig.height = 6, fig.alt = "patriot_act"} # A df for statistical models summary_df <- patriot_act |> group_by(T1_content, pid_3, sample_label) |> reframe(tidy(lm_robust(PA_support ~ 1))) # A df for direct labels label_df <- summary_df |> filter(sample_label == "Original Study", T1_content == "Control") |> mutate( PA_support = case_when( pid_3 == "Democrat" ~ conf.low - 0.15, pid_3 == "Republican" ~ conf.high + 0.15 ) ) ggplot(patriot_act, aes(T1_content, PA_support, color = pid_3, group = pid_3)) + # the data geom_point(position = position_sunflowerdodge(width = 0.5, density = 50, aspect_ratio = 0.5), size = 0.1, alpha = 0.5) + # the statistical model geom_line(data = summary_df, aes(x = T1_content, y = estimate), position = position_dodge(width = 0.5), linewidth = 0.5) + geom_point(data = summary_df, aes(x = T1_content, y = estimate), position = position_dodge(width = 0.5), size = 3) + geom_linerange(data = summary_df, aes(x = T1_content, y = estimate, ymin = conf.low, ymax = conf.high), position = position_dodge(width = 0.5)) + # the direct labels geom_text(data = label_df, aes(label = pid_3)) + # the rest scale_color_manual(values = c("blue4", "red3")) + scale_y_continuous(breaks = 1:7) + coord_fixed(ratio = 0.5) + # ratio for coord_fixed is y/x rather than x/y facet_wrap(~sample_label) + theme_bw() + theme(legend.position = "none", strip.background = element_blank(), panel.grid.minor = element_blank()) + labs(y = "Do you oppose or support the Patriot Act? [1: Oppose very strongly to 7: Support very strongly]", x = "Randomly assigned information") ``` The figure shows the design, the data, and the analysis at once. Each point is one respondent, arranged in flowers so that the number of subjects sitting on each of the seven scale points stays visible. The lines and vertical bars are group means with their 95 percent confidence intervals. Republicans support the Patriot Act more than Democrats do, by about a point on the seven-point scale, in both the original study and the replication. The treatments move the two groups by similar amounts and in the same direction. In the original study, pro-Patriot Act statements raise support by 0.69 points (robust standard error: 0.27) among Democrats and by 0.57 (0.30) among Republicans, while anti-Patriot Act statements lower it by 0.61 (0.26) and 0.84 (0.32). The replication reproduces the pattern at slightly smaller magnitudes. The lines run roughly parallel.