--- title: "Box Plots and Relations" output: html_document: toc: yes --- ```{r global_options, include = FALSE} knitr::opts_chunk$set(collapse = TRUE) ``` ## Boxplots Boxplots, or box-and-whisker plots, provide a skeletal representation of a distribution. They are very well suited for showing distributions for multiple groups. There are many variations of boxplots: * Most start with a box from the first to the third quartiles and divided by the median. * The simplest form then adds a whisker from the lower quartile to the minimum and from the upper quartile to the maximum. * More common is to draw the upper whisker to the largest point below the upper quartile $+ 1.5 * IQR$, and the lower whisker analogously. * _Outliers_ falling outside the range of the whiskers are then drawn directly: ```{r} library(gapminder) library(ggplot2) ggplot(gapminder) + geom_boxplot(aes(x = continent, y = gdpPercap)) ``` * There are variants that distinguish between _mild outliers_ and _extreme outliers_. * A common variant is to show an approximate 95% confidence interval for the population median as a _notch_: ```{r} ggplot(gapminder) + geom_boxplot(aes(x = continent, y = gdpPercap), notch = TRUE) ``` * Another variant is to use a width proportional to the square root of the sample size to reflect the strength of evidence in the data: ```{r} ggplot(gapminder) + geom_boxplot(aes(x = continent, y = gdpPercap), notch = TRUE, varwidth = TRUE) ``` With moderate sample sizes it can be useful to super-impose the original data, perhaps with jittering and alpha blending. The outliers in the box plot can be turned off with `outlier.color = NA` so they are not shown twice: ```{r} p <- ggplot(gapminder) + geom_boxplot(aes(x = continent, y = gdpPercap), notch = TRUE, varwidth = TRUE, outlier.color = NA) p + geom_point(aes(x = continent, y = gdpPercap), position = position_jitter(width = 0.1), alpha = 0.1) ``` ## Violin Plots A variant of the boxplot is the _violin plot_: Hintze, J. L., Nelson, R. D. (1998), "Violin Plots: A Box Plot-Density Trace Synergism," _The American Statistician_ 52, 181-184. The violin plot uses density estimates to show the distributions: ```{r} ggplot(gapminder) + geom_violin(aes(x = continent, y = gdpPercap)) ``` By default the "violins" are scaled to have the same area. They can also be scaled to have the same maximum height or to have areas proportional to sample sizes. This is done by adding `scale = "width"` or `scale = "count"` to the `geom_violin` call. A comparison of boxplots and violin plots: ```{r} ggplot(gapminder) + geom_boxplot(aes(x = continent, y = gdpPercap)) + geom_violin(aes(x = continent, y = gdpPercap), fill = NA, scale = "width", linetype = 2) ``` A Combination of boxplots and violin plots: ```{r} ggplot(gapminder) + geom_violin(aes(x = continent, y = gdpPercap), scale = "width") + geom_boxplot(aes(x = continent, y = gdpPercap), width = .1) ``` There are other variations, e.g. _vase plots_. ## Beeswarm Plots Beeswarm plots, also called violin scatterplots, try to show both the full data and the density of the data distribution. The [`ggbeeswarm`](https://github.com/eclarke/ggbeeswarm) package provides one implementation with two variants: * `geom_quasirandom` * `geom_beeswarm` The `quasi_random` version seems to work a bit better on many examples, e.g. ```{r} library(ggbeeswarm) ggplot(gapminder) + geom_quasirandom(aes(x = continent, y = gdpPercap), size = 0.2) ``` Combined with a width-scaled violin plot: ```{r} ggplot(gapminder) + geom_violin(aes(x = continent, y = gdpPercap), scale = "width") + geom_quasirandom(aes(x = continent, y = gdpPercap), color = "blue") ``` ## Effectiveness and Scalability * Boxplots are very simple and easy to compare. * Boxplots strongly emphasize the middle half of the data. * Boxplots may not be easy for a lay viewer to understand. * Box plots scale fairly well visually and computationally in the number of observations; over-plotting/storage of outliers becomes an issue for larger data sets * Violin plots scale well both visually and computationally in the number of observations. ```{r, fig.width = 10} library(gridExtra) p1 <- ggplot(diamonds) + geom_boxplot(aes(x = cut, y = price)) p2 <- ggplot(diamonds) + geom_violin(aes(x = cut, y = price)) grid.arrange(p1, p2, nrow = 1) ``` * Scalability in the number of cases for beeswarm plots is more limited. * The number of groups that can be handled for comparison by these plots is in the range of a few dozen. ```{r, fig.width = 10} library(lattice) p1 <- ggplot(barley) + geom_boxplot(aes(x = site, y = yield, fill = year)) p2 <- ggplot(barley) + geom_violin(aes(x = site, y = yield, fill = year)) grid.arrange(p1, p2, nrow = 1) ``` Axes can be flipped to avoid overplotting of labels: ```{r, fig.width = 10} library(lattice) p3 <- p1 + coord_flip() p4 <- p2 + coord_flip() grid.arrange(p3, p4, nrow = 1) ``` ## Ridgeline Plots [Ridgeline plots](https://blog.revolutionanalytics.com/2017/07/joyplots.html), also called _ridge plots_ or _joy plots_, are another way to show density estimates for a number of groups that has become popular recently. The package `ggridges` defines `geom_density_ridges` for creating these plots: ```{r, message = FALSE} library(ggridges) ggplot(barley) + geom_density_ridges(aes(x = yield, y = site, group = site)) ``` Grouping by an interaction with a categorical variable, `year`, produces separate density estimates for each level. Mapping the `fill` aesthetic to `year` allows the separate densities to be identified: ```{r} ggplot(barley) + geom_density_ridges(aes(x = yield, y = site, group = interaction(year, site), fill = year)) ``` Alpha blending may sometimes help: ```{r} ggplot(barley) + geom_density_ridges(aes(x = yield, y = site, group = interaction(year, site), fill = year), alpha = 0.8) ``` Adjusting the vertical scale may also help: ```{r} ggplot(barley) + geom_density_ridges(aes(x = yield, y = site, group = interaction(year, site), fill = year), scale = 0.8) ``` Sometimes reordering the grouping variable, `year` in this case, can help. The factor levels of `year` can be reordered to match the order of average yealds within each year by ```{r, eval = FALSE} reorder(year, yield) ``` Using `-yield` produces the reverse order. ```{r, message = FALSE} library(dplyr) ggplot(mutate(barley, year = reorder(year, -yield))) + geom_density_ridges(aes(x = yield, y = site, group = interaction(year, site), fill = year), scale = 0.8) ``` With some tuning ridgeline plots can scale well to many distributions. An example from [Claus Wilke's book](https://clauswilke.com/dataviz/): The `ggplot2movies` package provides data from [IMDB](https://imdb.com/) on a large number of movies, including their lengths, in a tibble `movies`: ```{r} library(ggplot2movies) dim(movies) head(movies) ``` A ridgeline plot of the movie lengths for each year: ```{r, message = FALSE} library(dplyr) mv12 <- filter(movies, year > 1912) ggplot(mv12, aes(x = length, y = year, group = year)) + geom_density_ridges(scale = 10, size = 0.25, rel_min_height = 0.03) + scale_x_continuous(limits = c(0, 200)) + scale_y_reverse(breaks = c(2000, 1980, 1960, 1940, 1920)) + theme_minimal() ``` This shown that since the early 1960's feature film lengths have stabilized to a distribution centered around 90 minutes: Another nice example: [DW-NOMINATE](https://en.wikipedia.org/wiki/NOMINATE_(scaling_method)) scores for measuring political position of members of congress over the years: ![](img/polarization.jpeg) [Original code]( http://rpubs.com/ianrmcdonald/293304) by Ian McDonald; another version is provided in [Claus Wilke's book](https://clauswilke.com/dataviz/).