Visualizing Tabular Data
Lesson 2 with Rachael Blake
- Meet the “layered grammar of graphics”
- Trust that ggplot2 is way better than base R’s
- Learn to layer visual elements on top of tidy data
- Glimpse the vast collection of ggplot2 options
- Create “aesthetic mappings” from variables to scales & geometries
- Build boxplots, scatterplots, smoothed lines and histograms
- Style plots with colors & annotate them with labels
- Repeat plots for different subsets of data
“A Layered Grammar of Graphics” is the title of an article by the author of ggplot2, Hadley Wickham. The package codifies the ideas presented in the article, especially the main idea that scientific visualization is all about assigning different variables to distinct visual elements. A plot is made up of several of these “aesthetic mappings”: for example, equating income to a linear scale on the y-axis, education to a ordinal scale on the x-axis, and displaying records about each person in a box-plot geometry.
The dataset you will plot is an example of Public Use Microdata Sample (PUMS) produced by the US Census Bureau. We’ll explore the wage gap between men and women.
The file to be loaded contains individuals’ anonymized responses to the 5 Year American Community Survey (ACS) completed in 2017. There are over a hundred variables giving individual level data on household members income, education, employment, ethnicity, and much more.
library(readr) person <- read_csv( file = 'data/census_pums/sample.csv', col_types = cols_only( AGEP = 'i', WAGP = 'd', SCHL = 'c', SEX = 'c'))
The readr package gives additional flexibility and speed over the
read.csv function. The CSV contains 4 million rows, equating to several
gigabytes, so a sample suffices while developing ideas for visualiztion.
The code to plot each invidual’s wage or salary income by their education
attainment calls three functions:
geom_histogram from the ggplot2 package.
ggplotcreates the foundation
aesspecifies an aesthetic mapping
geom_histogramadds a layer of visual elements
library(ggplot2) ggplot(person, aes(x = WAGP)) + geom_histogram()
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
Warning: Removed 1681 rows containing non-finite values (stat_bin).
ggplot command expects a data frame and an aesthetic mapping. The
function creates the aesthetic, a mapping between variables in the data frame
and visual elements in the plot. Here, the aesthetic maps
WAGP to the
x-axis; a histogram only needs one variable mapped.
ggplot function by itself only creates the axes, because only the
aesthetic map has been defined. No data are plotted until the addition of a
geom_* layer, in this example a
geom_histogram. Layers are literally added,
+, to the object created by the
Plotting histograms is always a good idea when exploring data. The zeros and the “top coded” value used for high wage-earners in PUMS are outliers.
library(dplyr) person <- filter( person, WAGP > 0, WAGP < max(WAGP, na.rm = TRUE))
geom_histogram aesthetic only involves one variable. A scatterplot
requires two, both an
x and a
ggplot(person, aes(x = AGEP, y = WAGP)) + geom_point()
aes function can map variable to more than just the
y axes in a
plot. There are several other “scales” that exist, although whether and how they
show up depends on the
geom_* layer. Commonly used arguments are
line or edge color and
fill for interior colors, but many more are
The aesthetic and the geometry are entirely independent, making it easy to
experiment with very different kinds of visual representations. The only change
needed is in the
ggplot(person, aes(x = AGEP, y = WAGP)) + geom_density_2d()
For a discrete x-axis, a boxplot is often beter than a scatterplot.
ggplot(person, aes(x = SCHL, y = WAGP)) + geom_boxplot()
To create a scatterplot, a boxplot, and even a 2d kernel density estimate, the
geom_* function takes no arguments. Every layer added on top of the foundation
generated by the call to
ggplot inherits the dataset and aesthetics of the
- What happens if you supply
x = AGEPto the aesthetic map in the boxplot?
- Boxplots aren’t designed for continuous x-axis variables, so the result is not useful. Fortunately, there’s a warning.
geom_* layers create a plot with multiple visual elements.
ggplot(person, aes(x = SCHL, y = WAGP)) + geom_boxplot() + geom_point()
geom_* object accepts arguments to customize that layer. Many arguments
are common to multiple
geom_* functions, such as changing the layer’s color.
ggplot(person, aes(x = SCHL, y = WAGP)) + geom_point(color = 'red') + geom_boxplot()
color specification was not part of aesthetic mapping between data and
visual elements, so 1) it applies to every record (or person) and 2) only the
elements in the scatterplot layer are affected.
stat parameter, in conjunction with
fun.y, provides the ability to
perform on-the-fly data transformations.
ggplot(person, aes(x = SCHL, y = WAGP)) + geom_boxplot() + geom_point( color = 'red', stat = 'summary', fun.y = mean)
stat = 'summary', the plot replaces the raw data with the result of a
summary function applied to whatever “grouping” is defined in the aesthetic. In
this case, it’s the ordinal x-axis that defines education attainment groups. The
fun.y argument determines what function, here the function
mean, with which
you want to summarize each group.
The true power of ggplot2 is the natural connection it provides between variables and visuals.
Associating color (or any attribute, like the shape of points) to a variable is
another kind of aesthetic mapping. Passing the
color argument to the
function works quite differently than assiging color to a
ggplot(person, aes(x = SCHL, y = WAGP, color = SEX)) + geom_boxplot()
- What sex do you think is coded as “1”?
- … Megan is skeptical about the answer!
Properties of the data itself are similarly independent of the aesthetic mapping and the visual elements, while still affecting the output.
person$SEX <- factor(person$SEX, levels = c("2", "1")) ggplot(person, aes(x = SCHL, y = WAGP, color = SEX)) + geom_boxplot()
There can be cases where you don’t want to or can’t modify the dataframe. Then, it is still possible to change properties of the data to get the plot you’d like within the
scale_* functions. More on modifying plots with
scale_* later in the lesson.
Storing and Re-plotting
The output of
ggplot can be assigned to a variable, which works with
schl_wagp <- ggplot(person, aes(x = SCHL, y = WAGP, color = SEX)) + geom_point( stat = 'summary', fun.y = 'mean')
The plot information stored in
schl_wagp can be used on its own, or with
Store additional layers by overwriting the variable (or creating a new one).
schl_wagp <- schl_wagp + scale_color_manual( values = c('black', 'red'))
Figures are constructed in ggplot2 as layers of shapes, from the
axes on up through the
geom_* elements. The natural file type for storing such
figures at “infinite” resolution are PDF (for print) or SVG (for online).
ggsave(filename = 'schl_wagp.pdf', plot = schl_wagp, width = 4, height = 3)
plot argument is unnecessary if the target is the most recently displayed
plot, but a little verbosity is not out-of-place here. When a raster file type
is necessary (e.g. a PNG, JPG, or TIFF) use the
dpi argument to specify an
geom_smooth layer used above can add various kinds of regression lines and
confidence intervals. A
method = 'lm' argument specifies a linear model.
Note, however, that with a categorical predictor mapped to an aesthetic element,
geom_smooth call would separately perform a linear regression (ANOVA)
within each group. The call to
aes must override the “group” aesthetic so the
regression is run once.
ggplot(person, aes(x = SEX, y = WAGP)) + geom_point() + geom_smooth( method = 'lm', aes(group = 0))
Is there really a confidence interval? Yes, it’s just pretty narrow and hard to
see. You could add a
size = 0.5 argument to
geom_smooth to see there is a
gray interval around the line. Or, as the next step shows, you could change
the size of the confidence interval for a better visual representation of the
level argument for
geom_smooth controls the limits of the confidence
interval, defaulting to 95%.
ggplot(person, aes(x = SEX, y = WAGP)) + geom_point() + geom_smooth( method = 'lm', level = 0.99, aes(group = 0))
Axes, Labels and Themes
aes and the
geom_* functions do their best with annotations and styling,
but precise control comes from
First, store a plot to simplify experiments with the labels.
sex_wagp <- ggplot(person, aes(x = SEX, y = WAGP)) + geom_point() + geom_smooth( method = 'lm', aes(group = 0))
Set the title and axis labels with the
labs function, which accepts names for
labeled elements in your plot (e.g.
title) as arguments.
sex_wagp + labs( title = 'Wage Gap', x = NULL, y = 'Wages (Unadjusted USD)')
For information on how to add special symbols and formatting to plot labels, see
Functions related to the axes, i.e. their limits, breaks, and any transformation
scale_* functions. To modify any property of a continuous y-axis, add
a call to
sex_wagp + scale_y_continuous( trans = 'log10')
“Look and feel” options in ggplot2, from background color to font
sizes, can be set with
sex_wagp + theme_bw()
theme_ on the console to see what themes are available in the
pop-up menu. The default theme is
theme_grey. A popular “minimal” theme is
theme_bw. Any option set by a
theme_* function can also be set by calling
theme itself with the option and value as an argument.
The options available directly through
theme offer limitless possibilities
Do be aware that if
theme comes after other custom specifications, it will overwrite
those customizations. Check the order if your plot isn’t looking how you’d like it to look.
sex_wagp + theme_bw() + labs(title = 'Wage Gap') + theme( plot.title = element_text( face = 'bold', hjust = 0.5))
?theme for a list of available theme options. Note that position (both
hjust for horizontal justification) should be given as a
proportion of the plot window (i.e. between 0 and 1).
To conclude this overview of ggplot2, we’ll apply the same plotting instructions to different subsets of the data, creating panels or “facets”.
facet_wrap function takes a
vars argument that, like the
relates a variable in the dataset to a visual element, the panels. The
facet_grid function works like
facet_wrap, but expects two variables to
facet by the interaction of a row variable by a column variable.
The gender wage gap apparent in the US Census PUMS data is probably not consistent across people who obtained different levels of education.
person$SCHL <- factor(person$SCHL) levels(person$SCHL) <- list( 'High School' = '16', 'Bachelor\'s' = '21', 'Master\'s' = '22', 'Doctorate' = '24')
The technical documenation for the PUMS data includes a data dictionary, explaining the codes used for education attainment, and everything else you’ld like to know about the dataset.
sex_wagp plot created above stored it’s own copy of the data, so create a
ggplot foundation using a cleaned up dataset.
ggplot(na.omit(person), aes(x = SEX, y = WAGP)) + geom_point() + geom_smooth( method = 'lm', aes(group = 0)) + facet_wrap(vars(SCHL))
- What wage gap trend do you think is worth investigating, and how might you do it?
- There are so many possibilities! For example, a scatterplot of wage against age colored by sex that includes a fitted regression model.
ggplotwith data and an
aesto pave the way for subsequent layers.
- Add one or more
geom_*layers, possibly with data transformations.
labsto annotate your plot and axes labels (not optional!).
- Optionally add
facet_*, or other modifiers that work on underlying layers.
- Data Visualization with ggplot2 RStudio Cheat Sheet
- Cookbook for R - Graphs Reference on customizations in ggplot
- Introduction to cowplot Vignette for a package with ggplot enhancements
ggplot to show how the mean wage earned in the U.S. varies with age,
showing males and females in different colors. (Hint: Baby steps! Start with a
scatterplot of wage by age. Then expand your code to plot only the means. Then
distinguish sexes by color.)
Create a histogram, using a
geom_histogram layer, of the wages earned by
females and males, with sex distinguished by the color of the bar’s interior. To
silence that warning you’re getting, open the help with
determine how to explicitly set the bin width.
facet_grid layer (different from
facet_wrap) requires an argument for
both row and column varaibles, creating a grid of panels. Create a plot with 8
facets, each displaying a single histogram of wage earned by women or men having
one of the four education attainment levels. Make the grid have 2 rows and 4
columns. Advanced challenge: add a second, partially transparent, histogram to
the background of each facet that provides a comparison to the whole population.
(Hint: the second histogram should not inherit the dataset from the
ggplot(person, aes(x = AGEP, y = WAGP, color = SEX)) + geom_line(stat = 'summary', fun.y = 'mean')
ggplot(person, aes(x = WAGP, fill = SEX)) + geom_histogram(binwidth = 10000)
ggplot(na.omit(person), aes(x = WAGP)) + geom_histogram(bins = 20) + facet_grid(vars(SEX), vars(SCHL))
For the advanced challenge, you must supply a dataset to a second gemo_histogram
that does not have the variable specified in
facet_grid. Note that
facet_grid affects the entire plot, including layers added “after faceting”,
as in the solution below.
ggplot(na.omit(person), aes(x = WAGP)) + geom_histogram(bins = 20) + facet_grid(vars(SEX), vars(SCHL)) + geom_histogram( bins = 20, data = na.omit(person['WAGP']), alpha = 0.5)
If you need to catch-up before a section of code will work, just squish it's 🍅 to copy code above it into your clipboard. Then paste into your interpreter's console, run, and you'll be ready to start in on that section. Code copied by both 🍅 and 📋 will also appear below, where you can edit first, and then copy, paste, and run again.
# Nothing here yet!