Effective data visualizations

Data visualization is one of the most important skills to develop as a data scientist. We use graphs, instead of tables for instance, to clearly communicate patterns, trends, and comparisons in data in a way that is inherently more interesting and informative than numbers in a grid.

Edward Tufte’s, “The Visual Display of Quantitative Information”, is probably the most famous text in all of data visualization (I’m guessing). In this text, Tufte outlines a theory of graphics and provides extensive detail regarding techniques for displaying data that maximizes clarity, precision, and efficiency. Tufte describes graphical excellence by the following points:

show the data

induce the viewer to think about substance…

avoid distorting what the data have to say

present many numbers in a small space

make large data sets coherent

encourage the eye to compare different pieces of data

reveal the data at several levels of detail, from a broad overview to the fine structure

serve a reasonably clear purpose: description, exploration, tabulation, or decoration

be closely integrated with the statistical and verbal descriptions of the dataset

These points describe how to make good visualizations. However, some visualizations are clearly better than others. Take a look at this presentation by Andrew Gelmen and Antony Unwin. Here, they describe what makes some visualizations effective vs. what makes others ineffective and distracting. I think the main conclusion comes down to design vs decoration. Effective data visualizations are intentionally designed to communicate a point in the clearest possible way. Ineffective visualizations often contain clutter in the form of decorations, unnecessary colors, patterns, lines, and unintuitive use of shapes and sizes.

Outside of the design choices that go into creating effective visualizations, there are also other expectations required of graphs. The data points should be labelled so that the viewer immediately knows what they’re looking at. The axes should be labelled in a way that’s easy to read and do not distort the message. If using colors, the color palette should augment the visualization in a way that enhances the display of information and not only ‘looks pretty’.

There are probably thousands (millions?, infinite?) of types of visualizations in use today. However, in scientific communication, there are basically 5 types of graphs that are most commonly used; line graphs, bar graphs, histograms, scatter plots, and pie charts. Most of these have persisted in the literature (with the exception of pie charts) because of their ability to clearly and quickly display visual information about quantitative data. Many other forms of these basic charts exist primarily as variations on a theme but the core display of information remains the same.

Plotting with `ggplot2`

We’ll use the ggplot2 R package to start creating effective visualizations. ggplot2 works in a way that can feel a little strange at first, especially if you’re used to creating plots in Python with matplotlib, for example. ggplot2 implements ideas from Leland Wilkinson’s, “Grammar of Graphics”. ggplot2, breaks down a graphic into several key components that can be combined in a layered fashion. These core components typically include:

Data: The dataset you want to visualize.
Aesthetics (aes): How columns in your data map to visual properties of the graphic. For example, mapping a ‘temperature’ column to the y-axis, a ‘time’ column to the x-axis, and a ‘city’ column to color.
Geometries (geoms): The visual elements used to represent the data, such as points, lines, bars, histograms, etc.
Scales: Control how the aesthetic mappings translate to visual output (e.g., how data values are converted to colors, sizes, or positions on an axis).
Statistics (stats): Transformations of the data that are performed before plotting (e.g., calculating a histogram, smoothing a line).
Coordinates: The coordinate system used for the plot (e.g., Cartesian, polar).
Faceting: Creating multiple subplots (a “trellis” or “lattice” plot) based on subsets of the data.
Theme: Non-data elements like background, grid lines, and font choices.

The power of this approach lies in its declarative nature. You specify what you want to plot by defining these components, rather than detailing how to draw each element step-by-step (like in matplotlib). This makes it easier to build complex visualizations and to switch between different representations of the same data with minimal changes to the code.

General design guidelines for plots

Don’t use a special color for the background color. Use the same color as your presentation background. Generally, either white or black backgrounds are best
Axes colors should be high contrast relative to the background. This means black axes on white backgrounds and white axes on black backgrounds.
Use tick marks that have whole-number intervals. e.g. don’t use axis ticks that are 1.33, 2.75, 3.38, …. Instead, use numbers that people use naturally when counting, e.g. 1, 2, 3 or 2, 4, 6, etc.
Axes need to be legible. The default font size is often too small. At least 12 point font.
Axes labels should clearly communicate what is being plotted on each axis. Always show units.
If showing a plot title, use the title to tell the user what the conclusion is, not simply describing the data on each axis.
If using multiple colors, they should contrast well with each other. Warm colors (reds, yellows) can be used to emphasize selected data.
Keep colors consistent across graphs. If you use the color “red” to represent treated samples in one plot, use the color “red” in all other plots where the treated samples are present.
Use colors to communicate. Hues are not emotionally neutral. A good color palette can affect the audience’s perception of the figure (just like a bad one can).
Only use gridlines if you need them, most of the time they’re not needed.
If using gridlines, be judicious. Their color should be subtle and not obscure the data. Only use grids that matter for the data.
Choose fonts that are easy to read. sans serif fonts are best. Helvetica or Arial are good default choices
When drawing error bars, don’t make them so wide/large as to obscure the data

A basic ggplot

ggplot() constructs a plot from data, what it calls aesthetic mappings, and layers. Aesthetic mappings describe how variables in the data are mapped to visual properties (aesthetics) of geoms. geoms then determine how the data is displayed. The other parts of the ggplot object have been handled automatically (i.e. scales, stats, coordinates, and theme). These, however, can be modified to enhance the plot. Check out the ggplot2 homepage for an overview or the ggplot2 book for details.

The code below demonstrates the most basic way of creating a plot with ggplot2.

ggplot(data = mtcars, mapping = aes(x = wt, y = mpg)) +
  geom_point()

This plot is okay but can be improved. Let’s improve this plot by removing the grey background and gridlines, increasing the font size of the axis ticks, improving the axis labels, and creating an informative title.

ggplot(data = mtcars, mapping = aes(x = wt, y = mpg)) +
  geom_point(size = 2) +        # increase the size of the points
  labs(                         # labs() can be used to modify axis label text
    title = "Fuel Efficiency Decreases with Increasing Weight",    
    x  = "Weight (1000 lbs)",
    y = "Fuel Efficiency (mpg)") +
  theme_classic() +             # Removes grey background and gridlines
  theme(                        # Adjust the plot and axes titles, and text
    plot.title = element_text(size = 18, face = "bold", color = "black"),
    axis.title = element_text(size = 14, color = "black"),
    axis.text = element_text(size = 12, color = "black")
  )

Saving themes

If you’re going to be using the same theme elements often, it can be helpful to save those as a new custom theme. You don’t have to understand exactly how this function works right now - simply modify the arguments to the theme() function and figure this out as you improve

theme_clean <- function(...) {
  ggplot2::theme_classic(...) %+replace%
  ggplot2::theme(
      text = ggplot2::element_text(family = "Helvetica"),
      plot.title = ggplot2::element_text(size = 18, face = "bold", color = "black", hjust = 0),
      axis.title = ggplot2::element_text(size = 14, color = "black"),
      axis.text = ggplot2::element_text(size = 12, color = "black")
    )
}

# The new theme can be applied to other plots
ggplot(data = mtcars, mapping = aes(x = wt, y = mpg)) +
  geom_point() +        
  theme_clean()

This plot is contains the same information but more quickly and clearly communicates the message by just modifying the design. What other element could be added to this plot to make the trend more apparent?

Scatter plots

We already showed a basic example of creating a scatter plot above. You can use that as a starting point for generating scatter plots. However, one common issue when designing scatter plots is overplotting, or, showing so many points that the data is cluttered. Below is an example of overplotting.

ggplot(diamonds, aes(carat, price)) +
  geom_point() +
  theme_clean()

One technique to overcome overplotting is to add transparency to the points

ggplot(diamonds, aes(carat, price)) +
  geom_point(alpha = 0.05) +
  theme_clean()

Another is to change the point type. Here, we plot each point as a single dot

ggplot(diamonds, aes(carat, price)) +
  geom_point(shape = ".") +
  theme_clean()

And another is random subsampling.

random_rows <- sample.int(nrow(diamonds), size = 500)

ggplot(diamonds[random_rows, ], aes(carat, price)) +
  geom_point() +
  theme_clean()

The density of the points could also be summarized. For example, explore geom_hex() or geom_density2d() geoms.

Let’s make a final version of this plot by cleaning up the background, axes, and titles. We’ll add a trendline to emphasize the relationship and we’ll also use a function to transform the y-axis to dollar format.

ggplot(diamonds[random_rows, ], aes(carat, price)) +
  geom_point() +
  geom_smooth(se = FALSE, color = "red") +   # Adds a smooth trendline
  labs(
    title = "Larger Diamonds are More Expensive",    
    x  = "Carat",
    y = "Price ($)") +
  scale_y_continuous(             # Formats the y-axis labels as dollar amounts
    labels = scales::dollar_format()
    )  +
  theme_clean()

`geom_smooth()` using method = 'loess' and formula = 'y ~ x'

Line graphs

Line graphs are meant to emphasize change in the y-variable over the x-variable. When designing line graphs:

Take advantage the range of the data. Data should take up about 3/4 of the y-axis. ggplot has pretty good defaults for this automatically.
Choose line weights that do not overshadow points (if present)
Dashed lines can be hard to read. Use contrasting colors instead
If presenting two graphs next to each other be sure to match the axes ranges

ggplot(economics, aes(date, unemploy)) + 
  geom_line() +
  labs(
    title = "Unemployed Individuals Over Time",
    x = "Year", 
    y = "Thousands of Persons"
    ) +
  scale_y_continuous(labels = scales::number_format(big.mark = ",")) +
  theme_clean()

We can also explore adding multiple colors to the plot to compare values. Another useful technique is to add the legend to the plot area.

econ_2 <- economics_long[!economics_long$variable %in% c("pce", "pop"), ]

ggplot(econ_2, aes(date, value01, colour = variable)) +
  geom_line() +
  labs(
    x = "Year", 
    y = "Variable",
    color = NULL
    ) +
  scale_color_brewer(                             # Add better colors and labels
    palette = "Set1", 
    breaks = c("psavert", "uempmed", "unemploy"),
    labels = c("Savings", "Duration", "Unemployed")
    ) +
  scale_y_continuous(labels = scales::number_format(big.mark = ",")) +
  guides(color = guide_legend(position = "inside")) +  # add the legend inside the plot
  theme_clean() +
  theme(
    legend.position.inside = c(0.5, 0.85),      # specify where to put the legend
    legend.text = element_text(size = 12)
    )

Bar graphs

People generally use bar graphs to display a mean and some variation around the mean. Filled bar plots can also be used to show proportions.

If there are relatively few data points, consider showing all points and a crossbar for the mean and error instead. For example, as a beeswarm plot. If there are many points, consider a boxplot.
In most cases, it’s usually best to start the y-axis at 0 if the data has a natural 0. The exception is for charts where 0 does not indicate ‘nothing’ but rather exists on a continuum, for example, temperature in Fahrenheit.
Don’t have bars with too thick of outlines
Avoid bars that are too thin. Aim for about 1/3 width of the bar as space between bars
Place extra space between categories if showing bars next to each other
If showing statistical information try not to overwhelm the data. Use subtle thin lines for comparisons and asterisks for significance.

Activity: Make this example plot better.

df <- data.frame(trt = c("a", "b", "c"), outcome = c(2.3, 1.9, 3.2))

ggplot(df, aes(trt, outcome)) +
  geom_col() +
  theme_clean()

Histograms

Histograms are used to show distributions of data with their relative frequencies.

The number of bins influences the interpretation of the data. Play around with the number of bins to ensure the best display
Don’t assign different colors to different bins - it doesn’t add any information
Don’t add spaces between bins, as in a bar plot
If displaying two datasets, you can either overlay each with a different fill and some transparency or split into two histograms with the same y-axis.

Activity: Make this example plot better.

ggplot(diamonds, aes(carat)) +
  geom_histogram() +
  theme_clean()

`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.

Saving images

Using ggplot2 you can save images with the ggsave() function. ggsave() can automatically detect the image format by the file extension. The ggsave() function works by saving the last plot created.

ggplot(data, aes(x, y)) + 
  geom_point()

ggsave("my-pretty-plot.pdf", width = 8, height = 6)

If using base R, you typically open the graphics device (png(), pdf(), jpeg(), etc.) first, depending on the file format you want to save, then create the plot, and close the plot device.

pdf("another-pretty-plot.pdf", width = 8, height = 6)
plot(x, y)
dev.off()

Image file formats

Save your images in .pdf format or .svg format. These formats are called vector graphics. Vector graphics are infinitely scalable. They never lose resolution no matter the zoom level.
Vector graphics don’t play nicely with Word documents. If you need to share images that will be used in Word docs, save your image as a high-quality .png. .png files are raster graphics. Raster based images store actual pixel values so they cannot be infinitely zoomed. They lose quality at high magnification. Consider at least a dpi of 300 when saving .pngs.
Vector graphics are also editable using a program like Inkscape. Often, you may need to add some custom color or annotations using an illustrator program. Vector graphics allow you to do this.

Plotting with ggplot2