When a scatterplot is the right choice, when another chart reads better, and how to draw one that does not overclaim. The same rules hold in every chart library. For picking between chart types in general, see Choosing a chart.
What it is for
A scatterplot shows how two numbers move together. Each point is one observation, such as a car, a customer or a day, placed by one measure on the horizontal axis and another on the vertical axis. The shape of the cloud shows the relationship, the clusters and the outliers.
When to use
The reader asks whether two measures are related, such as price and sales, or weight and fuel economy.
You look for clusters, gaps or outliers in a set of items.
A third measure matters for each item. Size the points by it to make a bubble chart.
You compare groups on the same two measures. Color the points by group, with three to five groups at most.
Use something else when
The horizontal axis is time and the reader wants the trend. Use a line chart.
One axis is a category, such as a region or a team. Use a bar chart or a dot plot.
There are tens of thousands of points and they cover each other. Use a hexbin or a binned heatmap, which shows where the points are dense.
You show how one measure is spread out. Use a histogram or a box plot.
You compare more than two measures per item. Use parallel coordinates, or one small scatterplot per pair of measures.
There are only a handful of items. Use a table, or a bar chart per measure.
How it reads
Both values are encoded as position on a common scale, the encoding people read most accurately. The reader sees the overall direction first: points rising to the right mean the two measures grow together, points falling mean one drops as the other grows. A loose cloud means a weak relationship.
In a bubble chart the third value is encoded as area, which people read far less accurately than position. Treat bubble size as a rough signal, not a value to compare.
Do it well
Put the measure you think drives the other on the horizontal axis, and label both axes with the unit.
Let the axes start where the data starts. A scatterplot compares positions, not lengths, so it does not need a zero baseline.
Make the points small and partly transparent, so overlapping points show as darker areas instead of one solid blob.
Size bubbles by area, not by radius. A square-root scale on the radius keeps area in proportion to the value. Draw the largest bubbles first, so small ones stay on top.
Add a trend line when the direction is the point, and say in the subtitle what kind of fit it is.
Use a log scale when the values span several orders of magnitude, and say so on the axis. The Axes guideline covers when a log scale is fair.
Label the few points the reader should notice, not all of them.
Pitfalls
Reading cause into the chart. Two measures that move together may both follow a third one, or match by chance. Write "related to", not "causes", unless you have other evidence.
Overplotting. Thousands of opaque points hide how many sit in the dense areas. Lower the opacity, shrink the points, or bin them.
Sizing bubbles by radius. A value twice as large then looks four times as large.
Trusting a summary number alone. Very different clouds can share the same mean, correlation and trend line. Always show the points.
A trend line through data with no trend, or through a curved relationship. A straight line suggests a pattern the points do not have.
Too many colors. More than about five groups turn into noise.
Variants
Plain scatterplot: two measures, one point per item.
Bubble chart: a third measure as the point's area.
With a trend line: a fitted line that shows the overall direction.
Connected scatterplot: points joined in time order, to show how two measures moved together over the years.
Lag plot: each value against the one before it, to show whether a series follows its own past.
With marginal histograms: a histogram along each axis, to show how each measure is spread out on its own.
Small multiples: one scatterplot per group, on shared axes, instead of one crowded chart.
Parallel coordinates: one vertical axis per measure and one line per item, for more than two measures.
Compared with similar charts
Against a line chart: a line joins values in order along one axis, usually time. A scatterplot places independent items and joins nothing.
Against a heatmap or hexbin: they count points per cell, which reads better than a scatterplot once points cover each other.
Against a bubble chart: the same chart with a third measure. Keep it for rough comparisons, since area is hard to read.
Against a correlation number: a scatterplot shows the shape behind the number, including outliers and curves the number hides.