A short guide to choosing the right chart

When it comes to presenting data, the choice of chart is often made by habit rather than by reasoning, and the result is a figure that technically shows the numbers but does not answer the question. In this guide we walk through a simple way to choose a chart type by starting from the question, rather than from the data, because the question is what determines which comparison the reader needs to see. The advice here is general, and it applies whether the chart is built by hand, by a script, or by an agent.

##Start from the question, not the data

The most useful first step is to name the comparison the chart is meant to make, because most questions reduce to one of a small number of comparisons. When we know the comparison, the chart type usually follows. The common cases are worth listing directly:

• Comparison between categories (e.g., revenue by region) is usually best served by a bar chart, ideally horizontal when the labels are long.
• Change over time (e.g., signups per week) is most clearly shown by a line chart, which makes the trend and its direction easy to read.
• Composition, or parts of a whole (e.g., share of spend by channel), is often better as a stacked bar or a simple table than as a pie chart, because angles are hard to compare precisely.
• Relationship between two variables (e.g., spend versus conversions) is the natural case for a scatter plot, which reveals correlation and outliers at a glance.
• Distribution of a single variable (e.g., order values) is best shown by a histogram, which makes the shape and the spread visible.

##Common mistakes worth avoiding

A few mistakes appear often enough that they are worth calling out. The pie chart is overused, because it looks friendly, however it makes precise comparison difficult once there are more than two or three slices, and a bar chart almost always communicates the same composition more clearly. Dual-axis charts, which place two different scales on the same plot, are another frequent source of confusion, because they can imply a relationship that the data does not support, and they are easy to read incorrectly. Lastly, truncating the vertical axis so that it does not start at zero can exaggerate a difference dramatically, which might be acceptable for a closely-watched metric, however it is misleading for a general audience and should be used with care.

##Make the chart readable, then stop

Once the chart type fits the question, a small amount of restraint usually improves the result more than additional decoration. This means labelling the axes and the series directly, ordering categories in a way that aids comparison (e.g., sorting bars by value rather than alphabetically), and removing elements that do not carry information (e.g., heavy gridlines, redundant legends, or three-dimensional effects). A useful test is to ask whether a reader could state the chart's main point within a few seconds, and if they could not, the problem is most likely the chart type or the clutter, rather than the data.

In summary, the chart should be chosen by working backwards from the comparison the reader needs to make, and then kept as plain as the question allows. When it comes to everyday business reporting, a well-ordered bar chart and a clearly-labelled line chart will answer the large majority of questions, and reaching for something more elaborate is worth doing only when the question genuinely requires it.

##Further reading

• Financial Times: Visual Vocabulary: which chart to use when
• Publications Office of the EU: Data visualisation guide: visual vocabulary
• Tableau: Truncating the y-axis: threat or menace?