How to Choose the Right Chart: One Question About Your Data

1. The one question, and the decision table

Before the explanation: two analysts get the same table of revenue by region. One makes a bar chart, one makes a pie. What question would decide who is right?

Not "what does the data look like". The deciding question is about the reader: what comparison must the reader make? A chart is a machine for making one comparison easy, and the chart type is chosen by that comparison, not by the shape of the table.

Here is the fork laid out fully, because it is the decision the whole page hangs on. The question: what must the reader compare? The four possible answers, and what each one means for you: values across categories, which needs lengths the eye can rank, so a bar. Change over time, which needs slope the eye can follow, so a line. Whether two measures move together, which needs one dot per record, so a scatter. Or shares of a total, which needs parts against a whole, and past a few slices that job also goes to a bar. What decides between them is only which comparison the reader must make first. It matters because the same table drawn with the wrong machine makes the right comparison hard and a wrong one easy.

If the reader must make two different comparisons, that is two charts, not one clever one. Section 7 covers the honest way to do that.

2. Categories: the bar chart, and how to make it readable

Before the explanation: here is quarterly revenue for five sales channels, in thousands. Which two are hardest to tell apart, and what would make that comparison instant?

The comparison is across categories, so this is a bar chart. The reason bars work is mechanical: every bar starts from the same baseline, so comparing values reduces to comparing lengths, and ranking lengths from a shared start is one of the judgements human eyes make most accurately. North against South is a 24 thousand dollar gap on 388, a 6.2% difference, and side-by-side bars from the same baseline make even that small gap visible.

Three habits turn a default bar chart into a readable one.

• Sort by value, not alphabetically, unless the categories have a real order. Sorted bars answer "which is biggest, which is next" before the reader asks.
• Go horizontal when labels are long. Category names read flat, nobody tilts their head, and ten categories fit without abbreviation.
• Put the number at the end of the bar when there are few bars. The chart carries the ranking, the label carries the value, and the axis can retire.

Total across the five channels is 1,705 thousand. Keep totals like that in the subtitle or a caption, because a bar chart of parts does not show its own sum.

3. Time: the line chart, and the first miscast

Before the explanation: monthly signups run 180, 176, 189, 197, 214, 222, 241, 238, 260, 271, 266, 290 from January to December. What is the one thing a reader should take from that series, and which visual feature carries it?

The takeaway is growth: 180 to 290 is a 61.1% rise across the year, climbing steadily with small monthly wobbles. The feature that carries it is slope. A line chart connects the months in order, and the eye reads the tilt of the line as speed of change directly. That is the whole reason time gets a line: the horizontal axis has a true order, so connecting the points draws the trend itself.

Which brings up the first of the two great miscasts. A line chart across categories implies an order that is not there. Draw North, South, East, West, Online as points and connect them with a line, and the line manufactures a journey: revenue "falls" from North to Online, as if the regions were stops on a route. The slope the eye reads so well is now reporting a relationship that does not exist, because nothing travels from North to South. Categories are not ordered, so nothing may connect them. The moment the horizontal axis stops being time or another true sequence, the line comes off and the bars go on.

The reverse swap is milder but real: bars across many time periods turn a trend into a picket fence, and the eye compares individual posts instead of following the path. Use bars for time only when the periods are few and the question is "this quarter against last quarter", a single comparison rather than a trajectory.

4. Relationship: the scatter plot

Before the explanation: ten stores, two measures each: staff on shift and weekly sales in thousands. Staff of 4, 5, 5, 6, 7, 8, 9, 10, 11, 12 against sales of 21, 24, 26, 27, 33, 34, 40, 42, 46, 49. What question does this table ask that neither a bar nor a line can answer?

The question is whether the two measures move together, and the scatter plot is the only chart in the set built for it. One dot per store, staff on one axis, sales on the other. No baseline, no time, just position. The eye reads the cloud's shape: drifting up and to the right means the measures rise together, and how tightly the dots hug a line is the strength of the relationship. For these ten stores the pattern is very tight: the correlation computes to 0.99, nearly a straight line.

The scatter is also the chart most likely to be over-read, because a clean rising cloud whispers "staff cause sales". A scatter shows association only. Busy stores may get more staff precisely because they sell more, or a third thing, like location, may drive both. The chart cannot pick between those stories, and correlation vs causation is the page for telling them apart before anyone acts on the cloud.

Picture your own data for a second: two numeric columns you have quietly wondered about, one dot per row. That mental image, and whether a shape appears in it, is the fastest analysis step you own.

5. Parts of a whole: why the bar beats the pie past a few slices

Before the explanation: nine product lines share revenue as 14, 13, 12, 11, 11, 10, 10, 10, and 9 percent, which sums to 100. On a pie, could you rank the top four slices by eye?

Almost nobody can, and the reason is worth owning because it explains half of chart design. A pie encodes each value as an angle, the width of the wedge at the centre. A bar encodes the same value as a length from a shared baseline. Human vision judges lengths against a common baseline far more accurately than it judges angles: in Cleveland and McGill's experiments, position and length judgements produced markedly smaller errors than angle judgements of the same values (Cleveland & McGill, 1984, Journal of the American Statistical Association, 79(387), 531–554). The pie is not ugly. It is asking the eye to do the measurement it is worst at.

Say in your own words why nine near-equal wedges are the worst case for angles before reading on. If you can say it, the pie rule stops being a rule and becomes a consequence.

Here is my version. Ranking wedges means comparing 14% of a circle with 13% of a circle, a difference of 3.6 degrees of arc, between wedges that start at different rotations. The same values as bars differ by a visible step in length from the same starting line. The comparison the reader must make, ranking parts, is a length job.

The honest scope of the pie: two or three slices, where the question is only "is this half, more than half, or less". "Support is about a third of tickets" survives as a pie. Nine slices never does.

6. The axis rules: when zero is mandatory and when zooming is fine

Before the explanation: two bars, 62 and 68. The chart's axis starts at 60 instead of 0. How big does the difference look, and how big is it?

Drawn from a baseline of 60, the first bar has length 2 and the second has length 8, so the second bar is four times as long. The real difference is 68 against 62: 9.7%. The chart shows 4x where the data says under 10%, and the reader's eye believes the chart, because in a bar chart length encodes value. Cut the bottom off the bars and you have cut the values themselves. So the rule is absolute for bars: the value axis starts at zero, every time, no exceptions for tight ranges. If the differences vanish at zero, that is the finding, or the wrong chart type is on duty.

A line chart lives under a different rule, and the reason is the encoding. A line encodes change as slope, not as distance from the floor. Zooming the axis of a line chart stretches the slopes so real movement becomes visible, and no length is being lied about because no length is being read. A line of daily temperatures plotted from zero kelvin would be perfectly honest and perfectly useless. Zoom the line, label the axis clearly, and keep zero for the bars.

Name the fork when you feel the urge to truncate: the question is which visual feature the reader is measuring. If it is length, zero is mandatory. If it is slope, zooming is legitimate. What decides is the chart type already on the page, and it matters because a truncated bar is the single most common way an honest analyst publishes a misleading chart.

7. The escape hatch: small multiples

Before the explanation: monthly trend lines for five regions on one chart is readable. For twelve regions it is a tangle. What could you change without dropping a single region?

Split one crowded chart into a grid of small identical ones: same chart type, same axes, same scale, one panel per region. The technique is called small multiples, and it is the standing answer whenever one chart is asked to hold too much. The reader learns to read one panel, then reads the other eleven for nothing, because every panel works identically. Comparison across panels stays honest because the scales match.

The trigger signs: more than about five lines on one line chart, a grouped bar chart three groups deep, or any legend you have to keep re-checking. Each of those is one chart trying to answer "compare across categories" and "compare over time" at once. Small multiples give each comparison its own chart and tile them.

The one discipline that makes them work: identical scales on every panel. The moment each panel scales itself, cross-panel comparison quietly breaks, and that comparison was the reason for the grid.

8. The full before and after

Same data both times: the nine product lines from section 5.

Before

Chart:  pie, nine slices, one colour per line
Order:  as the rows came out of the query
Title:  "Revenue Share by Product Line"
Legend: nine entries, matched to slices by colour

The reader's task: match nine colours between legend and wedges, then rank nine angles between 9% and 14% by eye. Nobody completes that task. They read the title, look at the circle, and leave with "revenue is split up", which was true before the chart existed.

After

Chart:  horizontal bar, sorted largest first, axis at zero
Labels: value at the end of each bar, no legend
Title:  "Top three lines hold 39% of revenue"

Licences      14
Support       13
Training      12
Consulting    11
Hosting       11
Hardware      10
Integrations  10
Templates     10
Other          9

Same nine numbers. The ranking is now instant because it is a length comparison from a shared baseline, the legend is gone because the labels sit on the bars, and the title states the finding: the top three lines, 14 plus 13 plus 12, hold 39% of revenue. Titling the claim instead of the column names is its own habit with its own page, label the claim, not the data. A chart titled with its finding gets that finding remembered. A chart titled with its column names gets skimmed.

9. Edge cases: the miscasts, named

Before the explanation: most bad charts are not exotic. They are one of about five repeat offenders. How many can you name before the list?

The line across categories. The first great miscast, from section 3. Connecting unordered categories manufactures a trend out of nothing. If the horizontal axis is not a true sequence, no line.

The nine-slice pie. The second great miscast, from section 5. Part-of-whole with more than about three parts is a length job, so it is a sorted bar.

The truncated bar axis. A 9.7% difference drawn as 4x. Bars start at zero because length encodes value.

The double axis. Two measures, two vertical scales, one chart. The crossing points and gaps between the lines are artifacts of two arbitrary scale choices, and the reader cannot help reading them as events. If the two measures are related, that is a scatter. If they share units, share one axis. Otherwise, two panels.

The spaghetti line chart. Eight or more lines, one chart, one legend doing all the work. This is the small-multiples trigger from section 7. Alternatively, grey out every line but the one or two the claim is about.

The decorated third dimension. Tilted 3D bars and pies change the drawn lengths and angles the reader is trying to measure, for zero information. Depth that does not encode data only distorts the encodings that do.