Choose and label a chart

Choose a chart type from the question being asked, measure how much a truncated axis exaggerates a difference, and recognise why two series in different units cannot share an axis.

A chart is the part of the work that most people will see and the only part most will remember, and it is the easiest place in a spreadsheet to mislead without writing down anything false.

The five months in this lesson vary by six percent. Drawn from zero, that is five bars of almost the same height, and a reader takes away that attendance has been steady, which is true. Drawn from an axis starting at 500 — every number correct, every label accurate — the tallest bar is nearly four times the shortest, and a reader takes away that December was far busier than September, which is not.

Nobody lied. There is no wrong number anywhere on the second chart. The difference between the two is a single number in the axis settings, and its effect is measurable, which is the useful thing about it: whether to truncate an axis does not have to be an argument about taste when the exaggeration can be put in a cell.

Choose the type from the question.

• How do these categories compare? A bar or column chart. Bars when the labels are long or there are many of them; columns when there are few and the axis is time.
• How has this changed over an ordered sequence? A line chart. The line asserts that the points are in order and that the space between them means something, so never use one for categories that have no order.
• Do these two measurements move together? A scatter chart, with one measurement on each axis. This is the only chart in the set where both axes are data.
• Is this one number the whole point? Then it is not a chart. Write the number, large, with the comparison it needs beside it.

The axis. Start it at zero for anything where the height of the bar is meant to represent the quantity — which is every bar and column chart. A truncated axis makes the difference visible at the cost of making the proportion wrong, and readers read bars as proportions whatever the axis says.

There are defensible truncations. A line chart of a stock index, a temperature series, a measure whose interesting variation is genuinely a fraction of a percent — in those, starting at zero hides everything and the differences are the whole point. What makes them defensible is that the reader is told, prominently, and that the audience is one that reads axes.

What is never defensible is truncating silently to make a change look bigger. And there is a test for which one you are doing: compute the exaggeration.

=(highest - axis start) / (lowest - axis start)

With months of 545 and 512 and an axis starting at 500, that is 45 over 12, or 3.75. A six percent difference drawn as a three-and-three-quarter-fold one. Once that number is in the sheet, the conversation about which chart to publish is a short one.

Two units cannot share an axis. Attendance in the hundreds and a fill rate around 0.72 on one scale means the fill rate is a flat line at the bottom, indistinguishable from zero. The chart claims to show two things and shows one.

The standard response is a second axis on the right, and it is worth understanding exactly what that does. It makes both series visible, and it makes every apparent relationship between them meaningless, because the two scales are chosen independently: you can slide one until the lines cross wherever you like, and readers interpret a crossing as an event. A dual-axis chart is a picture whose most salient feature is an artefact of two arbitrary decisions.

Two units means two charts. They can sit side by side, and each one can be read.

Every chart needs an equivalent data table. Not as an accessibility afterthought — although it is that, and a chart with no table is unusable to a screen-reader user — but because a table is the only way a reader can check the chart. Every chart in this course carries one, built from the same cells as the bars, so the two cannot disagree.

Label so the chart survives being separated from its page. A title that names the measure and the period, an axis label with its unit, and a caption saying what the chart does not establish. Charts get screenshotted into other documents, and the caption is what travels with it.

Five months at one site:

September   512      October   528      November  534
December    545      January   521

The spread is 33 people, which is 6.4% of the smallest month. Attendance has been steady.

Two charts of those five numbers appear under the grid in the exercise below. The first starts its axis at zero: five bars of nearly equal height, and the reader's takeaway is "steady", which matches the data.

The second starts at 500. September's bar is 12 units tall and December's is 45. The reader's takeaway is "December was much busier", which does not.

Both charts are labelled accurately. Both have correct numbers. One of them is a picture of the data and the other is a picture of the axis choice.

The third chart puts attendance and fill rate on the same axis, and it is worth looking at because it fails in a different way: the fill rate series is around 0.72 against numbers around 520, so it does not appear at all. A reader sees a chart with a legend naming two things and bars showing one.

November had 534 people attend across 25 sessions. October had 528 across 24. Places per session is 30.

Predict these three, then say which month a chart of attendance alone would suggest was better, and which month a chart of fill rate alone would suggest.

October fill rate   =528/(24*30)
November fill rate  =534/(25*30)

Three edits turn an unreadable chart into a readable one, in order of how much they matter.

Start the axis at zero. In this grid a chart declares baselineAtZero, and when it is turned off the chart is required to state where its axis starts, in a banner above it. In Excel and Google Sheets the equivalent is the axis minimum, which is set to Auto by default — and Auto will truncate on its own, without asking, whenever the data is clustered. Checking it is part of making a chart rather than an optional refinement.

Split the units. Instead of:

one chart:  Attended (hundreds) and Fill rate (around 0.7)

two charts, each with its own axis and its own label. The second one is a line chart because fill rate over ordered months is a change-over-time question.

Say what it does not show. A caption on the fill-rate chart that reads "November is slightly lower than October despite more people attending, because November ran more sessions" is worth more than the chart, and it is the sentence a reader would otherwise have to work out or, more likely, not.

One thing not to do: adding a data table is not one of the three edits, because it is not optional. Every chart gets one.

A slide says attendance is "up sharply". The chart supports it. The underlying numbers are the five months above, which vary by six percent.

1. Read the axis before the bars. Every time. If it does not start at zero, note where it does start, and treat the visual difference as uninformative until you have computed the real one.
2. Compute the real difference as a share. Highest minus lowest, over lowest. Six percent. That number goes in the sheet, not in your head.
3. Compute the drawn difference. Highest minus axis start, over lowest minus axis start. 3.75. Now you have both numbers and the conversation is about a factual gap rather than about design preference.
4. Check whether the axis was chosen or defaulted. This matters for how to raise it. Most truncated axes are not deceit; they are the default setting on clustered data, and the person who made the chart did not look. Saying "the axis defaulted to 500 and that overstates it fourfold" gets a chart fixed. Saying "this chart is misleading" gets an argument.
5. Ask what claim the chart invites. Not what it says — what a reader who spends four seconds on it will believe afterwards. That is the thing to be accountable for.

The general shape: a chart's honesty is a property of the impression it leaves, not of the accuracy of its numbers. Both charts in this lesson have perfectly accurate numbers.

Work out the fill rates, measure how much the figures actually vary, and measure how much a truncated axis would make them appear to vary.

Four charts are drawn under the grid from the cells you fill. Two of them are marked as examples of what not to do; look at all four and at the tables beneath them.

Collect three charts from the last week — a news site, a report at work, a product dashboard, a social post.

For each, in two sentences:

1. Where does the value axis start, and if it is not zero, roughly how much does that multiply the apparent difference?
2. What claim would a reader who spends four seconds on it come away with?
3. Is that claim supported by the numbers on the chart?

Then find one where the answer to 3 is no, and rewrite the chart's title so that it says what the data supports. A chart's title does more work than its axis, and it is the part most often written last and least.

Sign in to track your progress on this exercise.

Move on when you can do these.

• Match each of four questions — comparison, change over time, relationship, single headline — to what it should be drawn as, including the one that should not be a chart.
• Write the formula for how much a truncated axis exaggerates a difference, and compute it for an axis at 500 with values of 512 and 545.
• Give a case where a truncated axis is defensible, and say what has to accompany it.
• Explain what a second axis does to a chart of two units, and what to do instead.
• Say why every chart needs a data table, giving two separate reasons.
• Say what a chart's honesty is a property of.

Sign in to track your progress on this exercise.

A chart is honest when its type answers the question that was asked, its axis starts at zero unless the reader is told otherwise and the exaggeration has been measured, its units are not mixed on one scale, and it carries a table of its own numbers and a caption saying what it does not establish.

Next: A compact decision dashboard. One chart answers one question. The last lesson of this module puts several measures on one screen, where the hardest part is not the measures but saying what they cannot establish.