Write the sentence first. Not the title, not yet, and not the chart. Write the plain sentence you want a reader to believe when they look away, and put it above the data while you work.
Here are four sentences about the same table of state-wise revenue.
| Sentence | Type | What the reader must be able to do |
|---|---|---|
| Maharashtra is the largest state at 31.2 crore | comparison | rank the categories, read one value |
| Revenue has grown every year since FY22 | change over time | follow one series along an ordered axis |
| Six states make up the whole book, and two of them are 44.8% of it | composition | see parts against a whole |
| States with more stores grow faster | relationship | see two variables move together |
Four sentences, four types, four different charts, one dataset. Look at that table again, because most analysts arrive with the opposite instinct. They believe the data suggests the chart. It doesn't. The same numbers will support any of the seven message types, depending on which question you asked. Your form has to answer that question. Describing whatever table you happen to hold isn't the job.
Advertisement
What are the seven message types?
Seven types cover essentially everything an analyst claims.
| # | Message type | The claim, in shape | Typical verb |
|---|---|---|---|
| 1 | Comparison | A is larger than B across a set of categories | exceeds, leads, trails |
| 2 | Change over time | X moved from one value to another along an ordered axis | rose, fell, flattened |
| 3 | Composition | the parts of a whole are distributed in this way | makes up, accounts for |
| 4 | Relationship | two measured variables move together | is associated with, predicts |
| 5 | Distribution | the values are spread across a range in this shape | clusters, is skewed, is bimodal |
| 6 | Ranking | the ordered position is the point, not the level | first, top five, bottom quartile |
| 7 | Deviation | the gap from a reference is the point | against budget, versus the index |
Ranking and deviation are arguably special cases of comparison, and some textbooks fold them in. They're kept separate here for a working reason: they choose different forms. A ranking message wants a sorted dot plot or bar, with the order carrying meaning. A deviation message wants a diverging bar centred on zero. Run both through the general comparison rule and you'll get the wrong answer often enough for it to matter.
How do you check you've picked the right message type?
To test whether you've classified correctly, try to break your own sentence. If it can be argued for with a different chart form, you've probably written two sentences joined by a comma. Split them. One message per chart is a rule you'll keep meeting in the decluttering and action-title work that follows. It starts here, because a form that has to carry two claims will carry neither.
The shared form heuristic
Six inputs, one output. This is the chart-choice table the drills grade against and the checker runs, and it's short enough to memorise.
Six inputs go in. The message type comes from the seven above. Next, n, the number of categories or points, and whether the horizontal variable is ordered, meaning time or a scale with an intrinsic sequence. Then s, the number of series, and whether the parts sum to a meaningful whole with every value non-negative. Last, whether the reader's task is lookup or precision-critical.
How do you use the chart-choice table?
| # | Message | Condition | Primary form | Permitted alternate | Ruled out |
|---|---|---|---|---|---|
| 1 | Comparison | n ≤ 12, labels short | column, sorted | dot plot | pie, line, area |
| 2 | Comparison | n ≤ 12, any label over about 18 characters | horizontal bar, sorted | dot plot | rotated column labels |
| 3 | Comparison | n > 12 | dot plot, sorted | horizontal bar | column, pie |
| 4 | Comparison | s = 2, same units | paired bar or dot plot with two marks | slope chart | stacked bar |
| 5 | Change over time | ordered, points ≥ 5, s ≤ 4 | line | column if each period is a discrete event | area, slope, pie |
| 6 | Change over time | ordered, points ≥ 5, s > 4 | small multiples of lines, shared scale | line with one accent series | one crowded line chart |
| 7 | Change over time | exactly 2 periods, 3 ≤ n ≤ 15 | slope chart | diverging bar of the change | two pies |
| 8 | Change over time | points ≤ 4 and the message is a level | dot plot with a reference line | a sentence | bar with a truncated axis |
| 9 | Composition | 2 or 3 parts, one period | stacked bar, single | the numbers in text | pie |
| 10 | Composition | 4 or more parts, one period | sorted horizontal bar | treemap if a hierarchy exists | pie, donut |
| 11 | Composition | parts over few periods, shares are the point | 100% stacked bar, ≤ 5 parts | small multiples of lines | stacked area |
| 12 | Composition | a total decomposed into signed drivers | waterfall | table with a check column | pie, stacked bar |
| 13 | Relationship | two continuous variables, n ≥ 12 | scatter | scatter with a fitted line and stated r | bar of group means |
| 14 | Relationship | two variables plus a third magnitude | bubble, with area encoding | small multiples of scatters | 3-D scatter |
| 15 | Distribution | one variable, n ≥ 200 | histogram, bin width stated | density curve | box plot alone |
| 16 | Distribution | one variable, n < 200 | dot plot, every observation shown | strip plot | histogram |
| 17 | Distribution | several groups compared | box plot with the points overlaid | small multiples of histograms | box plot alone, bar of means |
| 18 | Ranking | the order is the message | sorted dot plot | sorted horizontal bar | alphabetical anything |
| 19 | Deviation | gap from a reference or a budget | diverging bar centred on zero | dot plot with a reference line | stacked bar |
| 20 | Any | lookup or precision-critical, n ≤ 20 rows | table | table with an in-cell bar | any chart |
Two tie-breaks settle the cases where more than one row applies. When a comparison rule and a composition rule both fit, ask whether the reader needs the whole. If the total isn't part of the claim, the comparison rule wins and you draw a sorted bar instead of a stacked one. And when another rule and the chart-choice table's lookup rule both fit, the lookup rule wins. A reader who has to read a number off a chart to two decimal places has been given the wrong object.
The four permanent fails sit outside the table and override everything in it. No three-dimensional effects. No dual axes. No truncated baseline where the mark encodes length. No pie with more than three slices, and no pie with a legend at any slice count. Section 9 gives the arithmetic behind each.
Running this costs you fifteen seconds. Write the sentence, name the type, count n, answer the four remaining questions, read the row. What you get back is that your answer and the drill's answer and the checker's answer are the same answer. So when two of them disagree, it tells you something about one of you. It's never a matter of taste.