Choosing the right chart for the question

TL;DR

Difficulty: Beginner

Say the question in one sentence first, and the chart nearly chooses itself.

  • Comparison and ranking: sorted bars from zero, or a table for exact lookup.
  • Change over time: lines, bars only for a few periods.
  • Part to whole: stacked bars, a pie only for two or three parts.
  • Distribution: a histogram, or a box plot to compare groups.
  • Relationship: a scatter plot, which shows that things move together, never why.
  • Deviation: diverging bars from zero. People judge position on a shared baseline far better than angle or area, which is why bars and lines win so often.

Most charts are chosen by habit, or by whatever the spreadsheet offers first. This chapter turns that round: start from the question, name its type, and pick the chart that answers it. One invented outdoor gear shop in Limerick, eight questions from its owner, and the research that explains why bars beat pies.

The chart follows the question

Most people choose a chart the way they choose a font: by habit, or by whatever the spreadsheet offers first. Then they wonder why the meeting goes sideways. The fix is to reverse the order. Before you pick bars, lines, a pie or a scatter, say the question out loud in one sentence. The kind of question decides the kind of chart, almost every time.

There are only a handful of question types in business reporting. The Financial Times built its well known Visual Vocabulary poster around nine of them (deviation, correlation, ranking, distribution, change over time, magnitude, part to whole, spatial and flow)1, and the UK Government Analysis Function, a handy reference on this side of the Irish Sea too, gives the same advice in plainer words: match the chart to the relationship you want to show, and if you cannot state the message in a sentence or two, the chart is not ready2. I work the same way with clients. Every analytics project I take on starts from the decision the numbers should support, not from a dashboard, and the chart question is that idea in miniature. If you want the longer case for it, read start with the decision, not the dashboard.

This article is the second in a series on charts that tell the truth. The first chapter covered why a chart at all, and when a table does the job better. Here we pick the right chart for the job.

The question sounds likeQuestion typeChart to reach for
Which is biggest? How do they compare?ComparisonBars, sorted
Which comes first, second, third?RankingSorted bars, or a table for exact lookup
How has it moved?Change over timeLines (bars for a few periods)
What is it made of?Part to wholeStacked bars, a pie only for two or three parts
What is typical, and how spread out?DistributionHistogram, box plot
Does one thing move with another?RelationshipScatter, with a trend line
Are we above or below the line?DeviationDiverging bars

Example: five questions from the owner of Shannon Outfitters

Shannon Outfitters is an invented outdoor gear brand in Limerick. It sells online at shannonoutfitters.example and through one shop on O'Connell Street. Its owner, Aoife, sends her analyst Sam five questions before the quarterly review. Every number in this article is invented for the example.

  1. "Which product category made us the most money last quarter?" That is a comparison.
  2. "How has online revenue moved over the last two years?" Change over time.
  3. "Where do our online orders come from?" Part to whole.
  4. "What does a typical online order look like?" Distribution.
  5. "Does spending more on paid search bring in more revenue?" Relationship.

Sam writes the question type next to each one before opening the spreadsheet. That takes a minute and saves the hour of fiddling that usually follows. The rest of this article walks through each question in turn, plus two more types Aoife will ask about sooner or later: ranking and deviation.

Why some charts are easier to read than others

Before the individual charts, one piece of research explains almost every rule that follows. In 1984 the statisticians William Cleveland and Robert McGill asked people to judge values from charts and measured how wrong they were. They ranked the ways a chart can encode a number, from the most accurate to the least3:

RankHow the number is shownTypical chart
1Position along a common scaleBars or dots from the same baseline
2Position along scales that are not alignedSmall charts side by side
3Length, direction, angleStacked segments, slopes, pie slices
4AreaBubbles, treemaps
5Volume, curvature3D shapes
6Shading, colour saturationHeatmaps

In 2010 Jeffrey Heer and Michael Bostock repeated the experiment online with many more people. The ranking of position and length held, and both angle and circular area were read significantly worse than position4. So the most accurate thing you can ask the eye to do is compare positions against one shared baseline. That is exactly what a sorted bar chart and a line chart ask for, and it is why they win so often.

Example: Sam compares two numbers two ways

Sam takes two of Aoife's numbers: online orders from email (330) and from direct visits (283). Drawn as two bars from the same baseline, the eye compares the tops of the bars and sees at once that email is a bit more than direct. Drawn as two slices of a pie, the eye has to compare two angles that start at different places on the circle. Most people can tell which slice is bigger only when the gap is large. Here it is 47 orders out of 2,360, about two percentage points, and the pie hides it. Nothing about the data changed. Only the encoding did, and with it the accuracy of every reader in the room.

Comparison: sorted bars

When the question is "which is biggest", use bars, and sort them from largest to smallest. Sorting does the reader's first job for them. The Analysis Function guidance says the same: rank bars by value unless the categories have a natural order, such as age bands or months2. Two more rules come straight from the perception research. Bars start at zero, because the length of the bar is the number, and a bar cut short tells a lie about its own length (the third chapter, how charts lie, shows how badly). And if one bar is the answer, give it the colour and let the others go grey.

A note on long labels. A common piece of advice is to turn the bars sideways when the category names are long. On my own website every chart sets its category labels at 45 degrees instead, so long names still fit on a phone, and every chart can open in a full screen view when a reader wants to look closer. Either approach works. What matters is that nobody has to tilt their head to read a label lying flat on its side.

Example: which category made the most money

Sam lists last quarter's revenue (July to September) by category: jackets €182k, footwear €146k, tents €97k, rucksacks €74k, sleeping bags €51k and accessories €38k. The spreadsheet's default puts the categories in alphabetical order, so accessories comes first and jackets sits somewhere in the middle. Sam sorts them, starts the axis at zero, highlights jackets, and writes a title that answers Aoife's question instead of naming the data.

Bar chart of invented revenue by product category for Shannon Outfitters in July to September 2026, sorted from largest to smallest: jackets 182 thousand euro, footwear 146, tents 97, rucksacks 74, sleeping bags 51, accessories 38. Jackets are highlighted and earned almost a third of the 588 thousand euro total. Tap or click to open full screen and move around

Aoife reads the answer in the title, checks it against the bars in a second, and can see the next question coming: footwear is not far behind.

Ranking: sorted bars, or simply a table

Ranking looks like comparison, and the chart is the same sorted bar. The difference is the question. A ranking asks about order ("what is in the top five"), so the exact positions matter more than the size of the gaps. When the reader needs to look up a precise figure, a ranked table is often better than any chart, which is the point the first chapter of this series makes at length. There is one trap: a ranking always ranks by something, and a different measure can give a different order. Say in the title which measure you ranked by.

Example: which five products earn a place on the homepage

Aoife wants the five bestsellers on the homepage. Sam ranks last quarter's products by units sold and puts the result in a table, because the homepage team needs the exact names and numbers rather than the shape.

Rank by unitsProductUnits soldRevenue
1Burren waterproof jacket412€61,388
2Galtee walking boots365€47,085
3Merino base layer340€16,660
4Trekking poles, pair298€11,622
5Lough two person tent118€28,202

Then Sam adds one line under the table: ranked by revenue, the tent moves up to third place and the trekking poles drop to fifth. If the homepage is there to sell volume, the units ranking is right. If it is there to earn money, the revenue ranking is. Aoife picks revenue, and the order changes. The data did not change. The question did.

Change over time: lines

When the question is "how has it moved", time runs along the horizontal axis, from left to right, and a line joins the points. A line shows the shape of the movement: the trend, the seasonal rhythm, the sudden break. The eye reads the slope between neighbouring points, which a set of separate bars makes harder.

Bars are fine for time when there are only a few periods, say four quarters or five years, and the reader cares about each period's total rather than the path between them. The Analysis Function adds two useful limits: use bars over time only when the periods are equally spaced, and keep a line chart to about four lines before it turns into spaghetti2. A line chart may also start its vertical axis above zero when the movement is the point, because the eye reads the slope, not the length from the baseline. Chapter three explains when that is fair and when it is not.

Example: how online revenue moved over two years

Sam plots monthly online revenue from November 2024 to October 2026 as one line.

Line chart of invented monthly online revenue for Shannon Outfitters from November 2024 to October 2026, in thousands of euro. Revenue peaks every December (112 in 2024, 131 in 2025) and dips every January and February (around 60 in 2025, around 70 in 2026). Every month in the second year is higher than the same month in the first year, from 88 to 104 in November and from 83 to 97 in October. Tap or click to open full screen and move around

Two things jump out that no table of 24 numbers would show as quickly. First, the rhythm: every December spikes and every January falls back, so a bad January is not news. Second, the growth: every month in the second year sits above the same month in the first. Sam writes that second point as the title, because it is what Aoife asked. If Aoife had asked "how did each year end", four yearly totals as bars would have been the better answer.

Part to whole: why the pie struggles

"What is it made of" is the question pie charts were invented for, and it is the one where they most often let the reader down. A pie asks the eye to compare angles and areas, which sit in the third and fourth rows of the Cleveland and McGill ranking3. Robert Kosara and Drew Skau later found that people do not even read pies purely by angle: area and arc length play a part, and pie variants that distort the shape (exploded slices, ellipses, 3D) made the errors larger5. With two or three slices that hardly matters. With eight it does, because neighbouring slices of similar size become impossible to rank, the labels crowd together, and the reader starts matching colours to a legend.

It is fair to say what pies do well. Stephen Few, one of their sharpest critics, grants that a pie announces "part of a whole" more directly than any other chart6. Claus Wilke's textbook concludes that no single chart for proportions wins every time, and that pies are good at showing simple fractions such as a half or a quarter in very small datasets, while bars are better for comparing the parts with each other7. The Analysis Function puts a practical limit on it: five categories or fewer, one of them dominant, sorted, starting at twelve o'clock, labelled directly2. My own rule of thumb is stricter: two or three parts, or a bar chart.

To read a share, you divide the part by the whole:

share=partwhole×100%\text{share} = \frac{\text{part}}{\text{whole}} \times 100\%

Here part\text{part} is the count or amount for one category, whole\text{whole} is the total of all categories, and share\text{share} is that category's percentage of the total. If percentages make you uneasy, Foundations: percentages builds them up from scratch.

Example: where online orders come from, shown wrong first

Last quarter Shannon Outfitters took 2,360 online orders. Sam's first draft is the default pie from the analytics tool, with eight traffic channels.

Pie chart of 2,360 invented online orders in July to September 2026 split across eight channels: organic search 637, paid search 496, email 330, direct 283, paid social 212, organic social 165, affiliates 142, referral 95. With eight slices of similar size it is hard to tell whether email beats direct, or paid social beats organic social, without reading the labels. Tap or click to open full screen and move around

Try it honestly: without reading the numbers, is email bigger than direct? Is paid social bigger than organic social? Now the same data as sorted bars, shown as shares so the "part of a whole" message survives.

The same invented data as a sorted bar chart of the share of 2,360 online orders by channel in July to September 2026: organic search 27 percent, paid search 21 percent, email 14 percent, direct 12 percent, paid social 9 percent, organic social 7 percent, affiliates 6 percent, referral 4 percent. Every comparison between neighbouring channels is now easy, and the two search channels together bring 48 percent. Tap or click to open full screen and move around

The share for email works out as:

shareemail=3302,360×100%=14.0%\text{share}_{\text{email}} = \frac{330}{2{,}360} \times 100\% = 14.0\%

So shareemail\text{share}_{\text{email}} is 14.0%, and every comparison that was a guess in the pie is now a glance: email (14%) beats direct (12%), paid social (9%) beats organic social (7%). The title states what Aoife needs to know: the two search channels bring almost half the orders, which matters the day someone suggests cutting the paid search budget.

Example: when the parts change over time, stack the bars

Aoife's follow up question is "is online taking over from the shop?" That is part to whole and change over time at once. A row of pies, one per year, would ask the eye to compare angles across separate circles, the worst of both. Stacked bars put each year's total on one common baseline and split it into its parts.

Stacked bar chart of invented yearly revenue for Shannon Outfitters from 2022 to 2025, in thousands of euro, split into online and shop. Online: 610, 720, 840, 981. Shop: 990, 940, 900, 860. The total grows from 1,600 to 1,841, and online overtakes the shop in 2025. Tap or click to open full screen and move around

There is a catch worth knowing, and it follows from the ranking. The bottom segment (online) sits on the shared baseline, so its growth is easy to read. The top segment (shop) floats on top of a moving base, so the eye has to compare lengths that are not aligned, which is harder. Put the part you care about most at the bottom. Wilke makes the same point and advises against long sequences of stacked bars with more than two segments for this reason7. If Aoife wanted to compare the shop's decline year by year in detail, Sam would draw it as its own line.

Distribution: histograms and box plots

"What does a typical order look like" is not a question about one number. It is a question about many numbers and how they spread. An average alone answers it badly, because one large tent order pulls the mean up while most orders stay small (the article on why averages mislead goes deep on this). The chart for a distribution is the histogram: cut the range of values into equal bins, count how many values fall in each, and draw the counts as touching bars.

How many bins? Too few and the shape vanishes, too many and it turns into noise. A classic starting point is Sturges' rule from 19268, which fits on one line:

k=1+log⁡2nk = 1 + \log_2 n

Here kk is the suggested number of bins, nn is the number of values, and log⁡2n\log_2 n is the power you would raise 2 to in order to get nn (how many times you can halve nn before you reach 1). Foundations: logarithms explains logarithms from the ground up. Treat the result as a first guess. Sturges built the rule with roughly bell shaped data in mind, and shop data rarely is: order values have a long tail of rare large orders. In practice, round bin widths (€10, €25, €50) read better than whatever the formula gives.

A box plot is the compact cousin of the histogram, popularised by John Tukey9. It draws five numbers: the smallest value, the first quartile (a quarter of the values sit below it), the median (half below), the third quartile (three quarters below) and the largest value, with unusually large or small values drawn as single dots. The middle box covers the middle half of the data, and its width is the interquartile range:

IQR=Q3−Q1,upper fence=Q3+1.5×IQR\text{IQR} = Q_3 - Q_1, \qquad \text{upper fence} = Q_3 + 1.5 \times \text{IQR}

Here Q1Q_1 is the first quartile, Q3Q_3 the third quartile, IQR\text{IQR} the interquartile range, and any value above the upper fence\text{upper fence} is drawn as a separate dot rather than inside the whisker. Foundations: spread covers quartiles and the interquartile range. The chart library on this site does not draw box plots, so when I need one here I give its five numbers as a table, and the histogram does the visual work. Box plots shine when you compare several distributions side by side, for example order values per channel.

Example: what a typical online order looks like

Sam takes the 2,360 online orders from last quarter. Sturges suggests:

k=1+log⁡22,360=1+11.2=12.2k = 1 + \log_2 2{,}360 = 1 + 11.2 = 12.2

So k≈12k \approx 12 bins. But the values run from €6 to €689, and 12 equal bins over that range would each be about €57 wide, which would crush most orders into the first two bars. So Sam does what the rule's caveat suggests and keeps round €25 bins up to €250, which covers 2,320 of the orders, and mentions the 40 larger orders in the title rather than stretching the axis for them.

Histogram of 2,320 invented online orders up to 250 euro in July to September 2026, in bins of 25 euro: under 25 euro 310 orders, 25 to 50 euro 640, 50 to 75 euro 520, 75 to 100 euro 330, 100 to 125 euro 190, 125 to 150 euro 120, 150 to 175 euro 80, 175 to 200 euro 55, 200 to 225 euro 40, 225 to 250 euro 35. The shape rises to a peak at 25 to 50 euro and then falls away in a long tail. A further 40 orders above 250 euro are not shown. Tap or click to open full screen and move around

The shape tells Aoife what one average could not: the typical order is a pair of socks, a base layer or a hat, somewhere between €25 and €75, and a long tail of jackets, boots and tents stretches out to the right. Sam adds the box plot's five numbers as a table:

SmallestFirst quartileMedianThird quartileLargest
€6€36€61€98€689

The interquartile range is €98 minus €36, so €62, and the upper fence is €98 plus 1.5 times €62, so €191. Every order above €191 counts as unusually large. The mean order value is €77, well above the median of €61, which is the long tail at work. If Aoife is setting a free delivery threshold, this chart is the one she needs, not the average.

Relationship: scatter plots

"Does one thing move with another" calls for a scatter plot. Each dot is one observation, here one week, placed by two measures at once: one along the horizontal axis, one up the vertical. If the dots climb from left to right, the two move together. A trend line drawn through them shows the general direction. The strength of a straight line relationship is summed up by Pearson's correlation coefficient:

r=∑i=1n(xi−xˉ)(yi−yˉ)∑i=1n(xi−xˉ)2  ∑i=1n(yi−yˉ)2r = \frac{\sum_{i=1}^{n} (x_i - \bar{x})(y_i - \bar{y})}{\sqrt{\sum_{i=1}^{n} (x_i - \bar{x})^2 \; \sum_{i=1}^{n} (y_i - \bar{y})^2}}

Here nn is the number of dots, xix_i and yiy_i are the two measures for dot ii, xˉ\bar{x} and yˉ\bar{y} are their means, and rr runs from −1-1 (a perfect falling line) through 00 (no straight line pattern) to +1+1 (a perfect rising line). Foundations: correlation walks through it slowly.

Then the warning that belongs next to every scatter plot: correlation is not causation. Two measures can move together because one drives the other, because a third thing drives both, or by coincidence. The scatter shows that they move together. It cannot tell you why. Chapter three of this series shows how charts manufacture correlations that are not there at all.

Example: does more paid search spend bring more revenue

Sam plots 12 weeks from the summer and autumn: weekly paid search spend against weekly online revenue.

Scatter plot of 12 invented weeks for Shannon Outfitters. Horizontal axis: weekly paid search spend from 1.0 to 4.6 thousand euro. Vertical axis: weekly online revenue from 17.5 to 25.9 thousand euro. The dots rise steadily from left to right close to a straight trend line, a correlation of about 0.97. The chart shows that the two moved together, not that the spend caused the revenue. Tap or click to open full screen and move around

Sam works out the three sums from the 12 weeks: the cross products add up to 31.94, the squared spend deviations to 14.16 and the squared revenue deviations to 77.34. So:

r=31.9414.16×77.34=31.9433.09=0.97r = \frac{31.94}{\sqrt{14.16 \times 77.34}} = \frac{31.94}{33.09} = 0.97

An rr of 0.97 is a very strong positive relationship. The trend line says that, across these weeks, each extra €1,000 of spend went with roughly €2,260 more revenue. Before Aoife doubles the budget, Sam points out the third thing: the budget was raised in the weeks before the autumn walking season on the Burren and the Galtees, when demand was rising anyway. Some of that revenue would have come without the ads. The honest answer to "does it bring more revenue" is a test, for example switching paid search off in a holdout period and comparing, and the article on statistical significance and error margins explains how to read the result. The scatter plot earned the question. It did not answer it.

Deviation: diverging bars

"Are we above or below the line" is a question about the gap, not the level. When Aoife asks how the shop did against target, a chart of the actual numbers and a chart of the targets makes her subtract in her head, month by month. A diverging bar chart does the subtraction for her: each bar is the gap, growing up from zero when the month beat target and down from zero when it missed.

The gap can be shown in euro or as a percentage:

deviation=actual−target,deviation%=actual−targettarget×100%\text{deviation} = \text{actual} - \text{target}, \qquad \text{deviation}_{\%} = \frac{\text{actual} - \text{target}}{\text{target}} \times 100\%

Here actual\text{actual} is what happened, target\text{target} is what was planned, deviation\text{deviation} is the gap in the original unit (euro) and deviation%\text{deviation}_{\%} is the same gap as a percentage of the target. Use euro when the size of the business is stable, a percentage when you compare units of very different sizes. Foundations: percentages has more on percentage change.

Example: online revenue against target this year

Sam takes this year's monthly online revenue (the same numbers as the line chart above) and the monthly targets from the plan, and draws only the gap.

Diverging bar chart of the invented gap between actual and target online revenue for Shannon Outfitters, January to October 2026, in thousands of euro: January minus 3, February minus 1, March plus 3, April plus 5, May plus 3, June minus 2, July plus 3, August minus 4, September minus 5, October plus 2. Bars above zero beat target, bars below missed it. Spring was strong, and June, August and September missed. Tap or click to open full screen and move around

Sam checks the title against the bars before sending it: from June to October the months go miss, beat, miss, miss, beat. Three misses in five months. A finding title is a claim, so it gets checked like one.

For August, actual revenue was €94k against a target of €98k:

deviation%=94−9898×100%=−4.1%\text{deviation}_{\%} = \frac{94 - 98}{98} \times 100\% = -4.1\%

A deviation%\text{deviation}_{\%} of −4.1%-4.1\% in one month is small. Two of them back to back in late summer, after a strong spring, is a pattern worth a conversation. The diverging bars make that pattern visible in a way two overlapping lines would not.

Which chart for which question: the whole decision

Here is the whole method as one decision. Start from the question, follow the branch, and finish every chart the same way: with a title that states the finding.

Tap or click to open full screen and move around

Two kinds of chart are missing on purpose. Maps answer "where", and flow diagrams answer "what leads to what". Both are real question types (the FT poster lists them as spatial and flow1), and the tenth chapter of this series covers technical diagrams that people understand. Uncertainty, the question of how sure you are, gets its own chapter too: showing uncertainty honestly.

Example: Sam's one page for the quarterly review

Sam puts the method to work on Aoife's five questions and the two that followed, and the review pack reads like this:

  1. Which category made the most money? Sorted bars, jackets highlighted, title "Jackets made the most money last quarter".
  2. Which products go on the homepage? A ranked table by revenue, with a note that the order changes by units.
  3. How has online revenue moved? One line over 24 months, title "Online revenue rose every month against the same month a year earlier".
  4. Where do online orders come from? Sorted bars of shares, both search channels highlighted, the pie retired.
  5. Is online taking over from the shop? Stacked bars per year, online at the bottom.
  6. What does a typical order look like? A histogram in €25 bins plus the five number table.
  7. Does paid search bring revenue? A scatter with a trend line, and a proposal for a holdout test.
  8. Are we on target? Diverging bars, with the late summer misses highlighted.

Eight questions, six chart types and two tables, and not one chart chosen because it looked nice. Aoife gets through the pack in ten minutes, says it's grand, and spends the rest of the meeting on decisions, which is the whole point.

Conclusion

Choosing a chart is not a matter of taste. It is a matter of the question. Say the question in one sentence, name its type, and the chart nearly chooses itself: sorted bars for comparison and ranking, lines for change over time, stacked bars for parts of a whole, a histogram for distribution, a scatter for relationships and diverging bars for gaps to a target. Behind every one of those rules sits the same finding: the eye judges position on a shared baseline far better than angle or area. Pies earn their place only with two or three parts. Scatter plots show that things move together, never why. And every chart ends with a title that states what the reader should take away, checked against the chart before anyone else sees it.

The next chapter, how charts lie, and how to catch them, takes the right chart and shows how a truncated axis, a second axis or a cherry picked window can still make it mislead.

If your reports are full of charts nobody acts on, book the free workshop: we take the questions your board actually asks and match each one to the chart that answers it, and the first two weeks of work are on me.

Sources

  1. 1.

    Financial Times Visual Journalism Team, Visual Vocabulary, The Financial Times, GitHub repository chart doctor, https://github.com/Financial-Times/chart-doctor/tree/main/visual-vocabulary

  2. 2.

    Government Analysis Function (Office for National Statistics), Data visualisation: charts, 19 May 2022, https://analysisfunction.civilservice.gov.uk/policy-store/data-visualisation-charts/

  3. 3.

    William S. Cleveland and Robert McGill, Graphical perception: theory, experimentation, and application to the development of graphical methods, Journal of the American Statistical Association 79 (387), 531 to 554, 1984, https://doi.org/10.1080/01621459.1984.10478080

  4. 4.

    Jeffrey Heer and Michael Bostock, Crowdsourcing graphical perception: using Mechanical Turk to assess visualization design, Proceedings of the ACM Conference on Human Factors in Computing Systems (CHI), 203 to 212, 2010, https://homes.cs.washington.edu/~jheer/files/2010-MTurk-CHI.pdf

  5. 5.

    Robert Kosara and Drew Skau, Judgment error in pie chart variations, EuroVis 2016 Short Papers, Eurographics Association, 2016, https://doi.org/10.2312/eurovisshort.20161167

  6. 6.

    Stephen Few, Save the pies for dessert, Visual Business Intelligence Newsletter, Perceptual Edge, August 2007, https://www.perceptualedge.com/articles/visual_business_intelligence/save_the_pies_for_dessert.pdf

  7. 7.

    Claus O. Wilke, Fundamentals of Data Visualization, chapter 10: Visualizing proportions, O'Reilly Media, 2019, free online edition, https://clauswilke.com/dataviz/visualizing-proportions.html

  8. 8.

    Herbert A. Sturges, The choice of a class interval, Journal of the American Statistical Association 21 (153), 65 to 66, 1926, https://doi.org/10.1080/01621459.1926.10502161

  9. 9.

    Robert McGill, John W. Tukey and Wayne A. Larsen, Variations of box plots, The American Statistician 32 (1), 12 to 16, 1978, https://doi.org/10.1080/00031305.1978.10479236