Why a chart, and when a table is better

TL;DR

Difficulty: Beginner

A table is for looking up exact values, a chart is for seeing patterns, trends and exceptions. Often the best answer is both.

  • Summary numbers can agree perfectly while the data tells four different stories (Anscombe's quartet). Draw the data before you trust a summary.
  • The eye picks up colour, length and position before you think, and judges position and length on a shared scale far more accurately than angle or area.
  • One chart, one message. Write the message as a sentence first, put it in the title, and check the title against the data.
  • Remove everything that does not carry data or help read it, grey the rest and give the one thing that matters the one strong colour.
  • Add one line on what the finding means for the decision, and keep the table in the appendix.

Every month the analyst sends a table, and every month the owner scrolls past the one thing that mattered. This chapter explains what a chart does that a table cannot, when a table is still the better choice, and how to turn a monthly table into one chart that tells one clear story, step by step, on an invented outdoor gear brand in Leeds.

The question behind this article

Every month Sam, the analyst at Pennine Outfitters, sends the owner a table. Pennine Outfitters is an outdoor gear brand in Leeds that sells through its web shop (pennineoutfitters.example), through one shop in the city and to a handful of other retailers. The business, Sam, the owner Helen and every number about them in this article are invented. The research, the books and the guidance quoted are real and listed in the footnotes.

Helen opens the table, scrolls, frowns at a few cells and closes it. Somewhere in those numbers is the most important thing that happened in the business this half year. She did not see it, and nobody told her.

So this article answers one question: why turn your numbers into a chart at all, and how do you make that chart tell one clear story? It also answers the opposite question, because sometimes the honest answer is "send the table".

This is the first chapter of a series on charts that tell the truth. It builds on one idea from What is an insight, and why most reports do not contain one: a number only matters when it changes a decision. A chart is how that number reaches the person who decides.

What a chart does that a table cannot

A table and a chart can hold exactly the same numbers. They are good at different jobs.

Stephen Few, who spent his career teaching business people to display numbers, puts the split plainly: a table works best when someone needs to look up an individual value or needs the value to be precise, and a graph works best when the message sits in the shape of the data, in its patterns, trends and exceptions1. Iris Vessey gave the same split a name in 1991, cognitive fit. She sorted questions into two kinds. Symbolic questions ask for exact values ("what was the figure in May?"), and spatial questions ask about relationships ("which channel grew fastest?"). People answer faster and more accurately when the format matches the kind of question: tables for symbolic ones, graphs for spatial ones2.

The reason is how you read each one. A table is read cell by cell, like text, so to see a trend you must hold several numbers in your head. A chart turns each number into a position or a length, and your eye compares them for you.

Example: two questions, one table

Here is Sam's table for the first half of 2026. The figures are monthly revenue in thousands of pounds, and they are invented.

MonthLeeds shopWeb shopWholesale
January£62k£49k£30k
February£58k£51k£34k
March£60k£68k£28k
April£66k£81k£22k
May£70k£93k£25k
June£72k£100k£31k

Question one, from the accountant: "What did the web shop take in May?" Find May, move across to the web shop, read £93k. Two seconds. A chart would make this harder, because you would have to read a bar against an axis.

Question two, from Helen: "When did the web shop start bringing in more than the Leeds shop?" Now you compare two columns row by row. January, 49 against 62, no. February, 51 against 58, no. March, 68 against 60, yes. You only find the answer because you knew what to look for. Helen did not. That is the whole problem.

The first question is symbolic, so the table wins. The second is spatial, so a chart wins. Same numbers, different job.

When a table is the better choice

Charts are the subject of this series, so it is worth saying clearly when not to draw one. A table is the better choice when:

  • the reader needs exact values to copy, check or reconcile (an accountant, an auditor, a supplier invoice)
  • there are only a handful of numbers, and a sentence or a small table says it faster
  • the numbers have different units that would need several axes (revenue in pounds, orders as counts, conversion as a rate)
  • the reader will look things up rather than read from start to finish, as in an appendix or a price list.

The UK Government Analysis Function, which sets the standard for official statistics, makes the same point in its chart guidance: if you expect people to read values off a chart, give them a table, and a well structured table can be as powerful as a chart3. A table next to a chart also gives people who use a screen reader the numbers the chart shows, which chapter 11, Accessible charts: colour, contrast and screen readers, covers in depth.

Often the best answer is both: the chart for the message, the table for the detail. The decision comes down to two questions about the reader.

Tap or click to open full screen and move around

Example: the accountant and the owner

Sam sends the same half year to two people. The accountant reconciles monthly revenue per channel against the bank and the VAT records: exact values yes, pattern no, so the accountant gets the table and nothing else.

Helen decides where the autumn marketing money goes. She will not copy any of these numbers anywhere, so exact values no. Pattern yes: something changed between the web shop and the Leeds shop. Helen gets one chart with a title that says what changed, and the table goes into the appendix.

Summary numbers can hide everything: Anscombe's quartet

There is a stronger reason to draw your data than speed. Summary numbers can agree perfectly while the data behind them tells completely different stories.

In 1973 the statistician Francis Anscombe published four small datasets to make this point. He was pushing back against the habit of trusting calculations and treating graphs as decoration4. Each dataset has 11 pairs of numbers, an xx and a yy. The four datasets share their summary statistics almost exactly:

StatisticDataset IDataset IIDataset IIIDataset IV
Mean of xx9999
Mean of yy7.507.507.507.50
Variance of xx11111111
Variance of yy4.134.134.124.12
Correlation0.8160.8160.8160.817
Best fitting liney=3.00+0.500xy = 3.00 + 0.500xy=3.00+0.500xy = 3.00 + 0.500xy=3.00+0.500xy = 3.00 + 0.500xy=3.00+0.500xy = 3.00 + 0.500x

Every figure in that table can be recalculated from the published data5. Two of the formulas behind it are worth knowing. The mean is the plain average (Foundations: mean):

xˉ=1n∑i=1nxi\bar{x} = \frac{1}{n}\sum_{i=1}^{n} x_i

where xˉ\bar{x} is the mean, nn is the number of values (11 here), xix_i is the iith value, and ∑i=1n\sum_{i=1}^{n} means "add up all nn of them".

The correlation coefficient measures how closely the points follow a straight line, from −1-1 (a perfect falling line) through 00 (no straight line at all) to +1+1 (a perfect rising line) (Foundations: correlation):

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}\,\sqrt{\sum_{i=1}^{n} (y_i - \bar{y})^2}}

where rr is the correlation, xix_i and yiy_i are the two values of the iith pair, xˉ\bar{x} and yˉ\bar{y} are their means, the top line adds up how far each pair sits from the means in the same direction, and the bottom line scales that by how spread out each variable is, so rr always lands between −1-1 and +1+1.

Now look at the data itself.

Scatter chart of Anscombe's first dataset, 11 points. The points rise from left to right in a loose cloud around a straight line, which is what the summary statistics suggest: mean of y 7.50, correlation 0.816, best fitting line y equals 3 plus 0.5 x. Tap or click to open full screen and move around
Scatter chart of Anscombe's second dataset, 11 points. The points lie on a smooth arch: they rise, flatten around x equals 11 and fall again. A straight line is the wrong description, yet the summary statistics are the same as dataset I. Tap or click to open full screen and move around
Scatter chart of Anscombe's third dataset, 11 points. Ten points lie exactly on a gently rising straight line. One point, at x equals 13 and y equals 12.74, sits far above it and drags the fitted line upwards. The summary statistics are again the same as dataset I. Tap or click to open full screen and move around
Scatter chart of Anscombe's fourth dataset, 11 points. Ten points sit in a vertical column at x equals 8 with y between 5.25 and 8.84. A single point at x equals 19 and y equals 12.5 creates the entire rising trend. Without it there would be no relationship to measure at all. Tap or click to open full screen and move around

Four datasets, one set of statistics, four different realities. Dataset I is what most people imagine when they read "correlation 0.816". Dataset II is a curve, so a straight line describes it badly. Dataset III is a perfect line spoiled by one odd point, perhaps a typing error. Dataset IV has no relationship at all, except for one point that manufactures the whole trend.

In 2017 Justin Matejka and George Fitzmaurice took the idea further with the Datasaurus Dozen: 12 datasets whose means, standard deviations and correlations match to two decimal places, one of them shaped like a dinosaur6. Only a picture tells them apart.

Example: four datasets that agree on every number

Take the mean of xx in dataset I and do it by hand, so you can see there is no trick. The 11 values are 10, 8, 13, 9, 11, 14, 6, 4, 12, 7 and 5. They add up to 99:

xˉ=9911=9\bar{x} = \frac{99}{11} = 9

Here xˉ\bar{x} is the mean of xx, 99 is the sum of the 11 values and n=11n = 11. Dataset IV uses ten eights and one 19, which is 80+19=9980 + 19 = 99 as well, so its mean is also exactly 9. The arithmetic is honest in both cases. It simply answers a different question from the one you care about.

Now the business version. Suppose Sam reported that over 11 weeks, social ad spend and weekly web shop revenue had a correlation of 0.82, and Helen heard "more ads, more revenue". If the weeks looked like dataset IV, ten ordinary weeks and one sale week with a big ad spend, the "relationship" is one event. The correlation is correct. A decision built on it would not be. One scatter chart shows that in a second. Statistical thinking for growth: why averages mislead you goes much deeper into how one summary number hides a whole distribution.

How the eye reads a chart before you think

Why can a chart show in a second what a table hides? Because part of seeing happens before attention.

Vision scientists call certain visual properties preattentive. Your visual system picks them up across the whole field of view in a fraction of a second, without you deciding to search. Christopher Healey and James Enns, reviewing decades of this research for visualisation designers, describe tasks done on large displays in under 200 to 250 milliseconds as preattentive. That is faster than the eye can even move to look somewhere deliberately7. The properties that work this way include colour (hue and intensity), size, length and width, orientation, shape features such as closure and curvature, and motion8.

This is what a good chart uses. A pink bar among grey ones, a line that suddenly turns down, one dot far away from the cloud: each of these jumps out before you read a single label. A table offers almost none of it. Every cell is the same font, the same size and the same colour, so your eye has nothing to grab and must read.

There is a catch. Pop out works for one property at a time. A target defined by a combination, say "the large pink one among large grey and small pink ones", needs slow, item by item search8. A chart that uses colour, size and shape for different things at once loses that speed.

Example: find the threes

Count how many times the digit 3 appears in this block. Notice how you do it.

7 5 9 1 8 2 3 6 4 0 9 7 1 5 8 2 6 3 4 0 9 1 7 2 5 8 4 6 0 3 1 9 5 7 2 8 6 4 0 3

You had to read, digit by digit, from start to end. Now the same block, with the threes in bold:

7 5 9 1 8 2 3 6 4 0 9 7 1 5 8 2 6 3 4 0 9 1 7 2 5 8 4 6 0 3 1 9 5 7 2 8 6 4 0 3

There are four, and this time you probably knew it before you had consciously counted. Nothing about the data changed. One visual property, the weight of the stroke, did the searching for you. That is the effect a single highlight colour has on a chart, and you will see it on Sam's chart in a moment.

Some visual encodings are read more accurately than others

Preattentive properties tell you what jumps out. A second body of research tells you which ways of drawing a number let people judge it accurately.

In 1984 William Cleveland and Robert McGill ran experiments in which people judged the relative size of two values shown in different ways: as positions on a scale, as lengths, as angles, as areas, as shades. They ranked these "elementary perceptual tasks" from most to least accurate9. Here is their ranking:

RankHow the number is drawnWhere you see it
1Position along a common scaleBars or dots against one shared axis
2Position along scales that are not alignedThe same measure in two separate small charts
3Length, direction, angleStacked bar segments, slopes, pie slices
4AreaBubbles, squares, treemaps
5Volume, curvature3D shapes, curves
6Shading, colour saturationHeatmaps, shaded maps

In 2010 Jeffrey Heer and Michael Bostock repeated these experiments with hundreds of people recruited online and found the same ordering, which makes it one of the better replicated findings in the field10. The practical rule is simple: when the comparison matters, put the numbers on a common scale and compare positions or lengths. Save area and shading for situations where rough magnitude is enough. Chapter 2, Choosing the right chart for the question, turns this ranking into a choice of chart for each kind of question.

Example: length against area

Pennine Outfitters' web shop took £100k in June and £49k in January, roughly double. Suppose a designer draws each month as a circle, with the circle's area matching the revenue. How much wider is the June circle?

The area of a circle grows with the square of its diameter, so the diameters relate as the square root of the areas (Foundations: ratios and rates):

d2d1=A2A1\frac{d_2}{d_1} = \sqrt{\frac{A_2}{A_1}}

where d1d_1 and d2d_2 are the diameters of the two circles and A1A_1 and A2A_2 are their areas, which here stand for January's and June's revenue. With June roughly twice January:

d2d1=2=1.41\frac{d_2}{d_1} = \sqrt{2} = 1.41

So the June circle is only about 1.411.41 times as wide as January's, even though it stands for twice the money. Readers tend to judge the width, not the area, and see growth of about 40% where there was about 100%. Two bars on a shared axis show the same pair as one bar exactly twice as long as the other, which is the comparison the eye judges best. Chapter 3, How charts lie, and how to catch them, shows how area and 3D effects are used to mislead on purpose.

One chart, one message

Once you know what the eye does well, the next question is what you ask it to do. The commonest failure in business charts is not a wrong number. It is a chart that tries to say four things and says none.

Claus Wilke, in his free textbook on data visualisation, argues that a figure should be made for a busy audience: strip it to what supports your point, help people see the pattern you see, and do not expect them to work it out alone11. Cole Nussbaumer Knaflic, writing for business presenters, goes a step further. Before you open any chart tool, she suggests, write down your main idea as one complete sentence that states your point of view and what is at stake12. If you cannot write that sentence, you are not ready to draw the chart.

This is a discipline, not a style. A chart with one message can be checked against the data. A chart with no stated message lets every reader find a different one.

Example: three messages fighting in one chart

Here is the chart most spreadsheet tools produce from Sam's table, with every channel and every month as bars, a legend and a title that names the data.

Grouped bar chart with 18 bars: three channels, Leeds shop, web shop and wholesale, for each month from January to June 2026, in thousands of pounds, invented figures. The title only names the data. Several stories compete: the web shop grows from 49 to 100, the Leeds shop rises gently from 62 to 72, wholesale wanders between 22 and 34, and the web shop passes the Leeds shop in March, but nothing points the reader to any of them. Tap or click to open full screen and move around

Ask three colleagues what this chart says and you will get three answers:

  1. "The web shop doubled in six months." True: £49k to £100k.
  2. "Wholesale dipped in the spring." True: £34k in February, £22k in April.
  3. "The web shop overtook the Leeds shop in March." Also true, and the one that matters for Helen's decision about where the autumn marketing money goes.

All three are in the chart, and none of them is said. Eighteen bars of equal visual weight ask the reader to find the story alone. Sam's job is to choose the one message that serves the decision, and then make the chart say only that. The worked example at the end does exactly this.

Titles that state the finding

The title is the part of a chart people read first and remember longest. Michelle Borkin and colleagues tracked where people's eyes went on hundreds of real charts and what they could recall afterwards. Among their conclusions: titles and supporting text should convey the message of the visualisation13.

The Analysis Function's guidance turns this into a rule of two titles3:

  • a headline title that states the main message, for example "The web shop passed the Leeds shop in March";
  • a statistical subtitle that states what is plotted, where and when, for example "Monthly revenue, Leeds shop and web shop, January to June 2026, £ thousands".

Compare that with "Revenue by channel". The old title tells you what the chart contains. The new one tells you what it means, and the subtitle keeps the facts so nothing is lost.

This power cuts both ways. Ha Kyung Kong, Zhicheng Liu and Karrie Karahalios showed people the same charts with titles slanted in opposite directions. The slant of the title changed what people took to be the chart's main message, with viewers drawing opposing messages from the same visualisation14. A finding title is only allowed when the finding is true and fair. If a title needs the data bent to support it, the title is wrong, not the data. Chapter 4, Same data, different truths: perspective and framing, is about exactly this choice.

Example: rewriting Sam's title

Sam starts from "Revenue by channel, January to June 2026". It is accurate and says nothing. A first rewrite, "Online is growing", is vague: growing compared with what? The decision Helen faces is the Leeds shop against the web shop, so the title has to compare those two. Final version:

  • Headline: "The web shop passed the Leeds shop in March, and the gap is growing."
  • Subtitle: "Monthly revenue, Leeds shop and web shop, January to June 2026, £ thousands, invented figures."

Check it against the table. March: web shop £68k, Leeds shop £60k, so passed in March is correct. The gap: March £8k, April £15k, May £23k, June £28k, so "growing" is correct. The title survives a check against the data, which is the test every finding title must pass.

Declutter: every mark has to earn its place

Edward Tufte, whose 1983 book on the visual display of quantitative information is still the starting point for most of this field, coined the term data ink. Data ink is the part of a graphic that would change if the numbers changed: the bars, the points, the lines. Everything else, frames, background shading, heavy gridlines, decorative 3D, is non data ink15. His ratio:

data ink ratio=data inktotal ink used to print the graphic\text{data ink ratio} = \frac{\text{data ink}}{\text{total ink used to print the graphic}}

where data ink\text{data ink} is the ink that shows the numbers themselves and total ink used to print the graphic\text{total ink used to print the graphic} is everything on the chart, so the ratio runs from 00 (pure decoration) to 11 (nothing but data). The idea links straight to Foundations: ratios and rates.

Tufte's advice is to raise this ratio, within reason. The "within reason" matters: axis labels, a unit and a clear title are non data ink, and essential. The aim is a chart where nothing distracts from the data and nothing needed to read it is missing.

The Analysis Function's practical list matches this well: light grey gridlines and only a few of them, no shaded backgrounds, no boxes around legends, no 3D, no unnecessary markers, and label lines directly instead of using a legend where you can3.

Example: auditing Sam's default chart

Sam's spreadsheet first produced the chart with every default switched on. Sam goes through it element by element and asks one question each time: does this help Helen see the message?

  • 18 coloured bars, each with its value printed on top. The values repeat what the bars already show, and 18 labels crowd the chart. Remove them. The exact figures live in the table in the appendix.
  • A 3D effect on the bars. It adds depth that represents nothing and makes bar heights harder to read against the axis. Remove.
  • A grey shaded background and a box around the plot. Pure decoration. Remove.
  • Ten dark gridlines. Keep three, in light grey, so a reader can still estimate a level.
  • A legend in a box, far from the bars. Replace with direct labels once there are fewer series.
  • The wholesale series. True and interesting, but not part of this message. Move it to its own chart for the wholesale conversation.
  • Axis labels and the unit, £ thousands. Keep. They are non data ink that the reader needs.

Before: 18 bars, 18 value labels, ten gridlines, a legend, a background, a frame and a 3D effect. After: six bars, three light gridlines, the axis labels, a headline and a subtitle. Nothing that helps Helen read the message has gone.

Focus with one highlight colour

Decluttering removes noise. Highlighting adds a signal. The tool is the preattentive effect from the threes exercise: make everything grey, then give the one thing that carries the message the one strong colour.

Knaflic describes this as directing attention: decide where you want your audience to look, then use colour, size or position to send them there, and use it sparingly, because if everything is highlighted, nothing is12. The Analysis Function calls the same pattern a focus chart: the series the story is about in a strong colour, everything else in light grey3.

On this website the strong colour is brand pink and the rest is grey. Every chart on my site follows the same rules: brand colours only, light and dark mode, labels at 45 degrees, a full screen view, diagrams drawn top to bottom for phones, and a text description for screen readers.

Colour must never carry the message alone, because not everyone sees colour the same way. The title says in words what the pink shows, so the message survives in black and white and in a screen reader.

Example: one pink bar among grey ones

Helen's question is "when did the web shop take over?" The month that answers it is March. So in the final chart, five months are grey and March is pink. The eye lands on March before it reads the axis, the title confirms in words what the colour already said, and the grey bars on either side give the context: lower before, higher after.

Had Sam highlighted June, the biggest bar, the chart would say "June was the best month": true, and not the message. The highlight always points at the same thing the title states.

Context, audience and the so what

A chart is not finished when it is correct and clean. It is finished when the person reading it knows what to do next.

Knaflic suggests answering three questions before designing anything: who is your audience, what do you need them to know or do, and how will you show it12. Wilke's version is to remember that a viewer will invent a story if you do not give one, and the invented story may be "this is boring" or "the author is wrong"11.

This is where I start every analytics project: from the decision the numbers should support, not from a dashboard. Start with the decision, not the dashboard explains that method in full. For a single chart it shrinks to one line under the chart, the "so what": what the finding means for the decision on the table.

Example: the same finding for two audiences

The finding is the same for everyone: the web shop passed the Leeds shop in March and the gap is growing. The so what differs.

  • Helen, the owner, decides where autumn marketing money goes. She gets the chart with the finding title and one sentence: "The gap was £28k in June and has widened every month, so web shop photography and paid search deserve the larger share of the autumn budget." Sam deliberately makes no claim about how big the gap will be in the autumn. That would be a forecast, and forecasts get their own chapter.
  • The Leeds shop manager worries about the shop. Same chart, but the so what is different, and it matters to say it: "The shop grew too, from £62k to £72k. The web shop grew faster. Nothing here says the shop is shrinking." This is the same data. Leaving that line out would let the manager read a threat that is not in it.

Worked example: Sam's monthly table becomes one chart

Now the whole method, start to finish, on Sam's numbers. The steps follow the next diagram.

Tap or click to open full screen and move around

Example, step 1: write down the decision

Helen decides each August how to split the autumn marketing budget between the Leeds shop (longer opening hours, window displays, local events) and the web shop (product photography, paid search). Sam writes that at the top of a blank page. Every later choice is tested against it.

Example, step 2: state the message in one sentence

From the table, the fact most relevant to that decision: "The web shop passed the Leeds shop in March, and the gap has grown every month since." Sam checks it against the numbers: the gap was £8k in March, £15k in April, £23k in May and £28k in June. The sentence holds.

Example, step 3: pick the chart that shows that one message

The message is about one channel's weight compared with another's, month by month. Sam could draw two lines, the web shop and the Leeds shop, crossing in March. That works well, and chapter 2 explains when lines are the natural choice for change over time. Sam chooses something even simpler for Helen: the web shop's share of the two channels' combined revenue, one bar per month, so the crossing becomes a single number passing 50% (Foundations: percentages):

sm=WmWm+Lm×100s_m = \frac{W_m}{W_m + L_m} \times 100

where sms_m is the web shop's share in month mm, in %, WmW_m is the web shop's revenue in that month and LmL_m is the Leeds shop's revenue in that month. Wholesale is left out on purpose, because the decision is between the two retail channels.

For March:

sMarch=6868+60×100=53.1s_{\text{March}} = \frac{68}{68 + 60} \times 100 = 53.1

so in March sms_m was 53.1%, with Wm=68W_m = 68 and Lm=60L_m = 60 in £ thousands. The same sum for the other months gives 44.1% in January, 46.8% in February, 55.1% in April, 57.1% in May and 58.1% in June. March is the first month above half.

Example, step 4: remove what does not carry data

Out go the value labels, the 3D, the shaded background, the frame, most gridlines, the legend and the wholesale series, as the audit decided. Six bars remain.

Example, step 5: grey everything, highlight the one bar

Five bars grey, March pink. The colour points at the month the title names.

Example, step 6: a title that states the finding, a subtitle for the facts

Headline: the web shop passed half of retail revenue in March. Subtitle: web shop share of Leeds shop plus web shop revenue, January to June 2026, invented figures. Here is the result.

Bar chart, one bar per month from January to June 2026, showing the web shop's share of Leeds shop plus web shop revenue, invented figures: January 44.1%, February 46.8%, March 53.1%, April 55.1%, May 57.1%, June 58.1%. March is highlighted as the first month above 50%; the other months are grey. The share has risen every month. Tap or click to open full screen and move around

Example, step 7: the five second test

Sam shows the chart to a colleague from the warehouse for five seconds and asks what it says. "Online took over from the shop in March, and it keeps growing." That is the message. Had the answer been "June was big" or "I am not sure", Sam would go back to step 4.

Example, step 8: send it, with the table behind it

The email to Helen carries the chart, one so what sentence ("the web shop deserves the larger share of the autumn budget, and the Leeds shop is growing too, just more slowly") and Sam's original table as an appendix. The table did not disappear. It moved to where it does its job best. The cluttered chart held more data, the clean one holds more meaning, and only the clean one helps Helen decide.

A short checklist before any chart leaves your desk

  • Decision: can you name the decision this chart supports?
  • Format: does the reader need exact values (table), a pattern (chart), or both?
  • Data first: did you draw the raw data before trusting a summary number?
  • Encoding: is the key comparison shown as position or length on a shared scale?
  • Title: does the headline state one finding, and does it survive a check against the data?
  • Clutter and focus: does every mark carry data or help read it, with one highlight pointing at the title's finding?
  • Words, not only colour: would the message survive in black and white?
  • Five seconds: did a colleague get the point in five seconds?

Conclusion

A table is for looking up, a chart is for seeing. When the reader needs an exact figure, send the table. When the message lives in the shape of the data, in a trend, a crossing or an outlier, draw it, because summary numbers can agree perfectly while the data tells four different stories, as Anscombe showed half a century ago.

Your eye reads some things before you think: colour, length, position. Use that. Put the key comparison on a shared scale, give the one thing that matters the one strong colour, and grey out the rest. Say one thing per chart, write that thing in the title, check that the title is true, and add one line on what it means for the decision.

Sam's table had every number Helen needed and told her nothing. One chart with a finding title and one pink bar told her the most important thing that happened in her business this half year, in five seconds. The next chapter, Choosing the right chart for the question, takes the next step: which chart for which kind of question.

If your monthly report is a table nobody reads, book the free workshop: we pick the decision it should support and turn it into charts that say what happened, and the first two weeks of work are on me.

Sources

  1. 1.

    Stephen Few, Effectively Communicating Numbers: Selecting the Best Means and Manner of Display, Perceptual Edge whitepaper, November 2005, http://www.perceptualedge.com/articles/Whitepapers/Communicating_Numbers.pdf

  2. 2.

    Iris Vessey, Cognitive Fit: A Theory Based Analysis of the Graphs Versus Tables Literature, Decision Sciences 22 (2), 219 to 240, Wiley, 1991, https://doi.org/10.1111/j.1540-5915.1991.tb00344.x

  3. 3.

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

  4. 4.

    F. J. Anscombe, Graphs in Statistical Analysis, The American Statistician 27 (1), 17 to 21, Taylor and Francis, 1973, https://doi.org/10.1080/00031305.1973.10478966

  5. 5.

    R Core Team, Anscombe's Quartet of 'Identical' Simple Linear Regressions, documentation of the anscombe dataset in the R datasets package, https://stat.ethz.ch/R-manual/R-devel/library/datasets/html/anscombe.html

  6. 6.

    Justin Matejka and George Fitzmaurice, Same Stats, Different Graphs: Generating Datasets with Varied Appearance and Identical Statistics through Simulated Annealing, Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems, 1290 to 1294, ACM, 2017, https://doi.org/10.1145/3025453.3025912 (project page: https://www.research.autodesk.com/publications/same-stats-different-graphs/)

  7. 7.

    Christopher G. Healey and James T. Enns, Attention and Visual Memory in Visualization and Computer Graphics, IEEE Transactions on Visualization and Computer Graphics 18 (7), 1170 to 1188, IEEE, 2012, https://doi.org/10.1109/TVCG.2011.127

  8. 8.

    Christopher G. Healey, Perception in Visualization, North Carolina State University, https://www.csc2.ncsu.edu/faculty/healey/PP/

  9. 9.

    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, Taylor and Francis, 1984, https://doi.org/10.1080/01621459.1984.10478080

  10. 10.

    Jeffrey Heer and Michael Bostock, Crowdsourcing Graphical Perception: Using Mechanical Turk to Assess Visualization Design, Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 203 to 212, ACM, 2010, https://doi.org/10.1145/1753326.1753357 (paper page: https://idl.uw.edu/papers/crowdsourcing-graphical-perception)

  11. 11.

    Claus O. Wilke, Fundamentals of Data Visualization, chapter 29, Telling a story and making a point, O'Reilly Media, 2019, free online edition, https://clauswilke.com/dataviz/telling-a-story.html

  12. 12.

    Cole Nussbaumer Knaflic, Storytelling with Data: A Data Visualization Guide for Business Professionals, Wiley, 2015, https://www.wiley.com/en-us/Storytelling+with+Data%3A+A+Data+Visualization+Guide+for+Business+Professionals-p-9781119002253

  13. 13.

    Michelle A. Borkin, Zoya Bylinskii, Nam Wook Kim, Constance May Bainbridge, Chelsea S. Yeh, Daniel Borkin, Hanspeter Pfister and Aude Oliva, Beyond Memorability: Visualization Recognition and Recall, IEEE Transactions on Visualization and Computer Graphics 22 (1), 519 to 528, IEEE, 2016, https://doi.org/10.1109/TVCG.2015.2467732 (abstract: https://vcg.seas.harvard.edu/publications/20150101-beyond-memorability-visualization-recognition-and-recall)

  14. 14.

    Ha Kyung Kong, Zhicheng Liu and Karrie Karahalios, Frames and Slants in Titles of Visualizations on Controversial Topics, Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, ACM, 2018, https://doi.org/10.1145/3173574.3174012 (author copy: https://mail.zcliu.cs.umd.edu/vistitles/CHI18-VisTitles.pdf)

  15. 15.

    Edward R. Tufte, The Visual Display of Quantitative Information, second edition, Graphics Press, 2001 (first edition 1983), https://www.edwardtufte.com/book/the-visual-display-of-quantitative-information