How charts lie, and how to catch them

TL;DR

Difficulty: Intermediate

A chart misleads when the picture and the numbers tell different stories, even if every number is correct.

  • Bars must start at zero. Lines may not, if the axis is labelled and the change matters.
  • Tufte's lie factor measures the distortion: the effect shown divided by the effect in the data. A 4.6% rise drawn on an axis from €400,000 has a lie factor of 34.
  • Dual axes can make any two lines meet. Index both to 100 instead.
  • Short time windows, running totals, scaled pictures and uneven time steps all bend the picture.
  • Run the thirty second check before you believe a chart.

A slide can carry only correct numbers and still send you out of the room believing something false. This article shows how, trick by trick, with each misleading chart drawn next to its honest twin. You get Tufte's lie factor to measure the damage, a thirty second check for any chart, and a rival's "sales doubled" slide taken apart.

Correct numbers, wrong impression

Someone puts a slide on the screen. The numbers in the corner are right, the source is real, the arithmetic checks out. And yet you walk out of the room believing something that is not true. That is the strange thing about charts: they can lie without a single false number in them.

This article answers one question: how can a chart with correct numbers still mislead you, and how do you catch it in the time it takes to turn a page? It builds on choosing the right chart for the question, which covers which chart fits which question. Here we look at what happens when the right chart is drawn the wrong way.

A chart is not a table with colours. It is a picture, and your eye reads pictures before your brain reads numbers. Lengths, heights, slopes and areas land first. The axis labels, the small print and the footnote land later, if at all. So a chart is honest when the picture and the numbers tell the same story, and it misleads when they part company. Most misleading charts are not forged. They are drawn so the picture says more, or less, than the data.

None of this is new. Darrell Huff gave a whole chapter of How to Lie with Statistics to what he called the "gee whiz" graph back in 19541, and Alberto Cairo's How Charts Lie updated the catalogue for a world of social media dashboards in 20192. What has changed is the volume: every tool now draws a chart in one click, with defaults that are not always on your side.

Throughout this article I use one invented business. Shannon Outfitters is an outdoor gear brand in Limerick that sells online at shannonoutfitters.example and through one shop in the city. Sam is its analyst. All of its numbers are made up for teaching, and they are labelled as such.

How we know charts fool people

Before the tricks, the evidence that they work. In 2015 a team of researchers in New York ran a crowdsourced experiment on four common distortions: a truncated axis, area used to show quantity, a squashed or stretched aspect ratio, and an inverted axis. People who saw the deceptive versions misread the message, often by a wide margin, compared with people who saw honest versions of the same data3.

Why does it work so reliably? Because the eye judges some visual quantities far better than others. In a classic set of experiments, William Cleveland and Robert McGill found that people judge position along a common scale most accurately, then length, and that judgements of angle and area are clearly worse4. Every trick in this article exploits that ranking. It either changes what a length means, or it swaps a quantity the eye reads well for one it reads badly.

Example: the same fact read two ways

Sam shows the owner two bars: shop sales and online sales for one month. In version A both bars stand on the zero line, so their lengths compare honestly. In version B the bars are drawn as pictures of rucksacks whose height and width both follow the number. The owner looks at version B and says "online is four times the shop". The data says twice. Nothing on the slide was false. The picture simply asked the eye to compare areas, which it does badly, and the eye overshot. I will put a number on exactly that overshoot later in the article.

Truncated axes: a bar has to start at zero

A bar chart encodes a number as the length of a bar. That only works if every bar starts at the same place, and that place is zero. Cut the axis at some higher value and you cut off the bottom of every bar by the same amount. The remaining stubs no longer have lengths in proportion to the numbers, but the eye still compares them as if they did.

Line charts are different. A line encodes values as positions and change as slope, not as the length of anything. So a line axis that does not start at zero is often fine, and sometimes necessary, because a conversion rate moving between 1.8% and 2.3% would be a flat line on an axis from zero to 100%. Michael Correll, Enrico Bertini and Steven Franconeri tested exactly this in 2020. Truncating the axis made people see bigger effects in every chart type they tried, and the effect persisted even when the chart carried a visible marker of the truncation. Their advice is not "always start at zero" but "decide which size of change matters, and choose the axis so that size looks like what it is"5.

So the rule is short. Bars and areas always start at zero. Lines may start elsewhere, as long as the axis is clearly labelled and the change you show is a change that matters.

Example: a 4.6% rise drawn as a near tripling

Shannon Outfitters' third quarter revenue was €412,000 last year and €431,000 this year (invented figures). Sam's reporting tool drew the first chart below. The axis starts at €400,000 because the tool picked a range that "fits the data".

Bar chart of Shannon Outfitters third quarter revenue, invented figures. Last year €412,000, this year €431,000. The value axis starts at €400,000, so this year's bar looks almost three times as tall as last year's, although revenue rose by 4.6%. Tap or click to open full screen and move around
The same two bars, invented figures, last year €412,000 and this year €431,000, now on an axis that starts at zero. The bars are almost the same height, which matches a rise of 4.6%. Tap or click to open full screen and move around

Step by step, here is what happened. Above the €400,000 line, last year's bar is €12,000 tall and this year's bar is €31,000 tall. The eye compares 12 with 31 and sees a bar more than two and a half times as tall. The data compares 412 with 431. Same numbers, two very different pictures. The honest chart looks boring, and that is the point: a 4.6% rise is a modest, real improvement, and the chart should look like one. If the 4.6% is the story, Sam writes it in the title ("Q3 revenue up 4.6% on last year") instead of stretching the bars to shout it.

Tufte's lie factor: putting a number on the distortion

Edward Tufte gave this problem a measurement. He called it the lie factor: how big an effect looks in the graphic, divided by how big it actually is in the data6. An honest chart has a lie factor of 1. Above 1, the chart exaggerates. Below 1, it plays the change down.

Lie factor=size of effect shown in graphicsize of effect in data\text{Lie factor} = \frac{\text{size of effect shown in graphic}}{\text{size of effect in data}}

Here size of effect shown in graphic\text{size of effect shown in graphic} is the percentage change in the drawn quantity (a bar's visible length, a symbol's area), and size of effect in data\text{size of effect in data} is the percentage change in the numbers themselves. Both are ordinary percentage changes, and the lie factor itself is a ratio, explained in Foundations: ratios and rates.

Each percentage change uses the same formula:

Δ%=xnew−xoldxold×100\Delta\% = \frac{x_{\text{new}} - x_{\text{old}}}{x_{\text{old}}} \times 100

Here Δ%\Delta\% is the percentage change, xoldx_{\text{old}} is the earlier value and xnewx_{\text{new}} is the later value. For the data you plug in the numbers, for the graphic you plug in what is drawn. The formula is worked through from scratch in Foundations: percentages.

Tufte's own textbook case was a newspaper chart of American fuel economy standards with a lie factor of 14.86. Charts in board packs rarely need to be that bad to do damage.

Example: the lie factor of Sam's truncated chart

First, the size of the effect in the data. Revenue went from €412,000 to €431,000:

Δ%data=431,000−412,000412,000×100=4.6\Delta\%_{\text{data}} = \frac{431{,}000 - 412{,}000}{412{,}000} \times 100 = 4.6

Here Δ%data\Delta\%_{\text{data}} is the real percentage change in revenue, 4.6%4.6\%, worked out from xold=412,000x_{\text{old}} = 412{,}000 and xnew=431,000x_{\text{new}} = 431{,}000.

Second, the size of the effect in the graphic. Above the €400,000 cut, the visible bars are €12,000 and €31,000 tall:

Δ%graphic=31,000−12,00012,000×100=158.3\Delta\%_{\text{graphic}} = \frac{31{,}000 - 12{,}000}{12{,}000} \times 100 = 158.3

Here Δ%graphic\Delta\%_{\text{graphic}} is the change the eye sees, 158.3%158.3\%, from a visible length of 12,00012{,}000 growing to 31,00031{,}000.

Third, divide one by the other:

Lie factor=158.34.6=34.3\text{Lie factor} = \frac{158.3}{4.6} = 34.3

Here 158.3158.3 is the effect shown in the graphic and 4.64.6 is the effect in the data, so the truncated chart exaggerates the change about 34 times.

There is a shortcut worth remembering. When both bars are cut at the same baseline bb, the two changes share the same numerator, and the lie factor reduces to:

Lie factor=xoldxold−b\text{Lie factor} = \frac{x_{\text{old}}}{x_{\text{old}} - b}

Here xoldx_{\text{old}} is the earlier value, 412,000412{,}000, and bb is the value where the axis starts, 400,000400{,}000. So 412,000/12,000=34.3412{,}000 / 12{,}000 = 34.3, the same answer. The closer the cut sits to the data, the bigger the lie. The ratio is explained in Foundations: ratios and rates.

Dual axes: two scales, one invented relationship

A dual axis chart puts two series with different units on one picture, one read against the left axis and one against the right. It looks efficient. The problem is that each axis can be stretched or shifted on its own, so whoever draws the chart decides whether the two lines run together, cross, or drift apart. Lisa Charlotte Muth of Datawrapper puts it plainly: the scales of a dual axis chart are arbitrary, and readers still read meaning into where the lines meet7.

The honest alternatives are simple. Index both series to the same starting point (both equal 100 in the first month), and they share one axis. Or draw two charts, one above the other, with the same time axis. Or, if the question is "does one drive the other", use a scatter chart with spend on one axis and revenue on the other. Indexing is covered in depth in same data, different truths.

Example: ad spend and online revenue moving "in step"

Sam's agency sends a monthly report with the first chart below. Shannon Outfitters' ad spend is on the left axis, online revenue on the right (both invented, in thousands of pounds). The lines sit almost on top of each other all year. The message, never said out loud, is "spend more, earn more, one for one".

Line chart with two value axes, invented figures for Shannon Outfitters. Monthly ad spend rises from €4,000 in January to €6,200 in December on the left axis. Monthly online revenue rises from €61,000 to €65,900 on the right axis. The axes were chosen so the two lines overlap all year, suggesting they grow together. Tap or click to open full screen and move around
The same invented figures with both series indexed so that January equals 100, on one shared axis. By December ad spend stands at 155, up 55%, while online revenue stands at 108, up 8%. The lines diverge steadily. Tap or click to open full screen and move around

Here is how Sam rebuilds it. She divides every month's spend by January's spend (€4,000) and multiplies by 100, and does the same for revenue with January's €61,000. December spend becomes 6.2/4.0×100=1556.2 / 4.0 \times 100 = 155, December revenue becomes 65.9/61.0×100=10865.9 / 61.0 \times 100 = 108. On one shared axis the story changes completely: spend grew 55% over the year, revenue grew 8%. The extra €2,200 a month in spend by December came with an extra €4,900 a month in revenue, which may or may not cover the margin on those sales. That is a real question for the owner, and the dual axis chart had hidden it.

Note that this line chart does not start at zero, and that is fine: it is a line, it is labelled, and the gap between the lines is the point.

Cherry picked windows: the start date is a choice

Every time series chart starts somewhere and stops somewhere. Those two dates are decisions, and they can turn a flat business into a rocket or a healthy one into a crisis. Seasonal businesses are the easiest to bend: start the chart in the quietest month and end it in the busiest, and any year looks like a breakthrough.

The defence is to ask for at least one full seasonal cycle, ideally two or three, and to compare like with like: this June against last June, not this June against this January. The existing article on forecasting demand with seasonality shows how to model the pattern itself.

Example: "up 42% in six months"

Shannon Outfitters sells tents, boots and rain jackets, so spring is busy and January is quiet, as anyone who has walked the Burren in January will understand. A new marketing lead shows the first chart below to the owner in July (invented figures).

Line chart of monthly revenue for Shannon Outfitters, invented figures, January to June of this year only. Revenue climbs from €49,400 in January to €70,000 in June, which looks like 42% growth in six months. Tap or click to open full screen and move around
Line chart of monthly revenue for Shannon Outfitters, invented figures, with one line per year: two years ago, last year and this year to June. All three lines follow the same seasonal shape, low in January and high in May. This year's line sits only about 1% above last year's. Tap or click to open full screen and move around

Sam works through it. January to June this year: 70.0/49.4=1.41770.0 / 49.4 = 1.417, so the "up 42%" is arithmetically right. But two years ago revenue rose from €48,000 to €68,000 over the same months, also about 42%, and last year it did the same. The rise is the season, not the marketing. The fair comparison is June against last June: 70.0/69.4=1.00970.0 / 69.4 = 1.009, a rise of under 1%. The first half as a whole is up about 0.9% on last year's first half. Grand and steady, not spectacular, and the honest chart, with one line per year on a January to December axis, shows it at a glance.

Cumulative charts: the line that can never go down

A cumulative chart shows a running total: orders since launch, customers ever acquired, downloads to date. Because every month adds something and nothing is subtracted, the line can only rise. It rises even when the business is shrinking, because a shrinking month still adds a positive number. The slope of a cumulative line does carry the truth (a flatter slope means fewer new orders that month), but the eye reads "up and to the right" long before it reads a subtle bend.

Running totals have honest uses: progress towards an annual target, for example. For "how are we doing right now", show the value per period.

Example: 10,730 orders and a quiet decline

Shannon Outfitters' web shop took 900 orders in January, rising to 1,000 in April, then falling every month to 740 in December (invented figures). A slide for the bank shows the first chart.

Line chart of the running total of online orders for Shannon Outfitters, invented figures. It rises every month from 900 in January to 10,730 in December, which looks like steady growth. Tap or click to open full screen and move around
Bar chart of monthly online orders for Shannon Outfitters, invented figures. Orders rise from 900 in January to a peak of 1,000 in April, then fall every month to 740 in December, 26% below the April peak. The months from May to December are highlighted. Tap or click to open full screen and move around

Sam turns the running total back into monthly numbers by subtracting each month's total from the next: 1,850−900=9501{,}850 - 900 = 950 for February, 2,830−1,850=9802{,}830 - 1{,}850 = 980 for March, and so on. The monthly bars show what the cumulative line buried: eight months of decline, with December at 740/1,000=0.74740 / 1{,}000 = 0.74 of the April peak, a fall of 26%. Notice the per month bars start at zero, as bars must, and the decline is still plain to see. An honest chart does not need a trick to show a real change.

Area, 3D and pictures: when doubling looks like quadrupling

When a picture of a rucksack, a coin or a shopping bag stands for a number, the designer has to decide how to scale it. Scale the height in proportion to the number and keep the shape, and the width grows too. Double the number and the picture's area becomes four times as large, because area grows with the square of the height. Huff called this the one dimensional picture: a quantity that changes in one dimension drawn as an object that changes in two or three1. Pandey and colleagues included exactly this "area as quantity" distortion in their study3.

Three dimensional bars and pies add their own problems. Perspective makes whatever is in front look bigger than what is behind, and the tilted base makes it hard to tell where a bar actually ends. A flat bar answers every question a 3D bar does, and answers it more accurately.

Example: the rucksack chart, measured

Back to the example from the start. Shop sales were 2,000 orders and online sales 4,000 (invented). The online rucksack is twice as tall and therefore twice as wide, so its area is 2×2=42 \times 2 = 4 times the shop's.

Using the lie factor from earlier: the effect in the data is (4,000−2,000)/2,000×100=100%(4{,}000 - 2{,}000) / 2{,}000 \times 100 = 100\%. The effect in the graphic, measured as area, is (4−1)/1×100=300%(4 - 1) / 1 \times 100 = 300\%. The lie factor is 300/100=3300 / 100 = 3. The chart below shows what the eye takes away next to what is true.

Grouped bar chart, invented figures. Real orders: shop 2,000, online 4,000. Orders implied by the area of a rucksack picture scaled in height and width: shop 2,000, online 8,000. The picture makes online look four times the shop when it is twice. Tap or click to open full screen and move around

The fix is to drop the picture and draw bars from zero. If the brand really wants icons, repeat a small icon (one rucksack per 500 orders, so four for the shop and eight online) instead of scaling one big one. Equal sized units let the eye count, which it does well, instead of comparing areas, which it does badly.

Uneven intervals: when the time axis skips years

Most chart tools treat the labels on a time axis as categories and space them evenly, whatever they say. If the data skips years, or mixes years and months, the slope between two points no longer means "change per period". A two year gap drawn the same width as a one year gap doubles the apparent speed of that stretch.

Example: growth that looked like it was slowing

Sam finds an old slide with Shannon Outfitters' annual revenue for 2018, 2020, 2022, 2024, 2025 and 2026. The six points are spaced evenly. The line climbs steeply from 2022 to 2024 and then flattens, and the slide says "growth has slowed since 2024".

Here are the invented numbers, with the years the slide left out filled in:

YearRevenueChange on the year beforeOn the old slide
2018€540kyes
2019€548k€8kno
2020€560k€12kyes
2021€578k€18kno
2022€600k€22kyes
2023€645k€45kno
2024€690k€45kyes
2025€735k€45kyes
2026€780k€45kyes

On the slide, the step from 2022 to 2024 covers two years (€90k) but takes one slot, and the steps after 2024 cover one year each (€45k a slot). So the line looks as if it halves its speed. In truth revenue has grown by €45k every single year since 2022. Nothing slowed. The fix is to plot every year, or use a real date axis that spaces points by time, and to say in the caption when a year is missing.

Spurious correlation: two lines that move together for no reason

Two series that rise and fall together are said to be correlated. The strength of that link is usually measured by the correlation coefficient rr, which runs from −1-1 to 11 and is explained in Foundations: correlation. A chart that puts two correlated lines side by side invites the reader to assume one causes the other. Often both are driven by a third thing, or the match is pure chance.

The most famous example was published in the New England Journal of Medicine. Franz Messerli plotted chocolate consumption per person against Nobel laureates per 10 million people for 23 countries and found a strong correlation, r=0.791r = 0.7918. Messerli himself called the finding hypothesis generating only. A plausible third factor is wealth: richer countries can afford both more chocolate and more research, and nobody became a laureate by eating a bar.

Example: social posts and tent sales

Sam's social media colleague puts two lines on one chart: posts per month and tents sold per month (invented figures). They rise in spring and fall in autumn together, and a spreadsheet's CORREL function gives r=0.76r = 0.76. The caption says "posting drives tent sales".

MonthJanFebMarAprMayJunJulAugSepOctNovDec
Posts66911121098771012
Tents sold4045901301601501209560507085

Sam asks one question: why did posting go up in spring? Because there was more to post about: new tents in stock, the camping season starting, customers sharing trips. The season drives both. To test whether posts sell tents, she would need posts that vary independently of the season, for example an A/B test on the same week, which the statistical significance article explains. Until then the honest caption is "posts and tent sales both follow the season".

The thirty second check for any chart

You will not always have the data behind a slide. You can still ask five questions, in this order, before you believe the picture. The flow below is the check I would run on any chart that asks me to make a decision.

Tap or click to open full screen and move around

Two more questions sit outside the flow because they apply to every chart. Are pictures or 3D shapes standing in for numbers? And are the time steps even? If the chart passes all of that, the remaining question is whether the data is right in the first place, which is a different job: is your dashboard lying to you covers the data quality audit.

Worked example: a competitor's "sales doubled" slide, taken apart

Shannon Outfitters' owner comes back from a trade event with photos of a slide from a rival brand, Atlantic Ridge Gear (invented, atlanticridge.example). The headline reads "Sales doubled. Join the brands that are growing." It has four charts. Sam takes them apart one by one with the thirty second check.

Chart one: the bar chart

Two bars, last year and this year, the second about twice as tall as the first. The axis labels, in small grey type, run from €800k to €1m. Reading the labels, last year was €900k and this year €990k.

  • Effect in the data: (990−900)/900×100=10%(990 - 900) / 900 \times 100 = 10\%.
  • Visible bars above the €800k cut: €100k and €190k, so the effect in the graphic is (190−100)/100×100=90%(190 - 100) / 100 \times 100 = 90\%.
  • Lie factor: 90/10=990 / 10 = 9, or with the shortcut, 900/(900−800)=9900 / (900 - 800) = 9.

Sales grew 10%. "Doubled" describes the bars, not the business.

Chart two: the line with a start date

A line of monthly revenue climbing steeply from January to May, labelled "momentum". For an outdoor brand, January to May is the seasonal climb. Without last year's line for the same months, the chart says nothing about growth. Sam notes it as "window covers less than one seasonal cycle, ask for the year before".

Chart three: "customers served"

A smooth curve sweeping up to 40,000. The small print says "customers served since 2019". It is a running total. It would rise if Atlantic Ridge Gear gained one customer a month. Sam notes "cumulative, ask for new customers per year".

Chart four: spend and sales

Marketing spend and sales on a dual axis chart, the two lines rising together. The left axis runs from €50k to €120k, the right from €800k to €1m, both chosen by the slide maker. Indexing both to last year: if spend went from €60k to €110k, that is an index of 110/60×100=183110 / 60 \times 100 = 183, against sales at 990/900×100=110990 / 900 \times 100 = 110. Spend rose 83% to buy 10% more sales.

What Sam tells the owner

Sam writes one paragraph: "Atlantic Ridge Gear grew sales about 10%, roughly in line with a good year in our market. They nearly doubled their marketing spend to do it. None of the four charts shows evidence of doubled sales. Nothing on this slide is a reason to change our plan." The owner had walked in worried. He walks out with a decision and a good deal less worry, which is what charts are for.

The lesson for Shannon Outfitters' own reports is the mirror image. Every one of Atlantic Ridge Gear's tricks is something a reporting tool can do by default. Sam now checks each of her own charts against the same flow before it goes into the board pack: bars from zero, one axis, full cycles, values per period, and a title that states the finding in numbers.

Rules that make the honest chart the default

Every chart on this site, including the ones in this article, follows a fixed set of rules: brand colours only, light and dark mode, axis labels at 45 degrees, a full screen view, top to bottom diagrams for phones, and a text description for screen readers. Fixed rules like these are worth setting for any business, whether you report to a board in Cork or a bank manager in Ennis: the time to decide how charts are drawn is before the deadline, not during it. Add "bars start at zero" and "one value axis per chart" to that list and most of this article's tricks become impossible by default.

When I start an analytics project for a client, I start from the decision the numbers should support, not from a dashboard. That order matters here too. A chart built to support a decision has to survive the question "would I decide differently if I saw the honest version?" If the answer is yes, the chart was persuasion, not information.

Conclusion

A chart lies when the picture and the numbers tell different stories. The usual ways are few and well documented: bars that do not start at zero, two axes that can be tuned to make any lines meet, time windows that start in the quietest month, running totals that cannot fall, pictures whose area grows faster than the number, time axes that skip years, and two lines that move together because a third thing drives both.

Tufte's lie factor turns the first and the fifth into a number you can quote. The thirty second check turns the rest into a habit. And the fix is nearly always the same: draw the honest version next to the persuasive one, and see which story is still standing.

If your board pack or your agency reports leave you unsure what the charts really say, book the free workshop: we go through your key charts together and redraw the ones that bend the truth, and the first two weeks of work are on me.

Sources

  1. 1.

    Darrell Huff, How to Lie with Statistics, illustrated by Irving Geis, W. W. Norton and Company, 1954 (reissued 1993), https://wwnorton.com/books/How-to-Lie-with-Statistics

  2. 2.

    Alberto Cairo, How Charts Lie: Getting Smarter about Visual Information, W. W. Norton and Company, 2019, https://wwnorton.com/books/9781324001560

  3. 3.

    Anshul Vikram Pandey, Katharina Rall, Margaret L. Satterthwaite, Oded Nov and Enrico Bertini, How Deceptive are Deceptive Visualizations? An Empirical Analysis of Common Distortion Techniques, Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems (CHI 2015), pages 1469 to 1478, https://doi.org/10.1145/2702123.2702608

  4. 4.

    William S. Cleveland and Robert McGill, Graphical Perception: Theory, Experimentation, and Application to the Development of Graphical Methods, Journal of the American Statistical Association, volume 79, number 387, 1984, pages 531 to 554, https://doi.org/10.1080/01621459.1984.10478080

  5. 5.

    Michael Correll, Enrico Bertini and Steven Franconeri, Truncating the Y Axis: Threat or Menace?, Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, 2020, https://arxiv.org/abs/1907.02035

  6. 6.

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

  7. 7.

    Lisa Charlotte Muth, Why not to use two axes, and what to use instead, Datawrapper blog, 8 May 2018, https://www.datawrapper.de/blog/dualaxis

  8. 8.

    Franz H. Messerli, Chocolate Consumption, Cognitive Function, and Nobel Laureates, New England Journal of Medicine, volume 367, number 16, 2012, pages 1562 to 1564, https://doi.org/10.1056/NEJMon1211064