How Data Visualization Can Mislead

A bar chart with its y-axis starting at 90 instead of 0 can turn a 2 percent difference into what looks like a tenfold gap. Every number on that chart can be accurate. The picture still is not. Chart designers, whether careless or deliberate, have a well-documented toolkit for bending how people read numbers that are technically correct, and researchers have spent decades measuring exactly how well those tricks work.

This is not a matter of opinion or design taste. Statistician Edward Tufte gave the field a name for it in 1983: graphical integrity, the principle that the visual representation of data should be proportional to the numerical quantities it represents (Tufte, The Visual Display of Quantitative Information, 1983). When that proportionality breaks, the chart lies even though the data behind it doesn't. This piece walks through the specific mechanisms, backed by cognitive science and real documented examples, and gives you a working checklist for catching them before they catch you.

Tufte's Lie Factor: measuring exactly how much a chart distorts

Tufte proposed a simple ratio for quantifying visual dishonesty: the Lie Factor, defined as the size of the effect shown in the graphic divided by the size of the effect in the underlying data. A Lie Factor of 1.0 means the picture matches the numbers exactly. Above roughly 1.05, or below roughly 0.95, the graphic is distorting the story, and the further from 1.0 it drifts, the worse the distortion.

Tufte's original example, still the field's canonical case study, involved a New York Times graphic on federal fuel economy standards. The actual policy change required average fuel economy to rise from 18 to 27.5 miles per gallon, a 53 percent increase. The graphic depicted that change using a bar whose visual length implied a 783 percent increase, driven by drawing the bars as roadway lanes narrowing in false perspective rather than as flat, proportional bars. Dividing 783 by 53 gives a Lie Factor of roughly 14.8, meaning the chart overstated the real-world change by nearly fifteenfold (Tufte, 1983).

What makes the Lie Factor useful is that it turns "this chart feels off" into a number you can actually calculate. If you know the real percentage change in the underlying data and can measure the visual change in the chart, whether that's bar height, bar area, or line slope, you can compute your own Lie Factor for any chart you're suspicious of. It's one of the few genuinely quantitative tools for catching visual manipulation, and it predates modern data journalism by decades.

Edward Tufte's Lie Factor for the *New York Times* fuel economy chart in his 1983 book was approximately 14.8: the graphic visually implied a 783 percent change against an actual 53 percent change in the underlying fuel economy standards.

What is the Lie Factor in data visualization?

The Lie Factor is a ratio Edward Tufte introduced in 1983, calculated as the visually depicted size of a change divided by the actual size of that change in the data. A value near 1.0 signals an honest chart. Values well above or below 1.0 indicate the visual is exaggerating or understating what the numbers actually show.

The axis is doing more work than the data

Truncating the y-axis, starting the vertical scale somewhere above zero, is the most common distortion in published charts, and it doesn't require anyone to alter a single data point. When an axis runs from 90 to 100 instead of 0 to 100 for values clustered between 92 and 98, small real differences get stretched to fill the entire visible height of the chart, and the eye reads that stretched height as the true magnitude of change.

A well-documented case from August 2012 illustrates the mechanism cleanly. Fox News aired a chart comparing the number of Americans receiving some form of government assistance against the number of full-time workers, with a y-axis that began at 94,000,000 rather than zero. The technique was flagged by multiple outlets at the time, including the statistics blog Simply Statistics and media watchdog Media Matters, both of which noted that starting the axis near the lower data value rather than at zero made a modest numerical gap between the two bars look enormous (Simply Statistics, November 2012; Media Matters, August 2012). Nothing about the underlying counts was fabricated. The distortion lived entirely in where the axis started.

This is not a partisan phenomenon confined to one outlet or one political direction; the same truncation technique appears across media covering nearly every topic, from stock indices to public health metrics, whenever a designer wants a modest change to read as dramatic. The fix is not complicated. A team of researchers led by Michael Correll at Tableau Research, working with Enrico Bertini and Steven Franconeri, ran controlled experiments specifically testing how truncated bar charts affect viewer judgment, publishing the results at the 2020 ACM CHI Conference on Human Factors in Computing Systems under the title "Truncating the Y-Axis: Threat or Menace?" Their findings confirmed that y-axis truncation reliably inflates the perceived magnitude of differences between bars, and that simply labeling the axis as truncated does not fully correct the effect (Correll, Bertini, and Franconeri, CHI 2020).

Why do truncated axes distort perception so effectively?

People read bar height as a direct, proportional signal: a bar that looks twice as tall is unconsciously assumed to represent twice the value. When the axis does not start at zero, that proportional relationship breaks down, but the visual shortcut people use to interpret bar charts doesn't know that, so the reader keeps applying the zero-baseline assumption to a chart that no longer satisfies it.

Is it ever legitimate to start a y-axis above zero?

Sometimes. Line charts tracking a value that fluctuates in a narrow band far from zero, such as body temperature, blood pressure, or a currency exchange rate, can lose all visible detail if forced to a zero baseline. The distinction between legitimate and misleading use comes down to disclosure: a non-zero baseline that is clearly labeled and chosen for a value that genuinely never nears zero is defensible; the same choice made silently, or made specifically because it exaggerates a comparison, is not.

Area, volume, and why humans are bad at comparing shapes

Bar charts and dot plots ask viewers to compare position and length, tasks the human visual system handles with reasonable accuracy. Pie charts, bubble charts, and treemaps ask viewers to compare angle, area, or volume instead, and decades of perception research show people are measurably worse at those comparisons. The foundational study here is William Cleveland and Robert McGill's 1984 paper in the Journal of the American Statistical Association, "Graphical Perception: Theory, Experimentation, and Application to the Development of Graphical Methods," which ranked the accuracy of different visual encodings by running controlled experiments where participants judged quantities shown through position, length, angle, area, and color (Cleveland and McGill, JASA, 1984). Position along a common scale ranked most accurate. Angle, area, and volume ranked well below it, with color saturation and volume performing worst of the encodings they tested.

That ranking has a direct, mechanical consequence for bubble charts. If a designer scales a circle's radius or diameter directly to a data value rather than scaling its area, the visual impression gets distorted geometrically: doubling a circle's diameter actually quadruples its area, since area scales with the square of the radius. A viewer comparing two circles by their apparent size, which is dominated by area, will perceive a fourfold difference where the underlying data only shows a twofold one. This mistake appears often even in professionally produced graphics, not because the designers don't know better, but because scaling by area correctly makes the visual difference between values look smaller and less dramatic than scaling by diameter does, which creates a quiet incentive to get it wrong.

Three-dimensional charts compound the same error further. A 3D bar or cone that scales all three dimensions in proportion to a single value doesn't just double or quadruple its visual footprint when the underlying number doubles, it can grow by a factor of eight, since volume scales with the cube of a linear dimension. This is one of the reasons data visualization educators, following Tufte's and Cleveland and McGill's work, generally recommend against 3D bar and pie charts for anything beyond decorative use: the extra dimension doesn't encode any additional information, but it does distort the encoding of the information that's already there.

Design choice	What it does to perceived magnitude	Honest alternative
Truncated y-axis	Small real differences appear stretched to fill the visible chart height	Zero-baseline bar chart, or a clearly labeled non-zero baseline with a stated reason
Diameter-scaled bubbles	A 2x data increase reads as a 4x visual increase (area scales with radius squared)	Scale circle area, not diameter, to the data value
3D bars, cones, pies	A 2x data increase can read as up to an 8x visual increase (volume scales with the cube)	Flat 2D bar or dot chart with labeled axis
Dual independently-scaled axes	Two unrelated variables can be made to visually "track" by choosing convenient scales	Single shared axis, or two separate charts stacked vertically for direct comparison
Cherry-picked time window	A short, favorable slice of a longer trend implies a pattern the full series doesn't support	Show the full relevant history, or clearly mark the shown window as partial

Five documented chart-design choices, ranked by the perceptual research that explains why each one distorts how a reader judges magnitude.

Dual axes, cherry-picked windows, and the manipulations that hide in framing

Not every distortion involves geometry. Dual-axis charts, where two different variables are plotted on the same chart but each gets its own independently scaled vertical axis, let a designer choose scales that make two unrelated trends appear to move together, or that make two related trends appear to diverge, purely through the choice of where each axis starts and how far it stretches. Because the reader sees only one shared horizontal position for both lines, the visual impression of correlation is driven almost entirely by axis scaling decisions that are invisible unless you specifically check both axis labels.

Cherry-picked time windows work on the same principle applied to the horizontal axis instead of the vertical one. A chart showing stock performance, temperature trends, or crime statistics over a short, carefully chosen window can imply a sustained pattern that the full historical series doesn't support, particularly if the window starts right after a low point or ends right before a reversal. This is a distinct failure mode from a truncated y-axis: the numbers plotted are entirely real, and the axis may even start at zero, but the reader is being shown a slice engineered to tell one story rather than the full series that would tell a more complete one.

Color scales introduce a subtler version of the same problem. A map using a diverging palette, running from one color through white or gray to a second color, is only honest if the midpoint of that palette sits at a value that's actually meaningful, such as zero for a change map or the historical average for a comparison map. When the midpoint is chosen arbitrarily, ordinary variation around that midpoint gets rendered as if it were an extreme split between two opposing categories, because the human eye reads a strong color contrast as a strong underlying difference regardless of whether the numbers justify it.

Choropleth maps, which color entire geographic regions by a data value, carry one more distortion that has nothing to do with color scale choice: land area, not population or magnitude, dominates the visual impression. A large, sparsely populated county colored bright red will visually overwhelm a small, densely populated city colored pale blue, even if the city represents a far larger share of whatever is being measured. This is why election-result choropleth maps are a frequent subject of criticism from data journalists: the map can accurately represent which candidate won the most land area while badly misrepresenting which candidate won the most votes, and readers scanning the map by eye tend to absorb the land-area impression first.

Why do dual-axis charts mislead even when both datasets are real?

A dual-axis chart lets the designer independently choose the starting point and scale of two separate vertical axes, one for each variable being plotted. Because those choices are made after seeing the data, a designer can select scales that make two unrelated trends appear to move together on the page, even though neither dataset was altered and no fabricated correlation exists in the underlying numbers.

What separates honest charts from deliberately misleading ones

The uncomfortable truth about most misleading charts is that no single design choice, taken alone, is obviously wrong. A truncated axis, a cherry-picked date range, a diverging color scale, or a pie chart with six slices can each be defended in isolation. What tips a chart from imperfect into misleading is usually a pattern: multiple choices stacked together, each one individually plausible, that collectively push the visual story further from what the full dataset actually shows.

Intent is genuinely hard to prove from the outside, and that's part of what makes this a persistent problem rather than a solved one. A journalist working under deadline pressure with an out-of-the-box charting tool might truncate an axis purely because the software defaults to it, with no intention of exaggerating anything. A political communications team might make the identical choice deliberately, for the identical visual effect. The chart itself often can't tell you which. That's precisely why frameworks like Tufte's Lie Factor and Cleveland and McGill's perceptual accuracy ranking matter: they let you evaluate the chart's honesty from its structure, independent of anyone's stated intent.

A few practical tells consistently separate more trustworthy charts from less trustworthy ones. Labeled axes with a stated starting point signal that the designer isn't hiding the scale decision. A visible source citation, with enough specificity to actually look up the raw numbers, signals confidence that the underlying data supports the visual. A chart type matched to the actual comparison being made (bars or dots for magnitude comparisons, lines for trends over time, and pie charts reserved for a genuinely small number of parts making up a whole) signals that the designer picked the format for clarity rather than for a specific rhetorical effect.

Cleveland and McGill's 1984 ranking of graphical perception tasks found that position along a common scale is the most accurately judged visual encoding, while angle, area, and volume, the encodings used by pie charts, bubble charts, and 3D charts, rank near the bottom.

How to check a chart for manipulation in under a minute

You don't need statistical training to catch most of the techniques above; you need a short, repeatable checklist. Before accepting a chart's visual message, work through these five questions.

• Where does the y-axis start? If it doesn't start at zero and there's no clear, stated reason tied to the nature of the data, treat the apparent magnitude of any difference with skepticism.
• What time window is shown, and is it the full relevant history? A chart covering the last three months of a five-year trend may be accurate and still misleading by omission.
• Is the chart type matched to the comparison? Pie charts beyond three or four similarly sized slices, and bubble or 3D charts for anything beyond a rough visual impression, are red flags for the underlying comparison being harder to judge accurately than the chart type suggests.
• Are there two independently scaled axes? If so, check whether the scales were chosen to make two lines visually converge or diverge, and ask whether the same data would still look related on a single shared axis.
• Is there a specific, checkable source? A chart citing "government data" with no link or date is harder to verify than one citing a named agency, report, and year.

None of these checks require special software or statistical expertise. They require the habit of pausing on a chart for roughly the same amount of time you'd spend reading the headline above it, which is usually enough to catch the distortions covered in this piece.

What people get wrong about misleading charts

The most common misconception is that a misleading chart requires fabricated or altered numbers somewhere in the underlying dataset. In nearly every documented case discussed here, from the 2012 Fox News welfare chart to Tufte's fuel economy example, the raw data was accurate. The distortion lived entirely in presentation choices: where an axis starts, how a shape is scaled, which window of time is shown, and where a color scale's midpoint sits. Treating "the numbers are real" as proof that a chart is honest skips the entire mechanism this piece has been describing.

A second, related misconception is that misleading visualization is always deliberate propaganda. Some of it clearly is, and the Fox News case is widely cited specifically because the pattern of choices was hard to explain as accidental. But a large share of misleading charts in circulation are the product of default software settings, rushed production, or a designer who genuinely didn't know that diameter-scaling a bubble chart quadruples the apparent difference between two values. Cleveland and McGill's research exists precisely because even trained analysts routinely reach for perceptually weak encodings like angle and area without realizing how much accuracy they're sacrificing.

A third misconception, common among people who've just learned to spot truncated axes, is that a zero-baseline chart is automatically honest. It is not automatically honest; it's just correct on that one specific dimension. A zero-baseline bar chart can still cherry-pick its time window, use a distorted color scale, or select a chart type poorly suited to the comparison being made. Graphical integrity, in Tufte's sense, is a property of the whole chart, not a single axis setting, and it's worth checking a chart against the full list above rather than stopping after the first thing that looks fine.

The takeaway

A chart can be built entirely from real, unaltered numbers and still lead a reader to a wrong conclusion, because the distortion lives in axis choices, shape scaling, color midpoints, and time-window framing rather than in the data itself. Tufte's Lie Factor and Cleveland and McGill's perceptual accuracy rankings give that distortion a measurable structure instead of leaving it as a vague sense that something looks off. The habit worth building is simple and doesn't require statistical training: before trusting a chart's visual message, check where the axis starts, what time window is shown, whether the shape encoding matches the actual comparison, and whether the source is specific enough to verify. That five-question habit catches the overwhelming majority of the techniques covered here.

---

Read next:

• How Correlation Differs From Causation
• How Survivorship Bias Hides Failures
• Why Outliers Can Change Averages

FAQ

Q: Why does a truncated y-axis mislead readers?

People instinctively read bar height as proportional to value: a bar twice as tall is assumed to represent roughly twice the quantity. When the y-axis does not start at zero, that proportional relationship breaks down invisibly, so a small real difference can appear several times larger than it actually is even though every plotted number is accurate. Researchers Michael Correll, Enrico Bertini, and Steven Franconeri confirmed this effect experimentally at the 2020 ACM CHI Conference.

Q: Is it always wrong to start a y-axis above zero?

Not always. Line charts tracking small fluctuations in a value that never approaches zero, such as body temperature or a narrow interest-rate band, sometimes need a non-zero baseline to show meaningful variation at all. The distinction is disclosure: a clearly labeled non-zero baseline chosen for a specific, stated reason is defensible, while the same choice made silently to exaggerate a comparison is not.

Q: What is Edward Tufte's Lie Factor?

The Lie Factor, introduced by Edward Tufte in his 1983 book The Visual Display of Quantitative Information, is the size of an effect shown in a graphic divided by the actual size of that effect in the underlying data. A Lie Factor near 1.0 indicates an honest chart; Tufte's own example, a New York Times fuel economy graphic, had a Lie Factor of roughly 14.8.

Q: Why do pie charts get criticized so often by data visualization researchers?

William Cleveland and Robert McGill's 1984 study in the Journal of the American Statistical Association found that people judge angle and area, the encodings pie charts rely on, far less accurately than they judge position or length, the encodings bar and dot charts rely on. Once a pie chart has more than three or four slices of similar size, most readers cannot reliably rank them by eye.

Q: How can I quickly check if a chart I'm seeing is misleading?

Check where the y-axis starts, whether the shown time window represents the full relevant history or a cherry-picked slice, whether the chart type matches the actual comparison being made, and whether a specific, checkable source is cited. If several of those checks come back negative at once, treat the visual impression with real skepticism until you can find the underlying numbers.