Take a single pair of numbers: a hospital's infection rate fell from 4.2 to 3.9 cases per thousand patients over one year. That is a decline of about 7 percent. Now hand those two numbers to two different designers and ask each for a chart.

The first draws bars starting at zero. The two bars look nearly identical, which is accurate: the improvement was modest. The second designer starts the vertical axis at 3.8. Now the first bar towers over the second, and the chart appears to show a collapse of 75 percent. Neither designer changed the data, and both charts are technically correct. Only one conveys the true size of the change.

Graphics are often treated as decoration added once the real analysis is finished. But a chart is itself an analytical claim, and the visual system reads it faster than any caption or footnote. Statisticians and designers have developed principles for keeping that claim honest. Three of them do most of the work: proportionality, appropriate encoding, and context.

In 1983, the statistician Edward Tufte proposed a simple test in The Visual Display of Quantitative Information. He called it the Lie Factor. You take the size of an effect shown in the graphic and divide it by the size of the effect in the data. A faithful graphic has a Lie Factor of 1. Tufte suggested that values outside roughly 0.95 to 1.05 indicate substantial distortion.

His best-known example was a 1978 newspaper chart of US fuel economy standards for cars. The required mileage rose from 18 to 27.5 miles per gallon between 1978 and 1985, which is an increase of 53 percent. The chart drew those values as lines receding along a road in perspective. Those lines grew from 0.6 inches to 5.3 inches, an increase of 783 percent. Divide 783 by 53 and you get a Lie Factor of 14.8. The graphic exaggerated the change nearly fifteenfold.

Distortion like this rarely requires bad intent. It often comes from encoding values in area or volume when length would be accurate. Suppose a designer shows a doubling of oil production by doubling the height of a barrel icon, then scales the width to keep the proportions. The barrel's area has quadrupled. Render it as a 3D object and the implied volume rises eightfold. The eye responds to the size of the object, not to the number printed beside it.

The same logic explains why bar charts should usually start at zero. A bar communicates its value through its length, so cutting off the bottom of the axis breaks the proportionality the reader assumes. Line charts work differently. They encode change through position and slope, so a non-zero baseline can be appropriate as long as it is clearly marked. In every case the test is the same: does the visual difference match the numerical one?

Takeaway

Before trusting a chart, estimate the visual ratio and compare it with the numerical ratio. When the two disagree, the picture is making a claim the data does not support.

In 1984, the statisticians William Cleveland and Robert McGill ran experiments in which people judged quantities from different graphical encodings. The results produced a ranking of perceptual accuracy. People judge position along a common scale most accurately. Length comes next, then angle and slope, then area. Volume and color saturation sit near the bottom. Later replications, including crowdsourced studies by Jeffrey Heer and Michael Bostock in 2010, broadly confirmed this ordering.

This hierarchy explains the persistent criticism of pie charts. A pie asks the reader to compare angles and areas, two of the weaker channels. Show five market shares of 18, 19, 20, 21 and 22 percent as a pie, and most readers will struggle to rank them. Show the same values as a sorted bar chart and the order is obvious at a glance. Add a 3D tilt and the slices nearest the viewer look larger than they are, so the chart now distorts the values as well as obscuring them.

Chart choice also shapes which patterns seem to exist. A dual-axis line chart plots two series against vertical axes that are scaled independently. The designer can stretch or squash either line until the two appear to move together, so almost any pair of upward trends can be made to look tightly linked. Similarly, connecting categorical data such as countries or product types with a line implies a continuity or ordering that the data does not contain.

The guiding question is which comparison the reader most needs to make. Comparing magnitudes across categories favors bars or dot plots. Showing change over time favors lines. Revealing a relationship between two variables favors a scatterplot. A pie can show parts of a whole when there are only two or three slices, but a stacked bar is often clearer. The data stays the same across these options, but what a reader can perceive changes.

Takeaway

Choosing a chart type determines which comparisons will be easy for the reader and which will be nearly invisible. Pick it based on the question, not on appearance.

A chart can be proportional and well chosen and still mislead, because a graphic only has meaning in relation to its frame. Axis labels, units, scales and sources are not clerical details. They are the information a reader needs to translate marks on a screen back into quantities in the world.

Consider money over time. A chart of US healthcare spending in nominal dollars across fifty years rises steeply, and part of that rise is simply inflation. Adjust to constant dollars and then divide by population, and the curve flattens considerably, though it still rises. Each version answers a different question. Tufte argued that monetary time series should usually be adjusted for inflation. The broader principle is that readers must know whether values are raw totals, per capita figures, inflation-adjusted amounts or rates, because the shape of the chart depends entirely on that choice.

Scales need the same explicit labeling. Logarithmic axes are invaluable for data spanning several orders of magnitude, such as income across countries or the early spread of an epidemic, because they turn constant growth rates into straight lines. But a reader who misses the log scale will badly misjudge magnitudes, since equal distances now represent equal multiples rather than equal differences. Studies during the COVID-19 pandemic found that many people misread log-scaled case charts. A clear label, and sometimes a short explanatory note, prevents this.

Annotation also supplies context that the data alone cannot provide. Marking when a measurement method changed, when a policy took effect, or how large the uncertainty is keeps readers from attributing a jump or dip to the wrong cause. Hans Rosling's animated bubble charts stayed with audiences partly because he kept explaining what each axis meant, where the data came from, and which country to watch. His words did as much work as his graphics.

Takeaway

Every chart answers some question, and labels, units and notes tell the reader which one. Without them, readers supply a question of their own, and it is often the wrong one.

Honest visualization is not restraint for its own sake. Its purpose is to keep the relationship between quantities in the data and impressions in the reader's mind intact, so that what people see matches what was measured.

These principles also work in reverse, as tools for evaluating charts made by others. When you meet a striking chart, check three things. First, compare the visual change with the numerical change. Next, ask whether the chart type makes the relevant comparison easy or hard. Finally, read the axes, units and source before forming an impression.

Most misleading graphics fail at least one of these tests, and you rarely need the underlying dataset to notice. A few seconds of scrutiny is usually enough to separate a chart that informs from one that merely persuades.