Prefer to click through the interactive model?
Charts as arguments — study guide
The same fragments the interactive model serves, read in order. One source, two views.
Every chart makes a claim
A chart is not a neutral picture of the data. It is an argument like a sentence: it states a claim that can be true, slanted, or false, rather than just displaying facts, and like any argument it can be sound, slanted, or flatly dishonest. The moment you turn a column of numbers into bars or a line, you have asserted something — this is bigger than that, this is rising, these two things move together — and the reader believes the assertion before they read a single label.
What makes the claim honest or not is the encoding: the rules that map each number onto something you can see, like the height of a bar, the position of a dot, or the size of a circle. Those rules are chosen by whoever drew the chart, and the same data under two different encodings can argue two opposite things. The data did not change; the claim did.
So reading a chart well means reading its encoding, not just its shape. The rest of this concept is a short catalog of the encoding decisions that decide honesty — which visual channels the eye can actually trust, and the classic tricks that make a modest difference look enormous.
Data becomes position, length, area, color
To draw a number you have to turn it into something visible, and there are only so many channels to choose from: position along a scale, the length of a bar, the area of a shape, an angle, or a shade of color. Every chart type is really just a choice of which channel carries the value. A bar chart uses length; a scatter plot uses position; a pie chart uses angle and area; a heatmap uses color.
The choice is not cosmetic, because the eye does not read every channel equally well. A reader can compare two bar lengths to within a few percent, but asked to judge which of two blobs has the larger area, or which of two squares is a slightly darker blue, they will guess — and guess wrong. The channel you pick decides how accurately your reader can recover the number you started with.
The honest move is to put the quantity that carries your argument on the channel the eye reads best, and reserve the weak channels for rough categories where precision does not matter. The dishonest move is the reverse: hide an uncomfortable comparison on a channel nobody can read.
Some channels are read more accurately than others
The visual channels form a rough ladder of accuracy, and it is one of the most useful results in the field. Cleveland and McGill measured how precisely people could read values off each channel and found a consistent order: position along a common scale is read most accurately, then length, then angle and slope, then area, and near the bottom, volume and color shading. Reading a value off aligned position is like measuring against a ruler; reading it off area or color is like eyeballing it in the dark like measuring against a ruler versus guessing a length by eye in the dark.
Treat this as an ordinal ladder, not a precise scale. The original work grouped several channels into shared ranks because the experiments could not cleanly separate them, so the lesson is the ordering, not a score for each rung. What survives is the direction: position beats length beats angle beats area beats color, top to bottom.
The practical rule falls out immediately. If a difference matters, encode it with position or length so the reader can actually see it. This is exactly why a plain bar chart, dull as it looks, out-argues a fancier chart built on area or color — it puts the comparison on the channel the eye trusts.
The truncated baseline
The most common chart lie is also the simplest: start the axis somewhere other than zero. A bar encodes its value by length, so a bar twice as long is supposed to mean a value twice as large. Cut the bottom off the axis and that promise breaks — the bars still differ in length, but the length no longer matches the numbers. A small real gap becomes a towering visual one like cropping a photo so the ground is out of frame, making a small step look like a cliff.
Because bars encode by length, a bar chart must include zero. This is not a style preference; it is the condition that makes the length mean anything. A famous worked example takes a real increase of about 1.08 times and, by truncating the axis, makes it look like a jump of roughly 2.7 times — the same data, inflated more than double by moving one number.
Lines are the exception that proves the rule. A line chart encodes value by the position of its points and the slope between them, not by area swept from the axis, so it may legitimately omit zero to show fine change over time. The test is always the encoding: if length or filled area carries the value, zero is mandatory; if position carries it, zero is optional and sometimes misleading to force.
When area lies about size
Sizing a circle by a value is a trap, and it is worth understanding exactly why. If you set a circle's radius proportional to the value, the area grows as the square of that value, because area is pi times the radius squared. Double the value, double the radius, and the circle covers four times as much ink. The reader judges the circle by its area, so a difference of two reads as a difference of four.
This is the principle of proportional ink: the amount of ink used to represent a value should be directly proportional to the value itself. Encode a quantity by radius and you break it — the ink outruns the number and every gap looks bigger than it is. Bubble charts, packed-circle graphics, and scaled icons all fall into this hole.
The fix is not to abandon circles but to size them honestly: make the area, not the radius, track the value. That means scaling the radius by the square root of the value, so a value twice as large gets a radius about 1.41 times larger and an area exactly twice as big. Simpler still, when a comparison actually matters, reach back down the ladder and use length — a bar cannot over-scale the way an area can.
Manufactured resemblance and cherry-picked range
Put two lines on one chart with two different y-axes, and you can make them appear to move together no matter what the numbers say. Each axis scales independently, so the designer chooses the ranges that make the lines cross, track, and rise in step. The apparent resemblance is an artifact of two freely chosen scales, not a fact about the data; slide either axis and the lines drift apart again. Two series that share nothing can be made to look like a matched pair.
The same freedom hides in the horizontal axis. Show only a short, hand-picked window and a long-run trend can vanish — a decline looks like a climb if you crop to the months that happened to be rising. Coarse or oddly chosen grouping does the same quietly: bucket the data one way and a pattern appears, bucket it another and it dissolves.
The defense is to ask what was left out. One shared axis instead of two, the full time span instead of a slice, the natural grouping instead of a convenient one. A chart that only holds up inside its chosen frame is arguing from the frame, not from the data.
Match the chart to the question; strip the junk
Before honesty comes fit: the chart type has to answer the question actually being asked. Comparing values across categories wants a bar or column chart. Showing change over time wants a line. Showing how a single variable is spread out — its center, its tails, its outliers — wants a histogram or a box plot. Showing the parts of a single whole wants a pie, and only when the slices are few. Pick the wrong type and even honest data argues badly, like answering a question nobody asked.
Then comes the junk. Edward Tufte coined chartjunk in his 1983 book on visual display for every mark that carries no data — the heavy gridlines, the drop shadows, the cartoon backgrounds, the gratuitous third dimension. He framed it as a ratio: the more of a chart's ink actually encodes data rather than decorating it, the better the chart. Junk is not just ugly; it competes with the data for the reader's attention.
Three dimensions are the worst offender because they do not merely distract, they distort. A 3D pie tilts the slices so the ones in front look larger than their share, and a 3D bar makes height ambiguous. The decoration has quietly changed the claim. The plain version almost always argues more honestly than the impressive one.
Reading a chart as an argument
Put the pieces together and reading a chart becomes an interrogation. Where does the axis start, and does anything with a length or a filled area fail to reach zero? What channel carries the important number — trustworthy position and length, or slippery area and color? Is there a second axis quietly doing the work of a claim? Is the time window the whole story or a flattering slice? A chart that survives those questions is probably arguing from its data; one that needs its frame to work is arguing from the frame.
The same discipline applies when you are the one drawing. Choosing an encoding is choosing what your reader will believe, and the honest defaults are humble ones: start bars at zero, put the load-bearing comparison on position or length, size areas by area and not radius, and cut the decoration until only the data is left.
This matters most when the chart is not just for you. A dashboard you build for a team is a stack of small arguments that other people will act on without checking your axes. Every encoding choice in it is a claim made on your behalf, so the cost of a misleading one is not your own confusion but someone else's decision. Charts persuade whether or not you meant them to; the only choice is whether they persuade honestly.