Charts and Graphs Didn’t Have to Be This Hard

A chart is an argument someone made with a picture, so read it from the words inward: the claim, what's measured and in what units, the axes, the comparison, the time window and source, and what's missing. Then read two numbers off it and work out the change yourself.

An axis is a ruler: find the units, any multiplier such as "thousands", and what one step is worth, then read any point as the label below it plus its fraction of the gap. Check that equal distances mean equal amounts before trusting a slope or a bar.

Pick a chart by the question it answers: bars to compare categories (sorted, from zero), a line for one thing over time, a pie for a few clearly different parts of one whole, dots for relationships, and a table for a few exact values. Eyes judge positions on a shared scale better than angles or areas, which is why bars beat pies for close comparisons.

A percent is always a percent of something. A change in a rate can be given in percentage points (the plain gap) or in percent (the gap compared with the start), and the honest description gives both; percent changes of different bases don't add, a fall and the rise that undoes it are different sizes, and a big percent of a tiny base is still tiny.

A count grows with the size of a place or group; a rate divides that size out (per person, per 1,000, per mile, per hour) so different sizes can be compared. Choose the denominator that matches your question, check the "per" and the time period on the axis, distrust the wild rates of tiny groups, and plan with totals.

A bar's length is its number, so bars must start at zero; when a bar chart's axis starts higher, the bars show only the leftovers above that start, and a small difference can look several times its real size. Lines and dots can zoom in, as long as the axis is labeled.

How steep a line looks is decided by the y-axis range, the chart's height and its width, not only by the data; zooming in turns small wobbles into drama and zooming out hides real change, so read the start and end values and say the change in numbers.

On a chart with two y-axes, each line's axis range is chosen separately, so where the lines sit, cross or run together is the chart maker's choice; compare each line with its own start, index both to 100, or give each its own chart.

A picture made twice as tall keeps its shape and so becomes twice as wide, covering four times the page (and a 3-D picture seems to hold eight times as much); scale the area to the number, size bubbles by area, avoid 3-D, or use bars and same-size icons.

On a log axis each equal step multiplies by the same amount, so steady percent growth is a straight line and its slope is the growth rate; a flatter line means slower growth, not shrinking, and equal-looking gaps are equal ratios, not equal amounts.

A change depends on where you start and stop, and a bumpy record offers a wide menu of true percentages; ask "from when, to when, and why those two?", compare a month with the same month a year earlier, and look at the whole record.

A running total adds each period to everything before, so it can only climb and a slowdown shows only as a bend; look at each period's own amount, the size of the latest step, and the raw numbers behind any rolling average or year-to-date figure.

An average stands in for a whole pile of numbers and hides its spread, its shape, its outliers and the sizes of its groups; ask to see the dots, ask whether it's the mean or the median, weight groups by size, and ask what the error bars measure.

A shaded map is read by area and color, so a map of counts mostly shows where people live, big empty areas get a voice far larger than their people, and cut points chosen by someone decide how dark it looks; map rates, read the legend's numbers and direction, and check the population of the smallest places.

Dollars shrink as prices rise, so money over time must be measured with one yardstick: real value = nominal value × (index now ÷ index then), per person when the population changed, and per year as a compound rate, not a simple division.

A relative change ("doubles", "cut by 75%") compares a chance with itself; only the absolute change, counted out of 100 or 1,000, tells you how much difference it makes, so always ask "out of how many?" and "over what time?".

A random sample's margin of error is about 1 ÷ √(number asked), so read every poll as a range, treat a lead smaller than about twice the margin as too close to call, and remember that no margin of error fixes the wrong people or a leading question.

A number means nothing until it's compared with something, so ask for its normal range, how long its record is, the size of the whole it came from, and how much it usually wobbles before you decide it's big.

A scatterplot and its correlation show how two things move together among the dots you have; they can't show which causes which, what happens past the last dot, or what a third factor or one outlier is doing, so always look at the dots.

A forecast is a range that should widen the further ahead it looks, so prefer fan charts to single lines, distrust "on track to" for anything seasonal, read best and worst cases as rare corners, and judge forecasters by how often their ranges caught what happened.

Claims in ads and reports are usually true but selected: "up to" is the single best case, "more" hides its comparison, "3 out of 4" hides its count, and "lasts up to" hides its test conditions. Find the best case, the comparison and the count, then work out the typical case and the price per unit.

A ranking keeps the order and throws away the gaps, so tiny score changes in a tight pack move ranks a long way, ranks built on small numbers reshuffle by luck, and composite scores depend on weights someone chose. Before trusting a rise or fall, find the scores, the count, the weights and the number of tries behind each score.

An honest chart says one thing in its title, uses the right kind of picture, starts its bars at zero, shares its scales, uses one axis per quantity, sorts and labels directly, points with color rather than decorating, shows its uncertainty, and names its source; every rule is a trick from this book turned around.

Ten questions, asked in order, catch nearly every trick in this book in about a minute: what's the claim; what's measured; where the axis starts; is it the right picture; percent or points, count or rate; who chose the window; what hides behind the average or the map; relative or absolute and how many asked; compared to what; and is it a forecast, an ad or a rank.