Misleading charts & cherry-picked statistics

How it works

A chart feels like evidence. Numbers feel like facts. That is exactly why the most durable manipulation does not fake the data at all: it keeps the numbers real and lies with the picture.

The oldest trick is the truncated axis. Unemployment goes from 4.1% to 4.4%. Plot that from zero and you see two nearly identical bars. Start the y-axis at 4.0% and the second bar towers over the first. Nothing was invented; your eye was simply told where to look. The second trick is the window. A line that has gone up and down for twenty years can be made to “soar” or “crash” by choosing which five years to show. The third is the naked percentage. “Violent crime up 40%” is a true description of going from five cases to seven. Medicine has names for this: relative risk (“doubles your risk”) versus absolute risk (“from 1 in 10,000 to 2 in 10,000”), and Gerd Gigerenzer’s research group has shown that even doctors are routinely misled by the relative version.

Then there is the pair of lines that rise together. Ice cream sales and drownings both peak in July; nobody thinks ice cream drowns people. Swap in “immigration” and “rent”, or “a policy” and “growth”, and the same picture gets shared as proof. Tyler Vigen’s Spurious Correlations site has hundreds of these, including cheese consumption tracking the number of people who died tangled in their bedsheets. A correlation is a reason to look closer, not a conclusion.

None of this is new. Darrell Huff’s How to Lie with Statistics catalogued most of it in 1954. The feed just made the charts faster and the sources rarer.

Example

@growth_report_daily, 6h Real wages SKYROCKET under this government. The numbers don’t lie. [chart: two bars, “2022” and “2025”, the second roughly three times taller; y-axis starts at 96 and ends at 101; no source]

@szabo_numbers, 5h Axis starts at 96. The actual change is from 97.4 to 100.1: under 3% over three years, most of it in 2025. The same chart drawn from zero is two bars of the same height.

@truth_about_streets, 4h Knife crime up 40% since the new police minister took over. Feel safe yet? [chart: red line shooting upward; starts in January, ends in March]

@szabo_numbers, 3h 40% is 5 incidents becoming 7. The same series over ten years is a slow decline. Three months is not a trend, it’s a weather report.

Notice that both posts use real numbers, and both posters probably believe them. The chart did the lying; the people just passed it on.

Why it works on you

• Pictures skip the checking. You read a sentence word by word; you take in a chart in one glance, and the glance is already a conclusion. Steep line: bad. Tall bar: big.
• Numbers borrow authority. “40%” feels like a measured fact even when the thing measured is seven people. See authority bias.
• You only check the charts you dislike. The chart that confirms your side gets a share; the one that contradicts it gets a “source?”. See confirmation bias.

What to do

When someone shows you a chart: three questions, in order. Where does the y-axis start? When does the time window start, and why there? What are the actual numbers behind the percentage? If the chart has no source, those questions have no answers, and neither does the chart. If it comes from a “study”, check who paid for it (see lobby-funded studies).

When you catch yourself nodding: look for the one thing the chart does not show. The years before the window. The absolute count. The third variable that could move both lines. If it took you ten seconds to find a plausible other explanation, the correlation proves nothing yet.

In a comment section: redraw it. A quick screenshot captioned “same numbers, axis from zero” does more than any argument. If you cannot redraw it, ask for the source table; a chart without one is a drawing. Keep it about the chart, not the sharer; most people have no idea the axis was cut.

Don’t confuse it with

• A non-zero axis is sometimes correct. Body temperature, exchange rates or approval ratings hovering around 50% genuinely need a zoomed axis to show anything. The test is whether the chart says so, and whether the zoom matches the size of the claim.
• Choosing a time range is unavoidable. Cherry-picking is choosing it because of the result, then hiding that you chose.