Visualization

In one line

A chart is not a neutral display of data. It is an argument, and every choice you make — axis, colour, binning, what you left out — is part of that argument. In health the argument changes what happens to a person.

Why plot at all: Anscombe

Anscombe's quartet is four datasets with identical means, identical variances, identical correlation, identical regression lines — and utterly different shapes. One is linear, one is curved, one is a perfect line with a single outlier, one is a vertical stack with one point dragging the fit.

The summary statistics cannot tell them apart. The plot tells you instantly.

That's the whole case for visualisation, and it's why "I ran the numbers" is not the same as "I looked at the data". Profile it, plot it, then model it.

The choices that are arguments

The truncated y-axis. Start a bar chart at 94% instead of 0 and a trivial difference becomes a cliff. This is the most common deception in health reporting and it's usually not deliberate — the tool did it automatically. That doesn't make it honest. Bar charts start at zero. Line charts of a trend legitimately may not, and should say so.

Colour. Red means bad. Choosing which end is red is choosing the conclusion. And roughly 8% of men have a colour vision deficiency, so a red/green clinical dashboard is unreadable to a meaningful slice of your users — that's an accessibility failure with clinical consequences.

Binning. Age groups of 0–18, 19–65, 66+ tell a different story from decades. Neither is wrong. Both are choices, and moving a boundary can create or erase a finding.

Dual axes. Almost always a way to manufacture a correlation that isn't there. Avoid.

Pie charts beyond three slices. Humans compare angles badly; that's why they exist and why they shouldn't.

Uncertainty is not optional

The failure that matters most in health: plotting the estimate without the uncertainty.

A rate of 12% from 8 patients and a rate of 12% from 8,000 patients render as the identical bar. One is noise; one is a finding. Without error bars or a confidence interval, the chart has concealed the single most decision-relevant fact.

A ward with 3 infections out of 20 looks catastrophic next to a ward with 30 out of 400 — and they're the same rate with wildly different certainty. Someone gets performance-managed over that chart.

If you plot a rate, plot the interval. If the denominator is small, say so in the chart, not in a footnote nobody reads.

Talking to clinicians and patients

Gigerenzer's work is the essential reading, and the finding is robust: natural frequencies beat probabilities, for clinicians and patients.

• Bad: "this test has a 90% sensitivity and 9% false positive rate; prevalence is 1%."
• Good: "of 1,000 people, 10 have the disease. 9 of those test positive. Of the 990 without it, 89 also test positive. So of ~98 positives, only 9 have the disease."

Same maths. Radically different comprehension — including among doctors, who get the first version wrong at rates that should alarm everyone. If your visualisation is for a clinical decision, use frequencies and show the whole 1,000.

The health-specific patterns

• Time series with the intervention marked. Did the thing you did change the thing you measured? Mark when it happened, and show enough before.
• Funnel plots for comparing institutions — they build the small-numbers problem into the shape, so a tiny hospital's extreme rate stops looking like a scandal.
• Survival curves for time-to-event, always with the numbers-at-risk underneath. The tail of a Kaplan-Meier is frequently three patients, and it's where the eye goes.
• Distributions, not means. The mean hides the inequity you're supposed to find.

The rule

Ask: would the person acting on this reach the same conclusion if I'd made different, equally defensible choices? If yes, the chart is showing them the data. If no, it's showing them your argument — and you owe them the caveat, in the chart, at the size of the headline.