The Broad Street Pump

Before any data, write down both explanations and what each one predicts. This is the step that does the work, and it is the step people skip. Bad air predicts that risk falls off with distance from the source of the smell, that it affects everyone in an area more or less equally, and that people who spend more time outdoors in the miasma fare worse. It is a theory about place. Waterborne transmission predicts that risk follows the water supply and nothing else. Two houses next door to each other should differ sharply if they draw from different pumps. A household far away that fetches water from the suspect source should be at high risk despite the distance. And an institution inside the affected area with its own supply should be spared. Notice that these two theories agree almost everywhere. In the middle of the cluster both predict lots of deaths, so the middle of the cluster is worthless as evidence. The theories disagree only at the exceptions — the workhouse, the brewery, the distant household with a preference. Those are the informative points, and there are only a handful of them. Write this down before you look at anything. Deciding in advance which observations would change your mind is the difference between analysis and decoration, and it is why Snow's map convinced people while other maps of the same outbreak did not.

The method is not about cholera and it is worth proving that to yourself on data you can gather. Pick something with a plausible point source and a spatial footprint. Damp in a building against the position of a leak. Corrosion on a fence against distance from a road that is salted. Which taps in a workshop are used by people who report a taste. Slug damage across a garden against the position of a compost heap. Any of these is a real inference problem with the same shape. Then work the four steps in order. Write both hypotheses first, and write what each predicts. Record exposure per subject, not per location, because the denominators are where these analyses die. Compute rates per head. And then go looking specifically for the cases where the two explanations disagree, because those are the only observations that will move you. Expect the exceptions to be rare and expect to have to go and find them deliberately — they almost never turn up in data you already had. That is the practical content of Snow's method and the reason it is still taught. Where this stops. This is an analysis technique, and it is genuinely yours to use. What it is not is a way to draw conclusions about anyone's health. Real outbreak investigation involves case definitions agreed before counting, laboratory confirmation, control of confounders, and ethical review of how the data is collected and held — and the same map that identifies a source can identify individuals, which is a serious matter. Use it on buildings, gardens, workshops and materials. If you ever think you are looking at a real cluster of illness, the correct action is not to publish a map. It is to tell a public health authority, who have the powers and the duty to investigate properly.