Astronomers cannot move stars. Geologists cannot rerun the history of a mountain range. And no ethics committee would let researchers assign people to smoke for thirty years. Yet astronomy, geology and much of medicine are thoroughly scientific. This level explains how science draws reliable conclusions when controlled experiments are impossible, too slow or wrong to perform.

Observational science

In an observational study, researchers do not assign treatments; they observe what already happens and compare groups. The danger is that groups may differ in more than one way. People who drink a little wine might be healthier, but they might also be wealthier, and wealth affects health. A hidden factor of this kind is a confounder. Much of the craft of observational science is finding ways to rule confounders out.

John Snow's map

In the 1850s most doctors believed cholera spread through “miasma,” bad air from filth. The London physician John Snow suspected contaminated water instead. When a severe outbreak struck Soho in late August 1854, he mapped the deaths and found they clustered around a public water pump on Broad Street. He also investigated the exceptions. Workers at a nearby brewery, who drank beer and had their own well, were largely spared. A woman who lived far away but had water from Broad Street brought to her because she liked its taste died of cholera. On 8 September the local authorities removed the pump's handle. The outbreak was already declining by then, so the removal alone proves little. The strength of Snow's case lay in how many different lines of evidence pointed the same way.

The natural experiment

Snow found an even stronger test. Two companies supplied water to the same South London neighborhoods, often to houses on the same street. One, the Lambeth Company, had moved its intake upstream to cleaner water; the other, Southwark and Vauxhall, still drew water from the Thames downstream of sewage outlets. Customers of the two companies were alike in wealth, housing and neighborhood. They differed mainly in their water. Deaths from cholera were many times higher among Southwark and Vauxhall customers. This was a natural experiment: chance and history had done the random assignment that Snow could not. Economists and epidemiologists still hunt for situations like this today.

Smoking and the Hill criteria

By 1950, studies by Richard Doll and Austin Bradford Hill in Britain, and others in the United States, showed that lung cancer patients were far more likely to be heavy smokers. Tobacco companies replied that correlation is not causation. How could causation be shown without an experiment? In 1965 Hill proposed a set of considerations, now called the Bradford Hill criteria, that strengthen a causal case. Among them:

• Strength: heavy smokers had many times the risk of non-smokers.
• Consistency: the link appeared in many studies, in different countries and populations.
• Temporality: smoking came before the cancer, not after.
• Dose-response: the more people smoked, the higher their risk.
• Plausibility and coherence: tobacco smoke contains substances that cause cancer in laboratory animals.

No single criterion is decisive. Together they built a case that convinced the scientific world and, eventually, governments.

Predicting the past

Historical sciences test hypotheses by predicting what evidence we have not yet found. In 1980 the physicist Luis Alvarez and his son, the geologist Walter Alvarez, with colleagues, reported a thin layer of clay rich in iridium, an element rare on Earth but common in asteroids, at the boundary marking the extinction of the dinosaurs. They proposed that a huge asteroid had struck the Earth. That hypothesis predicted that a giant crater of the right age must exist somewhere. In 1991 scientists identified the buried Chicxulub crater in Mexico's Yucatán region, about 180 kilometers wide and of the right age. A prediction about the distant past had been tested by observation in the present.

Worked example: does the new bike lane work?

A city adds a protected bike lane on Main Street. A year later, cycling injuries on Main Street are down 40 percent. Did the lane work? You cannot run an experiment, but you can reason like Snow.

• Compare to similar streets: did injuries fall on comparable streets without new lanes? If they fell everywhere, perhaps it was a mild winter or fewer cyclists.
• Check the timing: did the drop start when the lane opened, or before?
• Look at exposure: did the number of cyclists change? Fewer injuries among more riders is stronger evidence than fewer injuries among fewer riders.
• Look for consistency: have protected lanes in other cities shown similar results?

No single check is proof. But if several independent lines of evidence point the same way, you can reach a reasonable, honest conclusion without a laboratory. Much of the knowledge that shapes public policy is built exactly this way.