Skip to content

← Thinking Clearly

Without an experiment

Chapter 21 · Cause

In this chapter 5 sections
  1. Many kinds of clue
  2. How strong, and how much
  3. The doubter
  4. Nature’s own coin
  5. When, and how

In 1922, 612 people in England and Wales died of lung cancer. In 1947 it was 9,287, about fifteen times as many. Some doctors thought the rise was only on paper, because they had got better at telling lung cancer from other chest diseases. Two men at Britain’s Medical Research Council thought it was too large and too steady for that, and set out to find the cause. Austin Bradford Hill was a statistician. Richard Doll was a young doctor, and he smoked.

There were two main suspects. One was the air: exhaust from the growing number of cars, smoke from coal fires and gas works, and dust from roads newly surfaced with tar. The other was tobacco. Doll’s own hunch was the tar on the roads.

Between April 1948 and October 1949, twenty London hospitals told the team about every patient with lung cancer. A trained interviewer visited each one, and for each she also found another patient in the same hospital, of the same sex and about the same age, who had something else wrong with them. Everyone was asked the same questions: about their families, their food and their past illnesses, whether they had ever worked on the roads, and how much they smoked. Of 649 men with lung cancer, two had never been regular smokers. Among the 649 men in the comparison group, 27 had never been. Doll later remembered what struck him most. Some patients were sent in as lung cancer and turned out to have something else, and “if the person was a non-smoker, the diagnosis would always change.” He gave up smoking straight away. “It wasn’t so difficult.”

But a finding like that is exactly the kind a fair test exists to doubt. Nobody had tossed a coin to decide who smoked. People chose for themselves, and smokers might differ from non-smokers in a hundred other ways: in their jobs, their homes, their drink, perhaps even their bodies. And there was never going to be a trial. Nobody can hand out cigarettes by coin toss to thousands of teenagers and wait thirty years to count the dead. So how did the world become sure that smoking causes lung cancer? It did become sure, about as sure as it is of anything in medicine.

Many kinds of clue

The short answer is that no single study did it. What did it was many different kinds of evidence, each with its own weak spot, all pointing the same way, until nothing but cause could explain them all at once.

A court can decide a case without a confession. One witness might be lying. A footprint might be someone else’s. But if the witness, the footprint, the timing and the stolen goods in the cupboard all point to the same person, then any innocent explanation has to account for every one of them together, and that gets harder with each new clue. Each piece is weak alone. Together they can settle it.

For a cause, the clues are of a few kinds. How strong the link is: a small difference can be made by something hidden, a huge one is much harder to fake. Whether more of the cause brings more of the effect. Whether the cause comes first, by the right amount of time. Whether the same link turns up in different places, found by different people in different ways. What happens when people stop. And whether anyone can see how the cause does its damage. In 1965 Bradford Hill set out nine such points of view in a famous lecture, and he was firm that they weren’t a checklist: none of them could prove a cause on its own, and none was a must. What they could do, he said, was help answer one question: “is there any other way of explaining the set of facts before us”?

Smoking was his main example.

How strong, and how much

Doll and Hill’s hospital study had weak spots, and the two men knew it. A patient who knows he is ill with a chest disease might remember his smoking differently, and an interviewer who knows the diagnosis might ask differently. The team had a neat check on the second worry. Of the patients sent in as lung cancer, 209 turned out not to have it. The interviewers had thought they did, so if the interviewers were pushing answers, these patients should have looked like heavy smokers. They didn’t. They smoked like the other patients without lung cancer.

The bigger answer was a new kind of study, one that couldn’t be fooled by memory. In October 1951 Doll and Hill wrote to the doctors of Britain, 59,600 of them, and asked how much they smoked. More than 40,000 wrote back. Then the two men simply waited, and watched the deaths come in. They chose doctors, people the medical profession might find more reliable than hospital patients. And nobody’s smoking could now be remembered differently because of an illness, since the question was asked before anyone fell ill.

After ten years the pattern was plain. Among the men who didn’t smoke, about 7 in every 100,000 died of lung cancer each year. Among those who smoked up to 14 cigarettes a day, it was 57. From 15 to 24 a day, 139. Twenty-five or more, 227. That is 8, 20 and 32 times the non-smokers’ rate, rising step by step with every cigarette.

Both things matter. A link that big is very hard to explain away, and the statistician Jerome Cornfield and his colleagues showed why with simple arithmetic in 1959. Suppose smokers get lung cancer nine times as often, and the real cause is something else smokers happen to have, a hormone, say. Then that hormone would have to be at least nine times as common among smokers as among non-smokers. A hidden difference that large wouldn’t stay hidden for long. Bradford Hill made the same point more plainly: anything else to blame would have to be “so intimately linked with cigarette smoking and with the amount of smoking” that it should be easy to find.

The steady rise matters too. If light smokers had died at a high rate and heavy smokers at a lower one, the case would have been much more puzzling. Instead the risk climbs with the dose, the way you would expect if the thing being counted were doing the harm.

Hill was careful to say this kind of strength doesn’t always come. Smoking was linked with deaths from heart attacks too, but at no more than twice the non-smokers’ rate, and with a link that small, he wrote, it is “surely much easier” to think of other things that go along with smoking, such as exercise or diet, that might be doing the work. He thought the evidence on hearts pointed to a cause as well. He just knew a doubling was easier to fake than a thirtyfold rise.

The doubter

The sharpest doubter was the man who had made the coin toss a rule of science. Ronald Fisher’s book The Design of Experiments, in 1935, set out random assignment as a basic part of a fair experiment. Now, in the late 1950s, he argued that the smoking evidence had made “an error … of an old kind, in arguing from correlation to causation.” Perhaps, he wrote, the same inborn nature that inclines some people to smoke also inclines them to lung cancer, and the cigarettes do nothing at all. He had evidence of a sort: twins from Germany, and more sent to him by a doctor in London, showed that identical twins were more alike in their smoking than non-identical ones, so genes seemed to play some part in who smoked. “There is nothing to stop those who greatly desire it,” he wrote in Nature in 1958, “from believing that lung cancer is caused by smoking cigarettes.”

Fisher was a consultant to the tobacco industry, and he is said to have smoked himself. Those facts don’t make his argument wrong, and the argument deserved an answer: a third thing behind both is always a real possibility, and sometimes it is the truth. The answer came from the evidence, and in the end from twins themselves.

Identical twins share all their genes. So find pairs of identical twins in which one smoked and the other never did, and any inborn tendency to lung cancer is shared between them. If Fisher was right, the smoking twin should be no worse off. A large study of twins in four Nordic countries, followed for nearly thirty years, did exactly this. Among identical twin pairs where one smoked and the other didn’t, the smoker was about five times as likely to get lung cancer. The pairs were few enough that the range around that answer is wide, from about twice to fourteen times, but it is well clear of “no difference”.

There was a simpler answer too. A gene can’t be given up. Cigarettes can, and when people gave them up, the risk fell. A large share of the smoking doctors stopped during the study’s first decades, and over the fifty years the men who had stopped died of lung cancer at about a quarter of the rate of those who kept going. Stopping at 30 avoided almost all of smoking’s extra risk of dying, and stopping at 50 halved it. Fisher had an answer ready even for this. People who choose to stop, he wrote in 1957, might be different people from those who carry on, so the fall was only the same link seen again. On its own that is fair, since stopping is chosen too. But one hidden nature now had to explain the dose, the quitting and the twins all at once, and the clues still to come made its job harder.

Nature’s own coin

Quitting is a kind of experiment that people run on themselves. Now and then the world runs a better one, by accident, and the classic case came a century before Doll.

In 1854 cholera swept through London. A doctor named John Snow thought it spread through water fouled by the sick. South of the Thames, two water companies piped water into the same neighborhoods. Their pipes went down the same streets, and which house got which company had been settled years earlier by its owner or tenant, when the companies were competing for customers, so that, Snow wrote, “a single house has a supply different from that on either side.” Both served rich and poor, large houses and small. But in 1852 one of them, the Lambeth Company, had moved its intake up the river, above the city’s sewage. The Southwark and Vauxhall Company still drew from the Thames inside London.

So, in Snow’s words, more than three hundred thousand people “were divided into two groups without their choice, and, in most cases, without their knowledge.” It was nearly as good as a coin. Nobody had chosen their water for reasons that had anything to do with cholera. Snow went from house to house where someone had died, finding out which company supplied it. In the first seven weeks of the outbreak, houses supplied by Southwark and Vauxhall had 315 deaths for every 10,000 houses. Houses supplied by Lambeth had 37. Between eight and nine times as many deaths, and Snow could find nothing else that told the houses apart.

A black iron hand pump with a spout, on a stepped stone base, in front of a pub with JOHN SNOW written above its doors.
Figure 21.1 A copy of the Broad Street pump in Soho, outside a pub named after Snow. In an outbreak there that same year, Snow went to the parish’s Board of Guardians, and the next day the pump’s handle was taken off. Photo: Whispyhistory, CC BY-SA 4.0

Situations like this, where an accident of history splits people into groups that nobody chose, are called natural experiments, and when they are good, they come close to the coin.

When, and how

A cause has to come before its effect, and a slow disease should come long after. Lung cancer takes decades to grow, so if cigarettes cause it, the deaths should follow the smoking by something like a generation.

They did. In the United States, sales of factory-made cigarettes climbed from almost nothing in 1900 to a peak of about eleven a day for every adult in 1963. Men’s deaths from lung cancer kept climbing for another quarter of a century after that, and peaked in 1990. Smoking among American women only began to rise in the 1920s, and their deaths rose later too, and leveled off later. In 1964 the American government’s first report on smoking said that cigarettes cause lung cancer in men. For women, it said, the evidence was “less extensive”, though it pointed “in the same direction”. That year women were dying of lung cancer at about a sixth of men’s rate.

Cigarettes sold per adult per day04812Lung cancer deaths per 100,000020406080Highest, 1963Men, 1990Women, 19981900194019802020Year (United States)
Figure 21.2 Cigarettes and lung cancer in the United States. Above, cigarettes sold per adult each day, 1900 to 2014. Below, lung cancer deaths per 100,000 men and per 100,000 women, adjusted for age, 1950 (when these records begin) to 2022. The dashed line marks the year cigarette sales were highest.

The last kind of clue is the one people often think should come first: how does smoke actually cause cancer? Bradford Hill said it was helpful but couldn’t be demanded, because what seems biologically possible “depends upon the biological knowledge of the day.” Snow couldn’t see what in the water was doing the killing. Robert Koch found the germ that causes cholera about thirty years later. Snow was right anyway, and his evidence didn’t become wrong because no one could yet see the germ.

For smoking, the how arrived piece by piece. In 1953 Ernst Wynder and Evarts Graham, in the United States, collected tar from cigarette smoke and painted it on the backs of mice three times a week, and many of the mice developed skin cancer. Then in 1996 a team in California showed something much more exact. A chemical in tobacco smoke, benzo[a]pyrene, sticks to DNA, and in a gene called p53 it sticks most of all at three particular spots. Those are the same three spots that are most often damaged in the p53 gene of human lung tumors. The fingerprint at the scene matched the suspect’s.

Put together, then: smokers get lung cancer many times as often as non-smokers, more so the more they smoke. The deaths follow the cigarettes by decades, later in women, whose smoking rose later. The link was found by many teams, in many places. In 1964 the American report counted 29 studies that looked back from patients and 7 that followed people forward, all with the same link. The risk falls when people stop. Twins who share every gene differ when one of them smokes. Smoke tar causes cancer in animals, and a chemical in smoke damages the very spots on a gene that are found broken in lung tumors.

Each of these could, on its own, have another explanation. A third thing might make a link look strong. Timing might be a coincidence. Mice are not people. But the other explanation would have to fit every clue at once, the dose and the delay and the twins and the quitting and the gene, and nobody has ever found one that does. Each clue alone moves the dial a little. Clues that are independent of each other, each one hard to explain any other way, can carry it nearly all the way to the top.

Nearly, not completely. Hill ended his lecture with that. “All scientific work is incomplete,” he said, and it can always be “upset or modified by advancing knowledge.” But that “does not confer upon us a freedom to ignore the knowledge we already have.” Who knows, he added, quoting the poet Robert Browning, whether the world will end tonight? “True, but on available evidence most of us make ready to commute on the 8.30 next day.”

Without a coin, then, the questions change. How strong is the link? Does more bring more? Did it come first, by the right amount? Has it been found by others, elsewhere, in other ways? What happens when it’s taken away? Has some accident of history split people into groups nobody chose? Is there a way it could work? And the one Hill asked first and last: what else could explain all of this at once? When a claim about cause arrives without a trial, those are the questions to put to it. A claim that can answer only one of them hasn’t done much. A claim that answers all of them, as smoking’s did, can be trusted almost as far as a trial.

Getting there took about fifteen years, and at first hardly anyone took notice. When the British government did hold a press conference on the danger, the minister who announced it was, as Doll remembered it, smoking a cigarette at the same time. Knowing and acting are not the same thing. Before either comes deciding what is known at all, and that is the next part of this book: how science decides, how it catches its own mistakes, and what happens when it doesn’t.

Found a mistake? Please tell us.