Number of words: 270<\/p>\n\n\n\n
In a case-control study, risk is estimated post hoc\u2014in Doll\u2019s and Wynder\u2019s case by asking patients with lung cancer whether they had smoked. In an often-quoted statistical analogy, this is akin to asking car accident victims whether they had been driving under the influence of alcohol\u2014but interviewing them after their accident. The numbers one derives from such an experiment certainly inform us about a potential link between accidents and alcohol. But it does not tell a drinker his or her actual chances of being involved in an accident. It is risk viewed as if from a rearview mirror, risk assessed backward. And as with any distortion, subtle biases can creep into such estimations. What if drivers tend to overestimate (or underestimate) their intoxication at the time of an accident? Or what if (to return to Doll and Hill\u2019s case) the interviewers had unconsciously probed lung cancer victims more aggressively about their smoking habits while neglecting similar habits in the control group?<\/p>\n\n\n\n
Hill knew the simplest method to counteract such biases: he had invented it. If a cohort of people could be randomly assigned to two groups, and one group forced to smoke cigarettes and the other forced not to smoke, then one could follow the two groups over time and determine whether lung cancer developed at an increased rate in the smoking group. That would prove causality, but such a ghoulish human experiment could not even be conceived, let alone performed on living people, without violating fundamental principles of medical ethics.<\/p>\n\n\n\n