Causal Reasoning: Cause, Correlation, and Confounds

You already know the slogan: correlation is not causation. The harder and more useful question is what would let you say one thing does cause another. Doctors, policymakers, and engineers cannot stop at "these move together"; they have to decide whether acting on one will change the other. Getting this wrong wastes money, delays cures, and blames innocent factors.

Establishing causation is rarely a single decisive test. It is a case built from many strands — ruling out alternatives, finding a mechanism, watching what happens when the cause changes. This lesson goes past the slogan into the machinery: confounds, the kinds of causes, and how a careful case for causation is actually assembled.

Confounding variables

A confounder is a hidden third factor that influences both the supposed cause and the supposed effect, creating a correlation with no direct link between them. Ice-cream sales correlate with drownings, but heat drives both. Until you can rule out plausible confounders, a correlation tells you two things vary together, not that one moves the other. Much of good causal reasoning is a hunt for the confounder you have not yet named.

Necessary and sufficient causes

A necessary cause must be present for the effect to occur: without the varicella virus, no chickenpox. A sufficient cause guarantees the effect on its own: decapitation is sufficient for death. Many causes are one without the other. Smoking is neither strictly necessary nor sufficient for lung cancer — non-smokers get it and some smokers never do — yet it powerfully raises the risk. Sorting a candidate cause into these categories sharpens what you are actually claiming.

Mechanism and manipulation

Two things turn a correlation into a credible cause. A mechanism — a plausible story of how the cause produces the effect, ideally at a physical or biological level — makes a link believable rather than coincidental. And manipulation: if changing the cause changes the effect, especially in a controlled experiment, the case grows much stronger. When experiments are impossible, we lean harder on mechanism and on ruling out confounders.

A town notices that neighbourhoods with more streetlights have less crime and proposes more lighting. Before spending, a careful analyst asks: could a confounder explain this? Wealthier areas may have both more lighting and more policing. So they compare similar neighbourhoods, add lights to some and not others, and track crime over time. If the newly lit streets see crime fall relative to the unlit controls, and there is a mechanism — better visibility deters offenders — the causal case strengthens from mere correlation toward something worth acting on.

Consider hormone replacement therapy. Observational studies once found women taking it had less heart disease, suggesting a protective effect. But users tended to be wealthier and healthier to begin with — a confounder. When randomised trials, notably the Women's Health Initiative reported in 2002, actually manipulated the treatment, the apparent benefit vanished and some risks rose. The correlation was real; the causal story drawn from it was wrong. It is a standing warning that even large, consistent associations can mislead until the cause is tested directly.

In 1965, the British statistician Austin Bradford Hill gave a presidential address to the Royal Society of Medicine, published as "The Environment and Disease: Association or Causation?" He set out nine viewpoints for judging whether an association is causal: strength, consistency, specificity, temporality, biological gradient (a dose-response relationship), plausibility, coherence, experiment, and analogy. Crucially, Hill insisted these were aspects to weigh, not a checklist to tick, and that only temporality — cause before effect — was close to essential.

Hill had helped build the case that smoking causes lung cancer. With Richard Doll he ran studies from around 1950, including a long-running survey of British doctors, showing the association was strong, consistent across studies, and graded: heavier smokers faced higher risk. That dose-response pattern, a plausible biological mechanism, and the drop in risk when people quit together made the causal case compelling without any single controlled experiment on humans. Hill's viewpoints remain a standard framework for reasoning toward causation from imperfect evidence.

• "A strong correlation proves causation." Strength is one clue, but a confounder can produce a strong correlation with no causal link at all.
• "A cause must be both necessary and sufficient." Most real causes are one, neither, or merely contributory; smoking raises risk without being either.
• "Hill's criteria are a checklist that proves cause." Hill himself warned against that; the viewpoints are judgements to weigh, not boxes that add up to proof.
• "If you can't run an experiment, you can't establish cause." Mechanism, dose-response, temporality, and ruling out confounders can build a strong case without a trial, as smoking research showed.

Take a causal headline you have seen recently, name the most plausible confounder that could explain the association, and decide what evidence would tell the two apart.

Think Like a Maester: Before you crown a cause, search for the third factor sitting quietly behind both things you see moving together.

1. What is a confounding variable, and how does it create a misleading correlation?
2. How do necessary and sufficient causes differ, with an example of each?
3. Why is a plausible mechanism important when arguing for causation?
4. What did Austin Bradford Hill's 1965 address propose, and what did he warn against?
5. How was smoking shown to cause lung cancer without a controlled human experiment?

Moving past the slogan, causal reasoning asks what would justify saying one thing changes another. The first task is to hunt for confounders — hidden factors that drive both variables. The second is to be precise about the kind of cause at stake: necessary, sufficient, both, or merely contributory. The strongest cases combine a plausible mechanism, evidence that manipulating the cause moves the effect, and the ruling out of alternatives. Hill's 1965 viewpoints and the smoking-cancer story show how, weighed judiciously rather than ticked off, such evidence builds a case for causation from an imperfect world.

• Correlation vs Causation
• Correlation vs Causation