A causal link is a relationship where one event or factor directly produces or influences another. It goes beyond two things happening together: it means that changing the first factor would actually change the outcome. If removing cigarettes from the equation would reduce lung cancer rates, that’s a causal link. If ice cream sales and drowning rates both rise in summer but neither one drives the other, that’s just a correlation driven by a shared factor (hot weather).
Causation vs. Correlation
Two things are correlated when they tend to occur together or move in the same direction. Causation is a much stronger claim: it means one thing actually affects the probability of the other happening. The distinction matters because confusing the two leads to bad decisions. If you noticed that people who carry lighters are more likely to develop lung cancer, banning lighters wouldn’t help anyone. The real cause is smoking, and lighter ownership is just along for the ride.
The ancient Romans had a name for this mistake: “post hoc ergo propter hoc,” meaning “after this, therefore because of this.” Just because one event follows another doesn’t mean the first caused the second. This reasoning error remains one of the most common sources of misleading conclusions in medical news and everyday thinking alike.
Three Things That Fake a Causal Link
When a relationship looks causal but isn’t, one of three problems is usually at work.
Confounding happens when a hidden third factor drives both the supposed cause and the outcome. In orthopedic research, for example, a type of hip implant called a dual mobility construct appeared no better at preventing dislocation than standard implants. But patients who received those implants were more likely to have prior spine disease, which independently raises dislocation risk. The spine disease was a confounder, masking the implant’s true benefit.
Reverse causality flips the direction of the relationship. Very low body weight is associated with higher short-term death rates after joint replacement surgery. But low weight doesn’t cause the deaths. Instead, serious underlying illness causes both the low weight and the higher risk. Similarly, studies on alcohol and surgical complications sometimes find that non-drinkers have the worst outcomes. The explanation: people in poor health often stop drinking before surgery, so the abstinence is a consequence of illness, not a cause of better health.
Coincidence is the simplest explanation. With enough data points, you can find striking correlations between completely unrelated things. The number of films Nicolas Cage appeared in per year once tracked closely with the number of people who drowned in swimming pools. No mechanism connects these. With millions of possible comparisons, some will line up by pure chance.
How Scientists Test for Causal Links
The gold standard is the randomized controlled trial. Researchers randomly assign people to either receive a treatment or not, then compare outcomes. Randomization is the key ingredient: because participants are sorted by chance, the two groups end up similar in every way except the treatment itself. Any difference in outcomes can then be attributed to the treatment rather than to some lurking variable. This is why drug approvals, vaccine rollouts, and major medical guidelines rely heavily on randomized trials.
But randomized trials aren’t always possible. You can’t randomly assign people to smoke for 30 years, or to live in poverty, or to skip childhood vaccinations. When experiments would be unethical or impractical, researchers turn to quasi-experimental methods that try to approximate randomization using natural variation in the real world.
- Regression discontinuity takes advantage of arbitrary cutoff points. If a government program kicks in at a specific income level or test score, people just above and just below that line are essentially similar, creating a natural comparison group.
- Instrumental variables use a factor that influences whether someone receives a treatment but has no direct effect on the outcome, isolating the treatment’s true impact.
- Difference-in-differences compares how trends change over time in a group that was affected by some intervention versus a similar group that wasn’t, subtracting out any shared background changes.
These approaches aren’t as airtight as randomized trials, but they’re far more convincing than simple observational comparisons.
The Bradford Hill Criteria
In 1965, epidemiologist Austin Bradford Hill laid out nine considerations for evaluating whether an observed association is likely causal. He developed them while building the case that smoking causes lung cancer, and they remain widely used today. They’re not a rigid checklist where every box must be ticked. They’re a framework for weighing evidence.
Temporality is the only criterion considered absolutely necessary. The cause must come before the effect. This sounds obvious, but in slow-developing diseases, untangling the sequence can be genuinely difficult.
Strength of association asks how large the effect is. Cigarette smokers didn’t have a slightly elevated lung cancer rate; they had a dramatically elevated one. A strong association doesn’t guarantee causation, but it makes confounding harder to explain away, because any hidden factor would need to be equally strongly linked to both the exposure and the outcome.
Dose-response relationship means more exposure leads to more effect. Lung cancer death rates rose in a straight line with the number of cigarettes smoked per day. When you see a clean dose-response pattern, it’s harder to dismiss the relationship as coincidence.
Consistency means the same finding shows up across different studies, populations, and research methods. Hill put particular weight on results that held up when studied in completely different ways, such as both forward-looking and backward-looking studies reaching the same conclusion.
Plausibility asks whether a biological mechanism could explain the link. Hill was careful to note this depends on current knowledge. A relationship that seems implausible today might make perfect sense once the science advances.
Coherence means the causal interpretation shouldn’t contradict what’s already known about how the disease works. Specificity looks at whether the exposure is linked to one particular outcome rather than a broad range of diseases. Experiment asks whether removing the exposure actually reduces the outcome (when people quit smoking, does their risk drop?). And analogy considers whether similar causal relationships are already established for related exposures.
The Counterfactual: How Causation Is Defined Formally
At its core, a causal effect is the difference between what actually happened and what would have happened under different circumstances. If you took a medication and recovered, the causal question is: would you have recovered without it? The hypothetical scenario where you didn’t take the medication is called the counterfactual.
This creates what’s known as the fundamental problem of causal inference. You can never observe both realities for the same person. You either took the medication or you didn’t. Scientists get around this by comparing groups. If a large group of treated patients recovers at a higher rate than a similar untreated group, the difference in recovery rates estimates the average causal effect of the treatment. The word “similar” is doing heavy lifting in that sentence, and it’s exactly why randomization, careful study design, and statistical adjustment matter so much.
Statistical Evidence and Its Limits
A p-value of 0.05 or lower is the conventional threshold for statistical significance in clinical research. It means that if no real association existed, there would be no more than a 1-in-20 chance of seeing results this extreme by random variation alone. A 95% confidence interval that doesn’t include the “no effect” value leads to the same conclusion.
But statistical significance alone doesn’t prove causation. A p-value tells you that a pattern is unlikely to be pure noise. It doesn’t tell you why the pattern exists. A statistically significant correlation between two variables could still be driven by confounding, reverse causality, or bias in how data was collected. Conversely, a study that falls short of statistical significance doesn’t prove there’s no causal link; the study may simply have been too small to detect it.
In fields like genetics, where researchers test millions of associations simultaneously, much stricter thresholds are used to avoid false positives. Genome-wide studies typically require a p-value below 0.00000005 before flagging a result as significant. The more comparisons you make, the more likely you are to find something by chance, so the bar rises accordingly.
Why Causal Links Are Hard to Prove
Establishing a causal link is difficult because real-world exposures rarely come in clean, isolated packages. People who exercise more also tend to eat better, earn more, and have better access to healthcare. Separating the effect of exercise itself from everything that accompanies it requires careful study design and often years of accumulated evidence from multiple angles.
Researchers sometimes use causal diagrams, called directed acyclic graphs, to map out all the variables that might influence an outcome and the arrows between them. These diagrams help identify which factors need to be accounted for to get a clean estimate of a causal effect, and which factors would actually introduce bias if you tried to control for them. A variable that sits on the causal pathway between your exposure and outcome, for instance, should not be adjusted for if you want to estimate the total effect.
The strongest causal claims in science don’t rest on any single study or method. They come from converging evidence: randomized trials where possible, well-designed observational studies where not, a plausible biological mechanism, a dose-response pattern, and consistency across populations and time periods. Smoking and lung cancer, for example, became an accepted causal link not because of one landmark study but because evidence from every angle pointed in the same direction.