Causation means that one event directly produces another event. When scientists say there is a causal relationship between two things, they mean that the first thing (the cause) is responsible for making the second thing (the effect) happen. This sounds simple, but proving causation is one of the hardest problems in science, medicine, and everyday reasoning.
Causation vs. Correlation
The most important distinction to understand is the difference between causation and correlation. Correlation means two things tend to occur together or move in the same pattern. Causation means one actually produces the other. Ice cream sales and drowning deaths both rise in summer, but ice cream doesn’t cause drowning. The shared cause is hot weather, which independently drives both.
This confusion shows up constantly in health news. A study might find that people who eat a certain food have lower rates of a disease. That’s a correlation. It doesn’t mean the food prevents the disease, because the people who eat that food might also exercise more, earn higher incomes, or have other habits that explain the difference. Establishing that the food itself is responsible requires a much higher bar of evidence.
Three Requirements for a Causal Claim
For scientists to accept that A causes B, three conditions generally need to be met. First, A must happen before B. This is called temporal precedence, and it’s the most basic requirement. If a supposed cause comes after its supposed effect, the relationship can’t be causal.
Second, A and B must reliably vary together. When A increases, B should increase (or decrease, depending on the relationship). If there’s no consistent pattern linking the two, there’s no basis for a causal claim.
Third, and most difficult, the relationship can’t be explained by something else. This is called non-spuriousness. If a hidden third factor is driving both A and B, the apparent link between them is an illusion. Ruling out these hidden factors is what makes causal research so challenging.
How Researchers Test for Causation
The gold standard for establishing causation is the randomized controlled trial. In these studies, participants are randomly assigned to either receive an intervention or not. Randomization balances participant characteristics, both the ones researchers can measure and the ones they can’t, between the two groups. Any difference in outcomes can then be attributed to the intervention itself. No other study design achieves this.
But randomized trials aren’t always possible. You can’t randomly assign people to smoke for 30 years to see if it causes cancer. In these situations, researchers rely on observational evidence and apply a set of criteria to judge whether a causal relationship is likely. The most widely used framework comes from the epidemiologist Austin Bradford Hill, who proposed it in 1965. His nine viewpoints include factors like the strength of the association, whether it’s consistent across different studies and populations, whether a dose-response pattern exists (for example, more cigarettes leading to higher lung cancer rates), and whether removing the suspected cause reduces the effect.
The 1964 Surgeon General’s report on smoking and lung cancer is a landmark example of this approach in action. No randomized trial had been conducted, but epidemiologists used large-scale, long-term surveys to link rising lung cancer deaths to smoking. Pathologists and lab scientists confirmed the statistical relationship to lung cancer and other diseases like emphysema and heart disease. The combined weight of evidence from multiple angles made the causal conclusion overwhelming.
The Counterfactual Test
One of the clearest ways to think about causation is the counterfactual approach: if the cause had not happened, would the effect still have occurred? If you flip a light switch and the light turns on, the counterfactual question is whether the light would have turned on without the flip. If the answer is no, flipping the switch caused the light to come on.
This framework now dominates discussions of causation in fields like epidemiology and economics. It forces researchers to think carefully about what would have happened in the absence of the cause, which is essentially what a control group in a randomized trial represents.
Deterministic vs. Probabilistic Causation
In everyday thinking, causation often feels absolute: flipping the switch always turns on the light. This is deterministic causation, where A always produces B. But in biology, medicine, and social science, causation is almost always probabilistic. Smoking causes lung cancer, but not every smoker develops it. What the causal claim means is that smoking raises the probability of lung cancer compared to not smoking.
This probabilistic nature is why causal claims in health can feel slippery. When a doctor says obesity causes heart disease, they don’t mean every person with obesity will develop heart disease. They mean it meaningfully increases the risk. Understanding this distinction helps make sense of why individual exceptions don’t disprove a causal relationship.
What Makes Causation Hard to Prove
Three types of hidden variables complicate causal claims. The first is a confounding variable, which independently influences both the suspected cause and the outcome. If you find that coffee drinkers have higher rates of heart disease, smoking could be a confounder, since smokers also tend to drink more coffee. Failing to account for the confounder leads to incorrect conclusions about the relationship between coffee and heart disease.
The second is a mediating variable, which sits in the causal chain between the cause and the effect. If exercise reduces depression, and the mechanism is that exercise improves sleep quality, then sleep is a mediator. Exercise causes better sleep, which causes less depression. Mediators help explain how a cause produces its effect.
The third is a moderating variable, which changes the strength or direction of a causal relationship. A medication might work well for younger patients but poorly for older ones. Age doesn’t sit in the causal chain; it modifies how strong the causal link is. Researchers use tools called directed acyclic graphs (essentially causal maps) to visually lay out these relationships before designing a study. These diagrams help identify which variables need to be controlled for and which ones, if controlled for incorrectly, would actually introduce bias rather than remove it.
The Most Common Causation Mistake
The post hoc ergo propter hoc fallacy, Latin for “after this, therefore because of this,” is the most frequent error in everyday causal reasoning. It assumes that because event A happened before event B, A must have caused B. Everyone turns 18 before graduating high school, but turning 18 doesn’t cause graduation.
This fallacy plays a significant role in health misinformation. Vaccines are typically given during the same developmental window when autism symptoms first become noticeable. The temporal overlap led some parents to conclude that vaccines caused autism, a textbook case of mistaking sequence for causation. Large-scale studies have since found no causal link.
How Scientists Rate Causal Evidence
Not all evidence for a causal claim carries equal weight. In medicine, the GRADE system categorizes evidence into four levels: high, moderate, low, and very low. At the high level, researchers are very confident the true effect is close to what studies estimate. At the very low level, the true effect could be substantially different from current estimates. This rating system helps policymakers and clinicians judge how much to trust a causal claim when making decisions about treatments or public health guidelines.
The strength of causal evidence typically increases as you move from anecdotes to case reports, from case reports to observational studies, and from observational studies to randomized trials. Systematic reviews that pool results from multiple randomized trials sit at the top. Each step up provides more confidence that the relationship is genuinely causal and not an artifact of bias, confounding, or chance.