A headline says people who drink more coffee live longer.
Another says people who spend more time on social media are more depressed.
A third says children who read more books perform better at school.
The pattern feels immediate.
Coffee protects you.
Social media harms you.
Books make children smarter.
Maybe.
But a correlation is not a verdict. It is a clue.
And clues are dangerous when they arrive wearing the costume of an answer.
What correlation actually tells you
A correlation says that two variables vary together.
When one tends to be higher, the other tends to be higher or lower in a systematic way.
That can be extremely useful.
It can reveal patterns worth investigating, improve prediction and expose relationships that would otherwise remain invisible.
But the correlation itself does not contain an arrow showing which variable causes which.
If X and Y are associated, several stories can produce the same pattern.
X may cause Y.
Y may cause X.
A third variable Z may cause both.
Selection into the sample may create the association.
Measurement error may distort it.
Or chance may have produced a pattern that will shrink when new data arrive.
The graph shows a relationship.
It does not automatically tell you the machinery underneath it.
The third-variable trap
Imagine a study finds that children who carry lighters are more likely to develop lung cancer later in life.
The relationship could be statistically real.
But giving children lighters would not be the causal mechanism.
Smoking would be an obvious common cause: smokers are more likely to carry lighters, and smoking raises lung-cancer risk.
That common cause is a confounder.
Real research is rarely this obvious.
Confounders can be socioeconomic status, age, prior health, personality, genetics, education, family environment, treatment history or dozens of other factors.
The dangerous version is not the confounder everyone sees.
It is the one nobody measured.
Reverse causation: when the arrow points backwards
Suppose loneliness is correlated with heavy social-media use.
One explanation is that social media increases loneliness.
Another is that lonely people seek more online interaction.
Both processes may occur simultaneously.
A single cross-sectional correlation cannot cleanly separate them.
This is why time matters.
A proposed cause must precede its effect, but even longitudinal order does not by itself eliminate every alternative explanation.
The arrow needs evidence, not intuition.
A dataset can reverse its own story
One of the strangest warnings in statistics is Simpson's paradox.
A trend can appear in the combined data and reverse when the data are separated into relevant groups.
That sounds like mathematics misbehaving.
It is usually a causal reasoning problem.
Imagine two treatments used in patients with different illness severity. Treatment A is chosen more often for the sickest patients. Treatment B is used more often in mild cases.
Overall, B may appear to have the better survival rate.
Yet within both mild and severe groups, A may perform better.
The aggregate comparison was mixing treatment effect with who received the treatment.
The numbers were not fake.
The causal story was incomplete.
Reviews of Simpson's paradox emphasize that there is no purely statistical rule that always tells you whether the combined or stratified result is the meaningful one. You need knowledge about the causal structure.
That is an important lesson:
data do not interpret themselves.
Why randomized experiments are so powerful
Randomization is one of science's most elegant tricks.
If participants are randomly assigned to treatment and control conditions, known and unknown background factors should, in expectation, be distributed without systematic preference between the groups.
That makes the groups more exchangeable.
If the groups then differ in outcome, the intervention becomes a much stronger causal explanation.
But randomization is not magic.
Trials can suffer from dropout, non-adherence, broken blinding, measurement problems, small samples or poor generalizability.
And many causal questions cannot be randomized at all.
You cannot ethically assign people to smoke cigarettes for thirty years to test whether smoking causes lung cancer.
You cannot randomly assign childhood poverty, bereavement, abuse or war exposure.
If “only randomized trials can support causation” were the rule, science would be forced to remain silent about many of the most important causes in human life.
The smoking lesson
The history of smoking and lung cancer is the perfect corrective to a simplistic version of “correlation is not causation.”
The causal case was built largely from converging observational evidence.
Researchers examined strength and consistency of association, timing, dose-response patterns, biological plausibility and alternative explanations.
Austin Bradford Hill's famous 1965 paper described several aspects of an association to consider when deciding whether causation was the most reasonable interpretation.
They were never a mechanical proof machine.
No checklist converts correlation into truth.
The point was to compare causal explanations against competing ones.
The evidence became compelling not because one correlation was large, but because many different lines of evidence pointed in the same direction.
When observational evidence fooled researchers
There is a reason caution remains necessary.
Hormone replacement therapy became a classic methodological case.
Observational studies had suggested lower coronary heart disease risk among women using hormone therapy. Later randomized evidence produced a very different picture, triggering intense debate about confounding and study design.
Subsequent analyses showed the disagreement was more complicated than “observational research was wrong.” Timing since menopause, follow-up and analytic design contributed to the differences.
That is precisely why the case matters.
A large dataset does not remove design problems.
A million observations can estimate the wrong quantity with extraordinary precision.
More data reduce random error.
They do not automatically remove bias.
Can observational studies estimate causes?
Yes, under assumptions.
Modern causal inference does not solve the problem by pretending confounding disappears.
It makes assumptions explicit.
Researchers may use causal diagrams to identify which variables should be adjusted for, compare exposed and unexposed groups more carefully, use natural experiments, instrumental variables, matching, inverse probability weighting or sensitivity analyses, and ask how strong an unmeasured confounder would need to be to erase the result.
None of these methods creates certainty from thin air.
They make the causal argument inspectable.
That is a major improvement over simply running a regression and using causal language afterward.
Why “controlling for everything” can also go wrong
A common reaction to confounding is simple:
Just control for more variables.
But causal inference is not a contest to include the longest list of covariates.
Some variables are consequences of the exposure rather than common causes.
Others are colliders, variables influenced by two other factors. Conditioning on them can create associations that were not there before.
This is another reason causal diagrams matter.
The goal is not maximum adjustment.
It is appropriate adjustment.
Prediction is not explanation
A model can predict without identifying causes.
Your phone may predict that you are likely to buy running shoes because of dozens of behavioral signals.
It does not need to know which signal caused your desire.
That is acceptable if prediction is the goal.
Problems begin when predictive relationships are quietly converted into causal stories.
“People with feature X are more likely to experience Y” is not the same sentence as “changing X will change Y.”
Interventions require the second question.
Many headlines report the first and imply the second.
The better version of the famous rule
“Correlation is not causation” is useful because it blocks a common reasoning error.
But it is incomplete.
A better version is:
Correlation alone does not establish causation. Causal inference requires a design, assumptions and evidence capable of defeating credible alternative explanations.
That sounds less catchy.
It is also much closer to how serious science works.
The DarkBrain test for causal claims
When you see a psychological or health headline claiming that X causes Y, ask:
Was X manipulated, or merely observed?
Could Y influence X?
What could cause both X and Y?
Were those variables measured before or after the proposed cause?
How were participants selected?
Does the result survive in different populations and methods?
Is there a plausible mechanism?
Would a different causal model explain the same data?
What evidence would make the researchers change their conclusion?
You do not need to become a statistician to ask better questions.
You only need to stop mistaking a pattern for an arrow.
Key Takeaways
- Correlation describes association, not causal direction.
- Reverse causation, confounding, selection, measurement and chance can all create misleading causal stories.
- Simpson's paradox shows how a real numerical pattern can reverse when a relevant third variable is considered.
- Randomization strengthens causal inference because it reduces systematic differences between groups, but trials still have limitations.
- Many important causal questions cannot ethically or practically be randomized.
- Observational evidence can contribute to causal inference when assumptions are explicit and multiple lines of evidence converge.
- Statistical adjustment is not automatically helpful; adjusting for the wrong variables can introduce bias.
- Prediction and causal explanation are different scientific goals.
- The useful lesson is not “correlation tells us nothing,” but “correlation alone is not enough.”