Skip to content
verifiedEvidence-Based Practice

Correlation and Causality

Two things moving together doesn't mean one produces the other — and mistaking association for cause comes at a price.

Translated from the Portuguese original. Read the original (Português)

boltIn short

A correlation says one thing only: two measures vary together. It is symmetric — it doesn't point to a direction, doesn't say which came first, doesn't say whether any cause exists at all. Causality is something else: it claims that changing A produces a change in B. The problem is that several quite different situations produce exactly the same correlation: A really does cause B; a third variable causes both; B is actually causing A (reverse causation); it was coincidence; or the association arose from how the sample was built (selection bias). No calculation done only on the observed association can tell these apart — what separates them are the assumptions the study design licenses, and the gold standard remains manipulating the supposed cause and seeing what happens. That doesn't make correlation useless: it serves to predict, screen, evaluate instruments, and raise hypotheses. The mistake is treating it as if it already authorized a treatment decision.

Technical definition

Correlation is a statistical measure of the degree to which two variables covary. The best-known coefficient, Pearson's r, ranges from −1 to +1: near +1, both rise together; near −1, one rises as the other falls; around 0, there is no detectable linear relationship. One property defines its limit: it is symmetric (the r of A with B is identical to that of B with A) and therefore non-directional. Causality is an asymmetric relationship: saying A causes B means saying an intervention on A changes B. In the counterfactual framework used today in causal inference (Hernán & Robins, 2020), A causes B in a person if the outcome that person would have under A differs from what they would have without A — a comparison that, by definition, is impossible to observe in the same person; that is the fundamental problem of causal inference, and every study design is a strategy for working around it. In behavior analysis, the equivalent concept is the functional relation: it is demonstrated that manipulating the independent variable reliably and repeatedly produces the change in the dependent variable (Cooper, Heron & Heward). Correlation is, then, necessary but not sufficient evidence — with one technical caveat: the trace a cause leaves is some association, not necessarily a linear one captured by r, and subgroups that cancel out can zero the number overall. The reverse, however, does not hold: association shows up often with no cause behind it at all.

What correlation says — and what it refuses to say

Suppose you read: "children who receive more hours of intervention have worse performance." The correlation is negative and probably real. What it does not say is why. It could be that the intervention is harmful. It could also be — and it's far more likely — that whoever needs more support receives more hours, so that the need causes the hours, not the other way around. The statistical calculation is the same in both cases: r has no way of knowing which is the cart and which is the horse.

This is where reasoning tends to slip. If A causes B, you expect to find a correlation between the two — so when the correlation shows up, it feels natural to read the path backward. But the backward path isn't valid: the correlation is compatible with several different causal stories, and the number is the same in all of them.

What you observed B A Correlation says one thing only: when A rises, B rises along with it. It is symmetric and doesn't point to a direction — so it doesn't say which causes which. Four worlds that produce the same correlation 1. A really causes B A B The intervention produces the change — the only world where the conclusion is correct. 2. A third variable causes both C A B Heat sells ice cream and sends people to the beach. A and B move together without either causing the other. 3. Reverse causation: B causes A A B Whoever is doing worse seeks more help: the arrow points the other way. 4. Coincidence (spurious correlation) A B Nothing links the two. With enough variables, some will line up by pure chance.
The four worlds produce exactly the same correlation. No calculation done only on the observed association distinguishes one from another — what distinguishes them are the assumptions the study design licenses.

The competing explanations for a correlation

Faced with a correlation, the reader's job is to list the stories that would explain it without the cause someone wants to claim. Explore each one — including a fifth that no three-node diagram can draw: selection bias.

A causes B — the hypothesis you want to test

It's the desired reading, and sometimes it's the true one. The point is that it enters the lineup alongside the other four, on equal footing, until the study design rules out the competitors. Calling the preferred hypothesis "the obvious explanation" is not an argument: in all five scenarios, the number you calculated is the same.

Confounding — a third variable causes both

Ice cream sales and drownings rise together every summer, and neither causes the other: heat causes both. A confounding variable is anything that influences both the supposed cause and the outcome at the same time, creating an association with no direct causal relation. A famous example, published as a deliberate provocation: in 2012, in the New England Journal of Medicine, Franz Messerli reported a correlation of r = 0.791 between per capita chocolate consumption and the number of Nobel laureates per 10 million inhabitants, across 23 countries. The r is extremely high, and no one proposes handing out chocolate to produce Nobel Prizes — because the obvious candidate is a third factor moving both things, such as a country's wealth, education, and investment in research. In health, the classic confounders are age, severity of the condition, education, and income, and the problem is that not all of them even get measured. Worse: Simpson's paradox (Simpson, 1951) shows that the very direction of an association can reverse when you look within subgroups instead of the total.

Reverse causation — the arrow points the other way

Whoever is more severely affected seeks more treatment; whoever sleeps poorly exercises less; whoever has more difficulty receives more hours of support. In each case the correlation exists, but it is the outcome that is producing the "cause." The question, again, is which is the cart and which is the horse. Data collected at a single point in time — a questionnaire applied once, for example — can almost never answer this, because it doesn't record what came before.

Coincidence — the spurious correlation

Correlations also arise by pure chance, and that's more common than it seems: the more pairs of variables someone compares, the more likely some of them are to line up well just by luck — this is the problem of multiple comparisons. A study that measures forty outcomes and reports the three that "turned out significant" is, in practice, showing off the coincidences and hiding the rest. There's also a particular case that is very misleading, through a different route: two time series that simply grow over the years — the consumption of anything and the diagnoses of anything else — correlate strongly only because both rise over time, with no connection between them at all.

Selection bias — the sample already bakes in the relationship

This is the subtlest of the five, and the one most likely to slip by unnoticed. The correlation can arise from the way the sample was assembled, rather than from anything linking the two variables in the world. If a service only sees patients who have both a diagnosis and a referral, studying only those patients can create an association that doesn't exist in the population — or erase one that does. The sample is not a neutral window: it is also part of the design, and whoever decides who enters the study decides, without meaning to, part of the result.

So what does establish a cause

Since it's never possible to observe the same person both with and without the treatment, every causal inference is a maneuver to build a good comparison. The three main ones:

  • Manipulate and randomize. In a randomized clinical trial, the researcher decides who gets what by a random draw. The random draw is what gives it its strength: it balances between groups, on average, not only the confounders someone thought to measure, but also the ones no one imagined. No statistical adjustment done afterward can achieve the same, because adjustment only reaches what was measured.
  • Manipulate and replicate within the same person. In single-case designs, the demonstration comes from continuously measuring behavior from a stable baseline and turning the independent variable on and off — or staggering it across several baselines — until the change tracks the manipulation repeatedly. The criterion is not a p-value: it is the functional relation demonstrated and replicated.
  • Triangulate, when manipulation is impossible or unethical. No one randomizes who is going to smoke, or who is going to suffer violence. Here, inference is built through triangulation: natural experiments, instrumental variables, difference-in-differences, propensity score matching, and Mendelian randomization — which uses genetic inheritance, randomized at conception, as if it were the study's coin toss. One example: Choi and colleagues (2019, JAMA Psychiatry) applied two-sample Mendelian randomization to more than 600,000 adults and found a signal of a protective effect of physical activity (measured by accelerometer) on the risk of depression, without evidence of the reverse path — something simple correlation between the two could never separate.

In observational territory, the classic reference for organizing judgment is the nine viewpoints that Austin Bradford Hill proposed in 1965, in a speech about how to draw causal conclusions from observational studies: strength of association, consistency across studies, specificity, temporality, dose-response gradient, plausibility, coherence with existing knowledge, experimental evidence, and analogy. (The case of cigarettes and lung cancer, which Hill investigated with Richard Doll, is the historical example tied to Hill's name.) They are remembered as the "Bradford Hill criteria," but Hill was explicit in rejecting that use: none of the nine, he warned, "can bring indisputable evidence for or against the hypothesis of cause and effect, and none can be required as sine qua non." It is not a checklist, and no number of checked items substitutes for the experiment.

verified

When the experiment proved the correlation wrongFor years, observational studies indicated that hormone therapy after menopause reduced the risk of coronary heart disease by 40% to 50%, and clinical practice followed that reading. In 2002, the randomized Women's Health Initiative trial, with 16,608 women, was stopped early: the arm receiving estrogen plus progestin showed an increase — not a reduction — in coronary heart disease, along with more strokes, pulmonary embolism, and invasive breast cancer. The most widely accepted explanation for the divergence is confounding: the women who used hormones were, to begin with, healthier and better cared for than those who didn't. The correlation was real; the causal reading was not.

When reading correlation as cause hurt people

The heaviest example of this confusion in the history of autism is the so-called "refrigerator mother." In describing his 1943 case series, Leo Kanner noted that many of those parents seemed distant and intellectualized to him. The causal reading — cold mothers would produce autistic children — was amplified by Bruno Bettelheim in The Empty Fortress (1967) and sustained decades of blaming families, with children removed from home in the name of treatment.

The competing explanations that were missing are among the ones in this entry. Reverse causation: raising a child who responds little to eye contact and hugs changes any adult's behavior — and the clinical interview took place with exhausted, frightened families. Confounding: autism is strongly heritable, so parents of autistic children have a higher chance, compared with parents in general, of showing similar traits (what is now called the broader autism phenotype); the "coldness" observed could have been the same inherited trait appearing in both parents and children. A meta-analysis of twin studies estimates the heritability of ASD at 64% to 91% (Tick and colleagues, 2016). The book Infantile Autism (1964), by Bernard Rimland, was a landmark in refuting the thesis.

warning

The same mistake, againThe claim that vaccines cause autism relied on a temporal coincidence: the MMR vaccine is given between 12 and 15 months of age, the age at which signs of autism usually become noticeable. Two things happening close together in time is an extremely weak indicator of cause — and, in this case, the study that gave rise to the idea was fraudulent and was retracted, while research with millions of children found no relationship at all. Families don't cause autism — and neither do vaccines.

"I got better after I did X"

Reports like this are sincere, and the improvement is usually real — what doesn't hold up is the bridge between one thing and the other. In any treatment, including the ones that work, the improvement observed after it has well-known competing explanations. Tap the cards:

That is why convincing testimonials pile up even around practices that don't work, and why evidence-based practice requires a controlled study before recommending something. See also the pseudosciences in mental health, which thrive on this mechanism, and the hierarchy of evidence, which organizes how much each type of study licenses you to conclude.

Correlation isn't the villain of the story

Saying that correlation doesn't imply causation isn't saying that correlation is good for nothing. It does three jobs very well:

  • Predict and screen. To flag who needs a more detailed evaluation, it is enough for the instrument to be associated with the outcome — it doesn't need to cause it. That's how screening questionnaires work: they point to where to look.
  • Evaluate instruments. A good part of psychometrics is correlation: agreement between observers, stability of a measure over time, the relationship of one test to another already-established one. See reliability and validity of measurement.
  • Raise hypotheses. Almost every experimental finding began as an association that someone decided to investigate. Correlation is a good starting point; the problem is treating it as the finish line.

In practice, when reading a news story or a study, four questions resolve most cases: was there manipulation, or only observation? Who decided who received what — the random draw or the circumstances? What came before what? And what third variable — or what entry criterion into the study — would explain both things at once? If the answer to the first is "only observation," the verb in the conclusion should be "is associated with," never "causes." A good part of the noise about mental health fits into that single swap of verb. See also internal and external validity.

confounding variable reverse causation selection bias spurious correlation regression to the mean Mendelian randomization

Test your understanding

quiz

Quick quiz

Three questions to lock in the concept. You'll see the explanation right away.

Further reading

  1. Hill, A. B. "The environment and disease: association or causation?". Proceedings of the Royal Society of Medicine, 58(5), 295–300, 1965 — presidential address to the Section of Occupational Medicine, where the nine viewpoints appear along with the warning that none of them is a sine qua non.
  2. Hernán, M. A. & Robins, J. M. Causal Inference: What If. Boca Raton: Chapman & Hall/CRC, 2020 — contemporary treatment of the counterfactual framework; text made freely available by the authors.
  3. Pearl, J. & Mackenzie, D. The Book of Why: The New Science of Cause and Effect. Basic Books, 2018 — accessible introduction to causal diagrams.
  4. Simpson, E. H. "The interpretation of interaction in contingency tables". Journal of the Royal Statistical Society, Series B, 13(2), 238–241, 1951 — origin of Simpson's paradox.
  5. Shadish, W. R.; Cook, T. D. & Campbell, D. T. Experimental and Quasi-Experimental Designs for Generalized Causal Inference. Houghton Mifflin, 2002 — competing explanations and quasi-experimental designs.
  6. Cooper, J. O.; Heron, T. E. & Heward, W. L. Applied Behavior Analysis. 3rd ed. Pearson, 2020 — functional relation and experimental demonstration of control in single-case research.
  7. Messerli, F. H. "Chocolate consumption, cognitive function, and Nobel laureates". New England Journal of Medicine, 367(16), 1562–1564, 2012. DOI 10.1056/NEJMon1211064 — r = 0.791 between chocolate consumption and Nobel laureates across 23 countries, published as a provocation about causal inference.
  8. Writing Group for the Women's Health Initiative Investigators. "Risks and benefits of estrogen plus progestin in healthy postmenopausal women: principal results from the Women's Health Initiative randomized controlled trial". JAMA, 288(3), 321–333, 2002. DOI 10.1001/jama.288.3.321.
  9. Tick, B.; Bolton, P.; Happé, F.; Rutter, M. & Rijsdijk, F. "Heritability of autism spectrum disorders: a meta-analysis of twin studies". Journal of Child Psychology and Psychiatry, 57(5), 585–595, 2016. DOI 10.1111/jcpp.12499 — heritability estimated at 64–91%.
  10. Choi, K. W. et al. "Assessment of bidirectional relationships between physical activity and depression among adults: a 2-sample Mendelian randomization study". JAMA Psychiatry, 76(4), 399–408, 2019. DOI 10.1001/jamapsychiatry.2018.4175.
info

This entry is educational and does not replace assessment, diagnosis or treatment by a qualified health professional. Every person is unique — for a specific case, seek individual guidance.

Looking for professional guidance?

Instituto Walden4 is a psychology and education center in Brasília, Brazil, offering evidence-based assessment and care for autism, psychotherapy and development — in person and online. Our team serves in Portuguese; get in touch and we will find the best way to help.

chat Message us on WhatsApp
chat

Conta Walden4

Entrar

Use o e-mail e a senha da sua conta.