How Do Researchers Separate Correlation From Causation?

Many findings about anxiety take the form “X is linked to Y.” People who sleep less report more anxiety; people who exercise more report less. Links like these are often real. The harder question is what produces them. Does X cause Y, does Y cause X, or does something else drive both? This article explains how researchers approach that question, and why the answer is often less certain than a headline suggests.

The question in plain terms

A correlation means two things tend to vary together, across people or over time. Causation means that changing one would change the other. A correlation is compatible with causation, but it is also compatible with several other explanations.

A hypothetical example

The following is an invented example for illustration, not a real finding.

Suppose a survey finds that people who report more evening phone use also report more anxiety. Several explanations fit that pattern:

• Phone use increases anxiety, perhaps through lost sleep, comparison with others or stressful news.
• Anxiety increases phone use. Anxious people might scroll to distract themselves, check messages more or search for reassurance. This is reverse causation.
• A third factor drives both. Loneliness, a stressful job or poor sleep for some other reason could raise both. This is confounding.
• The measurement creates the link. If the same people report both on the same day, a bad mood might colour both answers.
• Chance. Apparent links can appear by chance in a single sample, especially when many comparisons are made.

These explanations aren’t mutually exclusive. Several might operate at once, in both directions.

The real research on screens shows why size and analytic choices matter too. Orben and Przybylski analysed three large datasets covering 355,358 adolescents, testing a very wide range of reasonable ways to run the analysis rather than picking one. They found the association between digital technology use and well-being was negative but small, explaining at most 0.4% of the variation in well-being. They also noted that large datasets analysed flexibly can make small effects look statistically significant, producing false positives and conflicting results [1]. Their outcome was well-being rather than anxiety specifically, but the lesson carries: a link can be real, tiny and still not causal.

The tools researchers use

No single design settles causation on its own. Each approach rules out some alternatives and leaves others open.

Following people over time

Longitudinal studies measure the same people at several points. If one thing predicts later change in another, after accounting for where people started, that helps address reverse causation.

Sleep and anxiety show how this works. A systematic review of mostly longitudinal studies concluded that the best available evidence suggests insomnia and anxiety are related in both directions: poor sleep predicts later anxiety, and anxiety predicts later poor sleep. The authors cautioned that the small number of studies and their differences prevented firm conclusions for most kinds of sleep disturbance [2]. Longitudinal designs still can’t rule out confounding, because something unmeasured could drive both over time.

Statistical adjustment

Researchers can “control for” known confounders such as age, income or sleep. This helps, but only for factors that were measured, and measured well. Westfall and Yarkoni used simulations to show how badly this can go: when a confounder is measured with moderate reliability, standard regression methods produce very high false-positive rates, approaching 100% in some cases with large samples [3]. That is why “we adjusted for many factors” is not the same as “we ruled out other explanations.”

Randomised experiments

In a randomised controlled trial, participants are assigned by chance to different conditions. Because assignment is random, other factors tend to balance out between the groups on average, including factors nobody thought to measure. This is why randomised trials are widely regarded as the strongest single design for causal questions.

Returning to sleep: a 2021 meta-analysis gathered 65 randomised trials (72 interventions, 8,608 participants) in which an intervention improved sleep. Improving sleep led to a medium-sized reduction in anxiety (g = −0.51), and larger improvements in sleep went with larger improvements in mental health, a dose-response pattern [4]. Combined with the longitudinal evidence, this is a reasonably strong case that sleep affects anxiety, not just the other way round.

Trials have limits, too:

• Some exposures can’t ethically or practically be randomised. No one can be assigned a difficult childhood.
• People may not stick to the condition they were assigned.
• Expectations are hard to equalise. In drug trials, a placebo pill can look identical to the real thing. In psychological trials, people usually know what they’re getting. Boot and colleagues argued that even an “active” control group doesn’t rule out placebo effects unless both groups expect the same benefit, and that almost no psychological intervention studies check this [5].
• A trial in one group under particular conditions may not generalise to others.

Natural experiments, genes and families

Sometimes nature supplies something close to randomisation. Mendelian randomisation uses genetic variation as a natural experiment to test whether a modifiable risk factor causes a health outcome. Its conclusions depend on specific assumptions that careful studies test and report [6].

Twin and family studies help separate inherited from environmental contributions, which matters because family resemblance can masquerade as an environmental effect. A meta-analysis of family and twin studies estimated heritability at 0.43 for panic disorder and 0.32 for generalised anxiety disorder, with much of the remaining variation attributed to experiences not shared within families, although shared family environment may also matter for generalised anxiety in women [7]. So if anxious parents tend to have anxious children, that alone doesn’t show parenting caused it; shared genes are a competing explanation.

Converging evidence

In 1965 the statistician Austin Bradford Hill set out nine viewpoints for judging whether an association is likely to be causal, including the strength of the association, consistency across studies, the right order in time, a dose-response gradient, plausibility and experimental evidence. He was explicit that none of them could bring indisputable evidence for or against cause and effect, and none was strictly required [8]. The practical upshot is that agreement between methods with different weaknesses is more persuasive than any single study. A plausible pathway on its own isn’t enough, as we discuss in the difference between a plausible mechanism and a proven benefit.

Summary of the designs

Design	What it helps rule out	What it leaves open
Cross-sectional survey	Very little on its own	Reverse causation, confounding, measurement artefacts
Longitudinal study	Some reverse causation	Unmeasured confounding
Statistical adjustment	Measured confounders, if measured well	Unmeasured or poorly measured confounders
Randomised trial	Confounding, on average	Expectation effects, poor adherence, limited generalisability
Genetic and family designs	Some confounding by genes or environment	Violations of the design’s assumptions

Language to watch for

Words carry implicit claims about cause:

• “Associated with,” “linked to,” “correlated with” describe a relationship without claiming cause.
• “Predicts” usually means “comes before and is statistically related to,” not “causes.”
• “Reduces,” “triggers,” “leads to,” “improves” are causal claims, which should generally rest on experimental or strong converging evidence.

The slide from one to the other often happens before a study reaches you. An analysis of 462 press releases from 20 UK universities found 33% contained causal claims that went beyond the correlational research behind them. When a press release exaggerated in this way, 81% of the related news stories did too, compared with 18% when it didn’t [9]. A useful habit is to check whether the language in a headline matches the design of the study behind it.

What remains uncertain

Anxiety rarely has a single cause. Many relationships are probably two-way, as with sleep, and the balance may differ from person to person. Even well-designed studies usually estimate average effects across groups, which may not describe any one person’s experience. And the conclusion depends on what was counted as anxiety in the first place; see what an anxiety study actually measures.

Reading research about your own symptoms. A study showing that something is linked to anxiety on average can’t tell you what is driving yours. If poor sleep, low mood or anxiety are affecting your daily life, talk to a GP or mental health professional rather than making large changes on the strength of a headline, and don’t stop or change prescribed treatment without speaking to your prescriber. If you are in crisis or thinking about harming yourself, contact emergency services or a crisis line now: in the US, call or text 988; in the UK and Ireland, Samaritans on 116 123.