Does the Evidence Show a Cause?

Introduction

A pattern is not the same as a cause. Two things can rise and fall together because one causes the other, because the second causes the first, because both are driven by a third factor, because the data have been grouped in a misleading way, or because coincidence has been mistaken for meaning. Good analytical thinking therefore asks a sharper question than “Are these things linked?” It asks: “What mechanism would make this link causal, and what rival explanations would also fit the evidence?”

This matters because causal claims appear everywhere: a headline says a food “reduces risk”, a manager says a new policy “improved productivity”, a business dashboard shows that customers who use one feature spend more, or a person concludes that a habit changed their mood. Some of those claims may be right. The error is not noticing a correlation; it is stopping there. Modern causal inference treats causal conclusions as claims about what would happen under an intervention, not merely as descriptions of what happened alongside what. Observational evidence can be useful, but it is vulnerable to confounding, selection bias and measurement problems unless those rival explanations are actively tested.[PMC]pmc.ncbi.nlm.nih.govby G Hammerton · 2021 · Cited by 394 — The goal of much observational research is to identify risk factors that have a causal effect o…

Patterns versus causes

A correlation says that two variables move together in some way. A causal claim says that changing one would change the other, through some mechanism. The difference is small in wording and large in consequence. “People who sleep very little have higher mortality” is a pattern. “Sleeping too little causes higher mortality” is a stronger claim. To accept the second, the reader needs more than a graph: they need a plausible pathway, careful comparison groups, and reasons to think the association is not mainly due to illness, stress, work patterns, age, socioeconomic factors, or measurement error.

The classic warning is often summarised as “correlation does not imply causation”, but the phrase is sometimes misunderstood. It does not mean correlations are useless. Correlations are often the first clue that something deserves investigation. The point is that correlation alone is not enough to identify direction, mechanism or intervention value. A correlation may suggest a causal hypothesis; it does not by itself prove one.[Wikipedia]WikipediaCorrelation does not imply causationCorrelation does not imply causation

A simple example is the relationship between ice-cream sales and drownings. In warmer periods, both may increase, but it would be a mistake to infer that buying ice cream causes drowning. A better explanation is that hot weather increases both swimming and ice-cream consumption. In that case, temperature is a common cause: it creates a relationship between two variables that are not directly causing each other. Forecasting models may still use ice-cream sales to predict drownings, but prediction is not the same as explanation.[OTexts: Online, open-access textbooks]otexts.comOnline, open-access textbooks7.8 Correlation, causation and forecastingThe model can give reasonable forecasts, not because ice-c…

This distinction is especially important in dashboards and everyday decision-making. A business may observe that customers who attend onboarding sessions spend more money later. That could mean onboarding increases spending. It could also mean that already-motivated customers are more likely to attend onboarding and more likely to spend. A school may find that pupils using a revision app get higher marks. That could mean the app helps, or that conscientious pupils both use the app and study more in other ways. In each case, the practical question is not “Is there an association?” but “Would changing the suspected cause actually change the outcome?”

Rival explanations

A good causal thinker does not merely object that “there could be other factors”. They name the rival explanations clearly enough that they can be checked. The most common alternatives are direction reversal, confounding, selection effects, measurement problems, chance, and misleading aggregation.

Reverse causation occurs when the apparent effect may actually be the cause. A person might say that social isolation causes poor health, but poor health can also make people more isolated. A company might say that using an advanced feature causes customers to stay subscribed, but loyal customers may be the ones most willing to learn the advanced feature. The proposed arrow can point the other way, or both ways at once.

Confounding occurs when a third factor influences both the proposed cause and the outcome. In observational research, confounders can distort the estimated association so that the apparent effect is too large, too small, or even in the wrong direction. Health, income, education, baseline motivation and prior risk are common confounders because they influence many later outcomes and many exposures people choose or experience.[PMC]pmc.ncbi.nlm.nih.govby G Hammerton · 2021 · Cited by 394 — The goal of much observational research is to identify risk factors that have a causal effect o…

Selection effects occur when the people or cases being compared are not comparable because of how they entered the dataset. Suppose a workplace compares productivity among employees who voluntarily use a new tool with those who do not. The users may differ before the tool is introduced: they may be more engaged, more comfortable with technology, or working in teams where managers encourage experimentation. The difference in outcomes may partly reflect who selected into the group, not what the tool did.

Measurement problems occur when the variables do not capture what the claim says they capture. A study may measure “screen time” by self-report, “wellbeing” with a short survey, or “performance” through a narrow metric that employees can game. Even if the numbers correlate, the causal story may be about the measurement process rather than the real-world concept the reader cares about.

Chance and multiple comparisons matter because people often inspect many possible relationships and remember the striking ones. In a large enough dataset, some variables will appear to move together by accident. This does not make statistics useless, but it does mean that surprising patterns need replication, pre-specified tests, or independent confirmation before they should be treated as explanations.

Misleading aggregation can make a pattern appear at one level and disappear or reverse at another. Simpson’s paradox is the famous case: a relationship in a combined dataset can reverse when the data are split into meaningful subgroups. The University of California, Berkeley graduate admissions example is widely used because the overall admission rate appeared lower for women, yet department-level analysis showed that application patterns across more and less competitive departments changed the interpretation.[Philosophy at HKU]philosophy.hku.hkOpen source on hku.hk.

The mechanism test

The word “mechanism” is the bridge between a pattern and a plausible cause. A mechanism explains how the cause would produce the effect. Without it, a causal claim is often just a decorated correlation.

A useful mechanism does three things. First, it identifies the pathway: what happens after the cause and before the outcome? Second, it makes timing matter: the cause should come before the effect, and the relevant intermediate changes should appear in the expected order. Third, it creates testable expectations: if the mechanism is real, certain related patterns should appear, and others should not.

Take the claim that a new meeting format improves project delivery. The mechanism might be that shorter meetings force clearer ownership, which reduces duplicated work, which speeds delivery. That is much stronger than simply saying “teams using the format delivered faster”. It suggests what to inspect: Were responsibilities clearer? Did duplicated work fall? Did the effect appear first in teams that previously suffered from unclear ownership? Did teams that adopted the format only in name show the same improvement?

In epidemiology, Austin Bradford Hill’s 1965 discussion of association and causation remains influential because it treated causality as a judgement built from several considerations rather than one magic test. His famous viewpoints include strength of association, consistency, temporality, biological gradient, plausibility, coherence, experiment and analogy. Later researchers have warned against treating these as a mechanical checklist, but the core habit remains valuable: a causal claim becomes stronger when different forms of evidence point towards the same mechanism and weaker when the association has no plausible route from cause to effect.[PMC]pmc.ncbi.nlm.nih.govOpen source on nih.gov.

Mechanisms also protect against overfitting stories after the fact. Humans are good at inventing explanations for patterns once they have seen them. The discipline is to ask whether the mechanism was predicted in advance, whether it explains more than one pattern, and whether it survives comparison with simpler alternatives. A persuasive story is not the same as a tested mechanism.

What stronger causal evidence looks like

The strongest causal evidence usually comes from a design that makes the comparison fair. Randomised controlled trials are powerful because random assignment aims to make treatment and control groups comparable, including on unknown or unmeasured confounders. Randomisation does not make a study perfect, but when well implemented it reduces selection bias and gives a clearer basis for attributing outcome differences to the intervention.[PMC+2BMJ Global Health]pmc.ncbi.nlm.nih.govOpen source on nih.gov.

The Oregon Health Insurance Experiment is a useful concrete example. Oregon used a lottery to allocate the chance to apply for Medicaid coverage among low-income uninsured adults, creating a rare opportunity to study the effects of insurance access with a randomised design rather than only comparing people who did and did not have insurance. The design mattered because health insurance status is normally entangled with income, employment, health needs and other factors that make ordinary observational comparisons difficult.[PMC]pmc.ncbi.nlm.nih.govPMCTHE OREGON HEALTH INSURANCE EXPERIMENTPMCTHE OREGON HEALTH INSURANCE EXPERIMENT

Experiments are not always possible, ethical or sufficient. You cannot randomly assign people to smoke for decades, live in poverty, suffer sleep deprivation, or experience social isolation. In those cases, researchers use observational designs, natural experiments, longitudinal data, instrumental variables, regression discontinuities, matching, difference-in-differences, directed acyclic graphs and sensitivity analyses. These tools do not remove the need for judgement; they make the assumptions visible and testable.[PMC+2PMC]pmc.ncbi.nlm.nih.govby G Hammerton · 2021 · Cited by 394 — The goal of much observational research is to identify risk factors that have a causal effect o…

A directed acyclic graph, often shortened to DAG, is one practical way to make a causal claim clearer. It draws variables as nodes and proposed causal influences as arrows. The point is not to make a pretty diagram. The point is to force the analyst to state which variables are causes, consequences, confounders, mediators or colliders. That matters because adjusting for the wrong variable can create bias rather than remove it. For example, controlling for a mediator can hide part of the effect you are trying to measure, while conditioning on a collider can create a misleading association between variables that would otherwise be unrelated.[Biostatistics.ca+2Queen's University Belfast]biostatistics.caOpen source on biostatistics.ca.

For everyday reasoning, the same principle can be used without formal statistics. Before accepting a causal claim, sketch the story:

This turns vague scepticism into useful analysis.

When a correlation is still useful

Rejecting weak causal claims does not mean ignoring patterns. Sometimes a correlation is enough for prediction, triage or hypothesis generation even when causality remains uncertain. A hospital may use correlated warning signs to identify patients at risk. A business may use behaviour patterns to forecast churn. A weather-related correlation may help a beach resort plan staffing even if the measured predictor is not itself the cause.

The danger is using predictive patterns as if they were intervention advice. If ice-cream sales predict drownings, banning ice cream would not solve the problem. If customers who use a feature are more loyal, forcing every customer to use it may not improve retention if the feature is mainly a marker of pre-existing commitment. If long sleep is associated with poor health outcomes, simply telling everyone to sleep less would be reckless if long sleep is partly a symptom of illness rather than a cause of it.

Sleep research illustrates the distinction. Reviews and cohort studies have repeatedly found that both short and long sleep duration can be associated with higher mortality risk, often producing a U-shaped pattern. That pattern is important, but causal interpretation is difficult because sleep duration is linked with health status, depression, work schedules, medication, age, chronic disease and other factors. The association can inform further research and risk assessment, but it should not be flattened into a simple claim that every extra hour of sleep directly causes harm.[PMC+2Frontiers]pmc.ncbi.nlm.nih.govPMCSleep duration and all-cause mortality: a critical reviewPMCSleep duration and all-cause mortality: a critical review

The analytical skill is to match the claim to the evidence. A correlation may justify “this is worth investigating”, “this can help predict”, or “this group appears at higher risk”. It does not automatically justify “changing this variable will produce that outcome”.

What would change the conclusion

Better thinking improves when people decide in advance what evidence would alter their view. This prevents causal claims from becoming unfalsifiable stories that absorb every result.

For a causal claim, the most useful question is: “What would I expect to observe if this cause really matters, and what would I expect if a rival explanation is true?” The answer depends on the case, but several tests are widely useful.

Timing should fit the causal story. The proposed cause should precede the outcome. If productivity improved before the new policy began, the policy cannot explain the original improvement. If symptoms began before a supposed trigger, the trigger may still matter later, but it cannot be the initial cause.

The effect should change when the cause changes. If a company rolls out a training programme at different times across teams, and performance improves only after each team receives training, that supports the causal claim more than a single before-and-after comparison. If performance improves equally in teams that did not receive training, the claim weakens.

The pattern should survive fairer comparisons. If a diet appears linked with better health, the claim is stronger if the association remains after comparing people with similar age, baseline health, income, exercise and smoking status. It is weaker if the association disappears once those factors are considered.

The mechanism should leave traces. If a reading intervention is said to improve comprehension by increasing vocabulary, vocabulary should improve before or alongside comprehension. If it does not, the intervention may still work, but the proposed mechanism is doubtful.

A rival explanation should be actively tested, not merely mentioned. It is easy to say “maybe motivation explains it”. The better step is to ask how motivation could be measured or approximated, whether the pattern appears among people with similar baseline motivation, or whether a design can reduce self-selection.

The finding should replicate where the mechanism should apply. A causal claim becomes stronger when it appears across settings where the same mechanism should operate and weaker when it appears only in one noisy dataset. Hill’s emphasis on consistency and coherence remains useful here, provided it is used as judgement rather than as a rigid checklist.[PMC]pmc.ncbi.nlm.nih.govOpen source on nih.gov.

Everyday checks before believing a causal claim

Most readers do not need to run a causal model every time they see a claim. They need a compact habit that slows down overconfident inference. The following checks are practical because they apply to health headlines, workplace reports, product analytics, personal decisions and media claims.

Separate the observation from the explanation.

Observation: “People who do X have more Y.” Explanation: “X causes Y.” Do not let the second hide inside the first.

Ask what the comparison group is.

A causal claim is only as good as its comparison. Compared with whom? At what starting point? Under what conditions? A before-and-after comparison may be weak if many other things changed at the same time.

Look for common causes.[biostatistics.ca]biostatistics.caSource details in endnotes.

Ask what could plausibly influence both variables. In human behaviour, common causes often include age, money, education, prior ability, motivation, health, geography, season, access and institutional rules.

Check whether the arrow could run backwards.

If people with better outcomes are more likely to adopt a behaviour, the behaviour may be a consequence or marker of success rather than its cause.

Beware of “controlled for everything important”.

Statistical adjustment only helps when the right variables are measured well and included for the right reason. Unmeasured confounding remains a central problem in observational causal inference.[PMC]pmc.ncbi.nlm.nih.govby G Hammerton · 2021 · Cited by 394 — The goal of much observational research is to identify risk factors that have a causal effect o…

Distinguish prediction from intervention.[pmc.ncbi.nlm.nih.gov]pmc.ncbi.nlm.nih.govSource details in endnotes.

A variable can be a useful warning sign without being a useful lever. Grey hair predicts age better than many things, but dyeing hair does not make a person younger.

Ask what result would embarrass the claim.

If no possible evidence would make the claimant revise the explanation, the claim is not being treated analytically.

The better habit

The goal is not to become reflexively dismissive. “Correlation is not causation” can itself become lazy if it is used to wave away evidence without thinking. The better habit is more precise: treat correlation as an invitation to compare explanations.

A strong causal judgement usually has three features. It identifies a credible mechanism, rules out the most plausible rivals, and explains what evidence would change the conclusion. Sometimes the result will be confidence: an intervention has been tested well enough to act on. Sometimes it will be caution: the pattern is real but the cause is uncertain. Sometimes it will be a better question: not “Does X correlate with Y?” but “Under what conditions would changing X change Y, for whom, and through what pathway?”

That shift is one of the most useful upgrades in analytical thinking. It turns evidence from a pile of interesting patterns into a disciplined search for the best explanation.

Amazon book picks

Further Reading

Books and field guides related to Does the Evidence Show a Cause?. Use these as the next step if you want deeper reading beyond the article.

Book

The Book of Why

By Judea Pearl, Dana Mackenzie

Directly explains the difference between correlation and causation and how causal claims are established.

See on Amazon

Book

The Signal and the Noise

By Nate Silver

Explores interpreting data, uncertainty, and distinguishing meaningful relationships from noise.

See on Amazon

Book

How to Lie with Statistics

By Darrell Huff

Helps readers recognize misleading patterns, weak evidence, and flawed statistical reasoning.

See on Amazon

Book

Calling Bullshit

By Carl T. Bergstrom, Jevin Darwin West

Covers misleading statistical claims, causal errors, and evaluating evidence.

See on Amazon

Browse more on Amazon: The Book of Why The Signal and the Noise How to Lie with Statistics

As an Amazon Associate I earn from qualifying purchases.

eBay marketplace picks

Marketplace Samples

Example marketplace items related to this page. Use the search link to explore similar finds on eBay.

Shop location

UsingUSA USA USAUKCanadaAustraliaIreland

Example eBay listing

SUPER COOL SCIENCE KIT - Over 55 Experiments + Scientist Name tag! STEM Learning

SearcheBay.co.uk: science education kit

Browse similar oneBay.co.uk

Example eBay listing

MEL Chemistry Kits Science for Kids 4 PACK Age 10 plus Factory Sealed Learning

SearcheBay.co.uk: science education kit

Browse similar oneBay.co.uk

Example eBay listing

Science Kit for Kids + Lab Coat - Over 20 Chemistry Science Experiments Age 5+

SearcheBay.co.uk: science education kit

Browse similar oneBay.co.uk

Example eBay listing

DIY STEM Earth Model Day Night Cycle Wooden Globe Toy Science Kit Kids Education

SearcheBay.co.uk: science education kit

Browse similar oneBay.co.uk

Browse more oneBay.co.uk

Example items shown for inspiration; availability and pricing can change. Branchoria may earn a commission if you purchase through outbound eBay links.

Endnotes

1. Source: pmc.ncbi.nlm.nih.gov
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC8020490/

2. Source: Wikipedia
Title: Correlation does not imply causation
Link:https://en.wikipedia.org/wiki/Correlation_does_not_imply_causation

3. Source: Wikipedia
Link:https://en.wikipedia.org/wiki/Correlation

4. Source: otexts.com
Link:https://otexts.com/fpp3/causality.html

5. Source: Wikipedia
Title: Spurious relationship
Link:https://en.wikipedia.org/wiki/Spurious_relationship

6. Source: philosophy.hku.hk
Link:https://philosophy.hku.hk/think/stat/simpson.php

7. Source: pmc.ncbi.nlm.nih.gov
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC1898525/

8. Source: pmc.ncbi.nlm.nih.gov
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC8206235/

9. Source: pmc.ncbi.nlm.nih.gov
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC6235704/

10. Source: gh.bmj.com
Link:https://gh.bmj.com/content/7/6/e008597

11. Source: biostatistics.ca
Link:https://www.biostatistics.ca/common-dag-structures-confounding-collider-bias-and-mediation/

12. Source: pmc.ncbi.nlm.nih.gov
Title: PMCSleep duration and all-cause mortality: a critical review
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC3660511/

13. Source: Wikipedia
Link:https://en.wikipedia.org/wiki/Bradford

14. Source: Wikipedia
Title: Bradford Hill criteria
Link:https://en.wikipedia.org/wiki/Bradford_Hill_criteria

15. Source: Wikipedia
Title: Randomized controlled trial
Link:https://en.wikipedia.org/wiki/Randomized_controlled_trial

16. Source: Wikipedia
Title: RAND Health Insurance Experiment
Link:https://en.wikipedia.org/wiki/RAND_Health_Insurance_Experiment

17. Source: blogs.bmj.com
Title: jeff aronson when i use a word austin bradford hills viewpoints
Link:https://blogs.bmj.com/bmj/2018/03/29/jeff-aronson-when-i-use-a-word-austin-bradford-hills-viewpoints/

18. Source: mathcircle.berkeley.edu
Title: Simpson Paradox Lisa Goldberg BMC Dec 15 2019 0
Link:https://mathcircle.berkeley.edu/sites/default/files/handouts/2019/Simpson%20Paradox%20-%20Lisa%20Goldberg%20BMC%20Dec%2015%202019_0.pdf

19. Source: iris.who.int
Link:https://iris.who.int/bitstreams/5e1eadc6-7016-4829-be3a-77c593271129/download

20. Source: pmc.ncbi.nlm.nih.gov
Title: PMCTHE OREGON HEALTH INSURANCE EXPERIMENT
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC3535298/

21. Source: pmc.ncbi.nlm.nih.gov
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC12674023/

22. Source: qub.ac.uk
Title: Queen’s University Belfast
Link:https://www.qub.ac.uk/sites/space/Filestore/Filetoupload%2C1844821%2Cen.pdf

23. Source: frontiersin.org
Link:https://www.frontiersin.org/journals/public-health/articles/10.3389/fpubh.2022.880276/full

24. Source: pmc.ncbi.nlm.nih.gov
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC10010939/

25. Source: ncbi.nlm.nih.gov
Link:https://www.ncbi.nlm.nih.gov/books/NBK606119/

26. Source: pmc.ncbi.nlm.nih.gov
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC3755849/

27. Source: pmc.ncbi.nlm.nih.gov
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC11149352/

28. Source: pmc.ncbi.nlm.nih.gov
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC7081045/

29. Source: pmc.ncbi.nlm.nih.gov
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC4589117/

30. Source: pmc.ncbi.nlm.nih.gov
Link:https://pmc.ncbi.nlm.nih.gov/articles/PMC4291331/

31. Source: abs.gov.au
Title: correlation and causation
Link:https://www.abs.gov.au/statistics/understanding-statistics/statistical-terms-and-concepts/correlation-and-causation

Additional References

32. Source: mixtape.scunning.com
Link:https://mixtape.scunning.com/03-directed_acyclical_graphs

33. Source: youtube.com
Title: Correlation vs Causation: Why This Mistake Destroys School Improvement Efforts
Link:https://www.youtube.com/watch?v=HqPli7mifio

34. Source: youtube.com
Title: Correlation vs. Causation: Don’t Confuse Relationship with Cause!
Link:https://www.youtube.com/watch?v=QuafiIpGqPU

35. Source: youtube.com
Title: Correlation Doesn’t Equal Causation: Crash Course Statistics #8
Link:https://www.youtube.com/watch?v=8B271L3NtAw

36. Source: cdc.gov
Link:https://www.cdc.gov/field-epi-manual/php/chapters/analyze-interpret-data.html

37. Source: youtube.com
Title: Correlation vs Causation: A Brief Guide To Communicating Research
Link:https://www.youtube.com/watch?v=7bT17r_yIrw

38. Source: youtube.com
Title: Causal Inference: Potential Outcomes and Confounders
Link:https://www.youtube.com/watch?v=QnkD6b7Czng

39. Source: himsr.co.in
Link:https://www.himsr.co.in/wp-content/uploads/2020/03/Association-Causation_Dr-Farzana.pdf

40. Source: healthknowledge.org.uk
Link:https://www.healthknowledge.org.uk/public-health-textbook/research-methods/1a-epidemiology/association-causation

41. Source: scribd.com
Link:https://www.scribd.com/presentation/329381127/Association-Causation