Reading the evidence
Correlation vs Causation: How to Read an Observational Study
AR
· 10 min read
A summary of published research, with citations. It states no amount, schedule or route, and makes no claim about what this compound does in a person. Supplied for laboratory research use only.
On this page · 9 sections
- Why doesn't correlation prove causation?
- What is confounding?
- Thyroid cancer: when observational designs disagree
- What is reverse causation?
- What is the healthy-user effect?
- Can observational studies ever show causation?
- What is Mendelian randomisation, and what are its limits?
- A checklist for any observational finding
- Frequently asked questions
Correlation means two things move together; causation means changing one changes the other. An observational study can show the first and only suggest the second, because the people who took a drug differ from those who did not in ways nobody fully measured. Before accepting a causal reading, rule out four alternatives: chance, confounding, reverse causation and bias in who was selected or measured. Randomisation removes most of them at once, which is why randomised trials and observational studies so often disagree.

Why doesn't correlation prove causation?
An association between an exposure and an outcome has several possible explanations, and only one of them is that the exposure causes the outcome.
| Explanation | What is going on | Example discussed below |
|---|---|---|
| Chance | A pattern in a sample that is not in the population | Small cohorts with a handful of events |
| Confounding | A third factor drives both exposure and outcome | Why a drug is prescribed, and to whom |
| Reverse causation | The outcome, or its early stages, causes the exposure | Disease shortening telomeres |
| Healthy-user bias | People who take a preventive treatment are healthier in other ways | Hormone therapy and heart disease |
| Measurement or selection bias | Outcomes found more often in one group because it is examined more | Specialist-clinic cohorts |
| Causation | The exposure changes the outcome | What the study hopes to show |
In 1965 Austin Bradford Hill set out the considerations epidemiologists still use when weighing a causal reading. They include strength, consistency, specificity, temporality, a dose-response gradient, plausibility, coherence, experiment and analogy. Only temporality, cause before effect, is strictly necessary. None of the others is sufficient alone.
What is confounding?
A confounder is associated with both the exposure and the outcome without being on the causal path between them. The most stubborn form in drug research is confounding by indication: the reason a person is prescribed a drug is itself linked to the outcome.
Semaglutide and non-arteritic anterior ischaemic optic neuropathy (NAION), a sudden loss of blood supply to the optic nerve, show the problem and how better designs respond to it.
- The first signal. Hathaway and colleagues studied patients seen by neuro-ophthalmologists at one academic centre. Among people with type 2 diabetes, 17 NAION events occurred in those prescribed semaglutide against 6 in those on other diabetes drugs: a hazard ratio of 4.28 (95% CI 1.62 to 11.29). In people with overweight or obesity it was 7.64 (2.21 to 26.36). The authors stated plainly that an observational study cannot establish causality.
- A larger, better-controlled test. Tesfaye and colleagues used insurance claims from three databases to build 482,912 matched pairs of adults with type 2 diabetes starting either a GLP-1 receptor agonist or an SGLT2 inhibitor. That is an active comparator, so both groups had a reason to start a new diabetes drug. They used a target trial emulation framework. The hazard ratio was 1.85 (1.51 to 2.27), and the absolute difference was 0.29 extra cases per 1,000 person-years.
The association survived a stronger design, which is informative. But it shrank, and its absolute size turned out small, which is what happens when a first estimate comes from a specialist clinic with few events. A patient sent to a neuro-ophthalmologist is not a random patient. Even the large study remains observational; residual confounding by diabetes severity, sleep apnoea or other factors cannot be excluded by matching on what was recorded. How to turn hazard ratios into absolute numbers is covered in relative vs absolute risk, and the trial-based tirzepatide safety record in tirzepatide side effects in the trials.
Thyroid cancer: when observational designs disagree
The incretin drugs and thyroid cancer show what happens when designs differ.
- A French nested case-control study by Bezin and colleagues found that one to three years of GLP-1 receptor agonist use was associated with more thyroid cancer: adjusted hazard ratio 1.58 (1.27 to 1.95).
- A Scandinavian cohort study by Pasternak and colleagues compared new users of GLP-1 receptor agonists with new users of DPP-4 inhibitors. It found 76 cancers among 145,410 patients against 184 among 291,667. The hazard ratio was 0.93 (0.66 to 1.31), and the upper limit was consistent with no more than a 31% relative increase.
The two designs answer slightly different questions, and they are open to different biases. One general mechanism worth knowing is detection bias: if one group is examined more closely, more small, slow cancers are found in it whether or not the drug did anything. Comparing new users of one drug with new users of another, as Pasternak did, is designed to make the two groups' starting points and follow-up as alike as possible. Randomised data have not settled it either; a meta-analysis of trials, discussed in tirzepatide side effects, had too few events to rule an increase in or out.
What is reverse causation?
Sometimes the outcome, or the process leading to it, causes the exposure. Telomeres are the clearest case on this site. Shorter telomeres are associated with many age-related diseases, which has been read as "short telomeres cause disease, so lengthening them should help". That reading underpins much writing about Epitalon and telomerase. But inflammation, oxidative stress, smoking and disease itself shorten telomeres. A sick person may have short telomeres because they are sick.
What is the healthy-user effect?
People who start and stick with a preventive treatment tend to differ from those who do not. They are more likely to exercise, attend screening, avoid smoking and have the resources to do all of these. Shrank and colleagues described this healthy-user effect, alongside a healthy-adherer effect and confounding by functional status and by selective prescribing. They noted that numerous highly publicised observational studies of prevention were later contradicted by randomised trials.
Menopausal hormone therapy is the standard example.
| Study | Design | Coronary heart disease result |
|---|---|---|
| Nurses' Health Study, 10-year follow-up (Stampfer and colleagues) | Cohort of 48,470 women | Current estrogen users: relative risk 0.56 (0.40 to 0.80) |
| Women's Health Initiative (Rossouw and colleagues) | Randomised trial, 16,608 women, estrogen plus progestin vs placebo | Hazard ratio 1.29 (1.02 to 1.63) |
The cohort analysis adjusted for age and risk factors. It even repeated the analysis in a low-risk group and in women who had recently seen a physician, and the apparent protection held. The randomised trial, stopped early, found more coronary events, strokes and pulmonary emboli on treatment. The two studies used different preparations and populations, which explains part of the gap. The broader lesson, that adjustment for measured factors could not remove the advantage of being the kind of woman who took hormones, is why the episode is taught.
Can observational studies ever show causation?
They can come close, under conditions. Strong associations, consistency across populations and designs, a clear time order, a dose-response gradient and an active comparator all help. Target trial emulation, used by Tesfaye, forces a study to state the randomised trial it is imitating: who is eligible, when follow-up starts, what the comparison is. That avoids several self-inflicted biases.
Some questions can only be answered observationally, because a trial would be unethical, too large or too long. Rare harms are the main example, which is why pharmacovigilance exists; see case reports and adverse-event signals.
What is Mendelian randomisation, and what are its limits?
Mendelian randomisation uses genetic variants that influence an exposure as a natural experiment. Genes are allocated at conception, roughly at random with respect to lifestyle, and cannot be changed by later disease. So if variants that lengthen telomeres are associated with an outcome, confounding and reverse causation are far less likely explanations.
Haycock and colleagues did exactly that, using genome-wide association data for 35 cancers and 48 other diseases. Genetically longer telomeres were generally associated with more cancer, not less. Odds ratios per standard deviation included 5.27 for glioma, 3.19 for lung adenocarcinoma and 1.87 for melanoma. They were associated with less coronary heart disease (0.78). That is the opposite of what the simple observational correlation suggested for cancer. It is the reason Epitalon and cancer risk treats "longer telomeres" as a question rather than a goal. Epitalon is approved nowhere; its evidence is cell culture, animal work and small human studies.
Mendelian randomisation is a partial answer, not a final one. Davies, Holmes and Davey Smith set out its assumptions for clinical readers. The variants must genuinely affect the exposure. They must not affect the outcome through another route, which is called pleiotropy. And they must not share ancestry-related confounders with the outcome. A lifelong, modest genetic difference is also not the same as a drug or peptide acting for months in adulthood.
A checklist for any observational finding
- What was the comparison group, and did it have a similar reason to be treated?
- Was follow-up started at the same point in both groups, ideally at treatment start?
- How many events drove the estimate, and how wide is the interval?
- Could the outcome, or its early signs, have caused the exposure?
- Is there an absolute risk, not just a ratio?
- Has the finding been tested in a different design, or by randomisation?
Statistical uncertainty is a separate issue from bias; see p value vs confidence interval. For laboratory evidence, the equivalent question is whether a cell result reaches a whole organism, covered in in vitro vs in vivo. How pooled observational data are judged is in systematic review vs meta-analysis. Nothing here is medical, dosing, or therapeutic guidance.
Frequently asked questions
Why don't researchers just adjust for confounders? They do, but adjustment only handles factors that were measured, measured accurately and modelled correctly. Unmeasured differences, such as how health-conscious a person is or why a doctor chose one drug over another, remain. The hormone-therapy cohort adjusted carefully and still pointed the wrong way for coronary disease.
Is a large observational study more reliable than a small trial? Larger size reduces random error, not bias. A study of half a million people can be precisely wrong if the groups differ systematically. A small randomised trial is unbiased in expectation but imprecise. The two answer different weaknesses, which is why agreement between them is more persuasive than either alone.
What does "association" mean in a paper's conclusion? It is the careful word for a correlation that the authors are not claiming as causal. When an abstract reports an association and a headline reports that a drug "causes" or "prevents" something, the headline has added a claim the study did not make.
Does Mendelian randomisation prove causation? No. It removes some biases that affect conventional observational studies, especially reverse causation and much confounding. It relies on assumptions that cannot all be tested, particularly that the genetic variants affect the outcome only through the exposure studied. Its results are strongest when they agree with other kinds of evidence.
Selected Research References
- Hill AB. The environment and disease: association or causation? Proceedings of the Royal Society of Medicine. 1965;58(5):295-300. pubmed.ncbi.nlm.nih.gov/14283879
- Hathaway JT, Shah MP, Hathaway DB, et al. Risk of Nonarteritic Anterior Ischemic Optic Neuropathy in Patients Prescribed Semaglutide. JAMA Ophthalmology. 2024;142(8):732-739. pubmed.ncbi.nlm.nih.gov/38958939
- Tesfaye H, Paik JM, Wexler DJ, et al. GLP-1RA and the risk of non-arteritic anterior ischaemic optic neuropathy in patients with type 2 diabetes: A population-based study. Diabetes, Obesity and Metabolism. 2026;28(2):1517-1528. pubmed.ncbi.nlm.nih.gov/41104517
- Bezin J, Gouverneur A, Pénichon M, et al. GLP-1 Receptor Agonists and the Risk of Thyroid Cancer. Diabetes Care. 2023;46(2):384-390. pubmed.ncbi.nlm.nih.gov/36356111
- Pasternak B, Wintzell V, Hviid A, et al. Glucagon-like peptide 1 receptor agonist use and risk of thyroid cancer: Scandinavian cohort study. BMJ. 2024;385:e078225. pubmed.ncbi.nlm.nih.gov/38683947
- Shrank WH, Patrick AR, Brookhart MA. Healthy user and related biases in observational studies of preventive interventions: a primer for physicians. Journal of General Internal Medicine. 2011;26(5):546-550. pubmed.ncbi.nlm.nih.gov/21203857
- Stampfer MJ, Colditz GA, Willett WC, et al. Postmenopausal estrogen therapy and cardiovascular disease. Ten-year follow-up from the nurses' health study. New England Journal of Medicine. 1991;325(11):756-762. pubmed.ncbi.nlm.nih.gov/1870648
- Rossouw JE, Anderson GL, Prentice RL, et al. Risks and benefits of estrogen plus progestin in healthy postmenopausal women: principal results From the Women's Health Initiative randomized controlled trial. JAMA. 2002;288(3):321-333. pubmed.ncbi.nlm.nih.gov/12117397
- Haycock PC, Burgess S, Nounu A, et al. Association Between Telomere Length and Risk of Cancer and Non-Neoplastic Diseases: A Mendelian Randomization Study. JAMA Oncology. 2017;3(5):636-651. pubmed.ncbi.nlm.nih.gov/28241208
- Davies NM, Holmes MV, Davey Smith G. Reading Mendelian randomisation studies: a guide, glossary, and checklist for clinicians. BMJ. 2018;362:k601. pubmed.ncbi.nlm.nih.gov/30002074
All products sold by NovaEvum are for research and laboratory use only. They are not intended for human or animal consumption, diagnosis, treatment, cure, or prevention of any disease.
