PERIOMED ORIGINAL 0414 OCTOBER 2026

CLINICAL RESEARCH · EPIDEMIOLOGY · CAUSAL INFERENCE

Association and Causality in Clinical and Epidemiological Research.

How evidence moves from an observed relationship to a defensible causal interpretation.

FIRST AUTHOR · MEDICAL SCIENCESThami Ouazzani-Touhami, MDMorocco

SECOND AUTHOR · PUBLIC HEALTH & EDUCATIONJuhaina Aboaghlah, PhD, MPHSaudi Arabia

SCIENTIFIC REVIEWERHamail FazilMedical Sciences · Pakistan

ABSTRACT

An association is an observation.
Causation is an explanation.

Distinguishing association from causation is fundamental to clinical research, epidemiology and evidence-based healthcare. Although observational studies frequently identify relationships between exposures and health outcomes, these associations do not necessarily establish causal effects. Chance, confounding, selection bias, measurement error and reverse causation may influence observed findings.

This review examines causal inference, study designs, Bradford Hill's viewpoints, counterfactual reasoning, causal diagrams and triangulation. Historical examples show how causal interpretations may be strengthened or overturned. Particular attention is given to periodontitis and cardiovascular disease, where epidemiological associations must be distinguished from causal evidence.

Reliable causal interpretation requires clearly defined questions, appropriate designs, transparent assumptions, rigorous analysis and integration of complementary evidence.

THE CENTRAL QUESTION

What could produce an observed association?

Chance

Random sampling variability may produce an apparent relationship.

Bias

Selection or inaccurate measurement may distort what is observed.

Confounding

A common cause may influence both the exposure and the outcome.

Causal effect

Changing the exposure may genuinely change the outcome distribution.

These explanations are not mutually exclusive: a causal effect may coexist with residual confounding or other biases.

Statistical association does not establish direction or mechanism.

Clinical practice and public health policy depend on understanding whether an exposure genuinely contributes to an outcome or is merely associated with it. Statistical association alone cannot establish that changing one variable will change another. [1] [2]

An association occurs when the distribution of an outcome differs by exposure status. Relative risks, odds ratios, hazard ratios, risk differences and correlation coefficients describe patterns, but none establishes causal direction or mechanism by itself. Reverse causation is also possible: an existing or developing outcome may influence the apparent exposure.

The counterfactual framework asks how outcomes would differ under alternative exposure conditions. Because both potential outcomes cannot be observed for the same person at the same time, causal inference relies on assumptions including exchangeability, consistency and positivity. Randomisation supports exchangeability by design; observational research requires additional justification. [3] [4]

STUDY DESIGNS AND CAUSAL INFERENCE

Every design answers a different part of the question.

Randomised trials

Strong evidence for intervention effects.

Random allocation balances confounders on average, but non-adherence, missing outcomes, selective reporting, limited follow-up and generalisability remain concerns. CONSORT supports transparent reporting. [18]

Cohort studies

Time order can be examined.

Cohorts follow outcomes according to exposure status and support assessment of temporality, while confounding, selection bias and misclassification may remain. [17]

Case-control studies

Efficient for rare outcomes.

These studies are useful for rare outcomes and long-latency diseases but are sensitive to control selection, recall and historical exposure measurement. [17]

Cross-sectional and ecological studies

Useful for hypotheses and surveillance.

Cross-sectional designs often cannot resolve temporal order. Population-level ecological patterns cannot automatically be transferred to individuals. [17]

CONTEMPORARY APPROACHES

Better methods make assumptions visible.

No method eliminates uncertainty. Each one changes which assumptions must be defended.

Target trial emulation aligns eligibility, treatment strategies, time zero and follow-up in observational analyses with a hypothetical trial. [6]

Mendelian randomisation uses genetic variants as instruments, subject to assumptions of relevance, independence and exclusion restriction. [7]

Directed acyclic graphs make hypothesised causal structures explicit and guide adjustment decisions without inappropriately controlling for mediators or colliders. [5]

Triangulation combines methods with different potential biases. Convergence across them can strengthen causal confidence. [8]

MAJOR THREATS TO INTERPRETATION

Five ways a plausible relationship can mislead.

Confounding

Common causes distort exposure–outcome relationships; residual and unmeasured confounding can remain after adjustment.

Selection bias

Participation, loss to follow-up or conditioning on a collider can alter the observed association.

Information bias

Misclassification of exposures, outcomes or confounders can change the estimate.

Reverse causation

The outcome—or its early manifestations—may influence the exposure.

Misalignment of eligibility, exposure classification and follow-up can also create immortal-time and related biases, especially in observational treatment studies. [2] [3] [5] [6]

Bradford Hill's viewpoints structure reasoning; they do not prove causation.

Strength, consistency, specificity, temporality, biological gradient, plausibility, coherence, experiment and analogy can organise a causal assessment. They are not a scoring checklist. Temporality is necessary but insufficient; strong associations can still be biased, and modest effects can be causal. [1] [15] [16]

Counterfactual reasoning requires a defined intervention or exposure contrast and explicit identification assumptions. Directed acyclic graphs help identify adjustment sets. Causal confidence may increase when cohort, trial, genetic and mechanistic evidence converge despite having different potential biases. [3] [4] [5] [8]

ILLUSTRATIVE CLINICAL EXAMPLES

Evidence can strengthen—or overturn—an initial interpretation.

Smoking and lung cancer

Converging evidence supported causation.

Case-control findings were followed by cohort, dose-response, cessation and mechanistic evidence. [9]

Hormone therapy

Trials challenged apparent cardiovascular protection.

The Women's Health Initiative found no cardiovascular protection from combined oestrogen–progestin therapy and observed an early increase in coronary events. [10]

Beta-carotene

Dietary associations did not predict supplementation effects.

Trials in high-risk groups found no prevention of lung cancer and possible increased incidence. [11] [12]

Helicobacter pylori

Eradication trials supported a causal role.

The discovery of H. pylori challenged established assumptions; later intervention evidence strengthened the causal interpretation. [13]

APPLICATION TO ORAL–SYSTEMIC HEALTH

Periodontitis and cardiovascular disease show why proportional language matters.

Epidemiological studies show associations, and proposed pathways include systemic inflammation, endothelial dysfunction and microbial mechanisms. Both conditions also share risk factors such as smoking, diabetes, age and socioeconomic disadvantage. Biological plausibility alone does not demonstrate that periodontitis causes cardiovascular events.

A joint consensus report supports the epidemiological association while finding intervention evidence insufficient to establish that periodontal treatment prevents myocardial infarction, stroke or other major cardiovascular events. Periodontal care remains important for oral health, and integrated prevention may address shared risks; cardiovascular-event prevention should not be presented as a proven treatment benefit. [19]

CLINICAL AND PUBLIC HEALTH IMPLICATIONS

State what the evidence shows—and what it does not.

Communication must distinguish established clinical benefits, observed associations, plausible pathways and uncertain causal effects.

Clinicians should distinguish observed associations from demonstrated intervention effects. Researchers should define causal questions before analysis and examine temporality, confounding, selection, measurement and sensitivity to assumptions. [3] [4] [5] [14]

Public health professionals should consider the totality of evidence before issuing recommendations. In oral-systemic health, claims should remain proportionate to the study design, outcome and certainty of evidence. [15] [19]

KNOWLEDGE GAPS AND FUTURE DIRECTIONS

Move beyond repeated association.

Which relationships are causal, clinically meaningful and modifiable?

Measurement

Use longitudinal designs with standardised periodontal and systemic outcomes.

Shared risks

Improve control of behavioural, metabolic and socioeconomic determinants.

Clinical outcomes

Test outcomes that matter to patients, not only surrogate biomarkers.

Transparent inference

Use target trial emulation, DAGs, triangulation, protocols and sensitivity analyses.

Research should distinguish causal contribution from shared-risk markers and bidirectional relationships. [3] [5] [8] [17] [18]

CONCLUSION

Association is a starting point for discovery—not equivalent to causation.

Reliable causal inference requires clear questions, appropriate designs, explicit assumptions, assessment of bias and confounding, and converging evidence. Bradford Hill's viewpoints aid reasoning but do not prove causality. In oral-systemic research, especially periodontitis and cardiovascular disease, clinical and public-health statements must remain proportionate to the strength of the evidence.

Observe the relationship. Test the explanation. Communicate the difference.

REFERENCES

Association and Causality in Clinical and Epidemiological Research.

An association is an observation.Causation is an explanation.

Chance

Bias

Confounding

Causal effect

Statistical association does not establish direction or mechanism.

Randomised trials

Cohort studies

Case-control studies

Cross-sectional and ecological studies

Better methods make assumptions visible.

Five ways a plausible relationship can mislead.

Confounding

Selection bias

Information bias

Reverse causation

Bradford Hill's viewpoints structure reasoning; they do not prove causation.

Smoking and lung cancer

Hormone therapy

Beta-carotene

Helicobacter pylori

State what the evidence shows—and what it does not.

Move beyond repeated association.

Measurement

Shared risks

Clinical outcomes

Transparent inference

Association is a starting point for discovery—not equivalent to causation.

Explore the supporting literature.

An association is an observation.
Causation is an explanation.