13 Research methods, statistics, critical review and evidence-based practice13.2 Epidemiological studies
13.2.4 Approaches to causal inference
5 min read3 figures
Written by Resident Doctor Committee · Published
Association and causation
Association means two variables occur together. Causation means that changing the exposure would change the outcome.
The counterfactual is what would have happened to the same person without the exposure. Only one history is observed, so studies estimate it from a comparison group.
Exchangeability means the comparison group shows what the exposed group would have experienced without exposure. Randomisation achieves this on average.
Prediction needs useful information, not a cause. Previous admissions predict readmission without being a target for intervention.
What observational designs can show
Case-control, cohort and cross-sectional studies can show a significant association, positive or inverse, but cannot by themselves establish that an exposure causes or prevents the outcome.
An observational cohort can show that a treatment is associated with a better outcome but not that it causes it. Exposure is not randomly allocated, and unmeasured or poorly measured confounding survives adjustment.
Moving from association to causation needs a design that manipulates the exposure. The randomised controlled trial is the strongest single design.
Adjustment, dose-response, replication and precision do not substitute for manipulating the exposure.
Observational associations, such as a cross-sectional table of odds ratios, generate hypotheses and do not by themselves justify a specific clinical action.
Among high-income countries, national wealth or development correlates positively with the prevalence of mental disorder. This association is real and counterintuitive, but it holds only within that income range and cannot be extended beyond it.
Temporality
Temporality means a cause must precede its effect.
A prospective cohort recruits disease-free people and follows them forward, so it can show that exposure preceded the outcome.
Case-control and cross-sectional studies cannot establish temporal order, so aetiological inference needs a cohort design.
Reverse causation remains possible in a cohort, because exposure recorded before diagnosis may follow an early prodrome, as when prodromal psychosis prompts cannabis use.
Bradford Hill viewpoints
The Bradford Hill criteria (viewpoints, 1965) are the standard framework for judging whether an association in epidemiological data is causal. They organise judgement and are not a score with a threshold.
Viewpoint | What it asks | Limitation |
|---|---|---|
Strength | Would confounding have to be implausibly large? | Weak associations can be causal; strong ones can be wholly confounded |
Consistency | Does it recur across populations, designs and investigators? | Repeating a design repeats its bias |
Specificity | One exposure, one outcome | Largely obsolete; its absence argues nothing |
Temporality | Does exposure precede outcome? | Observed order may not be true order when the prodrome is long |
Biological gradient | More exposure, more outcome? | Supportive but not necessary |
Plausibility | Is a mechanism conceivable given current biology? | Bounded by present knowledge |
Coherence | Does a causal reading conflict with established facts about the disease's natural history and biology? | Overlaps plausibility |
Experiment | Does removing the exposure reduce the outcome? | Rarely available |
Analogy | Did a similar exposure produce a similar effect? | Weakest; an analogy fits any hypothesis |
Temporality is the only viewpoint that must be satisfied.
Statistical significance is not one of the nine viewpoints.
A genuinely causal exposure can lack a dose-response gradient because of threshold or saturation effects, measurement error, confounding or chance. A graded confounder can also mimic a gradient.
Susser's framework
Susser held three properties essential to any cause.
Essential property | What it requires |
|---|---|
Association | Shown beyond chance |
Time order | The cause comes first; whether this can be shown depends on study design |
Direction | Change in the cause produces change in the outcome, not the reverse; hardest to show |
Susser's five supportive criteria are strength, specificity, consistency, predictive performance and coherence.
Survivability asks whether a hypothesis withstands a study capable of refuting it, not merely another similar replication.
Time order is the criterion shared by Susser and Bradford Hill, and the only one widely treated as indispensable.
Rothman's sufficient-component cause model
A sufficient cause is a complete set of conditions that produces the outcome. Each condition is a component cause, drawn as a slice of a causal pie.
Different pies can produce the same disease. A component present in every pie is a necessary cause.
The strength of a component cause depends on how common the other components of its sufficient cause are, so it is population-specific and not a fixed attribute of the factor.
Component attributable fractions can sum to more than 100%, because one case depends on several components at once.
Ecological studies
Ecological studies use group summary data for both exposure and outcome, measured across groups or areas.
The ecological fallacy is wrongly inferring something about individuals from an association measured between groups. This cross-level inference is the characteristic weakness of ecological studies.
A site-level correlation between psychosis incidence and high-potency cannabis use (r = 0.7) contributes evidence but cannot establish individual causation.
Direct, indirect and confounded associations
Type | Structure |
|---|---|
Direct | The exposure acts on the outcome with no intervening variable on the pathway and no third factor producing the link |
Indirect | A genuine relationship that runs through a mediator, which the exposure causes and which in turn causes the outcome |
Confounded | A third variable outside the pathway causes both exposure and outcome |
A mediator lies on the causal pathway and is caused by the exposure; a confounder lies outside the pathway and causes both.
Mediation analysis
Mediation analysis, a form of structural equation modelling, splits an exposure's total effect into a direct path and an indirect path through an intermediate variable on the causal route.
To act as a statistical mediator, a variable must be associated with the outcome after the exposure is controlled for.
A mediator with its own direct path to the outcome acts on that outcome whatever the level of the upstream variable, so mediation is not dependence on, or moderation by, the upstream variable.
Inconsistent (suppressor) mediation is a direct effect whose sign is opposite to that of the indirect effect, so the pathways partly cancel in the total effect.
When a drug's beneficial direct effect is offset by a harmful indirect effect through a side effect, clinicians should monitor and manage the side effect while keeping the drug, not withhold it.
Other approaches
Koch's postulates link one organism to one disease. They fit multifactorial psychiatric disorders poorly, because syndromal diagnoses provide no organism to isolate.
Mendelian randomisation uses genetic variants as instruments. They must predict the exposure, be independent of confounders and affect the outcome only through the exposure.