Bradford Hill Considerations for Causation

What it is

The Bradford Hill considerations (often loosely called the "Bradford Hill criteria") are nine aspects of an association that Sir Austin Bradford Hill set out in his 1965 presidential address to the Section of Occupational Medicine of the Royal Society of Medicine, "The Environment and Disease: Association or Causation?". They are: (1) strength of the association; (2) consistency across persons, places, circumstances, and investigators; (3) specificity of the exposure-outcome link; (4) temporality (the exposure must precede the outcome); (5) biological gradient (dose-response); (6) plausibility given existing biological knowledge; (7) coherence with the natural history and biology of the disease; (8) experiment (does removing or reducing the exposure reduce the outcome?); and (9) analogy with established cause-effect relationships. Hill was explicit that these are viewpoints to aid judgment, not a checklist and not a set of hard-and-fast rules — "None of my nine viewpoints can bring indisputable evidence for or against the cause-and-effect hypothesis and none can be required as a sine qua non." The framework has no maintaining standards body (it is not an EQUATOR reporting guideline, a Cochrane risk-of-bias tool, an ISPOR good-practice report, or an agency guidance); it is a foundational piece of epidemiologic reasoning that has been re-read and formalized many times, most usefully through the modern counterfactual and directed-acyclic-graph (DAG) lens.

When to use

Reach for the Bradford Hill considerations at the interpretation and discussion stage of a non-interventional study, when you have a well-estimated association and must argue — transparently and against alternative explanations — whether it supports a causal claim. They are appropriate as a structuring device for the causal-argument section of a peer-reviewed manuscript, an HTA/payer dossier that leans on observational comparative effectiveness, an FDA/EMA submission where RWE is offered as causal evidence, a safety-signal causality assessment, and pharmacoepidemiology programs that triangulate across designs and data sources. Decision rule: use Bradford Hill to organize and stress-test a causal narrative across a body of evidence; do not use it as the primary tool for any of the upstream jobs that have purpose-built instruments. For reporting transparency use STROBE/RECORD(-PE) or HARPER; for risk-of-bias of a single non-randomized study use ROBINS-I; for certainty of a body of evidence use GRADE; for the design that licenses a causal estimate in the first place use a target-trial emulation with a pre-specified protocol. Bradford Hill complements these — it does not replace any of them.

What it requires (read for real-world data)

Treated rigorously, the nine considerations demand evidence that maps directly onto modern RWE practice. Temporality is the one non-negotiable consideration and forces explicit time-zero alignment, a clean lookback/washout, an incident (new-user) exposure definition, and attention to immortal-time and induction/latency windows so the exposure demonstrably precedes the outcome. Strength must be a well-confounded-adjusted estimate, not a crude one — so confounding control (active-comparator new-user design, propensity or high-dimensional propensity scores, DAG-guided covariate selection), a clearly stated estimand and handling of intercurrent events, and quantification of residual confounding (E-value, negative controls) all sit underneath "strength." Biological gradient requires defensible dose/duration measurement from fills or administrations. Consistency is the engine of triangulation: reproducibility across databases, designs (cohort, case-control, self-controlled), and populations, which in turn depends on validated phenotypes/algorithms (with PPV/sensitivity), fit-for-purpose data assessment, and honest accounting of attrition and missing data so that "consistency" is not just shared bias. Experiment in RWE maps to quasi-experiments and natural experiments (policy changes, formulary shifts) and to the as-if-randomized logic of target-trial emulation. Plausibility, coherence, specificity, and analogy are judgment-laden and weak as discriminators, but they are where mechanism, prior evidence, and sensitivity/quantitative-bias analyses are marshaled.

When NOT to use — limitations and common misapplications

The dominant failure mode is treating the nine considerations as a checklist or additive score: counting how many are "met" and declaring causation above some threshold. Hill warned against exactly this; the considerations are not independent, not equally weighted, and most are neither necessary nor sufficient. Specific traps: (1) Specificity is largely obsolete — most exposures have many effects and most diseases many causes, so its absence says little. (2) Consistency can reflect a bias shared across studies (e.g., the same misclassified claims phenotype reused everywhere), so replication is reassuring only if the studies do not share the same flaw. (3) Plausibility/coherence/analogy are bounded by the knowledge of the day and invite confirmation bias; absence of a known mechanism is not evidence of no effect. (4) Using Bradford Hill as a substitute for design — applying it to a crude, confounded association to launder it into a causal claim — is the cardinal misuse; it cannot repair confounding by indication, immortal time, or selection bias baked into the study. (5) It is not a risk-of-bias instrument (use ROBINS-I), not a certainty-of-evidence grading system (use GRADE), and not a reporting checklist (use STROBE/RECORD/HARPER); presenting a "Bradford Hill table" in lieu of those is checklist theater. (6) Reverse causation and collider/selection structures can mimic several considerations at once and are best surfaced with explicit DAGs, not narrative coherence.

How it maps to this catalog

Each consideration is implemented by specific concepts here. Temporality -> `time-zero-index-date-alignment-rwe`, `washout-clean-lookback-period-rwe`, `new-user-design`, `immortal-time-bias-handling`, `exposure-lag-induction-latency-window-rwe`. Strength (properly adjusted) -> `active-comparator-new-user`, `propensity-score-methods-psm-iptw`, `high-dimensional-propensity-score-hdps-rwe`, `dags-backdoor-criterion-drug-studies`, `estimands-ate-att-intercurrent-events-rwe`, with the strongest causal scaffold being `target-trial-emulation`. Biological gradient -> `time-updated-exposures-cumulative-dose-rwe`, `exposure-episode-construction-rwe`. Consistency / triangulation -> `meta-analysis-obs`, `self-controlled-case-series`, `case-control`, plus the validity substrate of `diagnosis-phenotype-algorithm-1ip-2op-time-window-rwe`, `claims-outcome-algorithm-ppv-sensitivity-rwe`, `algorithm-validation`, `fit-for-purpose-data-assessment-rwe`, and `attrition-and-loss-to-follow-up-rwe`. Robustness behind the judgment -> `e-value-sensitivity-analysis`, `negative-control-outcomes-rwe`, `empirical-calibration-negative-controls-rwe`, `quantitative-bias-analysis-toolkit-rwe`, `unmeasured-confounding-probabilistic-bias-analysis-rwe`. Applied note for claims/EHR/registry RWE: before a single Bradford Hill consideration is invoked, the underlying estimate must rest on a validated outcome phenotype (report PPV and, where feasible, sensitivity), a real (not missing) drug-free washout supported by continuous enrollment, time-zero set at initiation, and a pre-specified estimand. "Strength" then means the confounding-adjusted comparative estimate with an E-value and negative-control calibration attached; "consistency" means the result holds across at least one independent database or design that does not share the same phenotype/data weakness. Used this way, Bradford Hill is a disciplined argument layered on top of a defensible design — never a shortcut around one.