From Association to Causation
In 1965, epidemiologist Austin Bradford Hill proposed nine criteria for judging whether observed associations reflect true causal relationships. These criteria remain foundational for distinguishing correlation from causation in observational research.
(1) Strength: strong associations are more likely causal than weak ones. A relative risk of 20 (vs 1.2) suggests causation. Smoking increases lung cancer risk ~15-fold, supporting causation. However, even weak associations can be causal if mechanisms are strong.
(2) Consistency: findings replicated across diverse populations, settings, and time periods strengthen causal inference. Smoking's lung cancer link appears consistently worldwide, in different eras, across multiple study types. Conversely, findings appearing in one population only raise concern about confounding or chance.
(3) Specificity: a cause produces a specific effect rather than multiple unrelated outcomes. Smoking causes lung cancer specifically, though also causes other cancers. Asbestos specificity varies—causes mesothelioma (specific) and various cancers (less specific). Specificity strengthens causation but isn't required (causes can have multiple effects).
(4) Temporality: exposure precedes disease. For smoking-lung cancer, smokers develop cancer after years of exposure. Reverse causation (does disease cause smoking?) seems unlikely but must be excluded. In microbiome research, temporality is critical: does dysbiosis precede disease or result from it?
(5) Biological gradient (dose-response): greater exposure produces greater effect. Smoking 40 cigarettes daily increases cancer risk more than 5 cigarettes. Dose-response relationships support causation. However, thresholds can occur (no effect until exposure exceeds threshold).
(6) Plausibility: biological mechanisms explaining association strengthen inference. Why would smoking cause lung cancer? Tobacco smoke contains carcinogens; mechanisms are known. For microbiome associations, plausibility varies. Dysbiosis reducing short-chain fatty acid production (compromising epithelial barrier) seems plausible. Implausible mechanisms warrant skepticism.
(7) Coherence: association aligns with existing knowledge. Smoking causing cancer coheres with established toxicology. Findings contradicting all known biology raise suspicion. Though sometimes new evidence overturns prior assumptions, coherence provides a reality check.
(8) Experiment: experimental evidence (intervention modifying exposure alters outcome) strengthens causation. Randomized cessation trials (smokers randomized to quit vs continue) would provide strongest evidence. Such trials are unethical, so observational evidence predominates. Animal models and cell studies provide experimental support.
(9) Analogy: similar cause-effect relationships support new associations. If smoking causes lung cancer, might it cause mouth and throat cancers? Analogy to known patterns increases plausibility.
Applying Bradford Hill to microbiome-disease associations: Take Faecalibacterium deficiency and inflammatory bowel disease. (1) Strength: Faecalibacterium abundance is moderately lower in IBD (relative difference ~2-fold). (2) Consistency: findings appear across studies and populations. (3) Specificity: low Faecalibacterium appears in other conditions too (not specific to IBD). (4) Temporality: dysbiosis exists at disease onset; whether it precedes disease remains unclear. (5) Dose-response: weak evidence for linear associations. (6) Plausibility: Faecalibacterium produces butyrate, which strengthens epithelial barriers—mechanistically plausible. (7) Coherence: short-chain fatty acid hypothesis aligns with existing knowledge. (8) Experiment: Some animal models show Faecalibacterium supplementation improves outcomes. (9) Analogy: similar deficiencies appear across inflammatory conditions.
Overall, evidence for Faecalibacterium-IBD causation is suggestive but not definitive. Dysbiosis could be consequence of disease, medication, or diet rather than cause. Randomized trials of Faecalibacterium supplementation (weak outcomes to date) would provide experimental evidence.
Bradford Hill criteria aren't rules but guidelines. They're not equally weighted; temporality, for instance, is more essential than analogy. They don't prove causation—association can meet all criteria while remaining non-causal. Rather, they systematize thinking about causation in complex health research.