Discuss the concepts of disease causation.

Association, Causation and the Criteria

• When AN association IS observed, four explanations must be considered in order, and reciting them in order is what a good answer does:
• 1. Chance (random error) — assessed by tests of significance and confidence intervals
• 2. Bias (systematic error) — selection bias (in how subjects entered the study) and information or measurement bias (in how data were collected), including recall bias and observer bias
• 3. Confounding — the association is explained by a third variable associated with the exposure and an independent risk factor for the outcome, and not on the causal pathway between them
• The classical example — an association between carrying matches and lung cancer, confounded by smoking
• Confounding is controlled at the design stage by randomisation, restriction or matching, and at the analysis stage by stratification or multivariable adjustment
• 4. Causation — considered only after the first three have been excluded
• Reverse causation must also be excluded — the outcome causing the exposure rather than the reverse, which is a particular problem in cross-sectional and case-control studies
• The bradford hill criteria (1965) are the framework for judging causation, and they are viewpoints or considerations, not a checklist to be ticked — Hill himself insisted that none is indispensable except temporality
• 1. Temporal relationship — the cause must precede the effect. This is the only criterion that is absolutely essential
• 2. Strength of association — a large relative risk is less easily explained by confounding; but a weak association may still be causal
• 3. Consistency — the association is repeatedly observed by different investigators, in different populations and by different methods
• 4. Specificity — one cause producing one effect. This is the weakest criterion, since most exposures cause several diseases and most diseases have several causes — smoking causes many cancers, and lung cancer has many causes
• 5. Biological gradient (dose-response) — more exposure produces more disease, which is strong evidence
• 6. Plausibility — the association makes biological sense; but this is limited by the knowledge of the day, and a lack of plausibility must not be allowed to override good evidence
• 7. Coherence — it does not conflict with what is known of the natural history and biology of the disease
• 8. Experiment — removing the exposure reduces the disease, which is the strongest evidence of all where it is available
• 9. Analogy — a similar exposure is known to produce a similar effect
• The smoking and lung cancer story is the standard illustration, and satisfies every criterion — which is why it is used to teach them

Models of Causation and Their Uses

• KOCH’S postulates (1882) established causation for infectious disease: the organism must be found in all cases; isolated and grown in pure culture; reproduce the disease when inoculated into a susceptible animal; and be recovered from that animal
• Their limitations explain why they had to be superseded — they fail for carriers and asymptomatic infection; for organisms that cannot be cultured (leprosy, hepatitis B, prions); where there is no suitable animal model; where infection does not always produce disease; for multifactorial and non-communicable disease; and where one organism causes several diseases
• The epidemiological triad — agent, host and environment, with disease resulting from a disturbance of their equilibrium. It works well for infectious disease and is the basis of the classification of control measures
• Multifactorial causation — recognised by pettenkofer and developed as chronic disease came to dominate; disease has many contributory causes, none of which is sufficient alone
• Its practical value is great: it means disease can be controlled by acting on any of several factors, and the most accessible factor may be the best target even if it is not the most important cause
• The web of causation (MacMahon and Pugh) — suited to non-communicable disease, showing a complex network of interconnected antecedents in which each factor is itself the result of other factors
• Its key implication — breaking any strand of the web reduces the disease, and it is not necessary to know the whole web to intervene effectively
• Rothman’S causal PIES (the sufficient-component cause model) is the most rigorous formulation, and is worth knowing because it clarifies terms that are otherwise used loosely:
• A sufficient cause is a complete causal mechanism — a whole "pie" — which inevitably produces the disease
• A component cause is one SLICE of that pie
• A necessary cause is a component that appears in every sufficient cause — as Mycobacterium tuberculosis is necessary for tuberculosis, though not sufficient, since exposure alone does not produce disease
• The model explains several things that puzzle students — why removing any single component prevents the disease arising by that mechanism; why the attributable fractions for different causes can sum to more than 100%, since one case may be prevented by removing any of several components; and why the "strength" of a cause depends on how common its complementary components are
• The practical conclusion for public health — we need not identify the complete causal mechanism to prevent disease; it is enough to identify one removable component
• Snow and cholera remains the paradigm — the pump handle was removed in 1854, thirty years before the vibrio was identified, and the epidemic ended
• Similarly, LIND prevented scurvy with citrus fruit in 1747, long before vitamin C was known; and semmelweis prevented puerperal sepsis by handwashing before the germ theory
• These examples are worth citing because they make the central point: action need not wait for complete understanding
• Chance, bias, confounding, then causation is the order in which any observed association must be examined
• Only temporality is essential among the Bradford Hill criteria; the rest are considerations rather than requirements
• Specificity is the weakest criterion, since most exposures cause several diseases and most diseases have several causes
• Plausibility is limited by current knowledge, so its absence must not override good evidence
• A dose-response gradient is strong evidence, and is one of the most persuasive criteria in practice
• Confounders are associated with exposure and outcome and are not on the causal pathway, which is the definition to give
• Randomisation controls unknown confounders, which no analytical method can do
• Exclude reverse causation, particularly in cross-sectional and case-control designs where timing is unclear
• Koch postulates fail for carriers and for organisms that cannot be cultured, which is why they were superseded
• A necessary cause appears in every sufficient cause, as the tubercle bacillus does for tuberculosis
• Necessary is not sufficient; exposure to the bacillus does not by itself produce disease
• Attributable fractions can exceed one hundred per cent, which the causal pie model explains and simpler models cannot
• Removing one component prevents disease by that mechanism, which is the practical value of the model
• Action need not wait for understanding, as Snow, Lind and Semmelweis all demonstrated
• The most accessible factor may be the best target, even where it is not the most important cause
• Smoking and lung cancer satisfies every criterion, which is why it is the standard teaching example
• Experimental removal is the strongest evidence where it is available, since it is closest to a trial
• Consistency across populations and methods makes bias an unlikely explanation for an association
• A strong association resists confounding, but a weak one may still be genuinely causal
• Recall bias affects case-control studies particularly, since cases search their memories harder than controls
• Matching and restriction control confounding at design, while stratification and adjustment do so at analysis