In my last two blogs I have used Austin Bradford Hill’s lecture, “The Environment and Disease: Association or Causation?”, which he gave in 1965 to the Royal Society of Medicine’s then newly formed Section of Occupational Medicine, as a reason for discussing the words “criteria” and “guidelines”. Although Hill outlined what he called “nine viewpoints” that he suggested could help in analysing causation in associations between two variables, they have commonly been called criteria. However, only one of his viewpoints is actually a criterion, and I suggested that we might better refer to them as guidelines.
The two words “guide” and “view” come from the same IndoEuropean root, WEID, to see. “Guide” comes via the Gothic witan followed by consonantal shift of w to g and t to d. “View” comes via the Latin videre, to see. So “guideline” is a good modern substitute for “viewpoint”.
Just as a guideline was originally a line used to guide something, before the word started to be used figuratively, so a viewpoint originally had a literal meaning: “A place or position from which something is viewed, esp. one which offers a particularly extensive or attractive view of the surrounding area, a natural feature within it, etc” (Oxford English Dictionary). The word was a surprisingly late addition to English—the first example given in the OED dates from 1839—although it was not long before it began to be used figuratively (“A mental position or attitude from which a subject or question is considered”); the earliest example dates from 1856. The surprise is that the phrase “point of view” dates from the start of the 18th century in the literal sense of “the position from which something is seen or viewed”, although it was not until the end of the century that it started to be used in the figurative sense of “a mental position or attitude”.
Hill’s nine viewpoints were:
- the strength of the association;
- consistency of the observed association;
- the specificity of the association;
- the temporal relationship of the association;
- biological gradient, or dose-response curve;
- biological plausibility;
- coherence with other generally known facts;
- experimental, or semi-experimental, evidence;
Of these, only temporality can be regarded as a criterion—if the outcome precedes the putative cause there can be no cause and effect relation, although other time courses may afford poor evidence. The other items are guidelines. This is not a mere verbal quibble. It emphasises, as Bradford Hill did, that none of the items in his list specifies a cause and effect association.
This also underlines the fact that some of Bradford’s Hill’s items are asymmetrical, and that not only their presence but also their absence should be taken into account. The extreme case of this is temporality, whose absence is strong evidence against an association, while its presence is of little confirmatory value. Conversely, the presence of similar effects produced by analogous interventions provides strong evidence in favour of an association; its absence is unhelpful. However, some of the other guidelines are more equally balanced. For example, the presence of what Hill called “gradient” (i.e. dose responsiveness) supports an association, while its absence is strong evidence against. Similarly, consistency (i.e. the ability to replicate findings in different settings and at different times) provides strong evidence in favour of an association, while its absence is strong evidence against, although one must consider the possibility that different susceptibility factors may be operating in different circumstances. Women, for example, are more susceptible to QT interval prolongation in response to a medication than men are.
Hill’s guidelines are listed in Table 1, showing how their presence and absence can be interpreted in the special case of adverse drug reactions.
Table 1. Guidelines for assessing the strength of evidence in favour of or against a hypothesised association between a cause or an intervention and an effect, with examples from adverse drug reactions
|If present||If absent|
|Strength||The stronger the association, the more likely it is to be a causative one; typically (e.g. in English courts) an odds ratio of 2 represents the watershed||Absence of association is weaker evidence against cause and effect than its presence, since it may be due to lack of power||The association of third-generation contraceptive progestogens with thrombophlebitis was not considered proven by an English court because the odds ratio was less than 2|
|Consistency||Replicability in different settings and at different times provides strong evidence in favour of an association||Failure to replicate is strong evidence against an association||People with the HLA B*5701 genotype are consistently likely to have a rash in response to abacavir|
|Specificity||Demonstration of a specific group of individuals with a particular susceptibility is strong evidence in favour of an association||Absence of specificity is unhelpful||Adverse reactions with identifiable susceptibilities, e.g. thalidomide-associated teratogenicity|
|Temporality||The cause/intervention should precede the effect and the effect should occur after a plausible interval; in addition the effect should occur at the same site as the intervention if relevant||If the effect precedes the intervention there cannot be an association; if it occurs too soon or too long after it the association is not supported||Thrombophlebitis due to amiodarone follows soon after intravenous administration into a peripheral vein|
|Biological gradient||The presence of dose-responsiveness and reversibility provides strong evidence in favour of an association||Absence of evidence is unhelpful; evidence of no dose-responsiveness is against an association||The dose-related effect of paracetamol on the INR in people taking warfarin suggests an interaction|
|Biological plausibility||The presence of biological, chemical, or mechanical plausibility supports an association||The absence of biological, chemical, or mechanical plausibility provides evidence against an association||The paracetamol-warfarin interaction cited immediately above as being dose-related is also biologically plausible|
|Coherence||Consistency with the current scientific paradigm supports an association [note the relation to biological plausibility]||Absence of consistency with the current scientific paradigm provides strong evidence against an association||COX-2 inhibitors are less likely to cause gastrointestinal bleeding|
|Experiment: size of effect is not attributable to plausible confounding||The larger the effect and the fewer the identifiable confounding factors, the stronger the evidence in favour of an association||Clinical trials evidence of no association may be evidence against but is not necessarily conclusive||There is poor trials evidence of an association of metformin with lactic acidosis, but a strong signal from observational studies and case reports|
|Analogy: similarity||If similar effects are produced by analogous interventions, this provides strong evidence in favour of an association||Lack of analogous effects is unhelpful||
That no other drugs have been reproducibly reported to cause phocomelia does not vitiate the case against thalidomide
Jeffrey Aronson is a clinical pharmacologist, working in the Centre for Evidence Based Medicine in Oxford’s Nuffield Department of Primary Care Health Sciences. He is also president emeritus of the British Pharmacological Society.
Competing interests: None declared.