Austin Bradford Hill’s ‘Environment and disease: Association or causation’

Austin Bradford Hill’s 1965 paper [1] on disease and the environment is one of the most widely cited papers in public health and related fields [2]. Hill’s nine ‘criteria’ for causal inferences were used in 87% of recent reviews of evidence on causality in population health [3]. In the addictions field, Hill’s paper has been used to evaluate causal explanations of: associations between psychosis and the use of cannabis and nicotine [4]; the effects of naloxone distribution on opioid overdose deaths [5]; the effects of a minimum unit price for alcohol policy on alcohol-related harm [6]; and the role played by vitamin E acetate in the 2019 outbreak of lung injury from vaping cannabis oil in the United States [7]. Hill’s ‘criteria’ have also been used in courts in the United States to assess causal claims about injuries caused by chemical exposures [8, 9].

Why has a paper that was based on a lecture given to occupational physicians over 50 years ago come to be regarded as an authoritative method for making causal inferences from observational data? Its continued role as a foundational citation classic is surprising because in the more than half-century since its publication there have been major advances in thinking about causal inference [10-12] and in the development of advanced statistical methods and research designs to improve the quality of causal inferences from observational studies [10, 11, 13-15].

Hill assumed that we have observations that: ‘… reveal an association between two variables, perfectly clear-cut and beyond what we would attribute to the play of chance’ and asked: ‘What aspects of that association should we especially consider before deciding that the most likely interpretation of it is causation?’ ([1], p. 295).

Hill outlined nine factors that he thought needed to be considered in making this decision.

Hill stressed (p. 299) that it was not possible to ‘lay down some hard-and-fast rules of evidence that must be obeyed before we accept cause and effect. None of my nine viewpoints can bring indisputable evidence for or against the cause-effect hypothesis and none can be required as a sine qua non. What they can do, with greater or less strength, is to help us make up our minds on the fundamental question—is there any other way of explaining the set of facts before us, is there any other answer equally, or more likely than cause and effect?’

Hill was also clear that tests of statistical significance could not directly answer causal questions. At most they ‘can, and should, remind us of the effects that the play of chance can create, and they will instruct us in the likely magnitude of those effects’ (p. 299).

Finally, Hill asked: ‘what action may flow from a conclusion that an association is causal? In ‘occupational medicine our object is usually to take action. If this be operative cause and that be a deleterious effect, then we shall wish to intervene to abolish or reduce death or disease’ (p. 300).

He suggested that we should use ‘differential standards before we convict’ depending upon the social consequences of the action we propose. We may, for example, be prepared to restrict the use of a drug to treat morning sickness in pregnant women on ‘slight evidence’. We may be prepared to ‘remove a probable occupational hazard on the basis of fair evidence’. We should, however, ‘need very strong evidence before we made people burn a fuel in their homes that they do not like or stop smoking the cigarettes and eating the fats and sugar that they do like… [but] this does not imply crossing every ‘t’, and swords with every critic, before we act’ (p. 300).

Hill’s paper has several strengths. First, he made clear that different types of evidence from multiple types of study needed to be considered when deciding whether an association was likely to be causal or not [8]. He emphasized the need to judge the degree of consilience between these diverse types of evidence in judging whether causation provided the most plausible explanation of an association [8]. Secondly, he emphasized the provisional nature of causal inference while insisting that we often need to take action on the basis of incomplete evidence [2]. Thirdly, he argued that the strength of the evidence required to justify action varied with the costs and inconvenience of any changes required to produce its benefits [2].

The major limitation of Hill’s paper was its lack of an explicit justification for the guidelines. His analysis arose out of the debate in the United Kingdom and United States in 1950s and 1960s regarding whether cigarette smoking was a cause of lung cancer and other diseases, a debate in which Hill and Doll’s epidemiological research played an important role [10, 16]. Hill’s approach expanded upon the five criteria (strength, consistency, specificity, temporality and coherence) that were proposed by Stallones [17] and included (without explicit justification) in the 1964 US Surgeon General’s (USSG) report on cigarette smoking and health [18].

Hill [1] used the 1964 USSG report to illustrate the role of several of his considerations, but he did not discuss its causal criteria [16]. His response to an inquiry from the librarian at the London School of Hygiene and Tropical Medicine about the source of his guidance was unilluminating (see [16]). It is also unclear whether an approach to causal inference derived from the debate about cigarette smoking and disease will be useful in assessing causal claims more generally [10].

Another major weakness of Hill’s analysis was that that each of his considerations were highly qualified. Except for temporality, he acknowledged that each one may, or may not, indicate a causal relationship [19]. Critics have since raised doubts about the value of specificity, plausibility, coherence and analogy [10, 15, 19]. Most critically, Hill did not provide any rationale for deciding how his nine considerations should be used in arriving at conclusion on the plausibility of a causal inference [8]. Indeed, he seemed to disavow the possibility of any ‘hard-and-fast rules’ [8].

Despite these weaknesses, and Hill's cautionary concluding statement, his nine considerations have been widely used as ‘criteria for causal inference’ [2, 8]. Their application to controversial cases requires judgement and it is unclear whether researchers with different views on a specific debate about causation would agree in how they judge the evidence against each of Hill’s considerations [10, 15]. Most often, Hill’s ‘criteria’ seem to have been used like a diagnostic checklist on the assumption that the more that are satisfied, the more plausible a causal inference is [8].

Scholars have attempted to improve upon Hill’s approach to causal inference (e.g. [20-25]), but none of their suggested alternative approaches has been widely adopted. I briefly summarize three.

Shimonovich et al. [15] compared Hill’s guidelines with more formalized approaches to causal inference, such as the use of directed acyclic graphs, sufficient-component causes (causal pies) and grading of recommendations, assessment and evaluation (GRADE approach). They found that the four approaches were consistent in the value of: strength of association (including analysis of plausible confounding); temporality; plausibility; and evidence from experiments. They argued that consistency was better assessed by heterogeneity of effect sizes and suggested that specificity was rarely met and should be replaced by the evidence that other plausible causes did not explain the association. They did not regard a dose–response relationship as strong evidence for causation because it could arise from confounding. They also concluded that Hill’s concepts of coherence and analogy were of limited utility in causal inference.

Bird argued that Hill’s guidelines can rule out R and N but are unable to exclude C, because observational studies cannot exclude all possible common causes. Hill’s guidelines can help to eliminate a specific confounder, but they cannot exclude confounding by unmeasured variables in the way that a well-conducted experiment, such as a controlled clinical trial, can: e.g. by using randomization of cases to treatment or comparison condition to eliminate all possible Cs.

It is perhaps time for the addictions field to abandon the uncritical use of ‘Hill’s criteria’ as a talisman for making causal judgements, not least because Hill disavowed this way of using what he described as nine ‘considerations’. We can still acknowledge that Hill’s paper made an historically important contribution to emerging thinking about how to draw causal inferences from observational evidence while appreciating that it was necessarily ‘an early rough cut’ (2], p. 2). If a more formal approach to causal inference is needed, then a simpler approach would be to use Bird’s reformulation of Hill’s approach. This would evaluate the available evidence to decide between the three logical explanations of an association between A and B; namely, that it is due to chance (using evidence from statistical tests and systematic reviews) that we can be confident that A precedes B (if A is a cause of B); and that there is evidence that excludes credible common causes of both A and B, such as studies that use statistical methods to control for plausible confounders (e.g. regression-based approaches; propensity scores; instrumental variables; and formal methods based in directed acyclic graphs) [10, 11].

Wayne D Hall: Conceptualization (lead); data curation (lead); formal analysis (lead); writing—original draft (lead); writing—review and editing (lead).

None to declare.