The Epistemological Function of Hill's Criteria

Preventive Medicine (forthcoming)

Austin Bradford Hill's criteria have been discussed in great detail since the publication of his famous paper "The Environment and Disease: Association or Causation?" in 1965. But almost all of the discussion of the criteria in the medical literature focuses on the specifics of each 'criterion': why from an epidemiological point of view this criterion (e.g. biological gradient) is a good criterion (or not a good criterion, or could be improved etc.). There has been some more philosophical discussion of what is meant by 'causation' and by 'criterion'. But little work has sought to relate the specific criteria to a broader epistemological framework. That is the purpose of this paper.

The framework I propose is that of eliminative induction. In the case of a general causal hypothesis between two variables A and B that appear to be correlated, eliminative induction implies that in order to establish the hypothesis that A causes B, we must eliminate the alternative hypotheses: (i) A and B are causally independent (no causal relation); (ii) B causes A; (iii) A and B have a common cause. Thus each of the criteria can be assessed regarding their support for the hypothesis that A causes B by considering their contribution to eliminating the alternative hypotheses (i), (ii), and (iii). The central part of the paper carries out this assessment.

This assessment shows that with the right evidence, the criteria between them can indeed rule out (i) and (ii). But the evidence will not suffice to rule out (iii) in a systematic manner. Note that (iii) involves an existential claim. Hence refuting it involves confirming a general hypothesis about possible common causes (nothing is a common cause of A and B). Some of the criteria can be used to eliminate specific common cause hypotheses. But they cannot eliminate all of them in a systematic manner.

This confirms the widely perceived weakness of observational epidemiological studies, that they are subject to confounding. Since a properly conducted controlled trial (e.g. an RCT) can eliminate (iii), this is a reason to regard controlled trials as providing 'better' evidence of causation than observational studies, as maintained in the EBM hierarchy of evidence. But that does not entail that the evidence from observational studies is always weak. We can use the framework of inference to the best explanation to explain the strong justification provided by some observational studies.