The 24th Bradford Hill Memorial Lecture was delivered by Professor John Ioannidis at the London School of Hygiene & Tropical Medicine on 6 July 2015. The lecture this year was titled 'Exposure-wide epidemiology: revisiting Bradford Hill'.
Professor Ioannidis is probably best known for his 2005 PLOS ONE paper 'Why most scientific findings are false', which claimed that it is more likely for a published research claim to be false than it is to be true. In this highly stimulating Bradford Hill lecture, Ioannidis considered each of Bradford Hill’s nine criteria for causality in turn, and what the evidence from the 50 years of scientific endeavour that have occurred since suggest about their success as criteria for causality.
The first point made by Ioannidis was that the amount of empirical science that has been and is being conducted is vast, with over 120 million scholarly articles having been published. A large portion of these are based on examining associations between various quantities, with some based on highly focused analyses and statistical tests, while others involve very high dimensional analyses (eg genome-wide association studies).
The important and pervasive issue of publication bias and selective reporting of results was highlighted by an analysis conducted by Ioannidis and colleagues in which 50 cooking ingredient were selected at random from a cookbook. A subsequent literature search then found that 40 of the 50 ingredients had been found to be associated with either increased or decreased cancer risk.
Ioannidis then turned to each of Bradford Hill’s causality criteria. Whereas Bradford Hill argued that a large association is supportive of causation, Ioannidis argued that in fact the reverse may be true: that when large associations are observed, these are most plausibly due to effect exaggeration, such that the true association is small, and arguably therefore less likely to be causal.
Next, he proposed the idea that effect (association) sizes ought to be interpreted as proxies for the biases of a particularly scientific field. That is, those fields which tend to report the largest effects are probably those that are, for a multitude of reasons, suffering from the largest biases (eg selection, publication, confounding), such that in truth the associations under study are in truth far closer to the null.
Ioannidis confirmed the important role of Bradford Hill’s criterion of consistency, in the form of research findings being replicated. However, he noted how exact replication has in practice proved quite rare, with replicating studies often changing one or more aspects of the design or analysis, relative to the original study.
The criteria of specificity and temporality were also deemed to be important criteria, but that in practice they may be difficult to establish. For the former, Ioannidis described work illustrating the fact that a collection of exposures typically exhibits a complex correlational structure, rendering the ascertainment of which exposure is causally responsible difficult. Similarly, he argued that establishing the temporal ordering of exposures and outcomes is sometimes highly problematic, with studies of Parkinson’s disease, where the disease process may take place over many years, as a case in point.
Lastly, he argued that the criteria of plausibility, coherence and analogy are problematic in practice, given the changing nature of biological knowledge and understanding. He cited the case of APOE, a genetic variant which it has been argued on the basis of biological knowledge may be causally linked to the incidence of a vast array of diseases and conditions, by way of example.
In conclusion, Ioannidis argued that while Bradford Hill’s criteria have been of undoubted value in furthering our understanding and consideration of causlity, he believes several of them have little predictive ability for causality, with the presence of some of them (eg large associations) even decreasing the chance of an association being causal.