Jane E Ferrie
Arguments about causal inference in ‘modern epidemiology’ revolve around the ways in which causes can and should be defined. The potential outcomes approach, a formalized kind of counterfactual reasoning, often aided by directed acyclic graphs (DAGs), can be seen as too rigid and too far removed from many of the complex ‘dirty’ problems of social epidemiology, such as social inequalities and racism. If a potential ‘cause’ cannot be manipulated is it sensible to disregard it, relegating it to the ‘not suitable for epidemiology’ category? The use of properly constructed DAGs may aid causal thinking and help plan relevant analyses – Neil Pearce and Debbie Lawlor provide a simple, but excellent discussion of the use of DAGs in their essay review of Causal Inference in Statistics: A Primer by Judea Pearl and colleagues. However, increasingly, DAGs and analyses are constructed by computer programs, such as DAGitty, now available as an R package ‘dagitty’. Useful as such programmes are, the temptation to use evaluations of DAG-dataset inconsistency to generate purely data-driven, post-hoc modifications to DAGs, raises concern about overfitting and biased inference.
Calls for a consequentialist epidemiology have pointed to a discipline increasingly preoccupied with arcane aspects of aetiology at the expense of prevention and public health improvement. Further concerns, that the potential outcomes approach is edging towards disciplinary dominance at the expense of questions not amenable to its reasoning, are reflected in two papers on the topic of causal inference. Written by influential epidemiologists, already known for their contributions to epidemiological theory and practice, these papers espouse methodological pluralism. For Jan Vandenbroucke and colleagues the approach is ‘pragmatic pluralism’ – quietism about the nature of causation and pluralism about causal concepts. For Nancy Krieger and George Davey Smith it is ‘inference to the best explanation’; a flexible, multi-faceted, historically-informed approach that focuses on how well a particular causal story fits with the totality of the evidence, rather than privileging one particular approach above others.
Five sets of commentators take up cudgels on behalf of the potential outcomes approach and pick apart the arguments of the pluralists. James Robins and Michael Weissman argue that potential outcomes reasoning by definition is the way consequentialists answer the question ‘What is to be done?’, and contend that the illustrative examples used by Krieger and Davey Smith remain firmly within the counterfactual framework. For Rhian Daniel and colleagues the attacks on the potential outcomes approach are misconceived; its rigorous framework is an asset that can ensure reliable answers to a wide range of causal questions. Tony Blakely and colleagues argue that inference to the best explanation is not dissimilar to what we do already, or what we should be doing, but caution that the ‘loveliest’ explanation (as the leading causal account has been called in the inference to the best explanation school) may be hard to replicate if beauty is in the eye of the beholder. Tyler VanderWeele, in a thoughtful, seven-point appraisal of the potential outcomes approach, concludes it is the right tool for a sub-set of causal questions, but that “multiple perspectives are currently needed in our thinking about causal reasoning”, while Douglas Weed doubts whether its presumed increasing dominance will really be a critical issue for epidemiology.
Despite the pluralists going out of their way to ward off such accusations, those defending counterfactual approaches almost universally accuse them of having set up straw men to create space for their arguments. In their responses the pluralists deny these accusations, pick apart the counterarguments of the commentators, and mount a vigorous defence of their own approaches. (Krieger and Smith response; Vandenbroucke et al. response) On one side are accusations of restrictiveness, on the other, lack of specificity; both claim to support a consequentialist epidemiology. Misrepresentations are pointed at on both sides of the argument and the debate is continued in the columns of the letters section.
Will these pluralistic approaches to defining causes prove more useful than the potential outcomes approach, or is it a case of horses for courses? How will new proposals to develop a more formal approach to triangulation – an updating of the Bradford Hill criteria – defined as “the practice of strengthening causal inferences by integrating results from several different approaches, where each approach has different (and assumed to be largely unrelated) key sources of potential bias” – contribute to how we carry out epidemiology in the future? Put preconceptions aside and enjoy the productive tensions of a debate in which the protagonists ultimately have more in common than they seem to admit 😉.
Productive tensions around the potential outcomes approach arise again in a Symposium which centres on a synopsis by VanderWeele of his recently published book: Explanation in Causal Inference: Methods for Mediation and Interaction. The book attracts generous plaudits in all three commentaries, but, whereas Michael Oakes and Ashley Naimi, and Jay Kaufman use the opportunity to expand on the merits of counterfactual thinking in relation to mediation and interaction, Pearce and Vandenbroucke point to its limitations and question its practical applicability.
The use of genetic variants, which can proxy for a potentially modifiable exposure and are essentially unrelated to confounding factors, to explore causal pathways – Mendelian Randomization – has been widely championed by the IJE. Three papers in the Causality issue use Mendelian Randomisation methods to re-examine associations previously documented in observational studies. In two cases, (Nordestgaard and Nordestgaard and Raymond Noordam and colleagues) evidence from the observational studies is contradicted by the Mendelian Randomisation analyses, but in the remaining study by Michael Holmes and colleagues the two sources of evidence are complimentary. These empirical papers are accompanied by two theoretical papers which assess methods for Mendelian Randomisation meta-analyses by Jack Bowden and colleagues and AF Schmidt and colleagues.
Other methodological challenges in the quest for causality are examined in the issue, with a particular emphasis on interventions. Of particular interest is the review by Sophie Jullien and colleagues on the impact of deworming programmes on schooling and economic development. Deworming – the mass medication of children and women of child-bearing age in developing countries – is a vast initiative involving the World Health Organisation and the ‘Deworm the World’ campaign. Jullien and colleagues’ review of the unpublished economic studies, widely cited as claiming benefits for health and economic development by advocates of deworming, concludes that they “should not be considered reliable evidence of effects.” A ‘worm wars’, for and against continuing current deworming policy, has been declared with considerable media coverage. Unsurprisingly, there are points of difference and emphasis with regard to Jullien and colleagues’ analysis, many of them highlighted in eight lively commentaries that accompany the review. However, there is some agreement that the evidence base for deworming is sparse and further evaluations of the long-term developmental and economic effects would be helpful.
The Causality themed issue will appear as the December issue of the IJE for 2016 (Volume 45, Issue 6): https://academic.oup.com/ije
Dates for your diary
Tuesday 11th – Thursday 13th July 2017, Bristol, UK
Mendelian Randomization Conference 2017 – Mendelian randomization in the age of large-scale accessible genomics data (http://www.mendelianrandomization.org.uk/)
Friday 6th October 2017, Bristol, UK
International Corresponding Club for Epidemiological Discussion (ICCED) 2017 – Following the success of the IJE conference, George Davey Smith and Shah Ebrahim are planning to host an online forum, ICCED, for discussion of matters epidemiological. The first ICCED conference will take place on Friday 6th October in Bristol, UK. If you would like to be added to the ICCED mailing list please email email@example.com.