To the Editor:
Flanders and Garber1 have published a commentary on our commentary on causal inference in epidemiology. We are broadly in agreement with their comments, although we could correct some minor points. For example, our note that causal inference methods that mimic randomized controlled trials (RCTs) “can improve individual studies with individual-level exposures….” did not imply that such methods cannot be used for group-level exposures. Also, we were clearly referencing our own work, to explain the reasons for our strong insistence on the usefulness of “older methods”, and the integration of different types of evidence to judge causality, and were not attempting to summarize the whole debate. The original 2016 debate about causal inference in epidemiology (see the December 2016 issue of the International Journal of Epidemiology) has moved forward in various forms, including personal discussions and e-mail exchanges. In the published record, we have seen comments on the nature of causality from VanderWeele2 that are broadly consistent with our own views, and those of Pearl.3 Greenland4 also published a thoughtful commentary that is critical of both sides of the debate. So, our understanding of these issues has progressed, and there have been clarifications and changes, although the views of some remain close to their original 2016 positions.5
Our main concern is the implication that we are misrepresenting the views of the so-called causal inference movement. Flanders and Garber quote some sentences to indicate that “the other side” is more inclusive than we have represented. One is from the book by Hernán and Robins of which different versions appeared on the internet; in the latest 2019 internet version, we searched for the quote and found one sentence stating that if the analogy with RCTs breaks down one might have to use “instrumental variable” techniques (p. 26).6 The other quoted sentence is similar. Such single sentences, in our opinion, do not indicate inclusiveness of other views, but rather give the impression that most causal inference involves RCT-based methods, and that these other approaches (e.g., Bradford Hill considerations, triangulation, instrumental variables, biologic knowledge, knowledge of mechanisms, etc.) are just optional extras. For example, the focus on RCTs has led to the generation of “risk of bias” tools that may not adequately identify bias and confounding in these studies, and that perpetuate the notion that observational studies should be assessed against the ideal RCT.7 Real causal inference is so much more than this.
Jan Vandenbrouckea,b,c, and Deborah A. Lawlora,d,e
From the aLondon School of Hygiene and Tropical Medicine, London, United Kingdom;
bLeiden University Medical Center, Department of Clinical Epidemiology, Leiden, The Netherlands;
cDepartment of Clinical Epidemiology, Aarhus University, Denmark;
dMRC Integrative Epidemiology Unit, University of Bristol, Bristol, United Kingdom; and ePopulation Health Science, Bristol Medical School, Bristol, United Kingdom.
1. Flanders WD, Garber MD. Is the smog lifting?: causal inference in environmental epidemiology. Epidemiology. 2019;30:317–320.
2. VanderWeele TJ. On well-defined hypothetical interventions in the potential outcomes framework. Epidemiology. 2018;29:e24–e25.
3. Pearl J. Does Obesity Shorten Life? Or is it the Soda? On Non-Manipulable Causes. 2018.Los Angeles, CA: UCLA.
4. Greenland S. For and against methodologies: some perspectives on recent causal and statistical inference debates. Eur J Epidemiol. 2017;32:3–20.
7. Bero L, Chartres N, Diong J, et al. The risk of bias in observational studies of exposures (ROBINS-E) tool: concerns arising from application to observational studies of exposures. Syst Rev. 2018;7:242.