In this issue of AJPH, Hernán (p. 616) argues that when the aim of an investigation is to estimate a causal effect, researchers should be allowed to say so. I agree, and extend this conversation to argue that if we embrace causal questions we must simultaneously incorporate a causal roadmap approach. A causal roadmap helps avoid confusion and conflation of different parts of the scientific process—this becomes quite important once the “c-word” has been unleashed.
BRIEF INTRODUCTION TO THE CAUSAL ROADMAP
Researchers and editors are reticent about articulation of causal questions because it is easy to confuse what you want to do—estimate a causal effect—with what you can actually do—estimate a statistical parameter. The beauty of a causal roadmap is that it clarifies which parts of a research process are causal, which parts are statistical, and how they link. There are variations on the roadmap, but they broadly include the following: step 1—articulate the scientific question, including definition of the causal parameter of interest; step 2—link the causal and statistical parameters through assessment of the assumptions under which they are equal (known as identifiability); step 3—estimate the statistical parameter; and step 4—interpret the findings.1–3
In following the roadmap, the causal question starts the investigation, consistent with Hernán’s call. In the subsequent step of identifiability, the specific ways in which a particular application meets or falls short of assumptions required for causal interpretation will become clear. Estimation is a purely statistical step. Finally, interpretation ties the process together by weighing the extent to which assumptions are met (step 2) in considering how strongly to interpret the estimate of the statistical parameter (step 3) with respect to the original causal parameter of interest (step 1).
SEPARATE STATISTICS FROM INTERPRETATION
Hernán mentions that methods for causal effects are not the same as methods for associations, and this topic merits some expansion. I agree that an association question does not require consideration of confounding, whereas a causal question requires confounder adjustment, and thus the appropriate methods are different. However, I would caution against conflation of certain statistical methods (step 3) with interpretation of findings as causal effects (step 4)—this is a particular issue for statistical approaches called “causal inference methods” such as propensity weighting, parametric g-formula, and targeted maximum likelihood estimation (TMLE; bit.ly/2FVfdfj).4,5 As these newer statistical methods have become more widely used, traditional analysis approaches such as regression have begun to seem that they must fall short when it comes to causality.
This is one common conflation that the roadmap helps us avoid. A TMLE-based estimate of a clearly defined statistical parameter (step 3) that corresponds to an interesting and plausible intervention on an exposure (step 1), cannot be interpreted as a causal effect if the identifiability assumptions are not met (step 2). By contrast, a conditional odds ratio, estimated with simple logistic regression (step 3) can be interpreted as a causal conditional odds ratio for the effect of the exposure (step 1) if the identifiability assumptions are met (step 2).
Certainly, there are reasonable arguments (with which I would agree) that, as a causal parameter of interest, a conditional causal odds ratio is rarely the most interpretable choice (step 1), but whether it can be interpreted causally is a different issue (step 2). Statistical estimation approaches have different strengths and weaknesses—for example, in terms of flexibility in the parameters estimated and ability to handle time-dependent confounding6,7—and these can be considered (as part of step 3) in deciding which methods are worth considering and, finally, which one is strongest for estimating the statistical parameter of interest.
In sum, I agree that researchers should articulate causal research aims. I would add the critical importance of doing this together with a structured roadmap process that avoids confusion and conflation.
Footnotes
REFERENCES
- 1.Petersen ML, van der Laan MJ. Causal models and learning from data: integrating causal modeling and statistical estimation. Epidemiology. 2014;25(3):418–426. doi: 10.1097/EDE.0000000000000078. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Heckman JJ, Vytlacil EJ. Handbook of Econometrics. Vol 6b. Amsterdam, The Netherlands: Elsevier; 2007. Econometric evaluation of social programs, part I: causal models, structural models and econometric policy evaluation; pp. 4779–4874. [Google Scholar]
- 3.Ahern J, Hubbard A. A roadmap for estimating and interpreting population intervention parameters. In: Oakes JM, Kaufman JS, editors. Methods in Social Epidemiology. San Francisco, CA: Jossey-Bass; 2017. pp. 432–457. [Google Scholar]
- 4.Robins J. A new approach to causal inference in mortality studies with a sustained exposure period—application to control of the healthy worker survivor effect. Math Model. 1986;7(9–12):1393–1512. [Google Scholar]
- 5.Rosenbaum PR, Rubin DB. The central role of the propensity score in observational studies for causal effects. Biometrika. 1983;70(1):41–55. [Google Scholar]
- 6.Ahern J, Colson KE, Margerson-Zilko C, Hubbard A, Galea S. Predicting the population health impacts of community interventions: the case of alcohol outlets and binge drinking. Am J Public Health. 2016;106(11):1938–1943. doi: 10.2105/AJPH.2016.303425. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Robins JM, Hernán MA, Brumback B. Marginal structural models and causal inference in epidemiology. Epidemiology. 2000;11(5):550–560. doi: 10.1097/00001648-200009000-00011. [DOI] [PubMed] [Google Scholar]