To the Editor:
In the last issue of EPIDEMIOLOGY, Keyes and Galea1 encouraged us to pause for a moment and consider whether EPIDEMIOLOGY is moving in the right direction. Mediation analysis was one of the methods brought forward as a possible way out of what they call the risk factor trap. The recent methodologic developments in the mediation analysis literature have been welcomed with enthusiasm as a way to describe and understand mechanistic processes by statistical analysis of empirical data, see, e.g., the recent book by Vanderweele2 and the many references therein. We have, along with numerous co-authors and co-researchers, been involved in both the methodologic developments3–7 and in applications of this methodology,8–10 but we have been disappointed in the scientific insights obtained by mediation analysis.
DO WE WANT TO UNDERSTAND MECHANISMS OR QUANTIFY PATHWAYS?
Taking social Epidemiology as an example, there has been a long-standing scientific interest in understanding the mechanisms underlying social inequality in health, and mechanisms such as stress, adverse health-related behavior, and suboptimal living conditions have been suggested as key explanations.11 This understanding of mechanisms is well founded on a range of large well-conducted studies of social inequality in, e.g., smoking, which is a well-known risk factor for a range of diseases. So, unless we come up with completely new candidates for key mechanisms, the improved models for mediation analyses could not truly improve our “understanding” of which mechanisms underlie social inequality in health. Even so, there has been a major interest in using mediation analyses in social Epidemiology with a specific aim to “quantify” the effect mediated by one of these well-known mechanisms. Changing the perspective from understanding to quantification might seem like a minor distinction, but it may potentially have a major impact on the usefulness of mediation methods with the available data.
REFINED MODELS APPLIED TO CRUDE DATA
Mediation analysis was partly developed and is often applied as part of clinical experiments, where the aim is to understand the mechanisms through which a potential intervention or treatment works. The aim of understanding the underlying mechanisms is in principle the same in observational Epidemiology. The main difference is the availability of data. While most clinical experiments follow a smaller group of people closely (i.e., high temporal resolution data), Epidemiology studies often follow a large group of people rather infrequently (i.e., low temporal resolution data). With the available data, we therefore often need to limit ourselves to a single (or very few) measure(s) of the mediator. It is well known that health-related behavior fluctuates over time and ignorance of such time variations may severely affect our ability to quantify the mediated effect. Thus, we question the impact of mediation analyses in Epidemiology before high temporal resolution mechanistic data are collected and incorporated into mediation analysis.
We find that the high hopes originally attached to better tools for mediation analyses have not yet materialized in corresponding scientific insights. The insufficiently rich data have been the main barrier. The good news is that high resolution data based on repeated measures on potential mediators in Epidemiology are gradually becoming more common; examples include the French CONSTANCES cohort, which includes annual measurements on approximately 200,000 people.12 In order to utilize the important methodologic developments in mediation methodology, we need to ensure the collection of such high temporal data in the future.
Naja Hulvej Rod
Department of Public Health
University of Copenhagen
Department of Public Health
University of Copenhagen
Center for Statistical Science
1. Keyes KM, Galea S. Commentary: the limits of risk factors revisited: is it time for a causal architecture approach? Epidemiology. 2017;28:1–5.
2. Vanderweele TJ. Explanation in Causal Inference. 2015.New York; Oxford University Press.
3. Lange T, Hansen JV. Direct and indirect effects in a survival context. Epidemiology. 2011;22:575–581.
4. Lange T, Vansteelandt S, Bekaert M. A simple unified approach for estimating natural direct and indirect effects. Am J Epidemiol. 2012;176:190–195.
5. Lange T, Rod NH. Klein J, Scheike T, van Houwelingen H, Ibrahim J. Causal models. In: Handbook of Survival Analysis. 2013:Boca Raton, FL: Chapman & Hall/CRC; 135–151.
6. Vansteelandt S, Bekaert M, Lange T. Imputation strategies for the estimation of natural direct and indirect effects. Epidemiol Methods. 2012;1:131–158.
7. Lange T, Rasmussen M, Thygesen LC. Assessing natural direct and indirect effects through multiple pathways. Am J Epidemiol. 2014;179:513–518.
8. Nordahl H, Lange T, Osler M, et al. Education and cause-specific mortality: the mediating role of differential exposure and vulnerability to behavioral risk factors. Epidemiology. 2014;25:389–396.
9. Nordahl H, Rod NH, Frederiksen BL, et al. Education and risk of coronary heart disease: assessment of mediation by behavioral risk factors using the additive hazards model. Eur J Epidemiol. 2013;28:149–157.
10. Hvidtfeldt UA, Lange T, Andersen I, et al. Educational differences in postmenopausal breast cancer–quantifying indirect effects through health behaviors, body mass index and reproductive patterns. PLoS One. 2013;8:e78690.
11. Marmot M, Wilkinson RG. Social Determinants of Health. 2006.2nd ed. New York: Oxford University Press.
12. Zins M, Goldberg M; CONSTANCES Team. The French CONSTANCES population-based cohort: design, inclusion and follow-up. Eur J Epidemiol. 2015;30:1317–1328.