From the aDivision of Environmental Epidemiology, Institute for Risk Assessment Sciences, Utrecht University, Utrecht, The Netherlands; and bMRC/HPA Center for Environment and Health, School of Public Health, Imperial College, London, United Kingdom.
Correspondence: Roel Vermeulen, Utrecht University, Jenalaan 18d, 3584 CK, Utrecht, NL. E-mail: R.C.H.Vermeulen@uu.nl.
Risk of chronic disease may be governed not only by cumulative levels of exposure but also by dynamic aspects of the exposure history. These dynamic aspects include characteristics of the exposure history itself (such as duration of exposure or time-varying intensities of exposure) as well as aspects of age-related susceptibility. The latter has been formalized in concepts of the exposome1 and life-course epidemiology,2 in which the risk of chronic disease can vary according to exposures at critical time-periods in life.
The approach proposed by Richardson et al3 in this issue of Epidemiology refines well-established regression models to efficiently account for exposure increments and their potential impact on disease risk. The proposed refinements lead to a flexible model that can accommodate between-person variability in exposure as well as within-person variation in exposure over time, thereby disentangling the effects of cumulative exposure and exposure rate. The method presents the great advantage of being readily implemented using classical software, allowing estimation of familiar statistical measures of association and easily interpretable results. As the authors illustrate with data from a cohort study of radon exposure in uranium miners and lung cancer mortality, this model represents an efficient extension of classical regression models of disease risk, accounting for the impact of full exposure history at the individual level.
Results from the example suggest a positive association between lung cancer risk and cumulative radon exposure, with effect modification by an inverse-exposure-rate effect. These results suggest that dynamics in individual radon exposure histories are indeed important in modeling disease risk. Similar observations have been made for several other exposures, including arsenic, cigarette smoke, and alcohol—although these studies did not account for within-person variations in exposure. The model proposed by Richardson and co-workers3 is an elegant and efficient extension of this previous work.
There are a few methodological aspects that may need further consideration. For example, such models require high-quality longitudinal exposure data. Although exposure assessment methods have certainly improved over time, most exposure measures are undoubtedly inexact. Where measurement errors (or errors in important covariates) are differential by exposure intensity (rate) or over time, this could lead to biased risk estimates in dynamic models of exposure. Given the high quality of historical exposure data for radon and also for smoking, those exposure examples are potentially best-case scenarios (certainly better than most occupational exposures).
A second consideration is that the suggested models are strongly parametric. Although the authors mention the possibility of accommodating nonlinear relationships in their model (through spline functions, for instance), we can expect that allowing flexible parametric forms for each of the risk determinants would lead to complex models and potential computational difficulties.
Notwithstanding these methodological considerations, the proposed model provides an example of how such models can offer a richer description of epidemiologic associations. Such insights may be important when risk assessments are based on epidemiologic results that assess cumulative exposures without consideration of exposure patterns or age-related susceptibility. If the risk with two equal cumulative exposures differs with low intensity and long duration compared with high intensities and short duration, this could lead to different conclusions about exposure limits and health. The same would be true if exposures after a certain age were associated with higher risk than similar exposures at a younger age (observed for benzene and leukemia, for example).4 Models that encompass these dynamic aspects should therefore be encouraged in risk modeling.
Such models may also provide information about biologic pathways of disease, leading to better understanding of (for instance) the impact of metabolic saturation on the observed exposure–response curve, or the natural progression of the disease. In this regard, it might be useful to introduce into chronic disease epidemiology the types of dynamic models developed for infectious disease (for example, Susceptible-Infectious-Recovered [SIR] models, which aim at the estimation of the parameters driving the transition dynamic between these three states). These approaches focus on the reconstruction of the natural history of disease at the individual (or population) level. In practice, these would be based on the same type of data as those presented in the present study, and would complement the results from regression-based models by quantifying the effect of risk determinants in terms of their abilities to reconstruct the dynamics of disease progression in each subject. Such an approach would also allow the identification of features of the exposure history that contribute to the individual (longitudinal) risk profiles.
In summary, the proposed work constitutes a convincing, important, and timely extension of ways to model the risk of chronic disease, accounting for the role of dynamic patterns in the exposure history. We encourage further applications that utilize the rich information potentially hidden in individual exposure histories.
ABOUT THE AUTHOR
ROEL VERMEULEN is an Associate Professor of Molecular Epidemiology and Risk Assessment at Utrecht University, the Netherlands. His research focuses on quantitative exposure assessment, molecular epidemiology and risk modeling. MARC CHADEAU-HYAM is a Lecturer at The MRC/HPA Centre for Environment and Health, School of Public Health at Imperial College, London. He is working on the development of statistical models for the analysis of high dimensional biological data.
1. Wild CP. Complementing the genome with an “exposome”: the outstanding challenge of environmental exposure measurement in molecular epidemiology. Cancer Epidemiol Biomarkers Prev. 2005;23:1847–1850
2. Ben-Shlomo Y, Kuh D. A life course approach to chronic disease epidemiology: conceptual models, empirical challenges and interdisciplinary perspectives. Int J Epidemiol. 2002;31:285–293
3. Richardson DB, Cole SR, Langholz B. Regression models or the effect of exposure rate and cumulative exposure. Epidemiology. 2012;23:892–899
4. Richardson DB. Temporal variation in the association between benzene and leukemia mortality. Environ Health Perspect. 2008;116:370–374
5. Chadeau-Hyam M, Clarke PS, Guihenneuc-Jouyaux C, Cousens SN, Will RG, Ghani AC. An application of hidden Markov models to the French variant Creutzfeldt–Jakob disease epidemic. Appl Statist. 2010;59:839–853