To the Editor:
We read with interest the recent article by Cefalu and Dominici1 regarding a linked statistical model for the assessment of spatially-dependent exposures and health. Our concerns center around the suggestion that this model can be applied to any epidemiologic design, the authors’ definition of confounding, that no personal risk factors were included in their health model, and that area-wide predictors for the exposure model are also confounding factors in a health model.
As the health model used normally distributed errors, we assume that the authors were referring to a cross-sectional study of continuous outcomes. Except possibly for an ecologic study, we know of no design in which personal risk factors would not be included as potential confounding variables.
In terms of what is a confounder, the authors stated “We refer to confounding bias as the bias in the health-effect estimate from a health-effects regression model that fails to control for any confounding….” This is not the accepted definition of confounding: quoting Breslow and Day from 1980,2 “Confounding is intimately connected to the concept of causality. …if some exposure E is associated with disease status, then the incidence of the disease varies among the strata defined by different level of E. If these differences in incidence are caused (partially) by some factor C, then we say that C has (partially) confounded the association between E and the disease.” We note that any noncausal variable could be associated with exposure and health, and these variables should not be included in a model to control for bias unless they are surrogates of causal processes.
In our experience, there are very few area-level variables included in exposure models that are true causal variables for health outcomes. For example, in our land-use regression model of NO23 that was used in case–control studies on breast cancer,4 the predictors of traffic-related exposure included population density, counts of traffic, and distance to roads. None of these variables are causal risk factors for these cancers. Could any of these variables represent some complex causal process that can affect the incidence of these cancers? Possibly, but one would have to postulate the purported mechanism. For example, green space may lower pollution, so that the effects of greenspace on health could be due in part to lower levels of exposure to air pollution: this is a measurement issue and is not confounding. While one might contend that these contextual variables represent causal exposures, we suggest that these variables be modeled directly.
Mark S. Goldberg
Division of Clinical Epidemiology
Department of Medicine
McGill University Health Center – RVH
Montreal, QC, Canada
Department of Health Sciences
Ottawa, ON, Canada
Department of Sociology
University of New Brunswick
Fredericton, NB, Canada
1. Cefalu M, Dominici F. Does exposure prediction bias health-effect estimation?: the relationship between confounding adjustment and exposure prediction. Epidemiology. 2014;25:583–590.
2. Breslow NE, Day NE. Statistical Methods in Cancer Research. Volume I - The Analysis of Case-Control Studies. IARC Scientific Publications, 32 (Lyon). 1980.Lyon, France: International Agency for Cancer.
3. Crouse DL, Goldberg MS, Ross NA. A prediction-based approach to modelling temporal and spatial variability of traffic-related air pollution in Montreal, Canada. Atmos Environ. 2009;43:5075–5084.
4. Crouse DL, Goldberg MS, Ross NA, Chen H, Labrèche F. Postmenopausal breast cancer is associated with exposure to traffic-related air pollution in Montreal, Canada: a case-control study. Environ Health Perspect. 2010;118:1578–1583.