Ecological bias in environmental health studies: the problem of aggregation of multiple data sources

Abstract

Ecological bias may result from interactions between variables that are characterized by different spatial and temporal scales. Such an ecological bias, also known as aggregation bias or cross-level-bias, may occur as a result of using coarse environmental information about stressors together with fine (i.e., individual) information on health outcomes. This study examines the assumption that distinct within-area variability of spatial patterns of the risk metrics and confounders may result from artifacts of the aggregation of the underlying data layers, and that this may affect the statistical relationships between them. In particular, we demonstrate the importance of carefully linking information layers with distinct spatial resolutions and show that environmental epidemiology studies are prone to exposure misclassification as a result of statistically linking distinctly averaged spatial data (e.g., exposure metrics, confounders, health indices). Since area-level confounders and exposure metrics, as any other spatial phenomena, have characteristic spatiotemporal scales, it is naively expected that the highest spatial variability of both the SES ranking (confounder) and the NOx concentrations (risk metric) will be obtained when using the finest spatial resolution. However, the highest statistical relationship among the data layers was not obtained at the finest scale. In general, our results suggest that assessments of air quality impacts on health require data at comparable spatial resolutions, since use of data layers of distinct spatial resolutions may alter (mostly weaken) the estimated relationships between environmental stressors and health outcomes.