Abstract

Introduction In air-pollution epidemiology, measured or modelled surrogate exposure estimates, prone to measurement error (ME), are used to investigate the health effects of exposure to pollution of outdoor origin, potentially leading to biased effect estimates. We predicted the annual personal exposure from outdoor sources by using personal measurements, compared it with concentrations from surrogate metrics, and quantified the ME magnitude, type, and determinants. Methods We used measurements from four panel studies in London, UK, and predicted personal exposures to fine particulate matter (PM2.5), nitrogen dioxide (NO2), ozone (O3), and black carbon (BC). We compared those with surrogate exposures, including measurements from fixed-site monitors, modelled ambient concentrations, or hybrid methods accounting for people’s mobility. We estimated the exposure ME magnitude, correlations, and variance ratios between surrogate measures and personal exposure, and the percentages of classical/Berkson-type errors. Individual- and area-level characteristics, such as age, sex, socio-economic status, and time spent outdoors, were assessed as potential error determinants. Results Predicted annual personal exposures to PM2.5, NO2, O3, and BC from outdoor sources were overestimated by surrogate metrics, with mean differences of up to 10.1, 40.0, 61.7, and 2.6 μg/m3, respectively. The variance ratios and Pearson correlation coefficients between surrogate and predicted personal exposures ranged from 0.03 to 165.02 and –0.24 to 0.25. Time–activity adjustment reduced errors substantially. Berkson-type errors dominated the ME for PM2.5 and BC (43%–81% and 26%–98%, respectively), whilst classical errors characterized gases (>94% for both NO2 and O3). Time spent outdoors, house type, and deprivation were associated with exposure error. Conclusion The use of surrogate exposures to investigate the health effects of long-term exposure to air pollution from outdoor sources may bias the epidemiological estimates due to ME. Information about the error structures and their determinants can be used for correction and the identification of the true exposure–response functions.