Comparison of models for estimation of long-term exposure to air pollution in cohort studies

This study compared three spatio-temporal models for estimation of exposure to air pollution throughout the central part of Scotland during 1970–79 for approximately 21,600 individuals in 2 closely-related prospective cohort studies. Although 181 black smoke (BS) monitoring sites operated in this region at some point during 1970–79, a substantial amount of BS exposure data was missing at many sites. The three exposure estimation methods were: (i) area-based regression models to impute missing data followed by assignment of exposure by inverse distance weighting of observed BS at nearby monitoring sites (IDWBS); (ii) area-based regression models to impute missing data followed by a spatial regression additive model using four local air quality predictors (LAQP): altitude; distance to the nearest major road; household density within a 250 m buffer zone; and distance to the edge of urban boundary (AMBS); (iii) a multilevel spatio-temporal model using LAQP (MultiBS). The three methods were evaluated using maps of predicted BS, and cross validation using monitored and imputed BS at sites with ≥80% data. The use of LAQP in the AMBS and MultiBS exposure models provided spatial patterns in BS consistent with known sources of BS associated with major roads and the centre of urban areas. Cross-validation analyses demonstrated that the MultiBS model provided more precise predictions (R2 = 60%) of decadal geometric mean BS concentrations at monitoring sites compared with the IDWBS and AMBS models (R2 of 19% and 20%, respectively). ► Three approaches for modelling long-term exposure to black smoke were evaluated. ► Missing data require development of imputation procedures. ► Cross validation & GIS-based visualisation were used for model evaluation. ► Marked differences in performance were noted in evaluation of different models. ► Improved black smoke exposure estimates were observed using a multi-level model.