Using spatio-temporal land use regression models to address spatial variation in air pollution concentrations in time series studies

Time series studies are used to assess the effects of short-term exposures to PM10 and NO2 on mortality using an integrated pollutant series taken to characterize exposure over a large area. We propose using spatio-temporal land use regression (LUR) models by smaller geographical sectors within an area of interest to account for spatial variability in these studies. Based on model-estimated time series, we conducted a case-crossover analysis for each sub-sector within two larger areas of interest (Athens and Thessaloniki, Greece) separately to investigate heterogeneity and provide combined results if appropriate. As sensitivity analysis, we compared the case-crossover method to classical time series analysis and also to using fixed site measurements only. For PM10 exposures in Athens, we found consistent adverse effects which were larger when using spatio-temporal LUR modeled concentrations (total mortality RR 2.55 and 95% CI - 0.30 to 5.39) compared to measurements (RR 0.36 and 95% CI - 0.21 to 0.93). For NO2, we found a similar magnitude in the effects, when using measurements from fixed sites (RR 0.81 and 95% CI 0.39 to 1.22) and modeled levels (RR 0.71 and 95% CI 0.14 to 1.28). Analysis by geographical sector did not add information over a unified analysis for the whole area. The effect estimates using classical Poisson regression time series yielded consistently smaller size effects compared to the case-crossover method. Our analysis demonstrates the potential of using spatio-temporal models in time series analysis for short-term air pollution effects to account for spatial variability in addition to the temporal.