Changes in environments over 20 years and the assessment of individual long-term exposure to air pollution for epidemiology

• Geographic variables are key inputs for air pollution prediction models. • Temporal changes in geographic features can affect prediction. • The addition of temporal changes improves prediction ability. • The improvement is substantial for long-term exposure and at residences. • Inclusion of temporal changes is crucial in areas with rapid urbanization. To address the unavailability of air pollution measurements for individuals, epidemiological studies have estimated air pollution at people’s residences or workplaces using exposure prediction models consisting of numerous geographic characteristics. However, these characteristics that represent potential pollution sources are typically fixed in time, ignoring temporal changes. Such ignorance may affect the accuracy of exposure prediction and the validity in health analyses. This study compared the predictive performance of time-aligned models (TAM) including temporally-varying geographic characteristics with time-constant models (TCM) comprised of fixed-year characteristics over 20 years for 2001–2019 in South Korea. Using regulatory air quality monitoring data from 136 to 422 sites, we computed annual-average concentrations of particulate matter ≤ 10 and 2.5 μm per diameter (PM 10 and PM 2.5 ), nitrogen dioxide (NO 2 ), and ozone. We also computed about 300 geographic variables for 2001, 2005, 2010, and 2015. Then, we developed two types of prediction models in universal kriging including a few predictors estimated by partial least squares regression from geographic variables constructed in four years (TAM) and in 2010 only (TCM). TAM showed generally higher cross-validation (CV) R 2 s than TCM across all four pollutants. The outperformance of TAM was notable with 3–15 higher percentage points (TCM and TAM CV R 2 : 0.45 and 0.48 for PM 2.5 ; 0.35 and 0.50 for PM 10 ) in long-term average predictions. Moreover, predictions at residential locations gave higher accuracy with ∼ 52 percentage points higher R 2 s in TAM (CV R 2 : 0.52) than TCM (0.00), when evaluated using measurements at the nearest regulatory monitoring sites. Our findings suggest the significant role of temporal changes of geographic characteristics in estimating long-term exposure to air pollution of individuals for epidemiological studies.