Influence of air quality data estimation on short-term health effect estimates of air pollution in epidemiological studies

摘要

Background In time series studies, daily missing values of air pollutants are generally imputed, but how this affects epidemiological risk estimates is not clear. Aims To quantify the influence of missing data imputation methods on short-term health effects estimates. Materials and Methods Daily counts of natural mortality and PM10, NO2 and O3 daily concentration were collected for Milan, Rome and Palermo during 2001-2005. Simulation of missing data patterns on air pollutant were done in two scenarios by removing 10% and 20% existing values. Three different imputation methods were applied: Epiairl (E1), based on mean in the same day of the other city-specific monitoring stations, weighted by the relative annual mean between stations, Epiair2 (similar to E1, but with weights based on seasonal means) and average of previous and following day (for isolated missing values). The effect of pollutants on mortality was evaluated using Poisson regression models, and estimating percent increases of risk of natural mortality for ten unit increases in the pollutants. For each city and method, a set of 1000 simulations was performed. Results Effect estimates obtained with methods El and E2 remained close to the reference value for all the cities and pollutants investigated. The poorest result (NO2 in Palermo, under 20% missing value scenario) showed a very good agreement between the original dataset [3.32 (95% C.I. 1.33; 5.35)] and the E1 [3.24 (0.91; 5.54)] and E2 [3.29 (0.88; 5.45)] estimates (in brackets the minimum of lower 5% confidence limits and the maximum of the upper 95% of the confidence limits obtained from the simulations are reported). Results were much less precise using time series without imputing missing data or using the method of average - in the example 3.29(-1.80; 8.29) and 3.49(-0.69; 8.64). Conclusions This study adds new elements to the knowledge about the estimation of environmental missing values and mortality risk estimates. Methods currently used in environmental epidemiology are robust for different pollutants and the cities considered.