Abstracts: ISEE 21st Annual Conference, Dublin, Ireland, August 25–29, 2009: Oral Presentations
Many studies world-wide have demonstrated the acute health effects of air pollution. Time series approaches have been most favoured because of the relative ease of calculation and data availability. In time series studying daily mortality only time-varying factors can confound the results. Seasonal trends and indicators of short term weather variability (temperature, air pressure, and humidity) are the most important.
In recent years a lively debate has sparked over the best way to model these possible confounders since their association with mortality does not follow a simple linear form. The splines modelled under the S-Plus GAM (General Additive Model) routine have been criticised as leading to an overestimation of the effects of the linear terms (i.e. the pollutants' concentrations).
Various alternative models have been proposed using more stringent criteria and fully parametric models. Also the degrees of freedom reserved for the modelling of the seasonal fit have been lively debated.
The city of Vienna has run an air pollution monitoring system with consistent methodology for the gaseous pollutants since the early 1990s. In a shorter time series (since 2000) we have recently shown risk estimates for NO2 and particles that were in accordance with international findings.
In a sensitivity analysis we calculated risk estimates for NO2 and O3 for total daily mortality using GAM (with varying degrees of freedom for the seasonal fit and with inclusion of a varying number of possible confounding variables), General Estimation Equations (with and without auto-regression terms), Poisson regression, ARIMA model, and case-crossover design. Results not so much depended on the statistical model but on the list of confounders. Daily weather patterns influence mortality directly. Additionally there could be an indirect effect by their influence on pollutants' concentration. Consequently an overly exhaustive list of possible confounding variables might lead to an over-adjustment in the model.