Association entre l’exposition à la pollution atmosphérique et la santé : utilisation des séries chronologiques

During the past twenty years, short-term relationships between air pollution and health have been largely investigated, mainly through time series studies. Their aim is to estimate the association between daily levels of an air pollution indicator and daily numbers of a health event (death, hospital admission, etc.). In order to get a non-biased estimation of this short term association, all factors which can modify this relationship have to be taken into account (these factors are related to air pollution and health events). The current approach consists of using a Poisson regression based on a Generalized Additive Model (GAM). This model fits nonparametric functions to allow for nonlinear effects and provide a better estimation of the relationships between the different variables. The following factors are included in the model : long-term trend and seasonality, day of the week, holidays, vacations, temperature, influenza epidemics, pollen counts, and the indicator of air pollution. A quasi-Poisson distribution allows taking into account the frequently observed over dispersion of the health data series. The parameters of the smoothing function (penalized spline) used for trend and seasonality are chosen in order to minimize the partial autocorrelation of the model residuals. This model allows to estimate the parameter of the pollutant in this multivariate model and to estimate a relative risk. Repeating such a study on successive periods is very useful in order to provide a real epidemiologic surveillance on the health risks related to air pollution, based on routinely registered data.