You are here:
Over-fitting Time Series Models of Air Pollution Health Effects: Smoothing Tends to Bias Non-Null Associations Towards the Null.
Citation:
Neas, L., J. Sacks, AND A. Rappold. Over-fitting Time Series Models of Air Pollution Health Effects: Smoothing Tends to Bias Non-Null Associations Towards the Null. Presented at International Society for Environmental Epidemiology, August 26 - 30, 2012.
Impact/Purpose:
to be submitted to a conference
Description:
Background: Simulation studies have previously demonstrated that time-series analyses using smoothing splines correctly model null health-air pollution associations. Methods: We repeatedly simulated season, meteorology and air quality for the metropolitan area of Atlanta from cyclic functions and random variates. We next simulated 500 replicates of daily mortality based on fixed parametric models for season, meteorology and air quality as determinants of the mortality rate. Finally, we applied standard time-series models with smoothing splines for time and temperature to estimate linear pollutant effects on mortality. Results: The simulated environment has similar means, variance, autocorrelations, and inter-correlations to observed data for Atlanta from 2001-2005. When the true underlying functions linking mortality with season and meteorology were known and correctly parameterized, epidemiologic models correctly estimated both null and non-null associations of pollutants with mortality. Epidemiologic models using smoothing splines for time and temperature correctly estimated null-associations, but consistently underestimated non-null associations. The magnitude of this bias appeared to depend less on the degree of smoothing and more on the inherent association of a specific pollutant with season and meteorology. For example, prior non-null associations for ozone were, on average, reduced 50% while the associations for fine particulates were reduced only 20%. Conclusions: When the true underlying functions linking mortality with season and temperature are unknown, estimation of these functions with smoothing splines will tend to bias non-null associations between air pollutants and mortality towards the null. This abstract of a proposed presentation does not necessarily represent EPA policy.