Objective: To compare the performance of frequentist and Bayesian generalized additive models (GAMs) in terms of accuracy and precision for assessing the association between daily exposure to fine particles and respiratory mortality using simulated data based on a real time-series study. Methods: In our study, we examined the estimates from a fully Bayesian GAM using simulated data based on a genuine time-series study on fine particles with a diameter of 2.5 μm or less (PM2.5) and respiratory deaths conducted in Shanghai, China. The simulation was performed by multiplying the observed daily death with a random error. The underlying priors for Bayesian analysis are estimated using the real world time-series data. We also examined the sensitivity of Bayesian GAM to the choice of priors and to true parameter. Results: The frequentist GAM and Bayesian GAM show similar means and variances of the estimates of the parameters of interest. However, the estimates from Bayesian GAM show relatively more fluctuation, which to some extent reflects the uncertainty inherent in Bayesian estimation. Conclusions: Although computationally intensive, Bayesian GAM would be a better solution to avoid potentially over-confident inferences. With the increasing computing power of computers and statistical packages available, fully Bayesian methods for decision making may become more widely applied in the future.
Keywords: Bayesian statistics; fine particulate matter; generalized additive model; respiratory mortality; time-series analysis.