Bayesian analysis of PM10 concentration by spatio-temporal ARIMA and STS models

Po Valley is well known to be one of the most polluted area in Italy, due to its population density, shape and climate. In this work we focus on Emilia Romagna region and analyse the panel data of its daily PM10 concentrations collected at 49 monitoring stations in 2018. We use different Bayesian spatio-temporal models. Specifically, we model the data time series following two popular techniques: structural time series (STS) and autoregressive integrated moving-average (ARIMA) process. Then in both cases we complement the model with some geographical and topographical covariates, a trigonometric seasonal component and a latent spatial Gaussian process. Based on the posterior inference made with the Stan software, the estimates of the effects of the station features on PM10 concentrations are robust with respect to the time trend modeling choice, but the STS strategy performs better than ARIMA process in fitting PM10 data.