A hidden Markov model for pollutants exceedances counts
Francesco Lagona, Antonello Maruotti · Iris (Roma Tre University) · 2009
Pollutant exceedances at the sites of a monitoring network are modeled via a hidden Markov model (HMM), where state transition probabilities depend on meteorological covariates and the observations are modelled in a generalized linear model framework, whose parameters depend on the hidden states. The estimated hidden states summarize the shape of the multinomial distribution of the pollutants at each time; define a model-driven air quality index and provide some useful insights into the processes driving pollutant exceedances. Model estimation is carried out by using a recursive forward-backward procedure, by taking a maximum likelihood approach. Parametric bootstrap is exploited to compute variances of parameter estimates. We model sequences of several air pollutants from monitoring stations in Rome.