A multilevel Poisson regression model to detect trends in frequency of extreme rainfall events

This study relies on a systematic casual learning to investigate the time space structure of extreme events, over 115 years of rainfall data, in 1254 high quality stations across the United States. The model developed is a partial pooling Bayesian model which is hierarchically characterized by regional properties of each stations. The random variable applied in the model is the number of days with rainfall more than 95-percentile of the whole period of record, aggregated within each year. In this study, ranges of hypotheses tests are considered to explain the time trend using cyclical variability from climate. The results suggest that low-frequency climate variability can justify the temporal trends of extreme events with moderate confidence.