A Bayesian hierarchical model for monthly maxima of instantaneous flow

TL;DR

This paper introduces a Bayesian hierarchical model for monthly maxima of river flow, utilizing Gumbel distribution and Gaussian latent models, with PC priors for robust inference, applied successfully to Icelandic catchments.

Contribution

The paper presents a novel Bayesian hierarchical framework with PC priors for modeling monthly maxima of river flow, improving data utilization over traditional annual maxima approaches.

Findings

Good predictive performance in cross-validation

Effective modeling of seasonal dependence and covariates

Robust inference with PC priors

Abstract

We propose a comprehensive Bayesian hierarchical model for monthly maxima of instantaneous flow in river catchments. The Gumbel distribution is used as the probabilistic model for the observations, which are assumed to come from several catchments. Our suggested latent model is Gaussian and designed for monthly maxima, making better use of the data than the standard approach using annual maxima. At the latent level, linear mixed models are used for both the location and scale parameters of the Gumbel distribution, accounting for seasonal dependence and covariates from the catchments. The specification of prior distributions makes use of penalised complexity (PC) priors, to ensure robust inference for the latent parameters. The main idea behind the PC priors is to shrink toward a base model, thus avoiding overfitting. PC priors also provide a convenient framework for prior elicitation based on simple notions of scale. Prior distributions for regression coefficients are also elicited based on hydrological and meteorological knowledge. Posterior inference was done using the MCMC split sampler, an efficient Gibbs blocking scheme tailored to latent Gaussian models. The proposed model was applied to observed data from eight river catchments in Iceland. A cross-validation study demonstrates good predictive performance.