Flow frequency analysis (FFA) is a widely used statistical method in riverine engineering and science applications. While the standard approach relies on annual-maximum series (AMS) of river discharge, the partial-duration series (PDS) offers advantages for certain applications, such as seasonal flooding, mixed flood populations, or analyzing lower-magnitude events. PDS-based FFA, however, is less frequently used due to the subjective choices involved in selecting thresholds and determining when flood events are independent.

This study develops a Bayesian model for flood frequency that addresses those limitations by incorporating the uncertainty associated with threshold selection into the final flood magnitude estimates. We first developed and validated the model using synthetic flood records with known properties. We then compared its performance to three common approaches in FFA: the method of moments, L-moments, and maximum likelihood estimation. Finally, we applied our model to real-world discharge records from several USGS gage stations.