Continuous integration of data into ground-motion models using Bayesian updating

The development of empirically constrained ground-motion models has historically followed a cyclic process in which every few years, existing models are updated to reflect new data and knowledge that has become available. Ground-motion developers make use of their prior knowledge to identify appropriate functional forms for the models, but the actual regression analysis and model calibration is effectively performed from a fresh start with each update. With the anticipated increase in data availability coming in the future, this traditional approach will become increasingly cumbersome. The present article presents a framework in which Bayesian updating is used to continuously update existing ground-motion models as new data becomes available. This new approach is shown to provide similar results to the more traditional approach, but is far less data-intensive and will scale well into the future. The approach is demonstrated through an example in which an initial regression analysis is conducted on a portion of the NGA-West2 dataset representative of the information available in 1995. Model parameters, variance components and crossed random effects are then updated with data from every other event in the NGA-West2 dataset and the results from Bayesian updating and traditional regression analysis are compared. The two methods are shown to provide similar results, but the advantages of the Bayesian approach are subsequently highlighted. For the first time, the article also demonstrates how prior distributions of model parameters can be obtained for existing ground-motion models that have been derived using both classical, as well as more elaborate multi-stage, procedures with and without constrained parameters.