Department of Biostatistics Seminar/Workshop Series

A Poisson Bayesian Kernel Model for the Prediction of Disruptions

Hiba Baroud, PhD

Assistant Professor, Civil and Environmental Engineering, Vanderbilt University

The protection of critical infrastructures has recently garnered attention with an emphasis on analyzing the risk and improving the resilience of such systems. With the abundance of data, risk managers should be able to better inform decision making under uncertainty. It is important, however, to develop and utilize the necessary methodologies that bridge between data and decisions.

In this talks, I present a data-driven analysis of the risk critical infrastructure systems face. I have developed a new Bayesian kernel model to predict the frequency of failures. The integration of Bayesian and kernel methods allows for a classification algorithm which provides probabilistic outcomes as opposed to deterministic outcomes. These models were developed for Gaussian distributions and later extended to other continuous probability distributions. My research develops a Poisson Bayesian kernel model to accommodate count data.

The model is applied in the analysis of an inland waterway, the Mississippi River Navigation System, to predict the number of lock outages. Inland waterways are critical elements in the nation’s civil infrastructure and the world’s supply chain. They allow for a cost-effective flow of approximately $150 billion worth of commodities annually across industries and geographic locations in the U.S., which is why they are called “inland marine highways.” Aging components (i.e., locks and dams) combined with adverse weather conditions, affect the reliability and resilience of inland waterways. Frequent disruptions and lengthy recovery times threaten regional commodity flows, and more broadly, multiple industries that rely on those commodities. The accurate prediction of future disruptive events assists policymakers and risk managers in the effective allocation of resources for preparedness and recovery strategies. The proposed Poisson Bayesian kernel model resulted in a better prediction accuracy than other classical statistical tools for modeling count data.