Comparison of Methodologies to Predict Bridge Deterioration

Abstract

Two methods for the estimation of bridge condition states were examined, one based on the Bayesian approach and the other using a binary probit model. The Bayesian approach was considered so that experts' opinions could be combined with observed data. Prior transition probabilities, based on bridge inspectors' experiences, were assumed to follow Dirichlet distribution. Observed data followed a multinomial distribution. The updated transition probabilities were used to predict bridge condition states. In the second approach, deterioration models were developed for each condition state. The dependent variable is a zero/one indicator variable for condition switching state. The binary probit models developed considered the discreteness of condition states and they explicitly linked deterioration to relevant explanatory variables. This approach also incorporated heterogeneity and state dependence due to the use of panel data. An application of these methodologies was demonstrated for substructure element using the Indiana bridge data base.