A probabilistic methodology for estimating transportation project costs using Bayesian models
The purpose of this study is to enhance the estimation of the aggregate planning phase costs of transportation facilities. Generally, during the planning phase of transportation facility development, cost-related data are limited and the project scope is often not well defined. As a result, the cost estimation of transportation projects at this phase has been largely deterministic and based on data from historical similar projects. However, the variability associated with project conditions leads to significant uncertainty and error in project cost estimates that are developed in this manner. Thus, actual project costs often end up very different from the planning phase predictions, thus posing a serious finance-related risk in highway management. Unfortunately, the current risk assessment and analysis tools that are intended to address such problems are impractical and are seldom used in the industry because they rely on historical data, require detailed project information often unavailable during the planning phase, and fail to exploit the available expert knowledge of cost factors that can often explain cost deviations. ^ To address this issue, this dissertation shows how agencies could conduct a risk-based probabilistic analysis of highway project costs using Bayesian statistical techniques with limited data. Four stochastic cost models based on Bayesian statistics are introduced here: the One-State Bayesian model, the Hierarchical Bayesian model, the General Bayesian Regression model, and the Random-Effect Bayesian Regression model. Each of these models is applicable at different levels of data availability and can estimate the point estimate, range estimate, and full probability distribution of project costs. Unlike the traditional Frequentist modeling structures, the Bayesian cost models can amalgamate information from historical data and expert opinion into the analysis simultaneously using probability theory. This approach thereby helps in reaping the benefits associated with each of these two information sources, particularly where each source alone is inadequate to develop a reliable cost estimate. This dissertation demonstrates the application of a new cost estimation approach using a case study involving functional hot-mix asphalt overlay data from Indiana. The Bayesian models are used to fit and predict the cost of this highway maintenance treatment. In contrast to traditional (Frequentist) cost estimation that yields a fixed value for each cost parameter, the Bayesian approach describes all such parameters using probability distributions, thus allowing the assignment of probabilities, rather than fixed values, to model parameters, confidence intervals, cost elasticity, and other model outputs. This dissertation also studies the effects of incorporating expert opinion into cost estimation analysis using data from the case study and is able to show that expert opinion can and does help improve the precision of model estimation and prediction. The application of the developed framework for Bayesian cost estimation is not only limited to transportation planning but also can be used for cost estimation at other phases of transportation project development and also in other sectors.^
Samuel Labi, Purdue University.