Can We Reliably Use Project-Level Pavement Performance Models in Place of Family Models?

Abstract: Performance modeling is a fundamental process in pavement management, and inaccurate performance models can lead to inefficiencies in many important pavement management practices. Given the significant number of pavements in any given pavement network, decisions must be made about how to define the subsets of condition data to use in developing each performance model. On one extreme, all data can be combined into a single performance model, but that would result in significant biases and poor prediction accuracy. For the other extreme, individual models can be developed for individual pavement segments, but that has not been historically done due to sparsity of data and significant measurement errors. Thus, some intermediate solution is most often followed where pavements are aggregated into families that are expected to perform similarly, and models are developed for those families. However, with many agencies collecting high quality data on an annual basis for the past several years, the question of how best to use individual project data to inform performance models should be revisited. This paper details many important considerations for modeling performance data at both the project and family level. For example, the same statistical model cannot be used to model project and family curves, and the same quality of fit metric cannot be used. Biases can also be significant when extrapolating project-level performance models. However, combining project and family models shows potential for significantly improving the prediction accuracy of performance models. Examples are demonstrated using historical condition data from multiple agencies and for multiple condition metrics.