Quantifying Uncertainty with Pavement Performance Models: Comparing Bayesian and Non-Parametric Methods

An important part of pavement management systems is accurately estimating the performance-time-degradation relationship. One common approach to establishing this relationship is to use performance family curves. These curves are developed by collecting performance data at specific points in time and collectively shifting pavements of various ages to identify the probable underlying function. This paper compares two alternative methods for characterizing such a family curve function. First, a Bayesian method (Method-A) is used, which fits both the family curve and the shift factor function in parallel by assuming a Beta distribution for pavement performance condition rating (PCR). Second, a non-parametric method (Method-B) is developed, which fits the model in two steps; (1) by fitting the family; and (2) by horizontal shift to minimize the error. PCR values from flexible pavements in North Carolina (NC-PCR) are used for this comparison. These data include a total of 30,988 pavement sections segregated according to surface type and traffic level. Data from 2013 to 2015 are used for model calibration, and data from 2016 are used for model validation. The root means square error and k-fold cross-validation test are used to conduct the comparison, and Method-A is found to be preferred. The uncertainty in both models is quantified and compared. On the basis of this uncertainty, the Bayesian method is preferred, but in cases with large data sets, a non-parametric method does result in lower uncertainty.