Explaining Heterogeneity in Pavement Deterioration: Clusterwise Linear Regression Model

Abstract

A clusterwise linear regression model of pavement deterioration is presented. The model provides a framework to simultaneously segment a population and to describe performance with a set of regression models, one for each segment. Instead of relying on observed criteria, the objective in the segmentation is to maximize within-segment variation explained by a set of commonly specified regression models. To illustrate the methodology, performance models were estimated for a panel of 131 pavements from the American Association of State Highway Officials (AASHO) road test. Pavements in different segments display systematic but unobserved differences in their responses, i.e., unobserved heterogeneity, which manifests itself in segment-level coefficients that differ in their magnitude and sign. This is radically different than other approaches in the literature that explain such differences with individual-level error/intercept terms, but that rely on the assumption of constant and homogeneous coefficients capturing the effect of explanatory variables across the population. How segment-level effects can be used to support tailored management policies, e.g., maintenance and repair, or setting weight restrictions is discussed. Finally, a rigorous assessment is conducted of the proposed model that includes comparison with a population-level regression model and to a clusterwise model that relies on observed factors from the original experimental design: the loop-lane, i.e., the design-loading, configuration.