A statistical methodology for reliable evaluation of calibrated dynamic modulus regression models

ABSTRACT

Predictive models are useful tools for reliable estimations of key mechanical properties when properly calibrated. Several research efforts have compared calibrated models in the literature, but the sampling techniques adopted in the model calibration, selection, and evaluation were not the focus of these studies. This work reviews different sampling techniques and employs hold-out and repeated k-fold cross-validation (CV) to evaluate three empirical dynamic modulus equations calibrated using a database containing 1,806 records from 65 asphalt mixtures. The results indicate that hold-out can induce unrealistic conclusions about the estimations, while repeated k-fold CV is a reliable methodology.

Keywords
Asphalt mixtures; dynamic modulus; empirical predictive models; resampling methods; repeated cross-validation