Abstract: We need structural models of traffic congestion to answer a wide variety of questions, but the standard models fail to match the data on travel times across the day. I establish the nature and magnitude of the problem, and show its source lies in how we model agent preferences, not in the specifics of the congestion technology. The poor fit of the models suggests that we are abstracting away from features with a first-order impact on model predictions, which limits our ability to use these models to evaluate counterfactuals quantitatively and---when agents are heterogeneous---qualitatively as well. I explore several ways of improving the fit of these models, concluding with recommendations for tractable and intuitive ways of doing so.
Keywords: Structural model; Congestion; Model fit; Calibration; Dynamic; Bottleneck Model; Traffic
JEL Classification: R4; H4