A statistical model has been developed to decompose the overall estimation error in link loads predicted with the usual traffic assignment procedures. This decomposition is derived by comparing load estimates from assignments performed under varying input conditions, that is with different assignment models and levels of detail. The absolute and relative contibution of the potential error sources to the total error-sum-of squares has been estimated for the primary roads of the car network of Eindhoven. Special attention is given to the bias in the load predictions. From the empirical findings it can be learned that assignment load outcomes are very sensitive to the assignment model type as well as level of spatial detail, whereas the trip input (if unbiassed) is less important.
Abstract