A fundamental input to transport planning is the representation of different economic agents' behaviour. Stated choice methods have become an accepted and widely used standard when attempting to represent behaviour of these agents. One of the more exciting recent developments in this field is a re-think of the basis of experimental designs, with the previous method of orthogonal designs losing favour to a range of new approaches. These new methods offer the opportunity to reduce sample sizes, whilst supporting the estimation of statistically robust models. A drawback of these approaches to experimental design is that it is difficult to know how close a particular design is to the best achievable design(s) for the specific experiment. This paper reports a study that set out to explore how efficient a design was by generating a large set of random designs to see how efficient the best design was, and to provide a frequency distribution of a widely used measure of efficiency. Against this pool of random designs, a comparison was made with the designs generated using a genetic algorithm. The initial phase of the study reported in this paper was restricted to exploring multi-nominal logit (MNL) models with fixed prior estimates of model parameters. (a) For the covering entry of this conference, please see ITRD abstract no. E216058.
Abstract