Efficient Designs for Alternative Specific Choice Experiments

Abstract

In the past, research on the construction of efficient designs for stated choice experiments has been limited to unlabeled experiments with generic parameter estimates. In this paper, by deriving the asymptotic (co)variance matrix for the alternative-specific MNL model, the authors are able to generate efficient alternative-specific experiments. The authors show that D-error assuming prior parameter values equal to zero is unable to explain statistical efficiency in orthogonal designs and that wide attribute levels are likely to yield more reliable parameter estimates than using narrow attribute levels. The authors also show that the D-optimality criterion may yield inefficient parameter estimates for some design attributes given that trade-offs are made between the efficiencies of different parameter estimates.