Constrained stated choice experimental designs

Abstract

While a significant literature exists on the generation of experimental designs for stated choice experiments, little work has been done to effectively accommodate constraints on these designs. Constraints may include the prevention of dominated alternatives, the exclusion of specific combinations of attribute levels on plausibility and realism grounds, and the imposition of attribute level balance. We argue that most such constraints are fundamental to the success of the experiment, and should always, at least where feasible, be respected. The notable exception is level balance, which can instead be imposed as a soft constraint, using a level balance measure introduced herein. Various rule structures for specifying the constraints are proposed, as are two new algorithms that can conform to any such rules, whilst allowing the analyst to control the level of importance placed on level balance. From the two case studies investigated, it is determined that the type of constraints specified may influence which algorithm performs best, or even if the algorithm finds a solution at all. Excluding specific combinations of attribute levels reduces the statistical efficiency of the design but is nonetheless warranted. By contrast, the reduction in efficiency resulting from level balance may be excessive, and the analyst should test the trade-offs between these two design properties, by varying the weight placed on level balance in the objective function. The proposed techniques allow researchers to focus not just on the statistical efficiency of these experiments, but behavioural realism and plausibility as well.