Semi-Bayesian D-Optimal Choice Designs for Mixtures

In order to acquire insight into consumer preferences for products and services that are described by certain attributes, choice experiments are employed. For e_ciency, this should be done by means of an optimal experimental design, which gives the most precise estimates for the parameters in the corresponding statistical model. Sometimes attributes of products and services can be mixtures of ingredients. Although mixture models are commonly used in industrial experiments, they have never been introduced in choice modeling. This master thesis aims at introducing mixtures in the choice context, since often consumer products and services can be described as mixtures of ingredients. An algorithm to construct semi-Bayesian D-optimal experimental designs is presented for the multinomial logit model when choices are based on a mixture of ingredients. The resulting designs are D-optimal and based on a mix- ture coordinate-exchange algorithm. Further, some features of them are discussed. It is shown that designs, when prior parameter values required for choice models are not assumed to be zero, di_er from the utility neutral designs, where such an assumption is made. We also show that semi-Bayesian designs di_er from and perform better than locally optimal ones (and the utility neutral designs) for most of the time. As often it is di_cult to obtain accurate prior parameter values, parameter misspeci_cation is also investigated. It is demonstrated that monotonous misspeci_cations in true parameters do not distort the outcome, and might help to design more robust designs.