Using an ANA data-based dynamic design to improve choice experiment design efficiency : a simulation analysis

Attribute non-attendance (ANA) has been gaining increasing attention in the field of choice modelling. While the modelling issues, effects on parameter estimation, and (less so) causes of ANA have been the main concern of research in this area to date, little attention has been paid to the efficiency of experimental design in the face of ANA. Conventional modelling of ANA involves restricting coefficients on the ignored attributes to zero. This is largely equivalent restricting the levels of the ignored alternatives to zero among choices at the task level. Such practice, however, disturb the initial efficiency of design. This issue is largely unaddressed in the literature, and this research proposes integration of dynamic design using a pre-set set of designs based on choice-level ANA characteristics. As part of this study, simulation are carried out to examine the effects of choice ANA on parameter estimation, under the conditions of static and dynamic choice designs. First, choices are simulated based on a static set of choice tasks assuming that ANA is present at various frequencies for each parameter. Additional specification includes a (negative) correlation structure between the relative level of ANA and relative magnitude of parameters, since parameters of lower magnitude (i.e. less importance) are more likely to be ignored.  Next, conventional Multinomial Logit Model  are used for parameter estimation and they confirm previous studies that found that ignoring attributes significantly biases all parameter estimates (those that were ignored and also those that were not ignored at the choice level). However, incorporating information about ANA in the model specification during estimation resolves this problem.  Next, another set of simulations are run, where the experimental design is adjusted after each simulated choice task completion, depending on which attribute was ignored. Rather than continuously optimizing the design, choice tasks are drawn from a pre-set pool of permutations that incorporate various choice level ANA options. It is argued that such ANA data based dynamic design provides a methodologically superior way of conducting stated choice experiments.