The multi-attribute elimination by aspects (MEBA) model

University of Technology, Sydney. Faculty of Business.Our research proposes a new, multi-attribute, parameterisation of Tversky’s Elimination- By-Aspects (EBA) model. The EBA model conceptualises choice as a covert sequential elimination process with choice probabilities formulated over all consideration sets of the choice set. This specification attempts to capture the effect of context on choice behaviour. However, the EBA model has seen limited usage due to the large number of required parameters given the set of items under study. For a set of items T, it has 2|T| - 3 free parameters, which is infeasible for all but the simplest of contexts. To provide a practical operationalisation, we impose a set of a priori constraints on the parameter space. We define a generic multi-attribute structure to the set of aspects. This restricts the cardinality of the set of unknown scale values while retaining the functional (recursive) form of the model. The EBA hypothesis of a population of lexicographic decision-makers can therefore be tested in more market-realistic contexts, and inferences made over a large universal set of items described by the complete factorial. We call this model the Multi-attribute Elimination-By-Aspects (MEBA) model. The MEBA model reduces the set of unknown free parameters to a maximum of |T|-1. We develop a general algebraic expression for the MEBA choice probabilities as a function of the attributes of the options in the choice set. This enables the derivation of a likelihood function, and consequently maximum likelihood estimation. We also consider the form of optimal MEBA paired comparison designs. Using Monte Carlo simulation and a discrete choice experiment with consumers, we conduct an initial empirical test of the model against the special case of the MNL model (that assumes no context effects) and find the MEBA model to be a better approximation of observed choice behaviour. This is achieved on a common set of parameters, and so it is due solely to the difference in functional form of the two models. We conclude with a discussion on future research directions, in particular the introduction of heterogeneity into the model, and the description of optimal choice experiments for larger choice set sizes