Preference estimation under bounded rationality: Identification of attribute non-attendance in stated-choice data using a support vector machines approach

Abstract

Stated-choice experiments have been useful in helping to make a number of operations management decisions. Many recent advances in this area have raised questions about estimating consumers’ preferences when they partially ignore the information provided in discrete choice experiments, a problem introduced as attribute non-attendance (ANA). This line of research explores the consequences of assuming that consumers consider all available information concerning attributes to evaluate product alternatives, when in fact, they might ignore some attributes completely. Diverse choice models, such as latent class models, have been developed to accommodate ANA using choice data. Due to the combinatorial nature of such an approach, researchers typically explore a limited number of specifications. Furthermore, although diverse modeling approaches have been proposed to accommodate ANA, no research has investigated the capability of these approaches to correctly identify ANA at the individual level. In this work, we propose the use of a machine learning approach based on support vector machines to identify ANA at the individual level and to predict consumer choices in conjoint experiments. We conduct an extensive simulation study varying the degree of non-attendance and the noise in the choice data to investigate the performance of the proposed approach. Our results with simulated data show good performance in terms of the identification of attended and non-attended attributes. We test our approach in two empirical applications and compare it to state-of-the-art benchmarks in the field. We demonstrate the usefulness and the alternative insights derived from our method.