Best-worst Discrete Choice Methods with Convex Optimization

Project: Research

View graph of relations



One of the critical challenges facing businesses in the manufacturing and service sectors is uncovering consumer preferences for products and services. Quantitative methods are being increasingly used in the analysis of consumer choice behavior with discrete choice methods emerging as the de facto standard. These methods are used to estimate the willingness to pay for goods and services and to develop new products and pricing strategies. Discrete choice methods can be used to estimate the effect of consumer and product attributes on consumer valuation. In the simplest choice design, a consumer is asked to state their most preferred alternative from a set of alternatives. However, even in this case, the evaluation of choice probabilities is a challenging task especially when the alternatives have correlated utilities due to shared attributes. Except in special models such as logit, simulation techniques remain the only way to analyze choice behavior. The computational challenges are compounded in more sophisticated choice designs such as asking consumers to select the best and the worst alternative. This approach known as Maximum difference scaling (Maxdiff) has become popular over the last decade due to consumers propensity to identify extreme options while revealing more about their preferences.In this project, we will build on advances in convex optimization to address such choice design problems. Currently, we have developed a convex optimization based approach to predict choices when consumers pick a single alternative from a set of alternatives. Our goal is to extend this to the Maxdiff setting. Unfortunately, the logit formula does not naturally extend to this case making the prediction problem a challenging one in itself. Furthermore, techniques that can capture the effects of correlations in this model are not known. Our objective is to develop a theoretically sound and a computationally tractable method to do prediction and estimation in this context. Then, with primary/secondary data, we will test this model and compare it with the state of art approaches. We believe that this project will be of interest to academics and practitioners due to the rising number of empirical applications that are based on Maxdiff methods. This is evidenced by commercial conjoint software such as Sawtooth integrating the Maxdiff approach into their codes due to demand for such tools to analyze sophisticated choice designs.


Project number9041712
Grant typeGRF
Effective start/end date1/01/121/01/12