Internal market structure analysis infers both brand attributes and consumer preferences for those attributes from preference or choice data. The authors exploit a new method for estimating probit models from panel data to infer market structures that can be displayed in few dimensions, even though the model can represent every possible vector of purchase probabilities. The result outperforms each of several other models, including Choice Map, SCULPTRE, and Chintagunta's latent class model in terms of goodness of fit, predictive validity, and face validity for a detergent data set. Because theirs is the only market structure model to outperform the structureless Dirichlet-multinomial stochastic brand choice model, the other methods cannot claim to have recovered market structure for these data.

This paper describes Choice Map, a model that infers a product-market map from panel data. Choice Map is a stochastic brand choice model, a random utility quantal choice model, and a multidimensional scaling procedure. Choice Map combines the parsimony of stochastic choice models, the rationale of random utility models, and the ability of multidimensional scaling procedures to simultaneously infer both brand positions and consumer preferences from preference (choice) data. Choice Map assumes that consumers have stationary purchase probabilities for the observational period, which makes the model best suited to frequently-bought mature product categories. Neither experimental nor survey data are required. Traditional mapping methods that scale such data may not yield the same structure for low-involvement products as will a procedure like Choice Map that scales behavioral data. Choice Map's descriptive and predictive validity exceeds that of an alternative stochastic brand choice model (the Dirichlet-multinomial) for a test data set.market structure analysis, multidimensional scaling, stochastic brand choice models, random utility models

The authors compare two approaches to conjoint analysis in terms of their ability to predict shares in a holdout choice task. The traditional approach is represented by three models fit to individual-level ratings of full profiles, whereas the other approach is represented by four multinomial logit models fit to choice shares for sets of full profiles. Both approaches predict holdout shares well, with neither the ratings-based nor the choice-based approach dominant, though some models predict better than others. Particularly promising is a new aggregate model that captures departures from independence of irrelevant alternatives (IIA).

No abstract

No abstract