Health-Related Product Attributes and Purchasing Behavior in the Ready-to-Eat Cereal Market: An Application with Household-level, Censored Data

This is the authors' abstract. We don't add key points for this paper.

but differentiated products.Brand-level demand estimation in these markets needs to handle large number of parameters in estimating the substitutions among numerous differentiated brands.Empirically, two types of models attempt to deal with this dimensionality issue.The most commonly used one is discrete choice modeling, including the logit, nested logit, or random coefficient logit models.They map consumers' purchasing decision from product space to characteristic space, which significantly helps to reduce the number of model parameters.However, while these models overcome the dimensionality problem, they still have several limitations.Both the logit and nested logit models suffer from the restrictive IIA assumptions and limit the flexibility of reflecting the substitution patterns among differentiated products.Although random coefficient logit model can solve IIA problems, its ability to model a large number of brands is limited due to significant computational burden of estimating the nonclosed form integration of market shares.Moreover, since discrete choice models assume that consumer choose only one unit of product that brings the highest utility, this type of model cannot fit cases where consumers usually make multiple purchases or multiple choices.For example, a discrete choice model may not be a good fit for the ready-to-eat cereal case because many consumers purchase multiple cereal products in each shopping trip.The second type of modeling framework is the neoclassical demand model.Compared to discrete choice models, neoclassical demand models do not require single-unit choice assumption.A good example is Hausman, Leonard and Zona (1994) who adopt the multi-stage budgeting approach (Gorman, 1971) and use the AIDS model to help reduce the unknown parameters in the lowest stage.However, the dimensionality still exists because their model is still product-space based and there are still a large number of brands.Pinkse, Slade and Brett (2002) propose the Distance Metric (DM) method to handle these issues.In their model, the relationship between differentiated products is captured by the relative distance between products in characteristics space along multiple dimensions.The closer the two products are, the more likely that these two will become substitutes as price changes.As a result, the number of unknown parameters can be significantly reduced because the cross price elasticity is transferred to be a function of distances of key characteristics.Compared to previous modeling efforts, the DM method allows more flexible substitutions between products since it does not impose any prior assumptions of groupings or decision stages as in nested logit model or the multi-stage budgeting approach.It also has fewer computational issues due to its simple functional form compared to the random coefficient logit model.Because an individual consumer or household often purchases only a very small number of brands, consumer-level scanner data can exacerbate the dimensionality problem associated with differentiated products by recording so-called zero purchases.While this second type of empirical problem may not be an issue for discrete choice demand models, it is for neoclassical demand models, where the likelihood