Web appendix for : attribute-level heterogeneity

This web appendix has three main purposes. First, we provide a more or less 'stand-alone' technical appendix that describes the estimation algorithm for the proposed attribute model using Markov Chain Monte Carlo techniques (sections A1 and A2). The reversible jump (RJ) algorithm (Green, 1995) is also described in detail for the (vector) finite mixture regression model. We first give a discussion of priors and a general description of the reversible jump algorithm; then we present details of the estimation schema for the standard finite mixture regression model. We subsequently extend these details for the attribute model. As we will show, the algorithms and equations for the attribute model are similar to the results for the vector model due to a simple transformation. This similarity makes the coding of the attribute model straightforward once computer code for the vector model (with reversible jump steps) is developed. Furthermore, we discuss how the algorithms should be modified to estimate a standard choice model (e.g. a probit model). Second, we briefly discuss in section A3 the benchmark models for heterogeneity considered in the main document and their implementation, including the mixture of normals model (Allenby et al. 1998, Lenk and DeSarbo 2000) and the Dirichlet Process Priors (Ansari and Mela 2003, Kim et al. 2004). Third, we present the results of an additional simulation experiment where the traditional (vector) finite mixture model is used to generate the data in section A4, which augments the Monte Carlo experiment in the main document. The paper "Attribute-Level Heterogeneity" to which these Appendices apply is available at the following URL: http://ssrn.com/abstract=1687107