Allenby, Rossi, and McCulloch (2005) Hierarchical Bayes Models: A Practitioners Guide

Summary

A practitioner-facing survey chapter (44 pp., January 2005) introducing hierarchical Bayes (HB) models to a marketing audience. Develops a four-block Gibbs sampler for random-effects logit in conjoint analysis, illustrates it on a credit card study (N=946, 14,799 paired comparisons), and argues that the off-diagonal structure of the heterogeneity covariance matrix has direct managerial content — low annual fee dominates low interest as an out-of-state bank incentive precisely because its covariance with out-of-state tolerance is 8.5 ( $\rho$ =0.80), putting 7.5% of respondents above the profitability threshold vs. 4.5% for low interest despite a lower mean utility. Announces the forthcoming book Bayesian Statistics and Marketing (Rossi, Allenby, McCulloch 2005, Wiley) with R software.

Key Claims

HB frees researchers from computational constraints: Markov chain Monte Carlo (MCMC) replaces intractable analytic integration with simple, repetitive simulation; convergence is distributional (stable mean and variance of draws), not to a point estimate.
Three-level hierarchy: data model (logit or probit for choices), latent-parameter model ( $\beta_h = \Gamma z_h + \xi_h$ , with multivariate normal $\xi_h \sim MVN(0, V_\beta)$ ), and hyperprior ( $\Gamma, V_\beta$ ); conditional independence means given $\{\beta_h\}$ , blocks 2–3 do not depend on raw choices.
Four-block Gibbs (eqs. 16–19): (1) draw each $\beta_h$ one respondent at a time given data and hyper-parameters; (2) draw $\Gamma | \{\beta_h\}, V_\beta$ ; (3) draw $V_\beta | \{\beta_h\}, \Gamma$ ; (4) repeat.
Covariance matrix $V_\beta$ has managerial content: the low annual fee / out-of-state bank off-diagonal = 8.5 ( $\rho$ =0.80) produces a wider tail than the low interest / out-of-state diagonal ( $\rho$ =0.23); 7.5% vs. 4.5% of respondents above the target threshold — nearly 2× — despite lower mean utility.
Full posterior vs. plug-in: using posterior means of hyper-parameters as point estimates systematically overestimates market potential by ignoring parameter uncertainty; the full posterior is required for valid tail-area calculations.
Continuous > discrete heterogeneity: across dozens of marketing studies, the normal mixing distribution outperforms finite mixture models; the finite mixture constrains individual posteriors to the convex hull of mass points and misses tail structure (Allenby-Arora-Ginter 1998; Andrews-Ansari-Currim 2002).
Software gap: WinBugs fails at commercial data scale; Sawtooth Software is limited to standard choice-based conjoint (CBC) / ordinary least squares (OLS); practitioners need custom R code — provided in the forthcoming Rossi-Allenby-McCulloch (2005) book.

Concepts Introduced or Extended

Entities Mentioned

Quotes

"Hierarchical Bayes models free researchers from computational constraints and allow researchers and practitioners to develop more realistic models of buyer behavior and decision making."

"The potential demand for an offering can be defined as the proportion of consumers for whom it is profitable to design and market that configuration... Invariably, this framework leads to market potentials defined as extremes (tail-areas) of distributions of response potential."

My Take

This is a survey/introduction chapter, not a primary methodology contribution — the random-effects logit Gibbs sampler is not new here. Its value lies in (1) the unusually clear exposition of why the covariance matrix of heterogeneity matters for managerial decisions, not just the means; (2) the frank discussion of software limitations that motivated the R book; and (3) the annotated bibliography of 40+ Bayesian marketing applications, which serves as a comprehensive entry point into the field. The credit card application is a textbook-quality illustration: the comparison of low-fee vs. low-interest incentives via tail-area analysis is both analytically clean and directly actionable.