Allenby and Rossi (1999) Marketing Models of Consumer Heterogeneity

Summary

A pedagogical survey advocating hierarchical Bayes (HB) over finite mixture models and classical random effects for discrete choice panel data. The paper synthesizes Rossi-Allenby (1993), McCulloch-Rossi (1994), and Rossi-McCulloch-Allenby (1996) into one readable treatment, argues that marketing requires individual-level parameter estimates rather than just population-level hyper-parameters, and demonstrates decisively — via formal marginal-likelihood comparison and disaggregate posterior inspection — that continuous normal mixing dominates discrete finite mixture approximations. An optimal targeted-couponing application illustrates the decision-theoretic value of full posterior uncertainty propagation.

Key Claims

Marketing vs. econometrics framing: economists treat heterogeneity as a nuisance parameter; marketers need individual-level estimates for targeting, pricing, and product design. Fixed-effects estimation fails due to data sparsity at the household level (~12 purchases/year per household); the hierarchical Bayes framework resolves this via shrinkage.
Three-level hierarchical multinomial probit (MNP) model: latent utility $y_{h,t} = X_{h,t}\beta_h + \varepsilon_{h,t}$ , $\varepsilon_{h,t} \sim \mathcal{N}(0, \Lambda)$ ; $\beta_h \sim \mathcal{N}(\bar\beta, V_\beta)$ ; conjugate priors $\bar\beta \sim \mathcal{N}(\bar{\bar\beta}, aV_\beta)$ , $V_\beta^{-1} \sim W(v_0, V_0)$ (Wishart), $\Lambda_i \sim \text{IG}(\nu, s_i)$ (inverse-gamma). Full posterior via McCulloch-Rossi (1994) Gibbs sampler for the hierarchical MNP.
Shrinkage interpretation: individual posteriors $\pi(\beta_h|\text{data})$ emerge from the joint posterior marginalisation (eq. 5); sparse households borrow heavily from the population mean $\bar\beta$ ; data-rich households are identified from their own observations. This simultaneously ensures identification and controls estimation uncertainty.
Continuous vs. discrete heterogeneity: finite mixture posteriors for $\beta_h$ are constrained to the convex hull of $J$ mass points, dramatically understating tail heterogeneity. Ketchup scanner panel ( $N=1401$ households, 8191 purchases, 4 brands, Springfield MO, 1986): $\ln p(\text{data}|M=\text{finite mixture}) = -5998.3$ vs. $\ln p(\text{data}|M=\text{continuous HB}) = -3928.3$ — overwhelming superiority of continuous HB. 95% highest posterior density (HPD) interval for price sensitivity under HB ( $-9.22$ to $-0.22$ ) is more than twice as wide as finite mixture.
Normality diagnostic: predictive distribution $p(\beta_{h'}|\text{data}) = \int \phi(\beta_{h'}|\bar\beta, V_\beta)p(\bar\beta, V_\beta|\text{data})d\bar\beta\,dV_\beta$ (eq. 23) matches empirical pooled Gibbs distribution across all households (Fig. 3), supporting the normal mixing assumption.
Classical approach critique: unconditional marginal likelihood integrates out $\{\beta_h\}$ yielding a $(m-1)T_h$ -dimensional integral for MNP — computationally demanding; individual-level estimates require additional numerical integration and are only approximate; full Bayesian approach delivers both simultaneously and exactly (eq. 20).
Decision-theoretic application: optimal price reduction for Hunt's ketchup maximises $E_\text{post}[\pi] = E[\Pr(i)(p_i - c_i)]$ (eq. 24–25). Household-specific optima range $\$0$ – $\$0.34$ (mean $\$0.15$ ); profit rises $\$0.106 \to \$0.117$ /household. "Plug-in" approach using posterior means overshoots (mean $\$0.18$ , profit $\$0.129$ ) because nonlinear decision rules require the full posterior — point estimation introduces overconfidence in large price reductions.

Concepts Introduced or Extended

Entities Mentioned

Quotes

"In contrast to this emphasis on individual differences, economists are often more interested in aggregate effects and regard heterogeneity as a statistical nuisance parameter problem which must be addressed but not emphasized."

"The advantage of hierarchical Bayes models of heterogeneity is that they yield disaggregate estimates of model parameters. These estimates are of particular interest to marketers pursuing product differentiation strategies in which products are designed and offered to specific groups of individuals with specific needs."

My Take

This is a review and advocacy paper, not a methods paper — the technical heavy lifting is in McCulloch-Rossi (1994) and Rossi-McCulloch-Allenby (1996). Its value to the wiki is as a self-contained exposition of the HB-MNP framework: the three-level hierarchy, the Gibbs sampler structure, and the shrinkage interpretation are all clearly laid out. The formal model comparison numbers ( $\ln p$ difference of 2070 log-units) are striking and worth recording. The decision-theoretic section is a clean illustration of why full posterior propagation matters for nonlinear objectives. One limitation: the normality diagnostic (pooled Gibbs draws vs. predictive) is informal and would be rejected by a more flexible nonparametric alternative; McCulloch-Rossi (1996) address this with mixture-of-normals heterogeneity.