Rossi-Allenby (2003) Bayesian Statistics and Marketing

Summary

A programmatic review of Bayesian statistical methods as applied to marketing research, published in Marketing Science Vol. 22, No. 3 (Summer 2003), pp. 304–328. The paper synthesises the three-component framework (within-unit likelihood, across-unit heterogeneity, decision action) that characterises the Bayesian approach to consumer data, covering latent variable models via data augmentation, hierarchical heterogeneity distributions (normal, mixture of normals, observable covariates, spatial/network), and Bayesian decision theory for optimal marketing actions. It also diagnoses the failure of plug-in (point-estimate) methods and Bayesian Information Criterion (BIC) model selection, arguing for full posterior integration and proper marginal likelihood estimation.

Key Claims

Three-component framework: Every marketing Bayesian model has (1) a within-unit likelihood $p(y_i|\theta_i)$ , (2) a cross-unit heterogeneity distribution $p(\theta_i|\tau)$ , and (3) an action rule derived from a loss function and the posterior — these three must be jointly coherent.
Latent variable unification: Tobit, ordered probit, Multinomial Probit (MNP), and multivariate probit all follow from a single system $z = X\beta + \varepsilon$ , $\varepsilon \sim N(0,\Sigma)$ with appropriate censoring or thresholding; data augmentation (draw $z|y,\theta$ from truncated normal, then $\theta|z,y$ from standard posterior) avoids all multivariate normal integrals.
Identification via projection: Navigate in the unidentified parameter space and project onto the identified subspace at each Markov chain Monte Carlo (MCMC) step — no exact parameter restrictions required.
Normal hierarchical model (three-level): $\theta_i \sim N(\bar\theta, V_\theta)$ , $\bar\theta \sim N(\bar{\bar\theta}, A^{-1})$ , $V_\theta^{-1} \sim W(\nu, V)$ . Two-block Gibbs: $\{\theta_i | \tau, y_i\}$ and $\tau|\{\theta_i\}$ ; shrinkage weight $\propto$ ratio of within- to between-unit variation.
Observable heterogeneity: $\theta_i = Bz_i + u_i$ , $u_i \sim N(0, V_\theta)$ — covariate adjustment before shrinkage; $B$ and $V_\theta$ estimated jointly.
Mixture-of-normals heterogeneity: $p(\theta|\tau) = \sum_{k=1}^K r_k \phi(\theta|\bar\theta_k, V_k)$ — accommodates multimodality, avoids the finite-mixture convex-hull limitation; Gibbs with component assignments as auxiliary variables.
Structural heterogeneity: $p(y_{it}|\{\theta_{ik}\}) = \sum_{k=1}^K r_k p_k(y_{it}|\theta_{ik})$ — models intraindividual change between distinct behavioral regimes (the mixture is at the observation level, not the parameter level).
Spatial and network heterogeneity: Conditional Gaussian field (Ter Hofstede et al. 2002); simultaneous network model $\theta_i = \rho W\theta + u_i$ (Yang-Allenby 2002) — captures peer influence and geographic spillovers.
Diagnostic check (Allenby-Rossi 1999): Compare unit posterior means $\hat\theta_i$ to the predictive $p(\theta_{i'}|\text{data}) = \iint \phi(\theta|\bar\theta, V_\theta)\, p(\bar\theta, V_\theta|\text{data})\, d\bar\theta\, dV_\theta$ — if the posterior means are outliers relative to the predictive, the model underfits heterogeneity.
Bayesian decision theory: Optimal action $a^* = \arg\min_a \int l(a,\theta)\, p(\theta|\text{data})\, d\theta$ ; model selection via posterior model probabilities (Bayes factors), not BIC.
BIC critique: "Extremely inaccurate" for the correlated parameter models typical in marketing; recommends Laplace approximation or importance-sampling marginal likelihood.
Plug-in failure: Evaluating profit at the posterior mean only overstates true profit ("overconfidence"); the correct decision integrates profit over the entire posterior.
Valuation of disaggregate data: Compare $\Pi_{\text{disagg}}$ (unit-level targeting using posterior $\theta_i$ ) against $\Pi_{\text{agg}}$ (aggregate policy using pooled estimates) — the gap quantifies the economic value of individual-level data.

Concepts Introduced or Extended

Entities Mentioned

Quotes

"The Bayesian approach to marketing problems is characterized by three components: a model of consumer behavior (the likelihood), a model of heterogeneity (the prior), and a model of the action or decision to be taken." (p. 304)

"BIC can be extremely inaccurate and should not be used in correlated parameter models." (p. 321, paraphrase)

"The plug-in approach evaluates profit only at the posterior mean parameter estimates. This approach overstates profits because it ignores parameter uncertainty." (p. 323, paraphrase)

My Take

The paper is primarily a review and synthesis, not a source of new methods — almost every technique described had been published earlier (often by the same authors). Its value is in the unified framework: the three-component decomposition is pedagogically powerful and clarifies what is otherwise a scattered literature. The latent-variable unification is particularly clean. The BIC critique is direct and correct; the plug-in vs. full-Bayes point on profit overstatement is an underappreciated practical warning. The spatial/network heterogeneity section is light on technical detail. The "valuation of disaggregate information" framework is the most practically novel contribution — it provides a principled dollar-denominated metric for justifying data collection investment.