Rossi-McCulloch-Allenby (1996) The Value of Purchase History Data in Target Marketing

Summary

Rossi, McCulloch, and Allenby develop a Bayesian hierarchical random-coefficient multinomial probit (MNP) model in which household preferences $\beta_h$ depend on demographic covariates ( $\beta_h = \Delta z_h + v_h$ , $v_h \sim N(0, V_\beta)$ ) and use it to place monetary values on different types of household purchase history data via a target couponing exercise. They define five information sets ranging from "Base" (population distribution only) to "Full" (complete purchase history with causal variables) and show that even one purchase occasion with causal variables yields 56% more net couponing revenue than a blanket strategy. The paper establishes the Bayesian decision-theoretic approach to coupon optimization (full posterior, not plug-in estimates) as the correct framework for marketing customization under parameter uncertainty.

Key Claims

Information value ranking: Full info = 2.55 $\times$ blanket gain; Choices-Only = 1.93 $\times$ ; One Observation (choice + causal) = 1.56 $\times$ ; Demographics-Only = 1.12 $\times$ . Even one purchase occasion with price/display data outperforms demographics-only by 40%.
The trigger strategy value: The widely-deployed Catalina "one purchase occasion" trigger strategy already captures 56% incremental gain over blanket couponing — a striking practical validation of short purchase histories.
Choices-only inference problem: When causal variables (price, display, feature) are unobserved, the reduced model intercepts $\mu_h$ must be translated back to the full $\beta_h = (\gamma_h, \delta_h)$ via iterated expectations ( $\gamma + R\delta \equiv \mu$ ) using cross-household information from the full-data panel.
Plug-in vs. Bayesian decisions: Because net revenue $R(\beta_h, \Lambda|F)$ is nonlinear in parameters, $E[R] \neq R(E[\beta], E[\Lambda])$ . Plug-in estimates produce overconfident coupon strategies with too-extreme face values. The full posterior correctly propagates uncertainty to the decision.
Continuous beats finite mixture: Continuous normal heterogeneity achieves nearly 100% improvement in log-likelihood over the best 3-mass-point finite mixture ( $-1170$ vs. $-2036$ ). The finite mixture constrains household posteriors to the convex hull of mass points, drastically understating tail heterogeneity in brand intercepts and price sensitivities.
Demographics explain 7–33% of heterogeneity across $\beta_h$ components; price sensitivity $R^2 \approx 7\%$ — demographics have very limited value for targeting pricing decisions relative to behavioral data.

Concepts Introduced or Extended

Multinomial Probit — demographic regression mean $\beta_h = \Delta z_h + v_h$ ; five-block Gibbs with demographic prior on $\beta_h$
Bayesian Hierarchical Model — target couponing application; information-value framework for Bayesian decision theory
Mixture of Normals — continuous normal mixing strongly outperforms finite mixture (Kamakura-Russell 1989) for household-level parameter estimation
Conjoint Analysis — Bayesian decision-theoretic approach to customized marketing actions under posterior uncertainty

Entities Mentioned

Quotes

"Even rather short purchase histories can produce a net gain in revenue from target couponing which is 2.5 times the gain from blanket couponing."

"Surprisingly, even the information contained in observing one purchase occasion boasts net couponing revenue by 50% more than that which would be gained by the blanket strategy."

My Take

The paper's greatest contribution is framing the information-valuation problem correctly: by tying statistical inference directly to a decision-relevant metric (net couponing revenue), it converts abstract Bayesian model comparisons into dollar figures. The five information sets are well-chosen and the results are striking. The most practically important finding — that one purchase occasion already captures most of the gains from customization — explains the market success of Catalina's trigger strategy.

The choices-only inference algorithm is technically interesting but somewhat ad hoc in how it maps $\mu_h$ to $\beta_h$ . The full-data and one-observation results are more cleanly motivated. The paper predates the formal development of the Bayesian decision theory for marketing that Rossi-Allenby (2003) formalizes, but the core ideas are already here.