Blattberg-George (1991) Shrinkage Estimation of Price and Promotional Elasticities

Summary

This paper tackles the "aggregation dilemma" faced when estimating price and promotional elasticities from supermarket scanner data at the chain-brand level: a single pooled model is too restrictive because elasticities genuinely vary across chains and brands, yet a separate ordinary least squares (OLS) model per chain-brand yields noisy, unstable, and frequently wrong-signed estimates. The authors resolve it with shrinkage estimators derived from a hierarchical Bayes model that borrows strength across chain-brands — pulling each unit's coefficients toward a common mean while still permitting genuine heterogeneity. Applied to a large bathroom-tissue scanner dataset, the shrinkage estimators deliver more reasonable (correctly signed) elasticities than OLS with no loss — indeed an improvement — in out-of-sample predictive power.

Key Claims

• The aggregation dilemma. Full pooling (one model for all chain-brands) is too restrictive because elasticities vary; full disaggregation (separate OLS per chain-brand) produces coefficient estimates that "fluctuate too much" and often carry the wrong sign, making them managerially useless.
• A semilog chain-brand model. Log-sales are regressed on relative price, deal discount and lagged deal discount, feature advertising, display, a final-deal-week indicator, competitive deals, and lagged sales (Model 2.1), with signs predicted from marketing theory (price and lagged-deal-discount negative; advertising/display positive).
• Hierarchical (exchangeable) prior. The chain-brand coefficient vectors are treated as exchangeable draws $\beta^c\mid\theta,\Sigma\sim\mathcal N(\theta,\Sigma)$ around a common mean $\theta$ (Lindley-Smith 1972); combined with the OLS likelihood this yields a posterior-mean shrinkage rule that pulls each $\hat\beta^c$ toward $\theta$ by an amount governed by its precision.
• Four shrinkage estimators. Two motivated by empirical Bayes (with variance components estimated, e.g., via a Morris 1986 truncated-moment estimator) and two by hierarchical Bayes using a normal-Wishart prior and computed by the Gibbs sampler (Gelfand-Smith 1990); a $g$-prior variant is included for tractability.
• Borrowing strength works. On real scanner data the shrinkage estimators produce economically sensible, correctly signed elasticities, and match or beat OLS in forecasting — reducing variance at little bias cost.

Concepts Introduced or Extended

• Bayesian Hierarchical Model
• Seemingly Unrelated Regression
• Random Coefficient Model

Entities Mentioned

• Robert C. Blattberg
• Edward I. George

Quotes

"By borrowing strength across chains and brands, these procedures reduce variability while providing flexibility that allows for separate elasticity estimates."

"Modeling sales at the chain-brand level often leads to counterintuitive and theoretically unreasonable estimates of separate elasticities… frequently having the wrong sign."

My Take

The paper is a clean, early applied demonstration that a hierarchical prior is the right cure for the bias-variance bind of panel-of-regressions estimation — exactly the random-coefficient / SUR setting where each unit has too little data to stand alone but the units are not identical. Its lasting interest is methodological rather than substantive: the same "shrink OLS toward a common mean, estimate the shrinkage from the data" logic underlies the Minnesota prior for VARs and the Black-Litterman treatment of return forecasts, and the paper is notable for using the then-new Gibbs sampler to make the fully hierarchical-Bayes version computable. The caveats are the usual empirical-Bayes ones — the variance-component estimates that drive the shrinkage are themselves noisy, and the wrong-sign diagnosis presumes the theory-predicted signs are correct.