Kamakura-Kang (2007) Chain-wide and Store-Level Analysis for Cross-Category Management

Summary

Proposes a factor-regression model for estimating chain-wide price-promotion elasticities that decompose store-level and time-varying heterogeneity into low-dimensional latent factor structures. The model avoids the dimensionality explosion of full random-coefficients seemingly unrelated regressions (SUR), which would require 420+ covariance parameters for two categories of 10 brands each, by constraining store deviations and temporal drift to lie in a low-rank factor space. Estimated by simulated maximum likelihood (ML) via the expectation-maximization (EM) algorithm. Applied to 17 brands (toothpaste + toothbrush) across 66 stores, the factor-regression dominates all competitors — aggregate SUR, store-level SUR, random-coefficients, and Kalman filter — on both Bayesian information criterion (BIC) and out-of-sample root mean squared error (RMSE).

Key Claims

• Dimensionality problem in retail SUR. A full random-coefficients SUR for $J+1$ brands across $S$ stores requires a $(J+1)^2$ covariance matrix $\Sigma_\beta$ — 420 parameters for a two-category, 10-brand system. This makes the model infeasible even for moderate-sized applications.
• Factor decomposition. The brand-specific regression coefficients $\beta_{jst}$ decompose as $\beta_{jst} = \mu_j + \lambda_j V_s + \gamma_j W_t$, where $V_s$ is a $q$-vector of store latent scores (heterogeneity factors), $W_t$ is a $r$-vector of time latent scores (nonstationarity factors), and $\lambda_j$, $\gamma_j$ are brand-specific factor loadings. This reduces free parameters to $O(Jq + Jt)$ loadings plus $Sq + Tr$ scores.
• EM estimation with simulated ML. Factor scores $V_s$ and $W_t$ are treated as latent variables and integrated out via the E-step; parameters are updated in the M-step. No Markov chain Monte Carlo (MCMC) or Bayesian priors are used. Standard errors from the inverse Hessian.
• Model selection via BIC. BIC selects 2 heterogeneity factors ($q=2$) and 2 nonstationarity factors ($r=2$). The 4-factor model substantially dominates all alternatives: aggregate SUR (BIC 92167), store-level SUR (239490), random-coefficients stationary (81755), Kalman filter (83675) vs. factor-regression (71107).
• Partial pooling interpretation. The factor model achieves partial pooling between full pooling (aggregate SUR) and no pooling (store-level SUR): chain-wide estimates $\mu_j$ capture global effects, while store scores $V_s$ capture local deviations. Stores with similar factor scores respond similarly to promotions.
• Asymmetric cross-category effects. Colgate and Crest promotions increase sales of their "sister" toothbrush brands, but not vice versa. Private-label promotions cause less brand-switching than manufacturer-brand promotions — a cross-category complement effect for the retailer.

Concepts Introduced or Extended

• Factor Model
• Random Coefficient Model

Entities Mentioned

• (none with existing wiki pages)

Quotes

"Our proposed factor-regression model takes store differences and longitudinal market shifts into account, thereby providing the retail chain manager with unbiased global, chain-level estimates. It also provides stable local estimates of cross-category promotion effects at the store level."

My Take

A clean application of the factor-regression idea to retail marketing, solving a practical dimensionality problem in a non-Bayesian framework. The key contribution is showing that partial-pooling via factor scores dominates both full pooling (aggregate SUR) and no-pooling (store-level SUR), consistent with the bias-variance trade-off familiar from Bayesian hierarchical models. The non-Bayesian EM approach is unusual in this wiki's context; a natural extension would be Bayesian inference on the factor scores with priors on $q$ and $r$ (cf. Aguilar-West 2000 for Bayesian dynamic factor models). The asymmetric cross-category results are interesting but specific to the toothpaste/toothbrush domain.