Allenby-Ginter (1995) Using Extremes to Design Products and Segment Markets

Summary

Allenby and Ginter introduce a two-level hierarchical Bayes (HB) conjoint model for credit card attributes in which respondent-specific part-worth deviations are drawn from a multivariate normal (MVN) distribution whose covariance matrix $D$ encodes attribute co-preference structure. The paper's central methodological contribution is the argument that optimal product design and market segmentation require focusing on the extremes of the heterogeneity distribution rather than means: two attribute levels with identical mean utility can generate very different segment sizes because their tail probabilities are driven by $D$ 's off-diagonal structure. Estimation proceeds via a six-block Gibbs sampler with rejection sampling for latent binary utilities, and the model outperforms latent class (3 segments) and aggregate models on both same-respondent and new-respondent predictive accuracy.

Key Claims

The two-level model is $y_{i,h} = x_{i,h}'(\alpha + \beta_h) + \varepsilon_{i,h}$ , $\varepsilon_{i,h} \sim N(0, \sigma^2)$ , with respondent-specific deviations $\beta_h = \Gamma z_h + \zeta_h$ , $\zeta_h \sim MVN(0, D)$ .
$\alpha$ is a fixed-effects vector common to all respondents; $\beta_h$ captures individual deviations; $\Gamma$ maps demographics $z_h$ to mean part-worths.
The covariance matrix $D$ governs the joint distribution of part-worths: cov(low annual fee, out-of-state bank penalty) = 5.34, meaning consumers with a strong preference for low fees also strongly dislike out-of-state banks.
Low annual fee and low interest rate have nearly identical mean utility (~4.2 utiles each), yet 24.2% of respondents exceed net utility > 5.0 for low fee vs. only 19.2% for low interest, due to the fatter tail driven by $D$ 's off-diagonal structure.
111 respondents have >70% membership probability in the low-fee segment vs. 58 in the low-interest segment, despite equal means.
Demographics in $\Gamma$ : females strongly prefer low annual fee (coeff = 1.092, s.d. = 0.304); older respondents are less attribute-sensitive overall; higher-income respondents are more interest-rate-responsive.
Gibbs sampler uses six blocks: $\alpha$ , each $\beta_h$ , $\Gamma$ , $D$ , $\sigma^2$ , plus rejection sampling for paired-comparison latent utilities; priors: $D \sim IW(15, 15I)$ (inverse-Wishart), $\sigma^2 \sim \text{Inv-}\chi^2(1, 5)$ ; 2000 iterations, last 1000 used.
Model comparison on N=946 credit card respondents (14,799 paired comparisons, 7 attributes): HB same-respondent mean absolute deviation (MAD) = 0.235 vs. latent class (3 segments) 0.282 vs. aggregate 0.311; new-respondent MAD = 0.329 vs. 0.373 vs. 0.364.
Segments should be viewed as a firm's decision tool (which product configuration captures the largest segment of high-preference consumers) rather than an assumption about consumer homogeneity.

Concepts Introduced or Extended

Conjoint Analysis — introduces the two-level HB formulation with $D$ matrix and the "extremes" argument for product design
Bayesian Hierarchical Model — two-level normal hierarchical prior; Gibbs estimation
Gibbs Sampler — six-block Markov chain Monte Carlo (MCMC) for the conjoint model with rejection sampling for paired-comparison latent utilities
Market Segmentation — reframes segments as tail-area probabilities under the continuous heterogeneity distribution
Random Coefficient Model — respondent-level random effects $\beta_h$ with structured covariance $D$

Entities Mentioned

Quotes

"Rather than finding a set of customers who prefer a product feature, we suggest finding the product feature that will generate the largest most profitable segment."

"The key to understanding why extremes are important for product design is the covariance matrix D."

My Take

The central insight — that product design is a tail-area problem and that equal-mean attributes can differ dramatically in tail mass — is genuinely valuable and anticipates the more general treatment in Allenby-Rossi-McCulloch (2005). The paper uses a simpler paired-comparison setup (continuous normal rather than logit likelihood) which makes the Gibbs steps cleaner and more pedagogically transparent. The same N=946 credit card dataset is revisited a decade later in the 2005 practitioners guide, allowing direct comparison of specifications. The "extremes" framing is the paper's lasting contribution: segments are not assumptions about consumer homogeneity but managerial constructs defined by which product attracts the most extreme and numerous enthusiasts.