Peter E. Rossi

Overview

Peter E. Rossi is a professor at UCLA Anderson School of Management (formerly University of Chicago Booth). His research focuses on Bayesian econometrics with applications to marketing and finance, including hierarchical models, discrete choice, and stochastic volatility.

Key Contributions / Features

Allenby and Rossi (1991): "Quality Perceptions and Asymmetric Switching between Brands" (Marketing Science 10(3): 185–204) — nonhomothetic logit with rotating indifference curves; $\phi_i(u)=\exp(a_i-k_i u)$ ; $k_i$ is an objective quality rank from revealed preference; asymmetric switching arises because price cuts on premium brands cross more household utility thresholds; BIC=−4,944.5 (17 params) beats 58-param nested logit; correlated-errors evidence in standard logit reframed as nonhomotheticity misspecification. See Nonhomothetic Preferences.
Jacquier, Polson, and Rossi (1994): Bayesian MCMC for stochastic volatility; data augmentation over latent variance path; simulation evidence on dominance over MM and QML.
Gallant, Rossi, and Tauchen (1992): "Stock Prices and Volume" — nonparametric model linking return volatility to trading volume; introduced the data used in many subsequent SV empirical studies.
Rossi, McCulloch, and Allenby (1996): "The Value of Purchase History Data in Target Marketing" (Marketing Science 15(4): 304–321) — hierarchical MNP with demographic regression mean βₕ = Δzₕ + vₕ; five information sets (Base/Demographic/Choices-Only/One Observation/Full); target couponing revenue: full info 2.55× blanket, one purchase occasion 1.56× blanket; Bayesian decision theory vs. plug-in (posterior uncertainty matters due to nonlinearity of revenue); continuous normal heterogeneity beats 3-mass-point finite mixture by ~100% in log-likelihood. See Multinomial Probit and Rossi-McCulloch-Allenby (1996).
McCulloch, Polson, and Rossi (2000): ID prior for the MNP model; places prior directly on identified space (σ₁₁=1 w.p. 1); reparameterises Σ = [[1,γ'],[γ,Φ+γγ']] with conjugate Normal/Wishart priors; four-block Gibbs with no tuning. Analytical characterisation of NID prior marginals (Results 1–3). Key advantage: informative priors and hierarchical extensions are natural on identified parameters. Journal of Econometrics 99: 173–193. See Multinomial Probit.
McCulloch and Rossi (1994): Exact likelihood analysis of the multinomial probit model via MCMC ( $y_{h,t} = X_{h,t}\beta_h + \varepsilon_{h,t}$ , $\varepsilon \sim N(0,\Lambda)$ ); foundational Bayesian discrete-choice Gibbs sampler; all blocks conjugate; avoids the $(m-1)T_h$ -dimensional integral required by classical correlated probit MLE. See Multinomial Probit.
Allenby and Rossi (1999): Hierarchical Bayes MNP for consumer heterogeneity ( $\beta_h \sim N(\bar\beta, V_\beta)$ ); advocates continuous normal mixing over finite mixture models; demonstrates $\Delta \ln p = 2070$ in favour of HB on ketchup scanner panel ( $N=1401$ ); derives optimal household-level targeted price reductions using the full posterior. See Multinomial Probit and Bayesian Hierarchical Model.
Jacquier, Polson, and Rossi (2004): Extended JPR (1994) to fat-tailed SV and leverage effect; scale-mixing augmentation for t(ν) innovations; Bayes Factor comparison of four nested SV models. See Stochastic Volatility.
Allenby, Rossi, and McCulloch (2005): Practitioners guide to hierarchical Bayes models for marketing; introduces HB random-effects logit for conjoint analysis ( $\beta_h = \Gamma z_h + \xi_h$ , $\xi_h \sim MVN(0,V_\beta)$ ; four-block Gibbs); credit card conjoint study (N=946, 14,799 comparisons); key insight that $V_\beta$ off-diagonals determine tail-area market potential — low fee (7.5%) dominates low interest (4.5%) despite lower mean utility; announces Bayesian Statistics and Marketing (Wiley 2005) with R software. See Conjoint Analysis.
Rossi and Allenby (2003): Programmatic synthesis of Bayesian statistics for marketing (Marketing Science 22(3): 304–328); articulates the three-component framework (within-unit likelihood, cross-unit heterogeneity distribution, decision action); unifies Tobit/ordered probit/MNP/multivariate probit under a single latent-variable system $z=X\beta+\varepsilon$ with data augmentation; extends heterogeneity to observable covariates ( $\theta_i=Bz_i+u_i$ ), mixture-of-normals, structural heterogeneity (observation-level mixture), and spatial/network models; argues against BIC (extremely inaccurate for correlated-parameter models) and plug-in decision-making (overstates profit by ignoring posterior uncertainty). See Bayesian Hierarchical Model.

Peter E. Rossi

Overview

Key Contributions / Features

Related