55 Choice Modeling and Demand Estimation Seminar

Choice modeling and demand estimation are the empirical heart of quantitative marketing and of the industrial-organization tradition it borrows from. The field rests on a single, powerful idea: that observed purchases are the realized outcome of a latent, utility-maximizing comparison among alternatives, so that a model of how individuals choose can be aggregated into a model of how a market demands. This seminar traces the arc from the random-utility microfoundation, through the estimation of individual choice on disaggregate panel data, to the recovery of aggregate demand systems from market-level data—and on to the counterfactual and welfare questions those demand systems were built to answer. It is the doctoral reading-map companion to the technical chapters that develop the estimators in full: the choice and Bayesian-methods chapter (Chapter 41), and the structural-modeling and industrial-organization material, to which this seminar stands in the same relation that the analytical-modeling seminar stands to its methods chapters.

The science matters because demand is the object on which nearly every managerial and policy question ultimately rests. A price elasticity is a demand derivative; a merger simulation is a demand counterfactual; the welfare gain from a new product is an integral under a demand curve; the return on an advertising or assortment decision is a demand response. Commercially, the same machinery powers pricing engines, assortment optimization, and the regulatory review of mergers in differentiated-product industries. The intellectual payoff of the seminar is that it teaches a student to see these disparate applications as instances of one estimation problem and to know which assumptions each application is buying.

The central tension that animates every week is identification. Preferences are never observed; only choices are. The analyst must recover taste parameters—and, crucially, price sensitivity—from data in which the very prices a consumer faces are set by firms who observe demand shifters the analyst does not. Endogenous prices, unobserved product quality, and unrestricted preference heterogeneity all threaten to contaminate the elasticity estimates on which everything downstream depends. The history of the field is in large part a history of instruments, control functions, and distributional assumptions devised to break this contamination. A student who completes the seminar can read a demand paper and locate, immediately, the source of variation that identifies its price coefficient, the assumption that makes that variation exogenous, and the counterfactual that assumption licenses.

55.1 Semester arc

The fourteen-week arc moves along three nested levels. The first third (Weeks 1–4) builds the random-utility microfoundation and its disaggregate estimators: McFadden’s conditional logit and its psychological antecedents in Thurstone and Luce; the first marketing applications to scanner panels; the independence-of-irrelevant-alternatives (IIA) problem and its nested-logit and generalized-extreme-value (GEV) repairs; and the random-coefficients or mixed logit that dissolves IIA by integrating taste heterogeneity out of the choice probabilities. The middle third (Weeks 5–8) confronts heterogeneity and aggregation: hierarchical Bayes estimation of individual-level parameters, conjoint and stated-preference measurement, and then the pivotal transition from individual choice to aggregate demand in the Berry inversion and the Berry–Levinsohn–Pakes (BLP) random-coefficients demand model, together with the practitioner’s craft of estimating it.

The final third (Weeks 9–14) takes the static demand system dynamic and strategic and then closes the loop to welfare. Weeks 9 and 10 relax the static, full-information benchmark: consumers stockpile, wait for durable-good prices to fall, and search over an endogenous consideration set rather than evaluating the full assortment. Weeks 11 and 12 are method-forward, treating endogeneity in choice models (control functions, copulas, instruments) and the Bayesian and Markov-chain-Monte-Carlo (MCMC) machinery that makes rich heterogeneous models estimable. Week 13 brings the machine-learning frontier—flexible demand prediction and heterogeneous-treatment-effect estimation—into contact with structural demand. Week 14 returns to first principles: the demand system exists to compute welfare and counterfactuals, so the capstone reads the new-product and merger-welfare papers that justify the entire enterprise.

The reading map uses two tags. [F] = Foundational marks canon a choice-and- demand scholar is expected to know cold; [R] = Frontier/Recent marks an active research front, refreshed as the literature moves. Each week pairs at least one foundational anchor with a more recent or methodological counterpart. DOIs are reproduced as verified against the Crossref REST API; works without a verifiable DOI—chiefly the canonical books and a 1974 edited-volume chapter—are named without a link and flagged as such in the text.

55.2 Week 1 — Random-utility foundations

Topic. The random-utility model (RUM) as the microfoundation of all that follows: a chooser maximizes a latent utility that the analyst observes only up to a random component, and the distribution of that component generates the choice probabilities.

Subtopics. Deterministic vs. stochastic utility; the additive-error specification; the extreme-value distributional assumption and the logit closed form; the psychological roots of stochastic choice.

Methods. Maximum-likelihood estimation of the multinomial/conditional logit; identification up to scale and level; the log-likelihood and its concavity.

Key readings.

McFadden (1974), “Conditional Logit Analysis of Qualitative Choice Behavior,” in Zarembka (ed.), Frontiers in Econometrics, Academic Press — the founding derivation linking utility maximization with extreme-value errors to the logit choice probability; no Crossref DOI (edited-volume chapter), named without link. [F]
Luce (1959), Individual Choice Behavior: A Theoretical Analysis, Wiley — the choice-axiom (independence from irrelevant alternatives) that the logit operationalizes; canonical book, no DOI. [F]
Thurstone (1927), “A Law of Comparative Judgment,” Psychological Review — the original random-utility/discriminal-process idea that anchors the whole tradition. doi:10.1037/h0070288 — the probit-flavored ancestor of stochastic choice. [F]

Debate. Is the logit’s tractable extreme-value error a substantive behavioral claim or a mathematical convenience whose IIA implication must later be undone?

55.3 Week 2 — Brand choice on scanner data

Topic. The first great marketing application of the RUM: estimating brand-choice probabilities from supermarket scanner panels, where the same household is observed choosing repeatedly across many purchase occasions.

Subtopics. Loyalty and state dependence; the brand-loyalty and purchase-event-feedback variables; marketing-mix covariates (price, display, feature) in the utility; the incidence/quantity decisions surrounding brand choice.

Methods. Conditional logit on panel choice data; constructing exponentially smoothed loyalty regressors; modeling purchase timing and quantity jointly with brand choice.

Key readings.

Guadagni & Little (1983), “A Logit Model of Brand Choice Calibrated on Scanner Data,” Marketing Science. doi:10.1287/mksc.2.3.203 — the paper that brought the conditional logit into marketing and defined the loyalty-variable template. [F]
Gupta (1988), “Impact of Sales Promotions on When, What, and How Much to Buy,” Journal of Marketing Research. doi:10.2307/3172945 — decomposes promotional response into purchase-timing, brand-choice, and quantity components, extending the scanner-choice program beyond brand choice alone. [F]

Debate. Does the loyalty variable capture genuine structural state dependence or merely soak up persistent unobserved heterogeneity—an identification confound that later mixed-logit and hierarchical-Bayes work would target directly?

55.4 Week 3 — IIA, nested logit, and GEV

Topic. The Achilles’ heel of the simple logit—independence from irrelevant alternatives—and the family of generalized-extreme-value models built to relax it while preserving a closed form.

Subtopics. The red-bus/blue-bus problem; proportional substitution and its implausibility; nesting correlated alternatives; the GEV class and its inclusive-value structure; cross-elasticity patterns.

Methods. Nested-logit estimation; the inclusive value and dissimilarity parameter; testing IIA; GEV generating functions.

Key readings.

McFadden (1980), “Econometric Models for Probabilistic Choice Among Products,” The Journal of Business. doi:10.1086/296093 — develops the GEV framework and the modeling of substitution patterns among differentiated products, the direct ancestor of structural demand. [F]
Guadagni & Little (1983), “A Logit Model of Brand Choice Calibrated on Scanner Data,” Marketing Science. doi:10.1287/mksc.2.3.203 — revisited here as the canonical setting where IIA’s restrictive substitution is most visibly violated by close-substitute brands. [F]

Debate. Nested logit fixes IIA within a pre-specified tree—but who chooses the tree, and does an analyst-imposed nesting merely relocate the maintained assumption rather than test it?

55.5 Week 4 — Random coefficients and mixed logit

Topic. The mixed (random-coefficients) logit: letting taste parameters vary across individuals according to a mixing distribution dissolves IIA entirely and can approximate any random-utility model.

Subtopics. The mixing distribution over tastes; correlated random coefficients; flexible substitution patterns; panel mixed logit with repeated choices.

Methods. Simulated maximum likelihood and maximum simulated likelihood; Halton draws; the McFadden–Train approximation theorem.

Key readings.

McFadden & Train (2000), “Mixed MNL Models for Discrete Response,” Journal of Applied Econometrics. doi:10.1002/1099-1255(200009/10)15:5<447::aid-jae570>3.0.co;2-1 — proves that mixed logit can approximate any RUM choice probabilities arbitrarily well; the theoretical license for the whole approach. [F]
Revelt & Train (1998), “Mixed Logit with Repeated Choices: Households’ Choices of Appliance Efficiency Level,” The Review of Economics and Statistics. doi:10.1162/003465398557735 — the template for estimating mixed logit on panel data with within-household correlation. [F]
Train (2009), Discrete Choice Methods with Simulation, 2nd ed., Cambridge University Press — the standard graduate reference for simulation-based estimation; cited as a book without a single-work DOI. [F]

Debate. Is the freedom of the mixing distribution a virtue or a trap—does a mis-specified parametric mixing law (e.g., normal price coefficients implying positive price sensitivity) distort the very heterogeneity it claims to recover?

55.6 Week 5 — Heterogeneity and hierarchical Bayes

Topic. Estimating individual-level preference parameters by pooling information across consumers through a hierarchical prior—the Bayesian counterpart to mixed logit, and the workhorse of marketing’s heterogeneity tradition.

Subtopics. The random-effects/hierarchical prior; shrinkage toward the population mean; individual-level posteriors for targeting; the value of purchase history.

Methods. Hierarchical Bayes; Gibbs sampling and data augmentation; the multivariate-normal heterogeneity distribution; posterior means as individual estimates.

Key readings.

Allenby & Rossi (1998), “Marketing Models of Consumer Heterogeneity,” Journal of Econometrics. doi:10.1016/S0304-4076(98)00055-4 — the manifesto for treating heterogeneity as the object of interest rather than a nuisance, with the hierarchical-Bayes machinery to recover it. [F]
Rossi, McCulloch & Allenby (1996), “The Value of Purchase History Data in Target Marketing,” Marketing Science. doi:10.1287/mksc.15.4.321 — shows how individual-level posteriors sharpen targeting, the commercial payoff of heterogeneity estimation. [F]

Debate. Do individual-level posteriors recover real heterogeneity or merely project a parametric prior onto thin per-person data—and how much purchase history is enough before the prior stops doing the work?

55.7 Week 6 — Conjoint analysis and preference measurement

Topic. Eliciting preferences from designed choice experiments rather than market data—the stated-preference complement to revealed-preference demand estimation.

Subtopics. Full-profile vs. choice-based conjoint; experimental design and attribute balance; part-worth utilities; the bridge from conjoint to choice simulators and willingness-to-pay.

Methods. Designed factorial/fractional experiments; choice-based conjoint estimated by (hierarchical Bayes) logit; reliability and external validity.

Key readings.

Green & Srinivasan (1978), “Conjoint Analysis in Consumer Research: Issues and Outlook,” Journal of Consumer Research. doi:10.1086/208721 — the field-defining review that established conjoint as marketing’s preference-measurement method. [F]
Green & Srinivasan (1990), “Conjoint Analysis in Marketing: New Developments with Implications for Research and Practice,” Journal of Marketing. doi:10.1177/002224299005400402 — the decadal update that reoriented conjoint toward choice-based designs compatible with the RUM. [F]

Debate. Do stated preferences from hypothetical choices recover the same parameters as revealed preferences from market choices—and when does hypothetical bias make conjoint willingness-to-pay an unreliable demand input?

55.8 Week 7 — Aggregate demand and the BLP model

Topic. The pivotal transition from individual choice to aggregate demand: recovering structural demand parameters from market-level shares when individual data are unavailable and prices are endogenous.

Subtopics. The unobserved product-quality term; the market-share inversion; random coefficients at the market level; instrumenting for price.

Methods. The Berry inversion (shares to mean utilities); the BLP contraction mapping; generalized method of moments (GMM) with demand-side instruments.

Key readings.

Berry (1994), “Estimating Discrete-Choice Models of Product Differentiation,” The RAND Journal of Economics. doi:10.2307/2555829 — introduces the share-inversion idea that turns an aggregate-demand problem into a linear instrumental-variables problem in mean utilities. [F]
Berry, Levinsohn & Pakes (1995), “Automobile Prices in Market Equilibrium,” Econometrica. doi:10.2307/2171802 — the full random-coefficients aggregate-demand model with supply-side equilibrium; the single most influential demand paper of the era. [F]

Debate. The BLP instruments (rival-product characteristics, cost shifters) rest on the assumption that product characteristics are exogenous to the unobserved quality $\xi$—is that assumption credible when firms design products knowing $\xi$? The worked treatment in Section 55.17.1 develops the estimator and this exclusion restriction in full.

55.9 Week 8 — BLP in practice

Topic. The craft of actually estimating random-coefficients aggregate demand: the numerical, instrumental, and identification choices that separate a credible BLP estimate from a fragile one.

Subtopics. Choice of instruments and the weak-instrument problem; the inner contraction and outer GMM loops; numerical pitfalls and convergence; market power and markups recovered from demand.

Methods. Nested fixed-point GMM; optimal instruments; supply-side pricing moments; the cereal-industry application.

Key readings.

Nevo (2000), “A Practitioner’s Guide to Estimation of Random-Coefficients Logit Models of Demand,” Journal of Economics & Management Strategy. doi:10.1162/105864000567954 — the pedagogical guide that made BLP estimable by a generation of students. [F]
Nevo (2001), “Measuring Market Power in the Ready-to-Eat Cereal Industry,” Econometrica. doi:10.1111/1468-0262.00194 — the canonical applied BLP study, recovering markups and market power from estimated demand. [R]

Debate. How much of a BLP estimate is identified by the instruments versus by the parametric functional form and distributional assumptions—and does the practitioner’s toolkit paper understate how fragile the numerics can be?

55.10 Week 9 — Dynamic demand: stockpiling and durables

Topic. Relaxing the static benchmark: when consumers are forward-looking, today’s demand depends on expected future prices, so a price cut shifts the timing of purchases, not only their level.

Subtopics. Inventory and stockpiling on storable goods; forward-looking adoption of durables with falling prices; the distinction between intertemporal substitution and genuine demand expansion; dynamic discrete-choice solution methods.

Methods. Dynamic programming/Bellman value functions in demand; estimating expectations; conditional-choice-probability and nested-fixed-point estimators.

Key readings.

Erdem & Keane (1996), “Decision-Making Under Uncertainty: Capturing Dynamic Brand Choice Processes in Turbulent Consumer Goods Markets,” Marketing Science. doi:10.1287/mksc.15.1.1 — brings forward-looking learning and uncertainty into brand choice, founding the dynamic structural tradition in marketing. [F]
Hendel & Nevo (2006), “Measuring the Implications of Sales and Consumer Inventory Behavior,” Econometrica. doi:10.1111/j.1468-0262.2006.00721.x — shows that ignoring stockpiling biases static demand elasticities and long-run policy conclusions. [F]
Gowrisankaran & Rysman (2012), “Dynamics of Consumer Demand for New Durable Goods,” Journal of Political Economy. doi:10.1086/669540 — a tractable dynamic BLP for durables with falling prices and improving quality. [R]

Debate. Dynamic demand models need consumer expectations the analyst cannot observe—does the rational-expectations closure earn its keep, or does it smuggle in the conclusions it is meant to deliver?

55.11 Week 10 — Search and consideration sets

Topic. Relaxing full information: consumers do not evaluate the entire assortment but form a consideration set through costly search, so observed choice is a choice from an endogenous, latent subset.

Subtopics. Consideration-set formation; sequential vs. simultaneous search; advertising as an awareness shifter; identification of search costs from choice and search data.

Methods. Structural search models; consideration-then-choice two-stage models; identifying consideration from the pattern of choices and (where available) search data.

Key readings.

Roberts & Lattin (1991), “Development and Testing of a Model of Consideration Set Composition,” Journal of Marketing Research. doi:10.2307/3172783 — the early marketing formalization of consideration-set formation and its effect on choice. [F]
Mehta, Rajiv & Srinivasan (2003), “Price Uncertainty and Consumer Search: A Structural Model of Consideration Set Formation,” Marketing Science. doi:10.1287/mksc.22.1.58.12849 — embeds costly search for prices inside a structural choice model. [R]
Honka, Hortaçsu & Vitorino (2017), “Advertising, Consumer Awareness, and Choice: Evidence from the U.S. Banking Industry,” The RAND Journal of Economics. doi:10.1111/1756-2171.12188 — separately identifies awareness, consideration, and choice, with advertising acting on the awareness stage. [R]

Debate. Consideration sets are latent—can the analyst distinguish “not considered” from “considered and rejected” without search data, or does the identification rest entirely on functional form?

55.12 Week 11 — Endogeneity in choice models

Topic. The identification problem at the center of the seminar: prices (and other marketing-mix variables) correlate with unobserved demand shifters, biasing the price coefficient that every counterfactual depends on.

Subtopics. Price endogeneity at the disaggregate level; control functions vs. instrumental variables; copula-based correction without external instruments; unobserved product characteristics.

Methods. Control-function (two-stage residual-inclusion) estimation; the Gaussian copula method; instrument construction and validity testing.

Key readings.

Villas-Boas & Winer (1999), “Endogeneity in Brand Choice Models,” Management Science. doi:10.1287/mnsc.45.10.1324 — documents that ignoring price endogeneity biases brand-choice elasticities in scanner data. [F]
Petrin & Train (2010), “A Control Function Approach to Endogeneity in Consumer Choice Models,” Journal of Marketing Research. doi:10.1509/jmkr.47.1.3 — the control- function alternative to BLP-style instruments at the individual level. [R]
Park & Gupta (2012), “Handling Endogenous Regressors by Joint Estimation Using Copulas,” Marketing Science. doi:10.1287/mksc.1120.0718 — corrects endogeneity without external instruments by exploiting non-normality of the endogenous regressor. [R]

Debate. The copula method buys identification from a distributional assumption (non-normal regressors, normal errors) rather than from an instrument—is trading an exclusion restriction for a distributional one a genuine advance or a relabeling of the maintained assumption?

55.13 Week 12 — Bayesian estimation and MCMC

Topic. The Bayesian computational toolkit that makes rich, heterogeneous choice models estimable: data augmentation, Gibbs sampling, and the Metropolis–Hastings algorithm applied to discrete-choice likelihoods.

Subtopics. Data augmentation for probit/logit; conjugate hierarchical priors; mixing and convergence diagnostics; demand for variety and corner solutions.

Methods. Gibbs sampling with latent-utility augmentation; random-walk Metropolis–Hastings for non-conjugate blocks; hierarchical priors over heterogeneous parameters.

Key readings.

Rossi, Allenby & McCulloch (2005), Bayesian Statistics and Marketing, Wiley — the standard reference uniting Bayesian computation with marketing choice models; cited as a book without a single-work DOI. [F]
Allenby, Arora & Ginter (1998), “On the Heterogeneity of Demand,” Journal of Marketing Research. doi:10.2307/3152035 — a worked demonstration of hierarchical-Bayes heterogeneity recovery and its managerial implications. [F]
Kim, Allenby & Rossi (2002), “Modeling Consumer Demand for Variety,” Marketing Science. doi:10.1287/mksc.21.3.229.143 — extends Bayesian choice modeling to multiple-discreteness and corner solutions beyond single-unit choice. [R]

Debate. Bayesian and classical (simulated-likelihood) estimators of the same mixed logit are asymptotically equivalent—so is the choice between them substantive or merely computational, and when does the prior actually change the conclusion?

55.14 Week 13 — Machine learning and demand

Topic. The machine-learning frontier: flexible prediction of demand and data-driven recovery of heterogeneous responses, in tension and dialogue with structural demand estimation.

Subtopics. Flexible/nonparametric demand prediction and model combination; heterogeneous-treatment-effect estimation by recursive partitioning; deep-learning models of product choice over large assortments; prediction vs. structural identification.

Methods. Regularized and ensemble prediction; causal/honest trees and forests for treatment-effect heterogeneity; neural choice models.

Key readings.

Bajari, Nekipelov, Ryan & Yang (2015), “Machine Learning Methods for Demand Estimation,” American Economic Review (Papers & Proceedings). doi:10.1257/aer.p20151021 — benchmarks machine-learning predictors against structural demand on out-of-sample fit. [R]
Athey & Imbens (2016), “Recursive Partitioning for Heterogeneous Causal Effects,” Proceedings of the National Academy of Sciences. doi:10.1073/pnas.1510489113 — the causal-tree method for recovering heterogeneous marketing responses with valid inference. [R]
Gabel & Timoshenko (2022), “Product Choice with Large Assortments: A Scalable Deep-Learning Model,” Management Science. doi:10.1287/mnsc.2021.3969 — a neural choice model that scales to assortments far larger than classical logit can handle. [R]

Debate. Machine learning excels at prediction but is silent on counterfactuals—can flexible predictors deliver the price elasticities and welfare integrals that structural demand exists to compute, or do they answer a different question?

55.15 Week 14 — Welfare, counterfactuals, and synthesis

Topic. The terminus of the arc: the demand system exists to compute welfare and counterfactuals, so the capstone evaluates new-product gains, merger effects, and the consumer surplus that estimated demand makes measurable.

Subtopics. Compensating variation from the logit inclusive value; the welfare value of new products and variety; merger simulation; the bias from ignoring preference heterogeneity in welfare.

Methods. Consumer-surplus computation from estimated demand; counterfactual equilibrium simulation; the log-sum welfare formula.

Key readings.

Petrin (2002), “Quantifying the Benefits of New Products: The Case of the Minivan,” Journal of Political Economy. doi:10.1086/340779 — uses estimated BLP demand to compute the consumer-welfare gain from a product innovation, the paradigmatic payoff of the whole enterprise. [F]
Trajtenberg (1989), “The Welfare Analysis of Product Innovations, with an Application to Computed Tomography Scanners,” Journal of Political Economy. doi:10.1086/261611 — an early discrete-choice welfare analysis of innovation that anticipates the new-goods literature. [F]
Nevo (2001), “Measuring Market Power in the Ready-to-Eat Cereal Industry,” Econometrica. doi:10.1111/1468-0262.00194 — revisited as the canonical counterfactual exercise: recovering markups and simulating mergers from estimated demand. [R]

Debate. Counterfactual welfare numbers inherit every maintained assumption of the demand model—how much should a regulator or manager trust a surplus estimate whose magnitude depends on an unidentified tail of the taste distribution?

55.16 Foundational vs. frontier at a glance

The foundational core—the canon a choice-and-demand scholar must know cold—runs from the random-utility microfoundation through the aggregate-demand revolution: Thurstone (1927), Luce (1959), and McFadden (1974, 1980) for the RUM and its GEV extensions; Guadagni & Little (1983) and Gupta (1988) for the scanner-choice application; McFadden & Train (2000) and Revelt & Train (1998) for mixed logit; Allenby & Rossi (1998) and Rossi, McCulloch & Allenby (1996) for hierarchical-Bayes heterogeneity; Green & Srinivasan (1978, 1990) for preference measurement; Berry (1994) and Berry, Levinsohn & Pakes (1995) for aggregate demand; Erdem & Keane (1996) and Hendel & Nevo (2006) for dynamics; Villas-Boas & Winer (1999) for endogeneity; and Petrin (2002) and Trajtenberg (1989) for welfare. These are the papers and books a dissertation in the area is expected to cite without prompting.

The frontier, refreshed each edition, tracks where the field is actively moving: Nevo (2001) and Gowrisankaran & Rysman (2012) for applied and dynamic BLP; Mehta, Rajiv & Srinivasan (2003) and Honka, Hortaçsu & Vitorino (2017) for structural search and consideration; Petrin & Train (2010) and Park & Gupta (2012) for endogeneity corrections; Kim, Allenby & Rossi (2002) for multiple-discreteness; and Bajari et al. (2015), Athey & Imbens (2016), and Gabel & Timoshenko (2022) for the machine-learning interface. The split is pedagogical, not chronological: a 1995 paper is foundational because the field still builds on its inversion and GMM, while a 2016 method is “frontier” because its integration with structural demand is still being worked out.

55.17 How this chapter expands

The weekly map is a backbone designed to grow along several axes.

A refresh cadence of two to three years on the frontier modules. The machine-learning, dynamic-demand, and search modules turn over fastest; frontier readings should be replaced or supplemented as new methods and applications appear, while the foundational anchors (the RUM, the Berry inversion, mixed logit) stay fixed.
Emerging modules as the field grows: machine-learning demand at scale (regularized and ensemble demand systems, double/debiased estimation of demand parameters); LLM- and embedding-based preference models that represent products and consumers in learned latent spaces and raise fresh identification questions; and demand estimation under privacy constraints and aggregated/differentially private data. Each should follow the template—foundational anchor, frontier paper, identification debate.
A parallel methods spine. Each module already names its identifying assumption (the extreme-value error, the nesting tree, the mixing distribution, the BLP exclusion restriction, the rational-expectations closure, the copula’s distributional assumption). A future edition should add a short companion per week that states the estimator, its moment or likelihood, and what breaks identification—turning the reading map into a methods course, of which the worked BLP section below is the model.
An applications track pairing each estimator with a canonical industry case (autos, cereal, durables, banking, retail assortments), so students see the same machinery re-identified across settings with different sources of exogenous variation.

The following section supplies the worked treatment the map points to.

55.17.1 BLP demand as an estimator

The Berry (1994) / Berry–Levinsohn–Pakes (1995) model is the hinge of the whole seminar: it carries the random-utility logic from individual choice to aggregate demand and confronts price endogeneity head-on. The treatment here states the estimator, its inversion, and the single assumption on which identification rests.

Consumer $i$ in market $t$ derives utility from product $j$, \[ u_{ijt} = \mathbf{x}_{jt}^{\top}\boldsymbol{\beta}_i - \alpha_i\, p_{jt} + \xi_{jt} + \varepsilon_{ijt}, \tag{55.1}\] where $\mathbf{x}_{jt}$ are observed characteristics, $p_{jt}$ is price, $\xi_{jt}$ is the unobserved (to the analyst) product quality in market $t$, and $\varepsilon_{ijt}$ is an i.i.d. extreme-value taste shock. The random coefficients $(\boldsymbol{\beta}_i, \alpha_i)$ vary across consumers with a parametric distribution, say $(\boldsymbol{\beta}_i,\alpha_i) = (\bar{\boldsymbol{\beta}},\bar\alpha) + \boldsymbol{\Sigma}\,\boldsymbol{\nu}_i$ with $\boldsymbol{\nu}_i \sim \Phi$. The outside good $j=0$ has utility normalized to $u_{i0t}=\varepsilon_{i0t}$.

Integrating the logit choice probabilities over the taste distribution gives the model’s predicted market share for product $j$, \[ s_{jt}(\boldsymbol{\delta}_t,\boldsymbol{\theta}_2) = \int \frac{\exp\!\big(\delta_{jt} + \mu_{ijt}(\boldsymbol{\theta}_2)\big)} {1 + \sum_{k} \exp\!\big(\delta_{kt} + \mu_{ikt}(\boldsymbol{\theta}_2)\big)} \, d\Phi(\boldsymbol{\nu}), \tag{55.2}\] where $\delta_{jt} = \mathbf{x}_{jt}^{\top}\bar{\boldsymbol{\beta}} - \bar\alpha\,p_{jt} + \xi_{jt}$ is the mean utility of product $j$ and $\mu_{ijt}$ collects the individual deviations governed by the heterogeneity parameters $\boldsymbol{\theta}_2=\mathrm{vec}(\boldsymbol{\Sigma})$. The integral has no closed form and is approximated by simulation over draws of $\boldsymbol{\nu}_i$.

The key result of Berry (1994) is that, holding $\boldsymbol{\theta}_2$ fixed, the mapping from mean utilities to shares is invertible: there is a unique vector $\boldsymbol{\delta}_t$ that equates predicted shares to observed shares $\mathbf{S}_t$. BLP compute it as the fixed point of the contraction mapping \[ \delta_{jt}^{(r+1)} = \delta_{jt}^{(r)} + \ln S_{jt} - \ln s_{jt}\!\big(\boldsymbol{\delta}_t^{(r)},\boldsymbol{\theta}_2\big), \tag{55.3}\] iterated to convergence. This inversion turns an aggregate-demand problem into a linear-in-$\xi$ model: once $\boldsymbol{\delta}_t(\boldsymbol{\theta}_2)$ is recovered, the structural error is \[ \xi_{jt}(\boldsymbol{\theta}) = \delta_{jt}(\boldsymbol{\theta}_2) - \mathbf{x}_{jt}^{\top}\bar{\boldsymbol{\beta}} + \bar\alpha\, p_{jt}. \tag{55.4}\]

Estimation proceeds by GMM. Because price $p_{jt}$ is correlated with the unobserved quality $\xi_{jt}$ (firms set higher prices for products with higher $\xi$), ordinary regression of $\delta$ on price is biased. Identification instead requires a vector of instruments $\mathbf{z}_{jt}$ that satisfy the population moment condition \[ \mathbb{E}\!\left[\, \xi_{jt}(\boldsymbol{\theta}_0)\,\big|\, \mathbf{z}_{jt} \right] = 0 \quad\Longrightarrow\quad \mathbb{E}\!\left[\, \mathbf{z}_{jt}\, \xi_{jt}(\boldsymbol{\theta}_0)\,\right]=0, \tag{55.5}\] and the estimator minimizes the empirical analogue, \[ \hat{\boldsymbol{\theta}} = \arg\min_{\boldsymbol{\theta}}\; \Big(\tfrac{1}{N}\textstyle\sum_{jt}\mathbf{z}_{jt}\,\xi_{jt}(\boldsymbol{\theta})\Big)^{\!\top} \mathbf{W} \Big(\tfrac{1}{N}\textstyle\sum_{jt}\mathbf{z}_{jt}\,\xi_{jt}(\boldsymbol{\theta})\Big), \tag{55.6}\] with weight matrix $\mathbf{W}$. The standard instruments are cost shifters and functions of rival products’ characteristics (the BLP instruments), which shift a product’s price and markup through competition without entering its own demand.

The entire causal content rests on one identifying assumption: instrument exogeneity, $\mathbb{E}[\xi_{jt}\mid \mathbf{z}_{jt}]=0$. Substantively, this requires that product characteristics and cost shifters be uncorrelated with the unobserved quality $\xi_{jt}$. That is precisely the assumption the Week 7 debate contests—firms that observe $\xi$ when they design and price products may choose characteristics correlated with it, in which case the instruments fail and the recovered price elasticity, along with every markup, merger, and welfare number built on it, is biased. As with the seminar’s other estimators, the BLP demand parameter becomes a finding only once its moment condition and its exclusion restriction are stated and defended in full.

55.18 Key Takeaways

The seminar’s spine is a single arc—random utility → individual choice → aggregate demand → welfare—and a student’s job is to see pricing, assortment, merger, and new-product questions as instances of one estimation problem with one recurring threat, endogenous prices correlated with unobserved quality.
The microfoundation runs from Thurstone (1927), Luce (1959), and McFadden (1974,
1. through the scanner-choice application of Guadagni & Little (1983); the IIA limitation of the simple logit is repaired first by GEV/nested logit and then dissolved by the mixed logit, which McFadden & Train (2000) show can approximate any random-utility model.
Heterogeneity is recovered either by hierarchical Bayes (Allenby & Rossi
1. or by the mixing distribution of mixed logit; the two are asymptotically equivalent, so the choice between Bayesian MCMC and simulated likelihood is largely computational.
The Berry inversion and the BLP random-coefficients model (Berry 1994; Berry, Levinsohn & Pakes 1995) carry the RUM to market-level data; as Section 55.17.1 shows, the contraction (Equation 55.3) linearizes the model in $\xi$ and the GMM moment (Equation 55.5) identifies demand only under instrument exogeneity.
Relaxing the static, full-information benchmark—stockpiling and durables (Erdem & Keane 1996; Hendel & Nevo 2006; Gowrisankaran & Rysman 2012) and search/consideration (Roberts & Lattin 1991; Honka et al. 2017)—buys realism at the cost of additional unobservables (expectations, latent consideration) that must themselves be identified.
Endogeneity corrections (Villas-Boas & Winer 1999; Petrin & Train 2010; Park & Gupta 2012) and the machine-learning frontier (Bajari et al. 2015; Athey & Imbens 2016; Gabel & Timoshenko 2022) extend the toolkit, but the demand system’s ultimate purpose remains the welfare and counterfactual computation that Petrin
1. and Trajtenberg (1989) exemplify—numbers that inherit every maintained assumption of the model that produced them.

# Choice Modeling and Demand Estimation Seminar {#sec-choice-demand-seminar} *Choice modeling* and *demand estimation* are the empirical heart of quantitative marketing and of the industrial-organization tradition it borrows from. The field rests on a single, powerful idea: that observed purchases are the realized outcome of a latent, utility-maximizing comparison among alternatives, so that a model of how individuals *choose* can be aggregated into a model of how a market *demands*. This seminar traces the arc from the random-utility microfoundation, through the estimation of individual choice on disaggregate panel data, to the recovery of aggregate demand systems from market-level data—and on to the counterfactual and welfare questions those demand systems were built to answer. It is the doctoral reading-map companion to the technical chapters that develop the estimators in full: the choice and Bayesian-methods chapter (@sec-choice-bayesian), and the structural-modeling and industrial-organization material, to which this seminar stands in the same relation that the analytical-modeling seminar stands to its methods chapters. The science matters because demand is the object on which nearly every managerial and policy question ultimately rests. A price elasticity is a demand derivative; a merger simulation is a demand counterfactual; the welfare gain from a new product is an integral under a demand curve; the return on an advertising or assortment decision is a demand response. Commercially, the same machinery powers pricing engines, assortment optimization, and the regulatory review of mergers in differentiated-product industries. The intellectual payoff of the seminar is that it teaches a student to see these disparate applications as instances of one estimation problem and to know which assumptions each application is buying. The central tension that animates every week is *identification*. Preferences are never observed; only choices are. The analyst must recover taste parameters—and, crucially, price sensitivity—from data in which the very prices a consumer faces are set by firms who observe demand shifters the analyst does not. Endogenous prices, unobserved product quality, and unrestricted preference heterogeneity all threaten to contaminate the elasticity estimates on which everything downstream depends. The history of the field is in large part a history of instruments, control functions, and distributional assumptions devised to break this contamination. A student who completes the seminar can read a demand paper and locate, immediately, the source of variation that identifies its price coefficient, the assumption that makes that variation exogenous, and the counterfactual that assumption licenses. ## Semester arc The fourteen-week arc moves along three nested levels. The first third (Weeks 1–4) builds the **random-utility microfoundation and its disaggregate estimators**: McFadden's conditional logit and its psychological antecedents in Thurstone and Luce; the first marketing applications to scanner panels; the independence-of-irrelevant-alternatives (IIA) problem and its nested-logit and generalized-extreme-value (GEV) repairs; and the random-coefficients or mixed logit that dissolves IIA by integrating taste heterogeneity out of the choice probabilities. The middle third (Weeks 5–8) confronts **heterogeneity and aggregation**: hierarchical Bayes estimation of individual-level parameters, conjoint and stated-preference measurement, and then the pivotal transition from individual choice to *aggregate* demand in the Berry inversion and the Berry–Levinsohn–Pakes (BLP) random-coefficients demand model, together with the practitioner's craft of estimating it. The final third (Weeks 9–14) takes the static demand system *dynamic and strategic* and then closes the loop to welfare. Weeks 9 and 10 relax the static, full-information benchmark: consumers stockpile, wait for durable-good prices to fall, and search over an endogenous consideration set rather than evaluating the full assortment. Weeks 11 and 12 are method-forward, treating endogeneity in choice models (control functions, copulas, instruments) and the Bayesian and Markov-chain-Monte-Carlo (MCMC) machinery that makes rich heterogeneous models estimable. Week 13 brings the machine-learning frontier—flexible demand prediction and heterogeneous-treatment-effect estimation—into contact with structural demand. Week 14 returns to first principles: the demand system exists to compute welfare and counterfactuals, so the capstone reads the new-product and merger-welfare papers that justify the entire enterprise. The reading map uses two tags. **[F] = Foundational** marks canon a choice-and- demand scholar is expected to know cold; **[R] = Frontier/Recent** marks an active research front, refreshed as the literature moves. Each week pairs at least one foundational anchor with a more recent or methodological counterpart. DOIs are reproduced as verified against the Crossref REST API; works without a verifiable DOI—chiefly the canonical books and a 1974 edited-volume chapter—are named without a link and flagged as such in the text. ## Week 1 — Random-utility foundations {#sec-cdem-week01} **Topic.** The random-utility model (RUM) as the microfoundation of all that follows: a chooser maximizes a latent utility that the analyst observes only up to a random component, and the distribution of that component generates the choice probabilities. **Subtopics.** Deterministic vs. stochastic utility; the additive-error specification; the extreme-value distributional assumption and the logit closed form; the psychological roots of stochastic choice. **Methods.** Maximum-likelihood estimation of the multinomial/conditional logit; identification up to scale and level; the log-likelihood and its concavity. **Key readings.** - McFadden (1974), "Conditional Logit Analysis of Qualitative Choice Behavior," in Zarembka (ed.), *Frontiers in Econometrics*, Academic Press — the founding derivation linking utility maximization with extreme-value errors to the logit choice probability; no Crossref DOI (edited-volume chapter), named without link. **[F]** - Luce (1959), *Individual Choice Behavior: A Theoretical Analysis*, Wiley — the choice-axiom (independence from irrelevant alternatives) that the logit operationalizes; canonical book, no DOI. **[F]** - Thurstone (1927), "A Law of Comparative Judgment," *Psychological Review* — the original random-utility/discriminal-process idea that anchors the whole tradition. [doi:10.1037/h0070288](https://doi.org/10.1037/h0070288) — the probit-flavored ancestor of stochastic choice. **[F]** **Debate.** Is the logit's tractable extreme-value error a substantive behavioral claim or a mathematical convenience whose IIA implication must later be undone? ## Week 2 — Brand choice on scanner data {#sec-cdem-week02} **Topic.** The first great marketing application of the RUM: estimating brand-choice probabilities from supermarket scanner panels, where the same household is observed choosing repeatedly across many purchase occasions. **Subtopics.** Loyalty and state dependence; the brand-loyalty and purchase-event-feedback variables; marketing-mix covariates (price, display, feature) in the utility; the incidence/quantity decisions surrounding brand choice. **Methods.** Conditional logit on panel choice data; constructing exponentially smoothed loyalty regressors; modeling purchase timing and quantity jointly with brand choice. **Key readings.** - Guadagni & Little (1983), "A Logit Model of Brand Choice Calibrated on Scanner Data," *Marketing Science*. [doi:10.1287/mksc.2.3.203](https://doi.org/10.1287/mksc.2.3.203) — the paper that brought the conditional logit into marketing and defined the loyalty-variable template. **[F]** - Gupta (1988), "Impact of Sales Promotions on When, What, and How Much to Buy," *Journal of Marketing Research*. [doi:10.2307/3172945](https://doi.org/10.2307/3172945) — decomposes promotional response into purchase-timing, brand-choice, and quantity components, extending the scanner-choice program beyond brand choice alone. **[F]** **Debate.** Does the loyalty variable capture genuine structural state dependence or merely soak up persistent unobserved heterogeneity—an identification confound that later mixed-logit and hierarchical-Bayes work would target directly? ## Week 3 — IIA, nested logit, and GEV {#sec-cdem-week03} **Topic.** The Achilles' heel of the simple logit—independence from irrelevant alternatives—and the family of generalized-extreme-value models built to relax it while preserving a closed form. **Subtopics.** The red-bus/blue-bus problem; proportional substitution and its implausibility; nesting correlated alternatives; the GEV class and its inclusive-value structure; cross-elasticity patterns. **Methods.** Nested-logit estimation; the inclusive value and dissimilarity parameter; testing IIA; GEV generating functions. **Key readings.** - McFadden (1980), "Econometric Models for Probabilistic Choice Among Products," *The Journal of Business*. [doi:10.1086/296093](https://doi.org/10.1086/296093) — develops the GEV framework and the modeling of substitution patterns among differentiated products, the direct ancestor of structural demand. **[F]** - Guadagni & Little (1983), "A Logit Model of Brand Choice Calibrated on Scanner Data," *Marketing Science*. [doi:10.1287/mksc.2.3.203](https://doi.org/10.1287/mksc.2.3.203) — revisited here as the canonical setting where IIA's restrictive substitution is most visibly violated by close-substitute brands. **[F]** **Debate.** Nested logit fixes IIA *within* a pre-specified tree—but who chooses the tree, and does an analyst-imposed nesting merely relocate the maintained assumption rather than test it? ## Week 4 — Random coefficients and mixed logit {#sec-cdem-week04} **Topic.** The mixed (random-coefficients) logit: letting taste parameters vary across individuals according to a mixing distribution dissolves IIA entirely and can approximate any random-utility model. **Subtopics.** The mixing distribution over tastes; correlated random coefficients; flexible substitution patterns; panel mixed logit with repeated choices. **Methods.** Simulated maximum likelihood and maximum simulated likelihood; Halton draws; the McFadden–Train approximation theorem. **Key readings.** - McFadden & Train (2000), "Mixed MNL Models for Discrete Response," *Journal of Applied Econometrics*. [doi:10.1002/1099-1255(200009/10)15:5<447::aid-jae570>3.0.co;2-1](https://doi.org/10.1002/1099-1255(200009/10)15:5<447::aid-jae570>3.0.co;2-1) — proves that mixed logit can approximate any RUM choice probabilities arbitrarily well; the theoretical license for the whole approach. **[F]** - Revelt & Train (1998), "Mixed Logit with Repeated Choices: Households' Choices of Appliance Efficiency Level," *The Review of Economics and Statistics*. [doi:10.1162/003465398557735](https://doi.org/10.1162/003465398557735) — the template for estimating mixed logit on panel data with within-household correlation. **[F]** - Train (2009), *Discrete Choice Methods with Simulation*, 2nd ed., Cambridge University Press — the standard graduate reference for simulation-based estimation; cited as a book without a single-work DOI. **[F]** **Debate.** Is the freedom of the mixing distribution a virtue or a trap—does a mis-specified parametric mixing law (e.g., normal price coefficients implying positive price sensitivity) distort the very heterogeneity it claims to recover? ## Week 5 — Heterogeneity and hierarchical Bayes {#sec-cdem-week05} **Topic.** Estimating *individual-level* preference parameters by pooling information across consumers through a hierarchical prior—the Bayesian counterpart to mixed logit, and the workhorse of marketing's heterogeneity tradition. **Subtopics.** The random-effects/hierarchical prior; shrinkage toward the population mean; individual-level posteriors for targeting; the value of purchase history. **Methods.** Hierarchical Bayes; Gibbs sampling and data augmentation; the multivariate-normal heterogeneity distribution; posterior means as individual estimates. **Key readings.** - Allenby & Rossi (1998), "Marketing Models of Consumer Heterogeneity," *Journal of Econometrics*. [doi:10.1016/S0304-4076(98)00055-4](https://doi.org/10.1016/S0304-4076(98)00055-4) — the manifesto for treating heterogeneity as the object of interest rather than a nuisance, with the hierarchical-Bayes machinery to recover it. **[F]** - Rossi, McCulloch & Allenby (1996), "The Value of Purchase History Data in Target Marketing," *Marketing Science*. [doi:10.1287/mksc.15.4.321](https://doi.org/10.1287/mksc.15.4.321) — shows how individual-level posteriors sharpen targeting, the commercial payoff of heterogeneity estimation. **[F]** **Debate.** Do individual-level posteriors recover real heterogeneity or merely project a parametric prior onto thin per-person data—and how much purchase history is enough before the prior stops doing the work? ## Week 6 — Conjoint analysis and preference measurement {#sec-cdem-week06} **Topic.** Eliciting preferences from designed choice experiments rather than market data—the stated-preference complement to revealed-preference demand estimation. **Subtopics.** Full-profile vs. choice-based conjoint; experimental design and attribute balance; part-worth utilities; the bridge from conjoint to choice simulators and willingness-to-pay. **Methods.** Designed factorial/fractional experiments; choice-based conjoint estimated by (hierarchical Bayes) logit; reliability and external validity. **Key readings.** - Green & Srinivasan (1978), "Conjoint Analysis in Consumer Research: Issues and Outlook," *Journal of Consumer Research*. [doi:10.1086/208721](https://doi.org/10.1086/208721) — the field-defining review that established conjoint as marketing's preference-measurement method. **[F]** - Green & Srinivasan (1990), "Conjoint Analysis in Marketing: New Developments with Implications for Research and Practice," *Journal of Marketing*. [doi:10.1177/002224299005400402](https://doi.org/10.1177/002224299005400402) — the decadal update that reoriented conjoint toward choice-based designs compatible with the RUM. **[F]** **Debate.** Do stated preferences from hypothetical choices recover the same parameters as revealed preferences from market choices—and when does hypothetical bias make conjoint willingness-to-pay an unreliable demand input? ## Week 7 — Aggregate demand and the BLP model {#sec-cdem-week07} **Topic.** The pivotal transition from individual choice to *aggregate* demand: recovering structural demand parameters from market-level shares when individual data are unavailable and prices are endogenous. **Subtopics.** The unobserved product-quality term; the market-share inversion; random coefficients at the market level; instrumenting for price. **Methods.** The Berry inversion (shares to mean utilities); the BLP contraction mapping; generalized method of moments (GMM) with demand-side instruments. **Key readings.** - Berry (1994), "Estimating Discrete-Choice Models of Product Differentiation," *The RAND Journal of Economics*. [doi:10.2307/2555829](https://doi.org/10.2307/2555829) — introduces the share-inversion idea that turns an aggregate-demand problem into a linear instrumental-variables problem in mean utilities. **[F]** - Berry, Levinsohn & Pakes (1995), "Automobile Prices in Market Equilibrium," *Econometrica*. [doi:10.2307/2171802](https://doi.org/10.2307/2171802) — the full random-coefficients aggregate-demand model with supply-side equilibrium; the single most influential demand paper of the era. **[F]** **Debate.** The BLP instruments (rival-product characteristics, cost shifters) rest on the assumption that product characteristics are exogenous to the unobserved quality $\xi$—is that assumption credible when firms design products knowing $\xi$? The worked treatment in @sec-cdem-worked develops the estimator and this exclusion restriction in full. ## Week 8 — BLP in practice {#sec-cdem-week08} **Topic.** The craft of actually estimating random-coefficients aggregate demand: the numerical, instrumental, and identification choices that separate a credible BLP estimate from a fragile one. **Subtopics.** Choice of instruments and the weak-instrument problem; the inner contraction and outer GMM loops; numerical pitfalls and convergence; market power and markups recovered from demand. **Methods.** Nested fixed-point GMM; optimal instruments; supply-side pricing moments; the cereal-industry application. **Key readings.** - Nevo (2000), "A Practitioner's Guide to Estimation of Random-Coefficients Logit Models of Demand," *Journal of Economics & Management Strategy*. [doi:10.1162/105864000567954](https://doi.org/10.1162/105864000567954) — the pedagogical guide that made BLP estimable by a generation of students. **[F]** - Nevo (2001), "Measuring Market Power in the Ready-to-Eat Cereal Industry," *Econometrica*. [doi:10.1111/1468-0262.00194](https://doi.org/10.1111/1468-0262.00194) — the canonical applied BLP study, recovering markups and market power from estimated demand. **[R]** **Debate.** How much of a BLP estimate is identified by the instruments versus by the parametric functional form and distributional assumptions—and does the practitioner's toolkit paper understate how fragile the numerics can be? ## Week 9 — Dynamic demand: stockpiling and durables {#sec-cdem-week09} **Topic.** Relaxing the static benchmark: when consumers are forward-looking, today's demand depends on expected future prices, so a price cut shifts the *timing* of purchases, not only their level. **Subtopics.** Inventory and stockpiling on storable goods; forward-looking adoption of durables with falling prices; the distinction between intertemporal substitution and genuine demand expansion; dynamic discrete-choice solution methods. **Methods.** Dynamic programming/Bellman value functions in demand; estimating expectations; conditional-choice-probability and nested-fixed-point estimators. **Key readings.** - Erdem & Keane (1996), "Decision-Making Under Uncertainty: Capturing Dynamic Brand Choice Processes in Turbulent Consumer Goods Markets," *Marketing Science*. [doi:10.1287/mksc.15.1.1](https://doi.org/10.1287/mksc.15.1.1) — brings forward-looking learning and uncertainty into brand choice, founding the dynamic structural tradition in marketing. **[F]** - Hendel & Nevo (2006), "Measuring the Implications of Sales and Consumer Inventory Behavior," *Econometrica*. [doi:10.1111/j.1468-0262.2006.00721.x](https://doi.org/10.1111/j.1468-0262.2006.00721.x) — shows that ignoring stockpiling biases static demand elasticities and long-run policy conclusions. **[F]** - Gowrisankaran & Rysman (2012), "Dynamics of Consumer Demand for New Durable Goods," *Journal of Political Economy*. [doi:10.1086/669540](https://doi.org/10.1086/669540) — a tractable dynamic BLP for durables with falling prices and improving quality. **[R]** **Debate.** Dynamic demand models need consumer expectations the analyst cannot observe—does the rational-expectations closure earn its keep, or does it smuggle in the conclusions it is meant to deliver? ## Week 10 — Search and consideration sets {#sec-cdem-week10} **Topic.** Relaxing full information: consumers do not evaluate the entire assortment but form a *consideration set* through costly search, so observed choice is a choice from an endogenous, latent subset. **Subtopics.** Consideration-set formation; sequential vs. simultaneous search; advertising as an awareness shifter; identification of search costs from choice and search data. **Methods.** Structural search models; consideration-then-choice two-stage models; identifying consideration from the pattern of choices and (where available) search data. **Key readings.** - Roberts & Lattin (1991), "Development and Testing of a Model of Consideration Set Composition," *Journal of Marketing Research*. [doi:10.2307/3172783](https://doi.org/10.2307/3172783) — the early marketing formalization of consideration-set formation and its effect on choice. **[F]** - Mehta, Rajiv & Srinivasan (2003), "Price Uncertainty and Consumer Search: A Structural Model of Consideration Set Formation," *Marketing Science*. [doi:10.1287/mksc.22.1.58.12849](https://doi.org/10.1287/mksc.22.1.58.12849) — embeds costly search for prices inside a structural choice model. **[R]** - Honka, Hortaçsu & Vitorino (2017), "Advertising, Consumer Awareness, and Choice: Evidence from the U.S. Banking Industry," *The RAND Journal of Economics*. [doi:10.1111/1756-2171.12188](https://doi.org/10.1111/1756-2171.12188) — separately identifies awareness, consideration, and choice, with advertising acting on the awareness stage. **[R]** **Debate.** Consideration sets are latent—can the analyst distinguish "not considered" from "considered and rejected" without search data, or does the identification rest entirely on functional form? ## Week 11 — Endogeneity in choice models {#sec-cdem-week11} **Topic.** The identification problem at the center of the seminar: prices (and other marketing-mix variables) correlate with unobserved demand shifters, biasing the price coefficient that every counterfactual depends on. **Subtopics.** Price endogeneity at the disaggregate level; control functions vs. instrumental variables; copula-based correction without external instruments; unobserved product characteristics. **Methods.** Control-function (two-stage residual-inclusion) estimation; the Gaussian copula method; instrument construction and validity testing. **Key readings.** - Villas-Boas & Winer (1999), "Endogeneity in Brand Choice Models," *Management Science*. [doi:10.1287/mnsc.45.10.1324](https://doi.org/10.1287/mnsc.45.10.1324) — documents that ignoring price endogeneity biases brand-choice elasticities in scanner data. **[F]** - Petrin & Train (2010), "A Control Function Approach to Endogeneity in Consumer Choice Models," *Journal of Marketing Research*. [doi:10.1509/jmkr.47.1.3](https://doi.org/10.1509/jmkr.47.1.3) — the control- function alternative to BLP-style instruments at the individual level. **[R]** - Park & Gupta (2012), "Handling Endogenous Regressors by Joint Estimation Using Copulas," *Marketing Science*. [doi:10.1287/mksc.1120.0718](https://doi.org/10.1287/mksc.1120.0718) — corrects endogeneity without external instruments by exploiting non-normality of the endogenous regressor. **[R]** **Debate.** The copula method buys identification from a distributional assumption (non-normal regressors, normal errors) rather than from an instrument—is trading an exclusion restriction for a distributional one a genuine advance or a relabeling of the maintained assumption? ## Week 12 — Bayesian estimation and MCMC {#sec-cdem-week12} **Topic.** The Bayesian computational toolkit that makes rich, heterogeneous choice models estimable: data augmentation, Gibbs sampling, and the Metropolis–Hastings algorithm applied to discrete-choice likelihoods. **Subtopics.** Data augmentation for probit/logit; conjugate hierarchical priors; mixing and convergence diagnostics; demand for variety and corner solutions. **Methods.** Gibbs sampling with latent-utility augmentation; random-walk Metropolis–Hastings for non-conjugate blocks; hierarchical priors over heterogeneous parameters. **Key readings.** - Rossi, Allenby & McCulloch (2005), *Bayesian Statistics and Marketing*, Wiley — the standard reference uniting Bayesian computation with marketing choice models; cited as a book without a single-work DOI. **[F]** - Allenby, Arora & Ginter (1998), "On the Heterogeneity of Demand," *Journal of Marketing Research*. [doi:10.2307/3152035](https://doi.org/10.2307/3152035) — a worked demonstration of hierarchical-Bayes heterogeneity recovery and its managerial implications. **[F]** - Kim, Allenby & Rossi (2002), "Modeling Consumer Demand for Variety," *Marketing Science*. [doi:10.1287/mksc.21.3.229.143](https://doi.org/10.1287/mksc.21.3.229.143) — extends Bayesian choice modeling to multiple-discreteness and corner solutions beyond single-unit choice. **[R]** **Debate.** Bayesian and classical (simulated-likelihood) estimators of the same mixed logit are asymptotically equivalent—so is the choice between them substantive or merely computational, and when does the prior actually change the conclusion? ## Week 13 — Machine learning and demand {#sec-cdem-week13} **Topic.** The machine-learning frontier: flexible prediction of demand and data-driven recovery of heterogeneous responses, in tension and dialogue with structural demand estimation. **Subtopics.** Flexible/nonparametric demand prediction and model combination; heterogeneous-treatment-effect estimation by recursive partitioning; deep-learning models of product choice over large assortments; prediction vs. structural identification. **Methods.** Regularized and ensemble prediction; causal/honest trees and forests for treatment-effect heterogeneity; neural choice models. **Key readings.** - Bajari, Nekipelov, Ryan & Yang (2015), "Machine Learning Methods for Demand Estimation," *American Economic Review (Papers & Proceedings)*. [doi:10.1257/aer.p20151021](https://doi.org/10.1257/aer.p20151021) — benchmarks machine-learning predictors against structural demand on out-of-sample fit. **[R]** - Athey & Imbens (2016), "Recursive Partitioning for Heterogeneous Causal Effects," *Proceedings of the National Academy of Sciences*. [doi:10.1073/pnas.1510489113](https://doi.org/10.1073/pnas.1510489113) — the causal-tree method for recovering heterogeneous marketing responses with valid inference. **[R]** - Gabel & Timoshenko (2022), "Product Choice with Large Assortments: A Scalable Deep-Learning Model," *Management Science*. [doi:10.1287/mnsc.2021.3969](https://doi.org/10.1287/mnsc.2021.3969) — a neural choice model that scales to assortments far larger than classical logit can handle. **[R]** **Debate.** Machine learning excels at prediction but is silent on counterfactuals—can flexible predictors deliver the price elasticities and welfare integrals that structural demand exists to compute, or do they answer a different question? ## Week 14 — Welfare, counterfactuals, and synthesis {#sec-cdem-week14} **Topic.** The terminus of the arc: the demand system exists to compute welfare and counterfactuals, so the capstone evaluates new-product gains, merger effects, and the consumer surplus that estimated demand makes measurable. **Subtopics.** Compensating variation from the logit inclusive value; the welfare value of new products and variety; merger simulation; the bias from ignoring preference heterogeneity in welfare. **Methods.** Consumer-surplus computation from estimated demand; counterfactual equilibrium simulation; the log-sum welfare formula. **Key readings.** - Petrin (2002), "Quantifying the Benefits of New Products: The Case of the Minivan," *Journal of Political Economy*. [doi:10.1086/340779](https://doi.org/10.1086/340779) — uses estimated BLP demand to compute the consumer-welfare gain from a product innovation, the paradigmatic payoff of the whole enterprise. **[F]** - Trajtenberg (1989), "The Welfare Analysis of Product Innovations, with an Application to Computed Tomography Scanners," *Journal of Political Economy*. [doi:10.1086/261611](https://doi.org/10.1086/261611) — an early discrete-choice welfare analysis of innovation that anticipates the new-goods literature. **[F]** - Nevo (2001), "Measuring Market Power in the Ready-to-Eat Cereal Industry," *Econometrica*. [doi:10.1111/1468-0262.00194](https://doi.org/10.1111/1468-0262.00194) — revisited as the canonical counterfactual exercise: recovering markups and simulating mergers from estimated demand. **[R]** **Debate.** Counterfactual welfare numbers inherit every maintained assumption of the demand model—how much should a regulator or manager trust a surplus estimate whose magnitude depends on an unidentified tail of the taste distribution? ## Foundational vs. frontier at a glance {#sec-cdem-ff} The **foundational core**—the canon a choice-and-demand scholar must know cold—runs from the random-utility microfoundation through the aggregate-demand revolution: Thurstone (1927), Luce (1959), and McFadden (1974, 1980) for the RUM and its GEV extensions; Guadagni & Little (1983) and Gupta (1988) for the scanner-choice application; McFadden & Train (2000) and Revelt & Train (1998) for mixed logit; Allenby & Rossi (1998) and Rossi, McCulloch & Allenby (1996) for hierarchical-Bayes heterogeneity; Green & Srinivasan (1978, 1990) for preference measurement; Berry (1994) and Berry, Levinsohn & Pakes (1995) for aggregate demand; Erdem & Keane (1996) and Hendel & Nevo (2006) for dynamics; Villas-Boas & Winer (1999) for endogeneity; and Petrin (2002) and Trajtenberg (1989) for welfare. These are the papers and books a dissertation in the area is expected to cite without prompting. The **frontier**, refreshed each edition, tracks where the field is actively moving: Nevo (2001) and Gowrisankaran & Rysman (2012) for applied and dynamic BLP; Mehta, Rajiv & Srinivasan (2003) and Honka, Hortaçsu & Vitorino (2017) for structural search and consideration; Petrin & Train (2010) and Park & Gupta (2012) for endogeneity corrections; Kim, Allenby & Rossi (2002) for multiple-discreteness; and Bajari et al. (2015), Athey & Imbens (2016), and Gabel & Timoshenko (2022) for the machine-learning interface. The split is pedagogical, not chronological: a 1995 paper is foundational because the field still builds on its inversion and GMM, while a 2016 method is "frontier" because its integration with structural demand is still being worked out. ## How this chapter expands {#sec-cdem-expands} The weekly map is a backbone designed to grow along several axes. 1. **A refresh cadence of two to three years on the frontier modules.** The machine-learning, dynamic-demand, and search modules turn over fastest; frontier readings should be replaced or supplemented as new methods and applications appear, while the foundational anchors (the RUM, the Berry inversion, mixed logit) stay fixed. 2. **Emerging modules as the field grows:** machine-learning demand at scale (regularized and ensemble demand systems, double/debiased estimation of demand parameters); LLM- and embedding-based preference models that represent products and consumers in learned latent spaces and raise fresh identification questions; and demand estimation under privacy constraints and aggregated/differentially private data. Each should follow the template—foundational anchor, frontier paper, identification debate. 3. **A parallel methods spine.** Each module already names its identifying assumption (the extreme-value error, the nesting tree, the mixing distribution, the BLP exclusion restriction, the rational-expectations closure, the copula's distributional assumption). A future edition should add a short companion per week that states the estimator, its moment or likelihood, and what breaks identification—turning the reading map into a methods course, of which the worked BLP section below is the model. 4. **An applications track** pairing each estimator with a canonical industry case (autos, cereal, durables, banking, retail assortments), so students see the same machinery re-identified across settings with different sources of exogenous variation. The following section supplies the worked treatment the map points to. ### BLP demand as an estimator {#sec-cdem-worked} The Berry (1994) / Berry–Levinsohn–Pakes (1995) model is the hinge of the whole seminar: it carries the random-utility logic from individual choice to *aggregate* demand and confronts price endogeneity head-on. The treatment here states the estimator, its inversion, and the single assumption on which identification rests. Consumer $i$ in market $t$ derives utility from product $j$, $$ u_{ijt} = \mathbf{x}_{jt}^{\top}\boldsymbol{\beta}_i - \alpha_i\, p_{jt} + \xi_{jt} + \varepsilon_{ijt}, $$ {#eq-cdem-utility} where $\mathbf{x}_{jt}$ are observed characteristics, $p_{jt}$ is price, $\xi_{jt}$ is the **unobserved** (to the analyst) product quality in market $t$, and $\varepsilon_{ijt}$ is an i.i.d. extreme-value taste shock. The **random coefficients** $(\boldsymbol{\beta}_i, \alpha_i)$ vary across consumers with a parametric distribution, say $(\boldsymbol{\beta}_i,\alpha_i) = (\bar{\boldsymbol{\beta}},\bar\alpha) + \boldsymbol{\Sigma}\,\boldsymbol{\nu}_i$ with $\boldsymbol{\nu}_i \sim \Phi$. The outside good $j=0$ has utility normalized to $u_{i0t}=\varepsilon_{i0t}$. Integrating the logit choice probabilities over the taste distribution gives the model's predicted market share for product $j$, $$ s_{jt}(\boldsymbol{\delta}_t,\boldsymbol{\theta}_2) = \int \frac{\exp\!\big(\delta_{jt} + \mu_{ijt}(\boldsymbol{\theta}_2)\big)} {1 + \sum_{k} \exp\!\big(\delta_{kt} + \mu_{ikt}(\boldsymbol{\theta}_2)\big)} \, d\Phi(\boldsymbol{\nu}), $$ {#eq-cdem-share} where $\delta_{jt} = \mathbf{x}_{jt}^{\top}\bar{\boldsymbol{\beta}} - \bar\alpha\,p_{jt} + \xi_{jt}$ is the **mean utility** of product $j$ and $\mu_{ijt}$ collects the individual deviations governed by the heterogeneity parameters $\boldsymbol{\theta}_2=\mathrm{vec}(\boldsymbol{\Sigma})$. The integral has no closed form and is approximated by simulation over draws of $\boldsymbol{\nu}_i$. The key result of Berry (1994) is that, holding $\boldsymbol{\theta}_2$ fixed, the mapping from mean utilities to shares is **invertible**: there is a unique vector $\boldsymbol{\delta}_t$ that equates predicted shares to observed shares $\mathbf{S}_t$. BLP compute it as the fixed point of the **contraction mapping** $$ \delta_{jt}^{(r+1)} = \delta_{jt}^{(r)} + \ln S_{jt} - \ln s_{jt}\!\big(\boldsymbol{\delta}_t^{(r)},\boldsymbol{\theta}_2\big), $$ {#eq-cdem-contraction} iterated to convergence. This inversion turns an aggregate-demand problem into a linear-in-$\xi$ model: once $\boldsymbol{\delta}_t(\boldsymbol{\theta}_2)$ is recovered, the structural error is $$ \xi_{jt}(\boldsymbol{\theta}) = \delta_{jt}(\boldsymbol{\theta}_2) - \mathbf{x}_{jt}^{\top}\bar{\boldsymbol{\beta}} + \bar\alpha\, p_{jt}. $$ {#eq-cdem-xi} Estimation proceeds by **GMM**. Because price $p_{jt}$ is correlated with the unobserved quality $\xi_{jt}$ (firms set higher prices for products with higher $\xi$), ordinary regression of $\delta$ on price is biased. Identification instead requires a vector of **instruments** $\mathbf{z}_{jt}$ that satisfy the population moment condition $$ \mathbb{E}\!\left[\, \xi_{jt}(\boldsymbol{\theta}_0)\,\big|\, \mathbf{z}_{jt} \right] = 0 \quad\Longrightarrow\quad \mathbb{E}\!\left[\, \mathbf{z}_{jt}\, \xi_{jt}(\boldsymbol{\theta}_0)\,\right]=0, $$ {#eq-cdem-moment} and the estimator minimizes the empirical analogue, $$ \hat{\boldsymbol{\theta}} = \arg\min_{\boldsymbol{\theta}}\; \Big(\tfrac{1}{N}\textstyle\sum_{jt}\mathbf{z}_{jt}\,\xi_{jt}(\boldsymbol{\theta})\Big)^{\!\top} \mathbf{W} \Big(\tfrac{1}{N}\textstyle\sum_{jt}\mathbf{z}_{jt}\,\xi_{jt}(\boldsymbol{\theta})\Big), $$ {#eq-cdem-gmm} with weight matrix $\mathbf{W}$. The standard instruments are **cost shifters** and functions of **rival products' characteristics** (the BLP instruments), which shift a product's price and markup through competition without entering its own demand. The entire causal content rests on one **identifying assumption**: instrument exogeneity, $\mathbb{E}[\xi_{jt}\mid \mathbf{z}_{jt}]=0$. Substantively, this requires that product characteristics and cost shifters be uncorrelated with the *unobserved* quality $\xi_{jt}$. That is precisely the assumption the Week 7 debate contests—firms that observe $\xi$ when they design and price products may choose characteristics correlated with it, in which case the instruments fail and the recovered price elasticity, along with every markup, merger, and welfare number built on it, is biased. As with the seminar's other estimators, the BLP demand parameter becomes a *finding* only once its moment condition and its exclusion restriction are stated and defended in full. ## Key Takeaways - The seminar's spine is a single arc—**random utility → individual choice → aggregate demand → welfare**—and a student's job is to see pricing, assortment, merger, and new-product questions as instances of one estimation problem with one recurring threat, endogenous prices correlated with unobserved quality. - The microfoundation runs from Thurstone (1927), Luce (1959), and McFadden (1974, 1980) through the scanner-choice application of Guadagni & Little (1983); the **IIA** limitation of the simple logit is repaired first by GEV/nested logit and then dissolved by the **mixed logit**, which McFadden & Train (2000) show can approximate any random-utility model. - **Heterogeneity** is recovered either by hierarchical Bayes (Allenby & Rossi 1998) or by the mixing distribution of mixed logit; the two are asymptotically equivalent, so the choice between Bayesian MCMC and simulated likelihood is largely computational. - The **Berry inversion** and the **BLP** random-coefficients model (Berry 1994; Berry, Levinsohn & Pakes 1995) carry the RUM to market-level data; as @sec-cdem-worked shows, the contraction (@eq-cdem-contraction) linearizes the model in $\xi$ and the GMM moment (@eq-cdem-moment) identifies demand *only* under instrument exogeneity. - Relaxing the static, full-information benchmark—**stockpiling and durables** (Erdem & Keane 1996; Hendel & Nevo 2006; Gowrisankaran & Rysman 2012) and **search/consideration** (Roberts & Lattin 1991; Honka et al. 2017)—buys realism at the cost of additional unobservables (expectations, latent consideration) that must themselves be identified. - **Endogeneity corrections** (Villas-Boas & Winer 1999; Petrin & Train 2010; Park & Gupta 2012) and the **machine-learning frontier** (Bajari et al. 2015; Athey & Imbens 2016; Gabel & Timoshenko 2022) extend the toolkit, but the demand system's ultimate purpose remains the **welfare and counterfactual** computation that Petrin (2002) and Trajtenberg (1989) exemplify—numbers that inherit every maintained assumption of the model that produced them.