flowchart TD
A[Economic theory:<br/>preferences, costs,<br/>information, game] --> B[Agent optimization<br/>+ equilibrium concept]
B --> C["Decision rule<br/>m(s, p; θ)"]
C --> D[Estimation:<br/>MLE / GMM<br/>recover θ]
D --> E[Re-solve under<br/>new policy p′]
E --> F[Counterfactual /<br/>policy simulation]
G[(Observed data)] -.->|reduced form| H[Environment-specific<br/>association]
G --> D
38 Structural Models
A structural model is an empirical model whose parameters are the primitives of an economic theory: the preferences of consumers, the technologies and costs of firms, the information agents hold, and the rules of the game they play. Rather than fit a flexible function from covariates to an outcome, the analyst writes down how agents decide—typically as the solution to an optimization problem under an equilibrium concept—and then estimates the deep parameters that rationalize the observed data. The payoff is that those parameters are, by construction, invariant to the policies the analyst wants to study, so the model can answer questions about worlds that were never observed: a price never charged, a product never launched, a compensation plan never offered.
This is a different ambition from prediction or even from causal inference. A reduced-form analysis asks what is the effect of X on Y? and answers it, ideally, with a research design that isolates exogenous variation. A structural analysis asks how does the system work?—and, having recovered the mechanism, can simulate the effect of any intervention the mechanism admits, including ones for which no experiment or natural experiment exists. The cost of that generality is steep: structural models impose strong assumptions (functional forms, equilibrium, the information environment), and when those assumptions are wrong the conclusions are wrong in ways that are hard to detect. Much of the craft lies in choosing assumptions that buy identification without buying bias, and in being honest about which results lean on which assumptions.
This chapter develops the structural approach for a reader who already understands regression and causal inference and wants to see what changes when economic theory is moved from the discussion section into the likelihood. We proceed from intuition to formalism: first the anatomy of a structural model and its relationship to causal inference (Section 38.1); then a worked supply-and-demand example that exposes the endogeneity problem at the heart of the field (Section 38.2); then the two workhorse estimators—maximum likelihood and the generalized method of moments—and the discrete-choice and dynamic models that dominate marketing applications (Section 38.3 through Section 38.5); and finally the identification threats that decide whether any of it can be believed (Section 38.6). Throughout, the running theme is the one stated above: structural parameters are valuable precisely to the extent that they are policy invariant, and they are policy invariant only to the extent that the model is correct.
38.1 What Makes a Model “Structural”
The defining feature of a structural model is that its parameters have an interpretation outside the data-generating environment in which they were estimated. Chintagunta et al. (2006), in a field review, characterize structural modeling as the practice of using economic theory to discipline the specification of an empirical model so that the estimated coefficients correspond to taste, technology, or strategic primitives rather than to environment-specific correlations. The distinction matters most when the goal is counterfactual: a demand elasticity estimated as a structural taste parameter can be carried into a pricing simulation, whereas a regression slope estimated in one price regime carries no guarantee of stability when the regime changes.
This is the longest-running methodological argument in empirical marketing, and a doctoral student will be asked to take a side in a seminar before understanding that the sides are not symmetric. They are answering different questions.
2006 — the field defines itself, and immediately checks itself. The same volume of Marketing Science that carries Chintagunta et al. (2006)’s review carries Mazzeo et al. (2006) under the title Marketing Structural Models: “Keep It Real”—an insider’s warning that structural assumptions can be undisciplined, that a model’s counterfactual authority is only as good as the primitives it claims to recover, and that elaborate machinery can dress up an untestable story.
2010 — the other camp’s manifesto. Angrist and Pischke (2010) describe the credibility revolution in economics as taking “the con out of econometrics” through better research design. The implication for structural work is pointed: if identification comes from functional form rather than from variation you can describe in a sentence, the reader is being asked for trust rather than evidence.
2014 — and the design camp is checked in turn. Rossi (2014) publishes Even the Rich Can Make Themselves Poor: A Critical Examination of IV Methods in Marketing Applications, arguing that instrumental-variables practice in marketing frequently rests on instruments that are weak or invalid, and that the appearance of design-based rigor can be worse than an honest structural assumption because it is harder to interrogate.
Why it does not resolve. Each camp’s strongest critique is aimed at the other’s weakest practice, and both critiques land. The useful reframing is that the two approaches answer different questions and the question should pick the method. If you need the effect of a policy that was actually run, design-based identification is the cleaner instrument. If you need the effect of a policy nobody has run yet—a price nobody charged, an assortment nobody stocked—no amount of quasi-experimental variation will produce it, and a structural model is the only route (Section 38.1.2). The failure mode is choosing the method first and retrofitting the question, which both literatures contain examples of.
It is useful to state the contrast formally. Suppose an outcome \(y\) is generated by agents who choose actions to maximize an objective indexed by a structural parameter vector \(\boldsymbol{\theta}\) (e.g., utility weights, marginal costs) given a state \(\mathbf{s}\) and a policy \(\mathbf{p}\) (prices, assortment, incentives): \[ y = m(\mathbf{s}, \mathbf{p}; \boldsymbol{\theta}) + \varepsilon , \tag{38.1}\] where \(m(\cdot)\) is the decision rule—the mapping from states and policy to behavior implied by the agent’s optimization and the market equilibrium. A reduced-form analysis estimates a projection of \(y\) on \((\mathbf{s},\mathbf{p})\) that is valid only on the support and policy regime observed. A structural analysis estimates \(\boldsymbol{\theta}\), then re-solves Equation 38.1 under a new policy \(\mathbf{p}'\) to predict \(y' = m(\mathbf{s}, \mathbf{p}'; \boldsymbol{\theta})\). The structural object is reusable; the reduced-form object, strictly, is not.
Causal inference estimates the effect of a treatment that occurred; structural modeling estimates the mechanism that would produce the effect of a treatment that did not. The first is a statement about a realized intervention, the second a statement about a class of interventions the model can simulate.
The two traditions are complements rather than rivals, and the modern literature treats them as such. Credible causal estimates discipline structural models—a quasi-experimental demand elasticity is a moment a structural demand system should reproduce—and structural reasoning, in turn, tells the causal analyst where the endogeneity is and what would constitute a valid instrument. Conlon and Mortimer (2013) illustrate the round trip: they build a structural demand model to handle endogenous product availability (stockouts that are correlated with demand shocks) and then use the structural estimates to run counterfactuals that no reduced-form design could deliver. The remainder of this section makes the three pillars of any structural exercise—behavioral assumptions, equilibrium, and counterfactuals—precise before we turn to estimation.
Figure 38.1 summarizes the workflow that distinguishes the structural pipeline from a reduced-form one.
38.1.1 Behavioral and Equilibrium Assumptions
Structural models earn their counterfactual power by committing to assumptions of four kinds, and the credibility of any application rests on how defensible those commitments are.
The first is functional form: the shape of utility, demand, or cost. Is demand linear in price, or constant-elasticity? Is utility linear in attributes, or does it exhibit diminishing marginal utility? These choices are not innocuous, because the counterfactual extrapolates along the assumed curvature.
The second is the decision protocol: the rule by which agents map information to actions. The canonical assumption is utility maximization—each consumer chooses the alternative delivering the highest utility—but the model must also say whether agents are myopic or forward looking, risk neutral or risk averse, and whether they learn.
The third is the information environment. Do consumers know all product attributes when choosing, or do they face uncertainty and update beliefs? Erdem, Keane, and Sun (2008), for instance, build a model in which advertising influences price sensitivity precisely because consumers are uncertain about quality in experience- good markets and advertising conveys information; the substantive conclusion is inseparable from the informational assumption.
The fourth is the equilibrium concept. In a market model the analyst must say how agents’ choices are mutually consistent: that the market clears, or that firms play a Nash equilibrium in prices, or that a dynamic game is in Markov-perfect equilibrium. Kadiyali, Sudhir, and Rao (2001) review how equilibrium assumptions let the analyst infer unobserved competitive conduct—whether firms behave more collusively or more competitively than Bertrand—from the comovement of prices and quantities, an inference impossible without the equilibrium restriction.
These assumptions are what make the model tractable and estimable, and they are simultaneously the model’s principal vulnerability: a misspecified information environment or an incorrect conduct assumption biases the deep parameters and, with them, every counterfactual. The discipline of the field is to state each assumption explicitly, to test those that are testable, and to flag the counterfactuals that hinge on the rest.
38.1.2 Counterfactuals: the Reason to Pay the Price
The justification for the structural apparatus is the counterfactual. Once \(\boldsymbol{\theta}\) is in hand, the analyst re-solves the model under a policy that was never implemented and reads off the predicted outcome—what would sales have been at a price we never charged? or what compensation plan maximizes profit? Gowrisankaran and Rysman (2020) provide a framework for exactly this kind of policy evaluation, and the counterfactual logic is what separates a structural paper from a descriptive one. The validity of a counterfactual is conditional on two things: the parameters are policy invariant (the Lucas-critique requirement), and the new policy lies within the class the model can represent. A model estimated under linear pricing cannot credibly simulate nonlinear tariffs; a static model cannot simulate a policy whose whole point is to change dynamic incentives. Stating the admissible counterfactual class is therefore part of specifying the model, not an afterthought.
38.2 A First Structural Model: Demand and Supply
The simplest model that exhibits the structural way of thinking is linear simultaneous supply and demand. It is worth working through in full because it already contains the central econometric problem of the entire field—the endogeneity of price—and the standard solution.
Let quantity demanded and supplied at price \(p\) be \[ Q^{d} = \alpha_0 - \alpha_1 p + u^{d}, \qquad Q^{s} = \beta_0 + \beta_1 p + u^{s}, \tag{38.2}\] with \(\alpha_1, \beta_1 > 0\) and structural shocks \(u^{d}, u^{s}\) (a demand shifter such as a taste shock; a supply shifter such as an input-cost shock). The structural parameters are the slopes \(\alpha_1\) (how strongly quantity responds to price along the demand curve) and \(\beta_1\) (the analogous supply response); these are the policy-invariant objects a manager would want for pricing. Market clearing, \(Q^{d} = Q^{s}\), is the equilibrium condition, and it determines the observed price \[ p^{\star} = \frac{\alpha_0 - \beta_0 + u^{d} - u^{s}}{\alpha_1 + \beta_1}. \tag{38.3}\]
Equation 38.3 is the crux. Because equilibrium price depends on the demand shock \(u^{d}\), price is endogenous: \(\mathrm{Cov}(p^{\star}, u^{d}) \neq 0\). A naive regression of observed quantity on observed price does not recover \(\alpha_1\); it recovers a hybrid of the demand and supply slopes—the textbook simultaneity bias. Intuitively, the observed price–quantity pairs are intersections of shifting demand and supply curves, so the scatter traces out neither curve. This single fact—that prices are set by agents who observe demand conditions the econometrician does not—motivates the instrumental-variable and control-function machinery that pervades structural demand estimation.
Identification comes from exclusion restrictions: a variable that shifts supply but not demand traces out the demand curve, and vice versa. A cost shifter \(z\) entering supply but excluded from demand instruments for price in the demand equation. Montgomery and Rossi (1999) exploit precisely this logic to estimate price elasticities at the store and household level, recovering structural demand slopes that a single-equation regression would contaminate. The simulation below makes the bias concrete and shows that an instrument recovers the truth.
Code
set.seed(29)
n <- 5000
a0 <- 10; a1 <- 1.5 # demand intercept, slope (truth)
b0 <- 2; b1 <- 1.0 # supply intercept, slope (truth)
z <- runif(n, 0, 4) # cost shifter: enters supply, excluded from demand
ud <- rnorm(n) # demand shock
us <- rnorm(n) # supply shock
# Equilibrium price and quantity (solve Qd = Qs with a cost shifter in supply)
# Qd = a0 - a1 p + ud ; Qs = b0 + b1 p + g*z + us
g <- 1.2
p <- (a0 - b0 - g * z + ud - us) / (a1 + b1)
q <- a0 - a1 * p + ud
# (1) Naive OLS of quantity on price: biased for the demand slope -a1
ols <- coef(lm(q ~ p))["p"]
# (2) Two-stage least squares using the cost shifter z as an instrument
phat <- fitted(lm(p ~ z))
iv <- coef(lm(q ~ phat))["phat"]
cat("True demand slope (-a1): ", -a1, "\n")
#> True demand slope (-a1): -1.5
cat("OLS estimate (biased): ", round(ols, 3), "\n")
#> OLS estimate (biased): -0.892
cat("IV / 2SLS estimate (consistent):", round(iv, 3), "\n")
#> IV / 2SLS estimate (consistent): -1.513The OLS slope is attenuated and unstable; the instrumented slope recovers \(-\alpha_1\). This linear example scales, conceptually, all the way to modern differentiated- products demand systems, where the same endogeneity reappears (price correlates with unobserved product quality) and is handled with the same idea—cost or competitor- based instruments—inside a far richer choice model, to which we now turn.
38.3 Estimation: Maximum Likelihood and GMM
Structural parameters are rarely obtained by ordinary least squares. The two dominant estimators are maximum likelihood (ML) and the generalized method of moments (GMM); the standard graduate references develop both in depth (Greene 2003; Cameron and Trivedi 2005; Hensher, Rose, and Greene 2005), and the Handbook of Econometrics chapters on the econometric evaluation of policy collect the identification theory (Heckman and Vytlacil 2007a, 2007b; Abbring and Heckman 2007). The choice between ML and GMM is chiefly a choice about how much of the data-generating process one is willing to specify.
Maximum likelihood specifies the full conditional distribution of the data. If the model implies a density \(f(y_i \mid \mathbf{x}_i; \boldsymbol{\theta})\), the estimator maximizes the log-likelihood, \[ \hat{\boldsymbol{\theta}}_{\mathrm{ML}} = \arg\max_{\boldsymbol{\theta}} \sum_{i=1}^{n} \log f(y_i \mid \mathbf{x}_i; \boldsymbol{\theta}), \tag{38.4}\] which is efficient when the distributional assumption is correct but inconsistent when it is not. ML is the natural estimator for discrete-choice models, where the choice probabilities are the likelihood contributions.
The generalized method of moments specifies only a set of population moment conditions implied by the theory—typically that the structural error is orthogonal to a vector of instruments \(\mathbf{z}_i\), \[ \mathbb{E}\!\left[\, \mathbf{z}_i \, g(y_i, \mathbf{x}_i; \boldsymbol{\theta}) \,\right] = \mathbf{0}, \tag{38.5}\] where \(g(\cdot)\) is the structural residual. GMM minimizes the empirical analog of these moments in a quadratic form, \[ \hat{\boldsymbol{\theta}}_{\mathrm{GMM}} = \arg\min_{\boldsymbol{\theta}} \bar{g}_n(\boldsymbol{\theta})^{\top}\, \mathbf{W}\, \bar{g}_n(\boldsymbol{\theta}), \qquad \bar{g}_n(\boldsymbol{\theta}) = \frac{1}{n}\sum_{i=1}^{n} \mathbf{z}_i\, g(y_i, \mathbf{x}_i; \boldsymbol{\theta}), \tag{38.6}\] with weighting matrix \(\mathbf{W}\); the efficient choice sets \(\mathbf{W}\) to the inverse of the moment covariance. GMM is the estimator of choice when the analyst trusts a few orthogonality conditions (a cost instrument, an equilibrium first-order condition) but does not wish to commit to a full likelihood—exactly the situation in differentiated-products demand, where the famous contraction-mapping estimator inverts market shares to demand shocks and stacks them against instruments in Equation 38.6. Table 38.1 contrasts the two.
| Dimension | Maximum likelihood | Generalized method of moments |
|---|---|---|
| Specifies | Full conditional density of the data | A set of moment/orthogonality conditions |
| Efficiency | Efficient if the distribution is correct | Efficient given the chosen moments; generally less than full ML |
| Robustness | Inconsistent under distributional misspecification | Consistent under far weaker assumptions |
| Endogeneity | Handled via control functions or joint likelihood | Handled natively through instrument moments |
| Typical use | Discrete choice, dynamic discrete choice | Demand systems, Euler-equation and conduct models |
| Failure mode | Wrong likelihood \(\Rightarrow\) biased everything | Weak or invalid instruments \(\Rightarrow\) weak identification |
38.4 Discrete Choice as the Workhorse
A large share of structural marketing models are built on random-utility discrete choice, because most consumer decisions—which brand, which channel, whether to buy at all—are discrete, and because the framework yields choice probabilities that are both behaviorally founded and computationally convenient. The reviews by Dubé et al. (2002) and Heiss (2002) lay out the family and its estimation; the model below is the canonical entry point.
Consumer \(i\) choosing among alternatives \(j = 1,\dots,J\) derives utility \[ U_{ij} = V_{ij} + \varepsilon_{ij} = \mathbf{x}_{ij}^{\top}\boldsymbol{\beta} - \alpha p_{ij} + \xi_j + \varepsilon_{ij}, \tag{38.7}\] where \(\mathbf{x}_{ij}\) are observed attributes, \(p_{ij}\) is price, \(\xi_j\) is an unobserved (to the analyst) product-quality term, and \(\varepsilon_{ij}\) is an idiosyncratic taste shock. The consumer picks \(j\) if \(U_{ij} \ge U_{ik}\) for all \(k\). When \(\varepsilon_{ij}\) is i.i.d. type-I extreme value, the choice probabilities take the closed-form multinomial logit, \[ P_{ij} = \frac{\exp(V_{ij})}{\sum_{k=1}^{J}\exp(V_{ik})}, \tag{38.8}\] which is the likelihood contribution maximized in Equation 38.4. Two structural warnings travel with Equation 38.8. First, the logit imposes independence of irrelevant alternatives (IIA): the relative odds of two alternatives are unaffected by any third, an implausible substitution pattern that the nested logit of Heiss (2002) relaxes by grouping close substitutes into nests. Second, the quality term \(\xi_j\) is correlated with price (better products are priced higher), so \(\alpha\) is endogenous—the same simultaneity as in Section 38.2, now inside a choice model—and is corrected with cost or competitor-price instruments.
A practical route to estimation, due to Berry, inverts the share equation. With an outside option \(j=0\) whose mean utility is normalized to zero, Equation 38.8 implies the linear relation \[ \ln s_j - \ln s_0 = \mathbf{x}_{j}^{\top}\boldsymbol{\beta} - \alpha p_{j} + \xi_j, \tag{38.9}\] so the left-hand side—an observed function of market shares—equals mean utility, and the demand parameters can be estimated by regressing it on attributes and price. The example below builds one market of many products, computes the inverted shares of Equation 38.9, and contrasts a naive regression (which ignores the correlation between price and \(\xi_j\)) with a two-stage least-squares estimator that instruments price with a cost shifter.
Code
set.seed(29)
J <- 200; N <- 4e5 # products in one market; consumers sampled
x <- rnorm(J) # observed product attribute
xi <- rnorm(J) # unobserved quality -> source of endogeneity
cost <- runif(J, 0, 2) # cost shifter -> instrument for price
beta_true <- 1.0; alpha_true <- 1.5
# Price rises in unobserved quality (endogeneity) and in the cost shifter
price <- 1.0 + 0.8 * xi + cost + rnorm(J, 0, 0.3)
delta <- x * beta_true - alpha_true * price + xi # mean utility of products 1..J
# Logit shares with an outside good (utility 0); add sampling noise via multinomial
num <- exp(delta)
s <- num / (1 + sum(num)); s0 <- 1 / (1 + sum(num))
counts <- rmultinom(1, N, c(s0, s))[, 1]
share <- counts[-1] / N; share0 <- counts[1] / N
y <- log(share) - log(share0) # Berry inversion: y_j = delta_j
ols <- coef(lm(y ~ x + price)) # ignores xi: endogenous, biased
phat <- fitted(lm(price ~ x + cost)) # first stage: cost instruments price
iv <- coef(lm(y ~ x + phat)) # 2SLS price coefficient
cat("True price coef (-alpha):", -alpha_true, "\n")
#> True price coef (-alpha): -1.5
cat("Naive logit price coef: ", round(ols["price"], 3), "\n")
#> Naive logit price coef: -0.718
cat("IV logit price coef: ", round(iv["phat"], 3), "\n")
#> IV logit price coef: -1.433The naive coefficient is badly attenuated toward zero—the unobserved quality \(\xi_j\) raises both utility and price, masking the true price sensitivity—while the instrumented estimate recovers \(-\alpha\). This is the lesson that recurs across every structural demand paper: the price coefficient is the parameter most contaminated by the unobserved quality term, and a valid instrument is the remedy.
38.4.1 From Estimated Demand to Optimal Design: The Exact Share-of-Choice Problem
Everything so far has been about estimating the logit—recovering \(\boldsymbol\beta\) and \(\alpha\) from observed choices. The mirror-image question is prescriptive: given the estimated logit, which product should the firm build? This is the share-of-choice product design (SOCPD) problem, and it is where a demand estimate earns its keep, because the same partworths recovered from choice-based conjoint (Section 41.2) become the input to a design optimization. Encode a product as a binary attribute vector \(\mathbf{x}\in\{0,1\}^{n}\) (each feature present or absent, each level chosen or not). With \(K\) customer types of weight \(\lambda_k\) and partworths \(\boldsymbol\beta_k\), each type buys the focal product against an outside option of utility zero with the binary-logit probability of Equation 38.8, and the firm chooses the design that maximizes their weighted sum, \[ \max_{\mathbf{x}\in\{0,1\}^{n}}\; f(\mathbf{x}) \;=\; \sum_{k=1}^{K} \lambda_k\, \frac{\exp(\boldsymbol\beta_k^{\top}\mathbf{x})}{1+\exp(\boldsymbol\beta_k^{\top}\mathbf{x})}. \tag{38.10}\]
The objective looks benign and is anything but. Each summand is a logistic sigmoid, and a sum of sigmoids is not concave, so the continuous relaxation carries multiple local optima and a gradient climb or a greedy attribute-flip has no optimality guarantee; the binary \(\mathbf{x}\) then layers a combinatorial search over \(2^{n}\) designs on top. Akchen and Mišić (2026) show the difficulty is intrinsic rather than an artifact of weak algorithms: the logit SOCPD is NP-hard even with only two customer types, and NP-hard even to approximate within a nontrivial factor, so no polynomial-time method can promise a near-optimal design in general.
Their contribution is that, the hardness notwithstanding, the problem admits an exact reformulation. Exploiting the convexity latent in the log-sum-exp structure of the logit, they recast Equation 38.10 as a mixed-integer convex program—concretely a mixed-integer exponential-cone program, given in three formulations of increasing tightness—that modern conic solvers drive to provable optimality, or to a certified optimality gap, on instances far larger than enumeration could reach, and with higher realized share than the share-of-choice heuristics and branch-and-price schemes that preceded them. The lesson for the structural modeler is that the logit is not only an estimation object but an optimization primitive: the same \(\exp(V)/(1+\exp(V))\) that makes shares tractable to compute is what makes them convex-representable and hence exactly optimizable. The brute-force enumeration below makes the objective—and the wall it runs into—concrete on a ten-attribute instance; the exact conic method is what carries the identical objective to the dozens of attributes a real design space spans.
Code
set.seed(29)
n_attr <- 10; K <- 5 # 10 binary attributes -> 2^10 designs; 5 segments
lambda <- rep(1 / K, K) # segment weights
beta <- matrix(rnorm(K * n_attr, 0, 1.2), K, n_attr) # partworths by segment
designs <- as.matrix(expand.grid(rep(list(c(0, 1)), n_attr))) # every 2^10 design
share <- as.vector(plogis(designs %*% t(beta)) %*% lambda) # share of choice, eq. sm-socpd
best <- which.max(share)
cat(sprintf("Attributes: %d Designs enumerated: %d Optimal share-of-choice: %.3f\n",
n_attr, nrow(designs), share[best]))
#> Attributes: 10 Designs enumerated: 1024 Optimal share-of-choice: 0.778
cat("Optimal design (attribute on/off):", designs[best, ], "\n")
#> Optimal design (attribute on/off): 0 1 1 0 1 0 0 0 0 1
cat(sprintf("At 30 attributes, the same brute force would face %.3g candidate designs.\n",
2^30))
#> At 30 attributes, the same brute force would face 1.07e+09 candidate designs.Ten attributes enumerate in a blink; thirty would face over a billion designs, and the non-concavity means the shortcuts that tame smooth problems do not apply. That gap between a toy instance and a real one is exactly what Akchen and Mišić (2026) close—turning a provably hard design problem into one a solver can certify—and it is the reason the humble logit of Equation 38.8 reaches past demand estimation into optimal product design.
38.4.2 Behavior in One Primitive: Assortment Optimization Under Choice Overload
The logit of Equation 38.8, used prescriptively in Section 38.4.1, carries a behavioral commitment that is easy to miss. With an outside option of fixed utility, the no-purchase probability \(P_0(S) = 1/\big(1+\sum_{j\in S} e^{u_j}\big)\) falls every time a product is added to the offer set \(S\), so a larger assortment always converts more shoppers. Every random-utility model whose outside option does not depend on \(S\) shares this property, which is why none of them can produce choice overload: the inverted-U relation between assortment size and purchase that Iyengar and Lepper (2000) made famous, and that the meta-analyses of Scheibehenne, Greifeneder, and Todd (2010) and Chernev, Böckenholt, and Goodman (2014) find to be real only under identifiable conditions (see the controversy callout in Chapter 13). Mirzaee, Fata, and Li (2026) show how to put the effect into the model without giving up the logit’s optimization machinery; the numbers below come from the working-paper version (Mirzaee, Fata, and Li 2025).
Their model changes a single primitive. Product \(i\) keeps its logit attraction \(a_i = e^{u_i}\), but the utility of not buying becomes a function of how many products are shown and of a golden number \(B\): \[ P_i(S) = \frac{e^{u_i}}{e^{u_0(|S|,B)} + \sum_{j\in S} e^{u_j}}, \qquad u_0(|S|,B) = \zeta\,(B-|S|)^2 . \tag{38.11}\] The general assumption is only that \(u_0\) is weakly decreasing in \(|S|\) below \(B\) (choice underload: too few options push shoppers to walk away), weakly increasing above it (choice overload), zero at \(|S| = B\) and positive elsewhere; the quadratic in Equation 38.11 with \(\zeta > 0\) is the authors’ parametric workhorse, and a piecewise-linear \(\zeta\,|B-|S||\) is their alternative. A U-shaped no-purchase utility produces a U-shaped no-purchase probability and hence an inverted-U in purchase. Within any given offer set, choice is still multinomial logit, IIA among the offered products included; only the comparison with not buying moves as the set grows or shrinks. That restraint is the reusable design lesson: isolate the behavioral effect in one primitive, keep everything else classical, and the classical results can be recovered by conditioning on the variable the primitive depends on.
The conditioning is needed because the effect breaks the best-known result in assortment optimization. Under the plain logit the revenue-maximizing assortment is revenue-ordered—the \(k\) highest-margin products for some \(k\)—so only \(n\) candidate sets need checking (Talluri and Ryzin 2004). Under choice overload this fails, as the first example below shows. What survives is a decomposition: fix \(|S| = k\), and \(u_0(k,B)\) is a constant, so each subproblem is an MNL assortment problem with an exact cardinality constraint. The authors adapt the geometric algorithm that Rusmevichientong, Shen, and Shmoys (2010) developed for the capacitated logit—lines \(h_i(\lambda) = a_i(r_i-\lambda)\) in the margins \(r_i\), whose \(O(n^2)\) intersection points split the real line into intervals within which the top-\(k\) lines do not change—with one twist: because the cardinality is imposed rather than bounded, the subproblem must take exactly the top \(k\) lines even when some are negative. That gives \(O(n^2)\) candidate sets per \(k\) and a polynomial algorithm overall. The structural results anchor the new optimum on the familiar one. Profit under choice overload never exceeds logit profit, with equality only when the optimal size, the golden number, and the size of the logit optimum \(S_{\text{MNL}}\) coincide. If \(|S_{\text{MNL}}| = B\), the logit optimum is also the overload optimum; if \(|S_{\text{MNL}}| < B\), the overload optimum contains \(S_{\text{MNL}}\) and adds at most \(B - |S_{\text{MNL}}|\) products; if \(|S_{\text{MNL}}| > B\), the overload optimum still has at least \(B\) products. When the objective is conversion rather than revenue (equal margins), the optimum is the \(k\) most attractive products for some \(k \ge B\).
The same fix-the-size trick carries the rest of the classical toolkit across. With prices chosen jointly, each fixed-\(k\) subproblem becomes a capacity-constrained joint assortment-and-pricing problem whose solution uses the full capacity, and the optimum is the fixed point of a decreasing function. Under a nested logit with a golden number for each nest, the problem becomes a linear program of the Gallego and Topaloglu (2014) type with \(m+1\) variables and \(O(mn^3)\) constraints for \(m\) nests; under a mixture of logits with segment-specific golden numbers, a fully polynomial-time approximation scheme exists when the number of segments is fixed.
Estimation follows the same logic. With transactions \(t\), offer sets \(S^t\), and choices \(c_t\) (where \(u_{c_t} = u_0(|S^t|,B)\) if nothing is bought), the log-likelihood is \[ \ell(\mathbf{u},\zeta; B) = \sum_t \Big[\,u_{c_t} - \log\Big(e^{u_0(|S^t|,B)} + \sum_{i\in S^t} e^{u_i}\Big)\Big]. \tag{38.12}\] Because \(u_0\) in Equation 38.11 is linear in \(\zeta\) once \(B\) is fixed, Equation 38.12 is the logit’s familiar log-sum-exp of linear indices and is concave in \((\mathbf{u},\zeta)\), so estimation is a concave program nested inside a one-dimensional search over the integer golden number. A nonparametric version that frees \(u_0(k,B)\) for every observed size, subject to the monotonicity in the assumption, is likewise convex for fixed \(B\), and an estimated \(\zeta \le 0\) signals that the U-shape is the wrong functional form for the data. Since \(u_0\) depends on the data only through \(|S|\), \(B\) and \(\zeta\) are identified only if transactions occur at several distinct set sizes; with a single size, \(u_0\) is a constant indistinguishable from a common shift in the \(u_i\). In the working paper’s numerical study (synthetic transactions from a four-segment mixture of logits with 10 or 20 products and two to four distinct offer-set sizes), the overload model improves the average absolute log-likelihood over plain MNL in every scenario, with gains that rise with the strength of the overload effect and shrink as the number of distinct offer sets and of products grows.
The chunk below is our toy re-implementation, not the authors’ code. Part (a) enumerates every subset in the paper’s two small examples; part (b) simulates choices from Equation 38.11 and fits plain MNL and the overload model by maximum likelihood, profiling over \(B\).
Code
# Our toy re-implementation of Mirzaee, Fata & Li's examples and estimator (not the authors' code).
# ---- (a) Exact enumeration: MNL vs. paradox-of-choice (PoC) optimal assortments ----
# Expected revenue of offer set S: a_i = exp(u_i) are attractions, r_i margins, and
# a0 the no-purchase attraction -- fixed at 1 under MNL, a function of |S| under PoC.
rev_S <- function(S, a, r, a0) sum(a[S] * r[S]) / (a0 + sum(a[S]))
enumerate <- function(a, r, a0_poc) {
sets <- unlist(lapply(seq_along(a), function(k) combn(length(a), k, simplify = FALSE)),
recursive = FALSE)
data.frame(set = vapply(sets, function(S) paste0("{", paste(S, collapse = ","), "}"), ""),
MNL = vapply(sets, function(S) rev_S(S, a, r, 1), 0),
PoC = vapply(sets, function(S) rev_S(S, a, r, a0_poc(length(S))), 0))
}
# Example 2: a = (1,2,4), r = (9,8,7), golden number B = 1, a0 = exp(0.3 (B - |S|)^2)
ex2 <- enumerate(c(1, 2, 4), c(9, 8, 7), function(k) exp(0.3 * (1 - k)^2))
# Example 4: a = (1, 1.01, 5), r = (1.001, 0.99, 0.6), B = 1, a0(|S|) = (1, 2.5, 2.9);
# the last column re-runs it with a0(3) = 2.99 to show how knife-edge the example is
ex4 <- enumerate(c(1, 1.01, 5), c(1.001, 0.99, 0.6), function(k) c(1, 2.5, 2.9)[k])
ex4$PoC_299 <- enumerate(c(1, 1.01, 5), c(1.001, 0.99, 0.6),
function(k) c(1, 2.5, 2.99)[k])$PoC
knitr::kable(cbind(ex2, ex4[, -1]), digits = 4,
col.names = c("Set", "Ex. 2 MNL", "Ex. 2 PoC", "Ex. 4 MNL",
"Ex. 4 PoC", "Ex. 4 PoC, a0(3) = 2.99"))| Set | Ex. 2 MNL | Ex. 2 PoC | Ex. 4 MNL | Ex. 4 PoC | Ex. 4 PoC, a0(3) = 2.99 |
|---|---|---|---|---|---|
| {1} | 4.5000 | 4.5000 | 0.5005 | 0.5005 | 0.5005 |
| {2} | 5.3333 | 5.3333 | 0.4975 | 0.4975 | 0.4975 |
| {3} | 5.6000 | 5.6000 | 0.5000 | 0.5000 | 0.5000 |
| {1,2} | 6.2500 | 5.7473 | 0.6648 | 0.4437 | 0.4437 |
| {1,3} | 6.1667 | 5.8269 | 0.5716 | 0.4707 | 0.4707 |
| {2,3} | 6.2857 | 5.9865 | 0.5706 | 0.4700 | 0.4700 |
| {1,2,3} | 6.6250 | 5.1356 | 0.6243 | 0.5046 | 0.5001 |
Code
best <- function(d, col) d$set[which.max(d[[col]])]
cat(sprintf("Ex. 2 optimum: MNL %s, PoC %s; PoC optimum earns %+.1f%% over the MNL optimum under PoC demand\n",
best(ex2, "MNL"), best(ex2, "PoC"),
100 * (max(ex2$PoC) / ex2$PoC[ex2$set == best(ex2, "MNL")] - 1)))
#> Ex. 2 optimum: MNL {1,2,3}, PoC {2,3}; PoC optimum earns +16.6% over the MNL optimum under PoC demand
cat(sprintf("Ex. 4 optimum: MNL %s, PoC %s (with a0(3) = 2.99: %s)\n",
best(ex4, "MNL"), best(ex4, "PoC"), best(ex4, "PoC_299")))
#> Ex. 4 optimum: MNL {1,2}, PoC {1,2,3} (with a0(3) = 2.99: {1})
# ---- (b) Estimation: plain MNL vs. MNL with quadratic PoC outside option ----------
set.seed(4345)
n <- 10; B_true <- 5; zeta_true <- 0.1
u_true <- runif(n) # product utilities ~ U(0,1)
sizes <- rep(c(3, 5, 7, 10), each = 1000) # four offer-set sizes, 4,000 choices
Tn <- length(sizes)
avail <- t(vapply(sizes, function(k) { z <- numeric(n); z[sample.int(n, k)] <- 1; z },
numeric(n))) # random offer set of each size
num <- avail * matrix(exp(u_true), Tn, n, byrow = TRUE)
u0_t <- zeta_true * (B_true - sizes)^2 # u0(|S|, B) = zeta (B - |S|)^2
prob <- cbind(exp(u0_t), num) / (exp(u0_t) + rowSums(num))
choice <- apply(prob, 1, function(p) sample.int(n + 1, 1, prob = p)) - 1 # 0 = no purchase
Y <- outer(choice, seq_len(n), "==") * 1
# Log-likelihood in theta = (u_1..u_n, zeta) for a given B, through q = (B - |S|)^2.
# u0 is linear in zeta, so for fixed B the log-likelihood is concave; plain MNL is zeta = 0.
negll <- function(theta, q) {
u <- theta[1:n]; z <- if (length(theta) > n) theta[n + 1] else 0
D <- exp(z * q) + drop(avail %*% exp(u))
-(sum(Y %*% u) + sum((choice == 0) * z * q) - sum(log(D)))
}
neggr <- function(theta, q) {
u <- theta[1:n]; z <- if (length(theta) > n) theta[n + 1] else 0
D <- exp(z * q) + drop(avail %*% exp(u))
g <- colSums(Y) - exp(u) * colSums(avail / D)
if (length(theta) > n) g <- c(g, sum((choice == 0) * q) - sum(q * exp(z * q) / D))
-g
}
fit_mnl <- optim(rep(0, n), negll, neggr, q = 0, method = "BFGS")
fits <- lapply(1:n, function(B) # profile the likelihood over integer B
optim(c(rep(0, n), 0), negll, neggr, q = (B - sizes)^2, method = "BFGS",
control = list(maxit = 500)))
ll_B <- -vapply(fits, `[[`, 0, "value")
B_hat <- which.max(ll_B); zeta_hat <- fits[[B_hat]]$par[n + 1]
cat(sprintf("LL plain MNL %.1f | LL PoC-MNL %.1f at B_hat = %d, zeta_hat = %.3f (truth: B = %d, zeta = %.2f)\n",
-fit_mnl$value, ll_B[B_hat], B_hat, zeta_hat, B_true, zeta_true))
#> LL plain MNL -7573.8 | LL PoC-MNL -7255.4 at B_hat = 5, zeta_hat = 0.100 (truth: B = 5, zeta = 0.10)
cat("Profile log-likelihood over B = 1..10:", sprintf("%.1f", ll_B), "\n")
#> Profile log-likelihood over B = 1..10: -7296.2 -7289.1 -7279.6 -7267.0 -7255.4 -7315.8 -7569.7 -7527.0 -7465.9 -7430.0
# Purchase probability by offer-set size: data vs. the two fitted models
pr_fit <- function(u, z, q) { s <- drop(avail %*% exp(u)); tapply(s / (exp(z * q) + s), sizes, mean) }
knitr::kable(data.frame(size = c(3, 5, 7, 10),
observed = as.vector(tapply(choice > 0, sizes, mean)),
mnl = as.vector(pr_fit(fit_mnl$par, 0, 0)),
poc = as.vector(pr_fit(fits[[B_hat]]$par[1:n], zeta_hat,
(B_hat - sizes)^2))),
digits = 3, col.names = c("Offer-set size", "Observed purchase rate",
"Plain MNL fit", "PoC-MNL fit"))| Offer-set size | Observed purchase rate | Plain MNL fit | PoC-MNL fit |
|---|---|---|---|
| 3 | 0.795 | 0.706 | 0.803 |
| 5 | 0.922 | 0.802 | 0.911 |
| 7 | 0.901 | 0.850 | 0.906 |
| 10 | 0.630 | 0.890 | 0.628 |
The first example has three products with attractions \((1, 2, 4)\), margins \((9, 8, 7)\), \(B = 1\), and \(e^{u_0} = \exp\{0.3(B-|S|)^2\}\). The logit offers everything (revenue 6.625), but under choice overload the optimum is \(\{2,3\}\), which drops the highest-margin product and so is not revenue-ordered. Under overload demand \(\{2,3\}\) earns 5.9865 against 5.1356 for the logit’s choice, a 16.6% gain, which reproduces the 16% gap the working paper reports. (The paper’s Table 1 prints 5.135 and 5.980 for these two cells; the first is truncation, the second appears to be a small typo that does not touch the conclusion.) The second example runs the other way. With attractions \((1, 1.01, 5)\), margins \((1.001, 0.99, 0.6)\), \(B = 1\), and no-purchase attractions 1, 2.5, and 2.9 for sets of one, two, and three products, the logit picks \(\{1,2\}\) (0.6648) while the overload optimum is the full set \(\{1,2,3\}\) (0.5046): the very attractive low-margin product 3 wins back enough shoppers from the outside option to outweigh the overload it adds. Choice overload can make the optimal assortment larger, and here the overload optimum contains the logit optimum. The enumeration matches every entry of the paper’s Table 2 up to truncation in the fourth decimal. One caution: the working paper’s text gives the three-product no-purchase attraction as 2.99, not 2.9, and that value does not reproduce its own table. At 2.99 the full set earns 0.5001 and the singleton \(\{1\}\) (0.5005) becomes optimal, as the last column shows. Every Table 2 entry reproduces with 2.9, so 2.99 appears to be a typo, and the example is knife-edge either way: recomputing a paper’s worked examples is a cheap check worth making a habit.
In the estimation exercise, 4,000 choices are drawn from ten products with utilities from \(U(0,1)\), offered in random sets of 3, 5, 7, and 10 products (the set sizes \(\lceil n/4\rceil\), \(\lceil n/2\rceil\), \(\lceil n/1.5\rceil\), and \(n\) of the paper’s design), with \(B = 5\) and \(\zeta = 0.1\). The profile likelihood peaks at \(\hat B = 5\) with \(\hat\zeta = 0.100\), and the overload model’s log-likelihood (\(-7255.4\)) beats plain MNL’s (\(-7573.8\)) by 318.4 points, about 4.2% of its absolute value. The table shows where the logit fails: it can only let purchase rise with set size (0.706, 0.802, 0.850, 0.890), while the data rise and then fall (0.795, 0.922, 0.901, 0.630), which the overload fit tracks closely (0.803, 0.911, 0.906, 0.628). The profile over \(B\) is also not unimodal—\(B = 6\) fits worse (\(-7315.8\)) than \(B = 1\) (\(-7296.2\))—which is why the golden number is searched over a grid: concavity holds for each given \(B\), not jointly. The broader lesson for structural work is that a behavioral regularity tied to a variable the firm controls need not force a new demand system; placed in one primitive and conditioned on, it lets decades of tractability results carry over.
38.4.3 Permutation Invariance and Deep Demand Estimation
Every model so far has bought tractability with a functional-form commitment, and the commitment shows up as a substitution pattern. The logit’s IIA is the crudest case; nested logit fixes the nests in advance; random coefficients let the data choose how much correlation to allow, but only within the family the analyst wrote down. The substitution matrix, which is the object every counterfactual in this chapter runs through, is therefore an assumption at least as much as a finding.
The alternative is to estimate the choice function itself. Let \(\mathbf{z}_j\) collect product \(j\)’s characteristics (attributes, price, and unobserved quality), and write market shares as \[ s_j = \sigma_j(\mathbf{z}_1, \dots, \mathbf{z}_J), \tag{38.13}\] with no parametric restriction on \(\sigma\). Estimating \(\sigma\) nonparametrically is hopeless as stated: its argument is the entire market, so the dimension grows with \(J\) and the curse of dimensionality bites immediately. Singh, Liu, and Yoganarasimhan (2026) supply the restriction that rescues it, and it is an economic restriction rather than a computational trick. A choice function should not depend on the labels we assign alternatives: relabeling the products permutes the shares correspondingly and changes nothing else, so for any permutation \(\pi\), \[ \sigma_{\pi(j)}\!\left(\mathbf{z}_{\pi(1)},\dots,\mathbf{z}_{\pi(J)}\right) = \sigma_j(\mathbf{z}_1,\dots,\mathbf{z}_J). \tag{38.14}\] This permutation invariance is not an approximation—it is a property every choice model in this chapter already satisfies—and it has a strong consequence. Any such function can be written in the sum-decomposable form \[ s_j = f\!\Big(\mathbf{z}_j,\ \textstyle\sum_{k=1}^{J} g(\mathbf{z}_k)\Big), \tag{38.15}\] in which rivals influence \(j\) only through a pooled summary \(\sum_k g(\mathbf{z}_k)\) that discards their order. The logit is the special case \(g(\mathbf{z}) = \exp(v(\mathbf{z}))\) with \(f(\mathbf{z}, m) = \exp(v(\mathbf{z}))/(1+m)\): Equation 38.8 is Equation 38.15 with particular \(f\) and \(g\) chosen a priori. The estimation strategy follows immediately— parameterize \(f\) and \(g\) as neural networks and learn them. The curse of dimensionality is broken because \(g\) maps one product’s characteristics, not the whole market, so the input dimension no longer grows with \(J\).
Two features make this a demand estimator rather than a demand predictor, which is the distinction Section 59.14 puts on trial. First, the framework extends to endogenous features: price still correlates with unobserved quality, and a flexible functional form does nothing to fix that, so instruments enter as they do in Equation 38.9. Second, the authors supply a formal inference procedure, so the deliverable is a confidence interval on an elasticity rather than an out-of-sample fit statistic. Applied to the automobile data of S. Berry, Levinsohn, and Pakes (1995), the estimator returns own- and cross-price elasticities in line with what the structural literature reports—a validation that matters precisely because a flexible model could have fit shares well while implying nonsense elasticities.
The chunk below makes the restriction earn its keep on a market where IIA is false by construction: two consumer segments with opposite brand tastes, so substitution within a segment is far stronger than across. It compares the true cross-price elasticities with three estimators—a logit, a flexible specification whose rivals enter through pooled sums as in Equation 38.15, and an equally flexible specification whose rivals enter by their position in the data file.
Code
set.seed(2029)
Tm <- 600; J <- 4 # markets; inside goods per market
grpv <- c("A", "A", "B", "B") # each product's segment: the source of non-IIA substitution
X <- matrix(rnorm(Tm * J), Tm, J) # observed quality
P <- matrix(runif(Tm * J, 1, 3), Tm, J)
Ai <- matrix(grpv == "A", Tm, J, byrow = TRUE)
# TRUTH: two consumer types with opposite segment tastes, so a price rise on a
# same-segment rival diverts far more demand than one on a cross-segment rival.
true_shares <- function(Pm, Xm = X, Am = Ai) {
s <- function(V) { E <- exp(V); E / (1 + rowSums(E)) } # outside good normalized
0.5 * s(Xm - 1.5 * Pm + 2 * Am) + 0.5 * s(Xm - 1.5 * Pm + 2 * (!Am))
}
# d ln s_j / d ln p_k for every (market, j, k), by numerical perturbation
elast <- function(fn, Pm, dl = 1e-4) {
base <- log(fn(Pm)); out <- array(0, c(nrow(Pm), J, J))
for (k in 1:J) { P2 <- Pm; P2[, k] <- Pm[, k] * (1 + dl)
out[, , k] <- (log(fn(P2)) - base) / dl }
out
}
# average cross-elasticity, split by whether the rival is in the same segment
cross <- function(E, gv) {
s <- d <- numeric(0)
for (j in 1:J) for (k in 1:J) if (j != k)
if (gv[j] == gv[k]) s <- c(s, mean(E[, j, k])) else d <- c(d, mean(E[, j, k]))
c(same = mean(s), diff = mean(d))
}
fit_all <- function(cols) { # cols = an arbitrary labeling of the products
Xm <- X[, cols]; Pm <- P[, cols]; Am <- Ai[, cols]; gv <- grpv[cols]
S <- true_shares(Pm, Xm, Am); y <- as.vector(log(S) - log(1 - rowSums(S)))
# (1) LOGIT: rivals enter only through the market's own normalization -> IIA
fitL <- lm(y ~ xo + po, data.frame(xo = as.vector(Xm), po = as.vector(Pm)))
sh_logit <- function(Q) {
D <- matrix(cbind(1, as.vector(Xm), as.vector(Q)) %*% coef(fitL), nrow(Q), J)
E <- exp(D); E / (1 + rowSums(E)) }
# (2) PERMUTATION-INVARIANT: rivals enter through pooled sums, as in eq-sm-deepsets
sym <- function(Q) data.frame(
xo = as.vector(Xm), po = as.vector(Q), ga = as.vector(Am) * 1,
rpa = as.vector(matrix(rowSums(Q * Am), nrow(Q), J) - Q * Am),
rpb = as.vector(matrix(rowSums(Q * !Am), nrow(Q), J) - Q * !Am),
rxa = as.vector(matrix(rowSums(Xm * Am), nrow(Q), J) - Xm * Am),
rxb = as.vector(matrix(rowSums(Xm * !Am), nrow(Q), J) - Xm * !Am))
fitS <- lm(y ~ (xo + po + rpa + rpb + rxa + rxb) * ga, sym(Pm))
sh_sym <- function(Q) { D <- matrix(predict(fitS, sym(Q)), nrow(Q), J)
E <- exp(D); E / (1 + rowSums(E)) }
# (3) ORDER-DEPENDENT: equally flexible, but rivals enter by storage position
cyc <- function(M, r) M[, ((seq_len(J) + r - 1) %% J) + 1, drop = FALSE]
ord <- function(Q) data.frame(
xo = as.vector(Xm), po = as.vector(Q), ga = as.vector(Am) * 1,
rp1 = as.vector(cyc(Q, 1)), rp2 = as.vector(cyc(Q, 2)), rp3 = as.vector(cyc(Q, 3)),
rx1 = as.vector(cyc(Xm, 1)), rx2 = as.vector(cyc(Xm, 2)), rx3 = as.vector(cyc(Xm, 3)))
fitO <- lm(y ~ (xo + po + rp1 + rp2 + rp3 + rx1 + rx2 + rx3) * ga, ord(Pm))
sh_ord <- function(Q) { D <- matrix(predict(fitO, ord(Q)), nrow(Q), J)
E <- exp(D); E / (1 + rowSums(E)) }
rbind(truth = cross(elast(function(Q) true_shares(Q, Xm, Am), Pm), gv),
logit = cross(elast(sh_logit, Pm), gv),
invariant = cross(elast(sh_sym, Pm), gv),
ordered = cross(elast(sh_ord, Pm), gv))
}
a <- fit_all(1:J) # one labeling of the products
b <- fit_all(c(2, 4, 1, 3)) # the SAME data, products relabeled
out <- cbind(a, b); colnames(out) <- c("same-seg", "diff-seg",
"same-seg (relabeled)", "diff-seg (relabeled)")
print(round(out, 3))
#> same-seg diff-seg same-seg (relabeled) diff-seg (relabeled)
#> truth 0.536 0.171 0.536 0.171
#> logit 0.333 0.333 0.333 0.333
#> invariant 0.580 0.152 0.580 0.152
#> ordered 0.366 0.259 0.580 0.152
cat(sprintf("\nIIA gap (same - diff): truth %.3f | logit %.3f | invariant %.3f\n",
a["truth", "same"] - a["truth", "diff"],
a["logit", "same"] - a["logit", "diff"],
a["invariant", "same"] - a["invariant", "diff"]))
#>
#> IIA gap (same - diff): truth 0.364 | logit 0.000 | invariant 0.428
cat(sprintf("Largest shift under relabeling: invariant %.4f | ordered %.4f\n",
max(abs(a["invariant", ] - b["invariant", ])),
max(abs(a["ordered", ] - b["ordered", ]))))
#> Largest shift under relabeling: invariant 0.0000 | ordered 0.2138Three results, in order of importance. The logit reports the same cross-elasticity for every rival—the IIA straitjacket, visible as an exactly zero gap between same-segment and cross-segment substitution where the truth has a large one. The permutation-invariant specification, using nothing but pooled sums of rivals’ characteristics, recovers the asymmetry: same-segment rivals divert several times the demand that cross-segment rivals do, as they should. And the order-dependent specification—identical in flexibility, differing only in that rivals enter by position rather than through a symmetric summary—gives one answer under the original labeling and a different answer after the products are relabeled, while the invariant estimator’s numbers do not move at all.
That last comparison is the whole argument for Equation 38.14. The order-dependent model is not merely less accurate; it is answering a question that has no economic content, because the ordering of rows in a data file is a bookkeeping convention and no counterfactual may depend on it. Under the second labeling it happens to land on the right answer, for the worst possible reason: that labeling accidentally aligns rival position with segment membership. Imposing invariance is what converts a flexible function approximator into an estimator of demand—the same lesson as the exclusion restriction in Equation 38.9, one level up. Structure is not the price you pay for tractability; it is what makes the output mean something.
38.5 Dynamics, Learning, and Forward-Looking Agents
Many marketing decisions are intertemporal—when to buy a durable, whether a salesperson exerts effort today for a bonus tomorrow, how a relationship evolves—and a static choice model cannot represent them. Dynamic structural models treat the agent as solving a sequential optimization, and they are the technically hardest but most rewarding members of the family.
The canonical object is a forward-looking agent who, in state \(s_t\), chooses action \(a_t\) to maximize the expected present value of a stream of payoffs, \[ V(s_t) = \max_{a_t}\; \Big\{ u(s_t, a_t; \boldsymbol{\theta}) + \delta\, \mathbb{E}\!\left[\, V(s_{t+1}) \mid s_t, a_t \,\right] \Big\}, \tag{38.16}\] the Bellman equation, where \(\delta\) is the discount factor, \(u(\cdot)\) the per-period payoff, and the expectation runs over the law of motion of the state. Estimation recovers the payoff primitives \(\boldsymbol{\theta}\) from observed state–action sequences, typically by nesting the solution of Equation 38.16 inside a likelihood. The substantive range is wide. Allenby, Leone, and Jen (1999) model purchase timing in direct marketing as a dynamic decision, recovering when a household will next buy. Gowrisankaran and Rysman (2012) let consumer preferences evolve and apply the machinery to durable-goods markets, where today’s purchase forecloses tomorrow’s. Hitsch (2006) study optimal product launch and exit under demand uncertainty, treating the firm as a forward-looking decision maker learning about an uncertain market. Netzer, Lattin, and Srinivasan (2008) model the dynamics of customer–firm relationships as a hidden Markov process, in which a latent relationship state migrates over time and governs observable purchase behavior—a structural account of relationship “states” rather than a reduced-form churn score.
Forward-looking firms appear on the compensation side of the same coin. Misra and Nair (2011) estimate a structural model of sales-force incentives in which the salesperson chooses unobserved effort given the compensation plan, the firm designs the plan anticipating that response, and the recovered primitives (the cost of effort, the agent’s risk attitude) support counterfactual plans the firm never tried—work that Kim, Sudhir, and Uetake (2022) extend to a multitasking sales force allocating effort across acquisition and retention. On the demand side over the customer lifecycle, Gupta et al. (2006) synthesize structural and other models of customer lifetime value, and Kamakura et al. (2005) embed choice models in customer-relationship strategy. Figure 38.2 organizes the marketing applications cited in this chapter by the decision they model.
flowchart LR
R[Structural marketing models] --> S[Static]
R --> D[Dynamic / forward-looking]
S --> S1[Demand & price elasticity<br/>montgomery1999estimating]
S --> S2[Discrete choice<br/>dube2002structural · heiss2002structural]
S --> S3[Competitive conduct<br/>kadiyali2001structural]
D --> D1[Purchase timing & durables<br/>allenby1999dynamic · gowrisankaran2012dynamics]
D --> D2[Customer relationships<br/>netzer2008hidden · gupta2006modeling]
D --> D3[Salesforce incentives<br/>misra2011structural · kim2022structural]
D --> D4[Product launch & exit<br/>hitsch2006empirical]
The dynamic apparatus also extends to settings where behavior and content co-evolve. Goh, Heng, and Lin (2013) model the co-evolution of user behavior in social media, and Rutz, Trusov, and Bucklin (2011) use a structural model to disentangle the indirect effects of paid search advertising—how a search click today shifts later behavior—an effect invisible to a single-period attribution model. Price promotions, too, have dynamic consequences: Elberg et al. (2019) study how promotions move purchases across time rather than merely expanding the category, a stockpiling response that only a dynamic model represents correctly.
38.6 Identification: What Can and Cannot Be Recovered
A model is identified if distinct parameter values imply distinct distributions of the observable data; without identification, no estimator—however clever—can recover the truth, and a perfectly converged optimizer may report a number that the data do not pin down. Because structural models lean on assumptions for identification, naming the threats is as important as writing the likelihood.
The first and most pervasive threat is endogeneity, already met twice: price is set by agents who observe demand shocks the analyst does not, so the price coefficient is identified only through an exclusion restriction—a cost or competitor instrument that shifts the choice set without entering utility directly. When no valid instrument exists, the elasticity is not identified, full stop.
The second is functional-form dependence. In a static logit the substitution pattern is dictated by the IIA assumption rather than estimated, so a counterfactual that turns on cross-elasticities is identified by assumption, not by data; relaxing to nested or random-coefficient logit buys realism at the cost of additional parameters that themselves require variation to identify.
The third is the discount factor in dynamic models. The forward-looking payoff in Equation 38.16 and the per-period payoff are notoriously hard to separate: a patient agent with low effort costs can mimic an impatient agent with high effort costs, so \(\delta\) is typically fixed rather than estimated, with the counterfactual’s credibility contingent on that choice.
The fourth is the equilibrium and conduct assumption: inferring competitive conduct from price–quantity comovement (Kadiyali, Sudhir, and Rao 2001) identifies conduct only if the maintained equilibrium concept is correct, and a misspecified conduct parameter contaminates the recovered costs.
The pattern across all four is the same and is worth internalizing as the chapter’s governing principle: structural identification trades data variation for assumptions, and every counterfactual inherits the assumptions that identified the parameters it uses. The honest structural paper, accordingly, reports which results are robust to weakening which assumption—and the practitioner reading one should ask, of every headline number, what assumption is this resting on?
38.6.1 Borrowing Identification from a Discontinuity
One way to reduce the assumptions a structural parameter rests on is to import its identification from a reduced-form design. Moehring (2026) does this for a social-media feed. The paper asks how Reddit’s news feed would change the credibility of the news people engage with if it ranked posts to maximize predicted engagement, and what it would cost to rank for credibility instead. Neither ranking was ever deployed, so the answer has to come from a model. The key parameter of that model, how much a post’s position governs whether anyone sees it, comes from a regression discontinuity built into Reddit’s own ranking formula (Section 44.7.3).
The model. User \(i\) visiting in period \(t\) sees the post in position \(r\) with probability \(p_t\,p_r\). Here \(p_t\) is the chance of logging on and \(p_r\) the chance of reaching position \(r\). Conditional on exposure, the user comments if a logit utility clears the outside option:
\[ \Pr(d_{ijt} = 1) = p_t\,p_{r_{jt}}\; \frac{\exp(\delta_{ijt})}{1 + \exp(\delta_{ijt})}, \qquad \delta_{ijt} = x_{jt}'(\bar\beta + \beta_i) + \xi_{jt}. \tag{38.17}\]
The vector \(x_{jt}\) holds the publisher’s political slant, slant squared, and a credibility rating. Every user gets their own \(\beta_i\), which the panel of comment histories makes estimable. \(\xi_{jt}\) is an article-period effect for latent commentability. Position does not enter utility: it changes who sees a post, not how much a reader who sees it wants to reply. That restriction comes from evidence that rankings move search but not purchase conditional on search (Ursu 2018).
Where identification comes from. The obstacle is that rank and \(\xi_{jt}\) are correlated, because the feed promotes commentable posts. Without \(p_r\), a post with few comments could be unappealing or simply unseen, and the data cannot tell which. Excluding rank from utility, with treatment effects of rank held constant across users and articles, makes the RD estimate at each boundary a ratio of exposure probabilities:
\[ \tau_r = \log\frac{\sum_i \mathbb{E}[d_{ijt}(r)]}{\sum_i \mathbb{E}[d_{ijt}(r+1)]} = \log\frac{p_r}{p_{r+1}}, \qquad p_1 = 1, \tag{38.18}\]
so the whole exposure curve is \(p_r = \exp\!\big(-\sum_{r'<r}\tau_{r'}\big)\). Setting \(p_1 = 1\) is a normalization, and the counterfactuals are robust to rescaling it. \(p_t\) is calibrated from the number of users online in each archived snapshot via Little’s law. With exposure fixed, each user’s \(\beta_i\) comes from a likelihood over their own comment decisions. The \(\xi_{jt}\) come from inverting observed comment shares (S. T. Berry 1994; S. Berry, Levinsohn, and Pakes 1995), and the individual estimates are shrunk by empirical Bayes to strip out sampling noise. The estimation sample is users who commented on at least 20 articles. They are 5% of commenters but write 44% of the comments.
The counterfactuals. A platform that knows \(\beta_i\) and predicts \(\xi_{jt}\) from post observables can rank by predicted comment probability. When exposure falls with position, sorting by that probability maximizes expected engagement. The credibility-aware alternative sorts by \(\big((1-\lambda) + \lambda\,\mathbf{1}[c_j \ge 0.65]\big)\hat P_{ijt}\). Here \(\lambda = 0\) is pure engagement maximization, \(\lambda = 1\) ranks every high-credibility publisher first, and the 0.65 threshold is the median score within a professional rater’s “medium” credibility band (Lin et al. 2023). Tracing \(\lambda\) from 0 to 1 gives an engagement-credibility frontier. The substantive results are in Section 70.6.
flowchart LR
A["Ranking code:<br/>hot score cutoff<br/>between adjacent posts"] --> B["RD estimates<br/>tau_r by position"]
B --> C["Exposure curve<br/>p_r = exp(-sum tau)"]
D["User comment panel<br/>+ publisher slant<br/>and credibility"] --> E["Comment-choice model<br/>beta_i, xi_jt"]
C --> E
E --> F["Counterfactual rankings:<br/>engagement max,<br/>non-personalized,<br/>credibility-aware lambda"]
F --> G["Engagement-credibility<br/>frontier"]
What the counterfactuals inherit. Each assumption that identified the parameters carries into the counterfactuals, as the identification section above warns. The constant-effect assumption turns a pooled RD into a single exposure curve applied to everyone. Only the 25 posts on the archived first page can be re-ranked, because no exposure or \(\xi_{jt}\) estimates exist below it. That rules out evaluating a reverse-chronological feed. Supply, voting, and users’ attention are held fixed. The paper relaxes the attention assumption in a search-cost version, in which search costs drive most of the position effect and re-optimized attention barely moves exposure. It also checks the personalization results against a collaborative-filtering recommender trained on over 500,000 users. Its predictions are validated by whether the RD effects are larger when the promoted post matches a user’s predicted taste.
The sketch below runs the pipeline forward. It takes an exposure curve calibrated to the smoothed RD estimates in the author’s April 2026 manuscript, simulates a comment-choice model in which some users prefer low-credibility publishers, and traces part of the frontier.
Code
set.seed(2906)
# --- Step 1: exposure by position, taken from an RD --------------------------
# Rank enters only through exposure, so each boundary's RD effect is a log
# ratio of exposure probabilities: tau_r = log(p_r / p_{r+1}), with p_1 = 1.
# Calibrated to the smoothed curve tau_r = 0.458 - 0.164 log r reported in the
# author's April 2026 manuscript (floored at zero from position 17 down).
R <- 25
tau_r <- pmax(0.458 - 0.164 * log(1:(R - 1)), 0)
p_r <- exp(-c(0, cumsum(tau_r)))
# --- Step 2: a comment-choice model, conditional on exposure -----------------
# Users differ in their taste for credible publishers (some prefer low-
# credibility outlets); xi is a post's latent commentability. Posts are drawn
# once so that every ranking rule re-orders the same feeds.
N <- 1500; Tn <- 30; cbar <- 0.65
beta <- rnorm(N, 0, 1)
feeds <- replicate(Tn, list(c = rbeta(R, 4, 1.6), xi = rnorm(R)), simplify = FALSE)
evaluate <- function(lambda, personalized = TRUE) {
eng <- hi <- numeric(N)
for (f in feeds) {
P <- plogis(-3.5 + outer(rep(1, N), f$xi) + outer(beta, 2.5 * (f$c - cbar)))
w <- (1 - lambda) + lambda * (f$c >= cbar) # credibility-aware weight
S <- if (personalized) sweep(P, 2, w, `*`) else
matrix(colMeans(P) * w, N, R, byrow = TRUE)
# Rank within each user's feed (1 = top), then expose by position.
o <- order(rep(seq_len(N), R), -S)
rk <- matrix(0L, N, R); rk[o] <- rep(seq_len(R), times = N)
E <- P * p_r[rk]
eng <- eng + rowSums(E); hi <- hi + drop(E %*% (f$c >= cbar))
}
low <- beta < quantile(beta, 0.25) # users who favour low credibility
c(engagement = mean(eng), hi_cred_share = mean(hi / eng),
hi_cred_share_low_taste = mean((hi / eng)[low]))
}
res <- rbind(`engagement max (lambda = 0)` = evaluate(0),
`lambda = 0.5` = evaluate(0.5),
`lambda = 0.8` = evaluate(0.8),
`credibility max (lambda = 1)` = evaluate(1),
`non-personalized, lambda = 0` = evaluate(0, personalized = FALSE))
res[, "engagement"] <- res[, "engagement"] / res[1, "engagement"]
round(res, 3)
#> engagement hi_cred_share hi_cred_share_low_taste
#> engagement max (lambda = 0) 1.000 0.648 0.368
#> lambda = 0.5 0.975 0.733 0.452
#> lambda = 0.8 0.906 0.833 0.618
#> credibility max (lambda = 1) 0.854 0.893 0.805
#> non-personalized, lambda = 0 0.952 0.657 0.471Engagement is indexed to the personalized engagement-maximizing feed. Setting \(\lambda = 0.5\) gives up 2.5% of engagement and raises the average high-credibility share from 0.648 to 0.733. Dropping personalization costs more engagement (4.8%) and buys almost no credibility (0.657). The last column shows why the distribution matters as well as the average. Among the quarter of users with the strongest taste for low-credibility outlets, the personalized engagement-maximizing feed leaves a high-credibility share of 0.368, below the 0.471 they get from the non-personalized feed. The personalized feed shows low-credibility content to the users who will comment on it. Credibility-aware weighting lifts this group the most. These magnitudes come from the calibration, not from estimates. The qualitative pattern, a flat frontier near the engagement maximum with gains concentrated among low-credibility users, matches the paper’s.
38.6.2 Checking a Structural Model Against a Field Experiment
The previous subsection imported a reduced-form estimate into a structural model. The stronger test runs the other way: estimate the structural model without the experiment, then ask whether it predicts what the experiment finds. Few structural papers can do this, because the experiment that would discipline the model is usually the one nobody ran. Gabel et al. (2026) ran it for the oldest parameter in scanner-data choice modeling, structural state dependence.
The parameter and why it is hard. Since Guadagni and Little (1983), brand-choice models have included a term for the brand bought last time. Its coefficient is structural state dependence: buying a brand today makes the household more likely to buy it next time, through a switching cost, habit, or the like. The trouble is that a household that keeps buying the same brand might simply like it. Persistent taste differences produce the same autocorrelation in choices with no causal link between purchases (Keane 1997). Structural models separate the two by assuming a distribution for taste heterogeneity and modelling the initial conditions. Dubé, Hitsch, and Rossi (2010) did this carefully and found that genuine state dependence survives once heterogeneity and alternative explanations are controlled. But the estimate still rests on those assumptions, which is the objection Angrist and Pischke (2010) raise against structural work in industrial organization in general (Section 38.1).
The experiment. In a large natural field experiment, randomly assigned coupons push some households toward a focal brand. The coupon is an encouragement in the sense of Section 44.6.1. It moves today’s brand choice, and as long as it affects next purchase only through today’s choice, the Wald ratio identifies the local average treatment effect (LATE) of buying the brand now on buying it again. That LATE is nonparametric. It needs no distributional assumption about preferences, and randomization removes the selection that confounds the observational comparison.
What they find. Across 223 brands, the authors find positive structural state dependence for most, using what they call a principled approach to aggregating the evidence. On average, buying a brand raises the probability of buying it again by 10 percentage points, conditional on a category purchase. That is a 74% lift and corresponds to relative switching costs of 21%. The result that matters for this chapter is the comparison: structural brand-choice models produce LATEs consistent with those from the encouragement design. The authors draw the practical implication. Structural models are easier to estimate and give more precise estimates, and only they are built for policy evaluation. Here the experiment shows that their extra assumptions do not distort the answer.
The result does not settle the question everywhere. Levine and Seiler (2023) use hurricane-driven stockouts of bottled water as a natural experiment that forces households to switch brands. In the working-paper version, purchases revert to their pre-hurricane path immediately and the authors cannot reject zero state dependence, while a discrete-choice model estimated from price variation in the same category finds a positive effect (Levine and Seiler 2022). Two quasi-random shocks to brand choice, two categories, and opposite verdicts on the structural model. A seminar should ask which differences explain the gap. Candidates are the category, the source of the shock (a coupon the household chose to redeem versus a stockout it did not choose), who the compliers are, and how much power each design has.
How to compare a model with an experiment. A structural coefficient and a LATE are different objects. The coefficient is a utility parameter for the whole population; the LATE is a probability difference for the households that the coupon moved. The comparison is valid only if the fitted model runs the same experiment internally. Simulate the coupon for each household and compute the complier-weighted change in repurchase probability that the model implies. Then compare that number with the Wald estimate. The sketch below does this on simulated data where the truth is known. It fits two structural models to a 12-period pre-experiment panel that never contains a coupon. One is a loyalty logit in the Guadagni–Little style without heterogeneity. The other adds a normal random intercept whose mean depends on the initial choice, which handles the initial conditions. Each model then predicts the coupon experiment.
Code
set.seed(2911)
# --- A market with true state dependence AND persistent taste heterogeneity ---
# Each period every household buys the focal brand (1) or the rival (0).
# v = alpha_i + gamma * [focal bought last time] - beta * (price gap)
N <- 20000; Tobs <- 12; burn <- 30
mu <- -1; sigma <- 1.5; gamma <- 0.6; beta <- 2
alpha <- rnorm(N, mu, sigma)
draw_dp <- function() rnorm(N, 0, 0.4) # focal minus rival price
step <- function(last, dp) rbinom(N, 1, plogis(alpha + gamma * last - beta * dp))
last <- rbinom(N, 1, 0.3)
for (t in seq_len(burn)) last <- step(last, draw_dp())
y0 <- last # initial condition
Y <- DP <- matrix(0, N, Tobs)
for (t in seq_len(Tobs)) { DP[, t] <- draw_dp(); Y[, t] <- last <- step(last, DP[, t]) }
LAG <- cbind(y0, Y[, -Tobs])
# --- The experiment: a randomized coupon in period 13, outcome in period 14 ---
coupon <- 0.5 # price cut on the focal brand
Z <- rbinom(N, 1, 0.5)
dp13 <- draw_dp(); dp14 <- draw_dp()
D <- rbinom(N, 1, plogis(alpha + gamma * last - beta * (dp13 - coupon * Z)))
Yn <- rbinom(N, 1, plogis(alpha + gamma * D - beta * dp14))
# Ground truth: complier-weighted effect of buying focal at 13 on buying it at 14
w_true <- plogis(alpha + gamma * last - beta * (dp13 - coupon)) -
plogis(alpha + gamma * last - beta * dp13)
d1 <- plogis(alpha + gamma - beta * dp14); d0 <- plogis(alpha - beta * dp14)
truth <- sum(w_true * (d1 - d0)) / sum(w_true)
lift <- truth / (sum(w_true * d0) / sum(w_true))
# Encouragement design: Wald estimator and its standard error
fs <- mean(D[Z == 1]) - mean(D[Z == 0]); rf <- mean(Yn[Z == 1]) - mean(Yn[Z == 0])
wald <- rf / fs
u <- Yn - wald * D
se_wald <- sqrt(var(u[Z == 1]) / sum(Z) + var(u[Z == 0]) / sum(1 - Z)) / abs(fs)
# Naive observational contrast in the control arm
naive <- mean(Yn[Z == 0 & D == 1]) - mean(Yn[Z == 0 & D == 0])
# --- Structural models, fit on the pre-period panel only (no experiment) ---
# (a) Loyalty logit with no heterogeneity
ca <- coef(glm(as.vector(Y) ~ as.vector(LAG) + as.vector(DP), family = binomial))
# (b) Random-intercept logit, alpha_i ~ N(a0 + a1 * y0, s^2) (initial conditions),
# integrated by 15-node Gauss-Hermite quadrature, with an analytic gradient
gh <- local({ n <- 15; J <- matrix(0, n, n); b <- sqrt(seq_len(n - 1) / 2)
J[cbind(1:(n - 1), 2:n)] <- b; J[cbind(2:n, 1:(n - 1))] <- b
e <- eigen(J, symmetric = TRUE)
list(x = e$values * sqrt(2), w = e$vectors[1, ]^2) }) # nodes/weights for N(0,1)
K <- length(gh$x)
nodes <- function(th) {
a0 <- th[1]; a1 <- th[2]; s <- exp(th[3]); g <- th[4]; b <- th[5]
ll <- matrix(0, N, K); sc <- vector("list", K)
for (k in seq_len(K)) {
p <- plogis(a0 + a1 * y0 + s * gh$x[k] + g * LAG - b * DP)
ll[, k] <- rowSums(Y * log(p) + (1 - Y) * log1p(-p))
r <- Y - p
sc[[k]] <- cbind(rowSums(r), y0 * rowSums(r), s * gh$x[k] * rowSums(r),
rowSums(r * LAG), -rowSums(r * DP))
}
m <- apply(ll, 1, max) # log-sum-exp for stability
post <- sweep(exp(ll - m), 2, gh$w, `*`)
list(ll = log(rowSums(post)) + m, post = post / rowSums(post), sc = sc)
}
nll <- function(th) -sum(nodes(th)$ll)
grad <- function(th) { z <- nodes(th)
-Reduce(`+`, lapply(seq_len(K), function(k) colSums(z$post[, k] * z$sc[[k]]))) }
th <- optim(c(-1, 1, 0, 0, 1), nll, grad, method = "BFGS",
control = list(maxit = 500))$par
# --- Run the coupon experiment inside each fitted model ---
late_a <- local({
w <- plogis(ca[1] + ca[2] * last + ca[3] * (dp13 - coupon)) -
plogis(ca[1] + ca[2] * last + ca[3] * dp13)
sum(w * (plogis(ca[1] + ca[2] + ca[3] * dp14) - plogis(ca[1] + ca[3] * dp14))) / sum(w)
})
late_b <- local({ # integrate each household's alpha over its posterior
a0 <- th[1]; a1 <- th[2]; s <- exp(th[3]); g <- th[4]; b <- th[5]
post <- nodes(th)$post; num <- den <- 0
for (k in seq_len(K)) {
a <- a0 + a1 * y0 + s * gh$x[k]
w <- plogis(a + g * last - b * (dp13 - coupon)) - plogis(a + g * last - b * dp13)
num <- num + sum(post[, k] * w * (plogis(a + g - b * dp14) - plogis(a - b * dp14)))
den <- den + sum(post[, k] * w)
}
num / den
})
res <- data.frame(
estimate = c("True complier effect (simulation)",
"Encouragement design (Wald LATE)",
"Naive repurchase gap, control arm",
"Structural (a): loyalty logit, no heterogeneity",
"Structural (b): random intercept + initial conditions"),
repurchase_effect = round(c(truth, wald, naive, late_a, late_b), 3))
print(res, row.names = FALSE)
#> estimate repurchase_effect
#> True complier effect (simulation) 0.106
#> Encouragement design (Wald LATE) 0.117
#> Naive repurchase gap, control arm 0.346
#> Structural (a): loyalty logit, no heterogeneity 0.363
#> Structural (b): random intercept + initial conditions 0.107
cat(sprintf("Wald SE %.3f | first stage %.3f | true lift %.0f%% | gamma: (a) %.2f, (b) %.2f, true %.2f\n",
se_wald, fs, 100 * lift, ca[2], th[4], gamma))
#> Wald SE 0.043 | first stage 0.154 | true lift 31% | gamma: (a) 1.71, (b) 0.61, true 0.60In this simulation the true effect on compliers is 0.106: buying the focal brand raises the chance of buying it again by about 11 points, a 31% lift. The coupon experiment recovers it (0.117). The naive comparison of households that did and did not buy the brand in the control arm (0.346) is more than three times too large, because loyal buyers differ in taste. The loyalty logit without heterogeneity makes the same mistake from inside a structural model (0.363), with an estimated loyalty coefficient of 1.71 against a true 0.60. Ignoring taste heterogeneity biases state dependence upward, as the classic literature warned. The random-intercept model, fitted to data that contain no coupon, predicts 0.107. That is the pattern Gabel et al. (2026) report: a correctly specified structural model matches the experiment it never saw.
Note the standard error of the Wald estimate, 0.043 with 20,000 households. The first stage is weak (the coupon shifts purchase for about 15% of households), so a single brand’s experiment is noisy. This is why the paper aggregates evidence across 223 brands, and why the structural model’s precision is worth having once the experiment has validated it. The parameter values are illustrative and are not calibrated to the paper, which we could read only in its published abstract. The paper’s data, coupon design, and model specifications are not described here.
38.7 Getting Started
For readers building toward this literature, the graduate econometrics texts of Greene (2003) and Cameron and Trivedi (2005) supply the estimation foundations; Hensher, Rose, and Greene (2005) is the standard reference for the discrete-choice models that anchor most marketing applications; Diamantopoulos, Fritz, and Hildebrandt (2013) surveys quantitative modeling in marketing more broadly; and the Handbook of Econometrics policy-evaluation chapters (Heckman and Vytlacil 2007a, 2007b; Abbring and Heckman 2007) connect structural estimation to the identification of treatment effects. A productive path is to reproduce a linear demand-and-supply system (Section 38.2), then a logit with endogenous price (Section 38.4), before attempting a dynamic model—each step adds one layer of structure, and the endogeneity intuition learned in the first carries unchanged into the last.
38.8 Key Takeaways
- A model is structural when its parameters are economic primitives—tastes, costs, information, conduct—and are therefore policy invariant; the payoff is counterfactual simulation of policies never observed (Equation 38.1).
- Structural modeling and causal inference are complements: causal estimates discipline structural models, and structural reasoning locates the endogeneity a causal design must address.
- Endogeneity of price is the central econometric problem, visible already in linear supply and demand (Equation 38.3) and recurring inside discrete-choice demand; it is resolved only by a credible exclusion restriction.
- MLE (full likelihood, efficient if correct) and GMM (moment conditions, robust) are the two workhorse estimators; the choice is about how much of the data-generating process one is willing to specify (Table 38.1).
- Dynamic models recover forward-looking primitives via the Bellman equation (Equation 38.16) but face hard identification of the discount factor.
- Every counterfactual inherits the assumptions that identified its parameters; the credible structural paper reports which conclusions survive weakening which assumption (Section 38.6).
- A behavioral effect can enter a structural model through one primitive: making the outside option’s utility depend on assortment size reproduces choice overload while keeping the logit, and fixing the set size recovers the classical assortment algorithms and a concave likelihood (Section 38.4.2).
Cite this section
Nguyen, M. (2026). Structural Models. In Marketing research: A technical and academic treatment. https://mikenguyen13.github.io/marketing_research/29-structural-models.html
BibTeX
@inbook{Nguyen2026Marketing-29-structural-models,
author = {Nguyen, Mike},
title = {Marketing Research: A Technical and Academic Treatment},
chapter = {Structural Models},
year = {2026},
url = {https://mikenguyen13.github.io/marketing_research/29-structural-models.html}
}
Use this book with your AI assistant
Ask questions about this book in Claude, Cursor, VS Code, or ChatGPT, with answers that cite the section.
https://research-methods-mcp.vercel.app/mcp?src=book-marketing-research