flowchart TB
Q["A marketing question"]
Q --> A["Analytical / game-theoretic<br/>(what should happen?)"]
Q --> E["Empirical / econometric<br/>(what is happening, and why?)"]
Q --> P["Predictive / machine learning<br/>(what will happen next?)"]
A --> A1["Primitives: payoffs, info, timing"]
A --> A2["Solution: equilibrium"]
A --> A3["Output: comparative statics"]
E --> E1["Structural: utilities, costs (invariant)"]
E --> E2["Reduced-form: response functions (variant)"]
E --> E3["Output: estimates + counterfactuals"]
P --> P1["Features and a loss function"]
P --> P2["Solution: regularized fit"]
P --> P3["Output: forecasts, scored by test error"]
30 Modeling in Marketing
A model is a deliberately incomplete description of a marketing system, built to answer a specific question. Models are not the world; they are arguments about the world, made precise enough that their assumptions can be stated, their logic checked, and their predictions confronted with data. Every quantitative chapter in the methodology part of this book builds a model of some kind, and the chapters differ less in subject matter than in the kind of question each model is licensed to answer. This chapter is the map. It distinguishes the three modeling traditions that organize quantitative marketing—analytical (game-theoretic), empirical (econometric and structural), and predictive (statistical and machine learning)—and gives the reader a principled way to decide which one a given problem calls for.
The distinction matters because the three traditions answer different questions and fail in different ways. An analytical model asks what should happen when rational firms and consumers interact under stated rules; it trades realism for logical transparency and yields comparative statics—signed predictions about how equilibrium outcomes move with primitives. An empirical model asks what is happening, and why; it recovers parameters of preferences, costs, or response functions from data and, when those parameters are invariant to the policies under study, supports counterfactual prediction. A predictive model asks what will happen next; it learns a flexible mapping from features to outcomes and is judged almost entirely by out-of-sample accuracy, with little claim on the mechanism that generated the data. Confusing these aims is the most common and most expensive error in applied marketing modeling: a model fit for prediction is routinely, and wrongly, read as if it identified a causal effect, and a stylized analytical result is routinely, and wrongly, taken as an empirical fact.
This chapter proceeds from the shared anatomy of a model to the three traditions in turn, formalizing each with its canonical object, its notion of a “solution” or “estimate,” and the assumption whose failure is fatal. It closes with a decision framework and a worked illustration that runs the same marketing question—how a price change affects demand—through all three lenses, so the reader can see exactly where they agree, where they diverge, and why. The goal is not to rank the traditions but to make the reader fluent in choosing among them. Subsequent chapters develop each in depth: analytical models in Chapter 31 and Chapter 39, empirical and structural models in Chapter 32, Chapter 34, and Chapter 53, and predictive methods in Chapter 65.
30.1 What a Model Is
Before partitioning models into traditions, it helps to fix the parts every model shares. Strip away the subject matter and a quantitative marketing model is a tuple of four objects: primitives, structure, a solution or estimation concept, and an empirical content.
The primitives are the objects taken as given and not themselves explained: consumer preferences, firm cost functions, the information each agent holds, the timing of moves, the data-generating distribution. The structure is the set of assumptions that link primitives to observable outcomes—a utility function, a demand system, a conduct assumption about how firms compete, a likelihood. The solution or estimation concept is the rule that turns structure into an answer: an equilibrium notion in an analytical model, an estimator in an empirical one, a learning algorithm and loss function in a predictive one. The empirical content is whatever the model says about data we can observe—signs of comparative statics, point estimates with standard errors, or a forecast with a quantified error.
A model is a representation that is simpler than the system it represents, built so that reasoning about the model is a reliable guide to reasoning about the system, within a stated scope. Its value lies entirely in the gap between the assumptions we impose and the conclusions we can then defend.
The decisive property that separates the traditions is what each model holds fixed when the environment changes. A policy-invariant parameter is one that does not shift when the policy under study shifts—a consumer’s taste for a product attribute, say, which should not change merely because a firm reprices. A policy-variant relationship is one that is itself a function of the policy environment—a reduced-form regression of sales on price confounds the demand the analyst wants with the supply behavior that set the price, so its coefficient changes if pricing conduct changes. This distinction, central to the structural tradition, is the thread we follow throughout: it explains why two models can fit the same data equally well yet give opposite answers to a counterfactual question.
Notation for this chapter
Following the book’s master notation (Chapter 29), scalars are italic (\(p\), \(x\)), vectors bold lowercase (\(\mathbf{x}\)), matrices bold uppercase (\(\mathbf{X}\)), estimators carry hats (\(\hat\beta\)), and \(\mathbb{E}[\cdot]\) denotes expectation. A parameter \(\theta\) is a feature of the data-generating process; an estimate \(\hat\theta\) is a function of a finite sample; a prediction \(\hat y\) is a model’s output for a new input.
30.2 A Taxonomy of Marketing Models
The three traditions can be placed on two axes that together explain most of their differences in practice. The first axis is the role of optimizing behavior: whether the model derives outcomes from agents solving stated optimization problems (analytical, and the structural branch of empirical work) or treats outcomes as the realization of a statistical process to be fit (reduced-form empirical and predictive work). The second axis is the target of inference: explanation and counterfactual prediction on one side, out-of-sample forecasting on the other. Figure 30.1 locates the traditions, and the table that follows gives the formal contrast.
| Dimension | Analytical | Empirical (structural) | Empirical (reduced-form) | Predictive (ML) |
|---|---|---|---|---|
| Core question | What should happen | Why does it happen | What is the effect | What happens next |
| Canonical object | Game \((N, S, u)\) | Likelihood from optimization | Conditional mean \(\mathbb{E}[y\mid \mathbf{x}]\) | Mapping \(\hat f:\mathbf{x}\mapsto \hat y\) |
| “Solution” concept | Equilibrium | (Quasi-)MLE / GMM | OLS / IV | Empirical-risk minimization |
| Data role | None (or calibration) | Disciplines parameters | Identifies an effect | Trains and tests \(\hat f\) |
| Key virtue | Logical transparency | Counterfactual validity | Credible causal estimate | Predictive accuracy |
| Fatal assumption | Wrong game / behavior | Misspecified structure | Confounding / weak instrument | Distribution shift |
| Generalizes via | Theory | Invariant parameters | Design (exogenous variation) | i.i.d. test data |
The remaining sections take each tradition in turn. The reader should keep Table 30.1 in view: nearly every methodological decision in the rest of the book is a choice among its columns.
30.3 Analytical Models
An analytical model derives marketing outcomes as the equilibrium of a game played by rational agents. It contains no data. Its purpose is to establish, with the certainty of a proof, that under stated assumptions a particular force operates in a particular direction—that, say, increased competition lowers equilibrium price, or that informative advertising can support a separating equilibrium in which only high-quality firms advertise. The payoff is a comparative static: the sign, and sometimes the magnitude, of the derivative of an equilibrium outcome with respect to a primitive.
Formally, a game is a triple \((N, \{S_i\}_{i \in N}, \{u_i\}_{i \in N})\) where \(N\) is the set of players (firms, consumers), \(S_i\) is player \(i\)’s strategy set (prices, advertising levels, qualities), and \(u_i: \prod_j S_j \to \mathbb{R}\) is \(i\)’s payoff as a function of the chosen profile. A pure-strategy Nash equilibrium (Nash 1950) is a profile \(\mathbf{s}^\* = (s_1^\*, \dots, s_n^\*)\) such that no player can improve unilaterally,
\[ u_i(s_i^\*, \mathbf{s}_{-i}^\*) \;\ge\; u_i(s_i, \mathbf{s}_{-i}^\*) \qquad \text{for all } s_i \in S_i,\ \text{all } i \in N, \tag{30.1}\]
where \(\mathbf{s}_{-i}\) denotes the strategies of \(i\)’s rivals. The equilibrium is the solution concept: it plays the role that an estimator plays in empirical work. To solve the model is to characterize \(\mathbf{s}^\*\) and then differentiate it with respect to a primitive to obtain the comparative static of interest.
A concrete instance fixes the idea. Consider two firms choosing prices for horizontally differentiated products located at the ends of a unit interval, with consumers uniformly distributed and incurring a quadratic “transport” cost \(t\) for distance from their preferred variety—the Hotelling spatial-competition setup (Hotelling 1929; Dixit 1980). Each firm \(i\) has constant marginal cost \(c\) and earns \(\pi_i = (p_i - c)\, D_i(p_i, p_j)\), where the demand \(D_i\) is the mass of consumers who prefer firm \(i\) at the posted prices. Solving the first-order conditions and imposing symmetry yields the equilibrium price
\[ p^\* = c + t, \tag{30.2}\]
a closed-form result whose comparative statics are immediate and exact: \(\partial p^\* / \partial t > 0\) and \(\partial p^\* / \partial c = 1\). Greater differentiation (larger \(t\)) softens competition and raises equilibrium margins; the result holds for every admissible parameter value, with no data and no estimation. This is the characteristic strength of the analytical tradition—universally valid, sign-definite conclusions—and its characteristic limitation: the conclusion is true of the model, and whether the model resembles any real market is a separate question the framework cannot answer. Marketing’s analytical literature has used this machinery to study product line design, channel structure, advertising, and pricing (Moorthy 1985, 1988, 1993; Narasimhan 1988; Villas-Boas and Winer 1999; Iyer, Soberman, and Villas-Boas 2005), and signaling games to rationalize how brands credibly convey unobservable quality (Spence 1973; Akerlof 1970; Nelson 1974).
The assumption whose failure is fatal to an analytical model is the specification of the game itself: the players’ rationality, the information they hold, the order of moves, and the equilibrium concept invoked. If real firms do not best-respond, or hold different beliefs, or the relevant equilibrium is not the one selected, the comparative statics need not describe the world even approximately. Analytical models therefore earn their keep as theories—sources of sharp, falsifiable hypotheses and of the behavioral content that disciplines structural estimation—rather than as descriptions to be read off as fact. Chapter 31 develops the solution techniques and Chapter 39 the equilibrium models of competition that structural work later takes to data.
30.4 Empirical Models
An empirical model confronts structure with data to recover parameters. The tradition splits along the policy-invariance line introduced above, and the split is the single most important methodological distinction in the chapter.
30.4.1 The Reduced-Form Branch
A reduced-form model specifies a response function relating an outcome to marketing and control variables, and estimates it without deriving that function from an explicit optimization problem. The workhorse is the regression of a sales or choice outcome \(y\) on marketing inputs \(\mathbf{x}\) and controls \(\mathbf{w}\),
\[ y_i = \mathbf{x}_i^\top \boldsymbol{\beta} + \mathbf{w}_i^\top \boldsymbol{\gamma} + \varepsilon_i, \tag{30.3}\]
estimated by ordinary least squares (OLS) when the regressors are exogenous, or by instrumental variables (IV) / two-stage least squares when they are not (Greene 2003; Cameron and Trivedi 2005). The marketing-mix and response-modeling literatures are largely reduced-form in this sense (Little 1970, 1976; Lambin 1970; Clarke 1976; Lilien et al. 1992), and Chapter 53 treats them at length.
The defining hazard of Equation 30.3 is endogeneity: a regressor correlated with the error makes the OLS estimator inconsistent, \(\mathbb{E}[\hat{\boldsymbol\beta}] \neq \boldsymbol\beta\). Price is the canonical offender—firms set higher prices where they observe (but the analyst does not) higher demand, so \(\mathrm{Cov}(p_i, \varepsilon_i) > 0\) and the naive price coefficient is biased toward zero, sometimes to the point of the wrong sign. The remedies are an instrument \(\mathbf{z}\) that shifts \(\mathbf{x}\) without entering the outcome directly, or a control function that conditions on the endogenous portion of the regressor (Petrin and Train 2010), or a research design that isolates exogenous variation. The marketing literature has matured on exactly this point: modal recent work addresses both sample selection and endogeneity rather than one in isolation (Frennea, Han, and Mittal 2018), and the rise of quasi-experimental designs—exploiting borders, weather, regulatory changes, and abrupt firm-policy shifts—gives reduced-form estimates a credible causal warrant (Goldfarb, Tucker, and Wang 2022). A reduced-form estimate, properly identified, answers “what is the effect of this intervention, here.” It does not, in general, answer “what would happen under a policy we have never observed,” because the estimated \(\boldsymbol\beta\) may itself be a function of the prevailing policy regime—it is policy-variant.
Reduced-form does not mean atheoretic
A reduced-form model can be every bit as causally credible as a structural one when its identifying variation is genuinely exogenous; “reduced-form” is a statement about where the estimating equation comes from, not about rigor. What reduced-form estimates lack is portability: their coefficients are valid for counterfactuals spanned by the variation that identified them, and silent about the rest.
30.4.2 The Structural Branch
A structural model imposes the optimization problem itself and estimates the parameters of its primitives—utilities and costs—rather than a response function. Following Tülin Erdem and Keane (1996), structural estimation “assumes and imposes a structure on the consumer’s maximization problem,” and the recovered parameters are those of the consumers’ utility functions, which are taken to be policy-invariant. Because the primitives do not move when the policy moves, the estimated model can be simulated under counterfactual policies the data never contained—the property that justifies the additional assumptions structural work requires (Reiss 2011; P. Chintagunta et al. 2006; Dubé et al. 2002).
The canonical structural object in marketing is the random-utility discrete-choice model (McFadden 1986, 2001). Consumer \(i\) obtains utility
\[ u_{ij} = \mathbf{x}_j^\top \boldsymbol{\beta}_i - \alpha_i p_j + \xi_j + \varepsilon_{ij} \tag{30.4}\]
from alternative \(j\), where \(\mathbf{x}_j\) are observed attributes, \(p_j\) is price, \(\xi_j\) is a demand shock observed by firms and consumers but not the analyst, and \(\varepsilon_{ij}\) is idiosyncratic taste. Under a Type-I extreme-value assumption on \(\varepsilon_{ij}\) the choice probabilities take the multinomial-logit form, and with random coefficients \((\boldsymbol\beta_i, \alpha_i)\) integrated over a population distribution one obtains the mixed-logit / BLP demand system estimated by maximum likelihood or GMM (Guadagni and Little 1983; S. T. Berry 1994; S. Berry, Levinsohn, and Pakes 1995; Rossi 2014; Dubé, Hortaçsu, and Joo 2021). The structural-marketing program has extended Equation 30.4 to consumers who learn about product quality through experience and forward-looking dynamics (Tülin Erdem and Keane 1996; Tulin Erdem 1998; Tülin Erdem, Keane, and Sun 2008; P. K. Chintagunta and Haldar 1998; Seetharaman, Ainslie, and Chintagunta 1999; Allenby, Leone, and Jen 1999; Gilbride and Allenby 2004; Bradlow, Hu, and Ho 2004; Sun and Shibo 2011), to firm conduct and equilibrium pricing (Kadiyali, Sudhir, and Rao 2001; Misra and Nair 2011), and to dynamic competition (Gowrisankaran and Rysman 2012; Borkovsky et al. 2017). The estimation machinery—mixed logit, hierarchical Bayes, GMM—is the subject of Chapter 34 and Chapter 41.
The crucial contrast between the two structural choice models in marketing is whether the consumer is myopic or forward-looking. A myopic model has the consumer choose to maximize current-period utility given current beliefs; a forward-looking model has the consumer choose to maximize the expected present value of utility, internalizing how today’s choice affects tomorrow’s information and options—relevant precisely when learning about uncertain product characteristics drives the choice process (Tülin Erdem and Keane 1996). The two can rationalize the same observed choices yet imply sharply different responses to, for example, a temporary promotion, which a forward-looking consumer partly anticipates.
The assumption whose failure is fatal to a structural model is specification: if the imposed utility, the error distribution, or the conduct assumption is wrong, the estimated “primitives” absorb the misspecification and the counterfactuals built on them are unreliable. Structural work thus buys counterfactual portability at the price of strong, and contestable, functional-form and behavioral assumptions—the exact trade reduced-form work declines to make. Neither dominates; the right choice depends on the question, as Section 30.6 argues.
30.5 Predictive Models
A predictive model learns a flexible mapping from features to an outcome and is judged by how well it forecasts data it has not seen. It makes minimal commitments about the mechanism that generated the data; in exchange it can fit relationships far richer than the linear-in-parameters forms of Equation 30.3, and it scales to the high-dimensional, unstructured data—text, images, clickstreams—that now pervade marketing (Wedel and Kannan 2016; Athey, Catalini, and Tucker 2017). The aim is not to estimate \(\boldsymbol\beta\) but to construct an \(\hat f\) that minimizes expected loss on future inputs.
Formally, given a loss \(L(y, \hat y)\) and a joint distribution \(\mathcal{P}\) over \((\mathbf{x}, y)\), the target is the function minimizing risk, the expected loss,
\[ f^\* = \arg\min_{f \in \mathcal{F}} \; \mathbb{E}_{(\mathbf{x}, y)\sim\mathcal{P}} \big[\, L(y, f(\mathbf{x})) \,\big]. \tag{30.5}\]
Because \(\mathcal{P}\) is unknown, the model minimizes empirical risk—average loss on a training sample—typically with a penalty that controls complexity,
\[ \hat f = \arg\min_{f \in \mathcal{F}} \; \frac{1}{n}\sum_{i=1}^{n} L\big(y_i, f(\mathbf{x}_i)\big) \;+\; \lambda\, \Omega(f), \tag{30.6}\]
where \(\Omega(f)\) penalizes complexity (e.g., the \(\ell_1\) norm of coefficients in the lasso, tree depth in a gradient-boosted ensemble) and \(\lambda \ge 0\) is a regularization strength chosen by cross-validation. The whole apparatus exists to manage the bias–variance trade-off: a model too simple to capture the signal is biased; one flexible enough to memorize the training noise has high variance and generalizes poorly. The discipline of the tradition is that performance is reported on a held-out test set, never on the data used to fit—so the relevant guarantee is i.i.d. generalization, not parameter recovery. Marketing applications span churn and response scoring, recommendation, and the extraction of structured signal from text and images for downstream marketing use (Netzer et al. 2008; Toubia and Stephen 2013; Ansari et al. 2018; Tirunillai and Tellis 2014; Büschken and Allenby 2016; Hauser, Tellis, and Griffin 2006); Chapter 65 develops the methods and Chapter 43 their application to language.
The assumption whose failure is fatal to a predictive model is distributional stability: Equation 30.5 presumes that future \((\mathbf{x}, y)\) are drawn from the same \(\mathcal{P}\) as the training data. When the deployment environment shifts—a new competitor enters, a price moves outside its historical range, a recommendation changes the very behavior it predicts—the i.i.d. assumption breaks and accuracy can collapse without warning. This failure mode is the deep reason a predictive model, however accurate, does not license a causal or counterfactual claim. A model that predicts churn superbly from a feature that includes a retention discount cannot be used to evaluate changing that discount: intervening on the feature changes the distribution, and Equation 30.5 no longer applies. The point generalizes—predictive accuracy and causal validity are distinct objectives, and optimizing one does not deliver the other (Athey, Catalini, and Tucker 2017).
Prediction policy versus parameter
A useful test of which tradition a problem belongs to: ask whether the decision requires knowing the effect of an action you will take, or only an accurate forecast given actions already in train. Choosing whom to target with a retention offer, conditional on the offer policy, is a prediction problem. Choosing whether to change the offer policy is a causal problem and needs an empirical or analytical model. Many marketing tasks are prediction-policy problems—who, not whether—and for those a predictive model is exactly right.
30.6 Choosing a Modeling Approach
The traditions are complements, not rivals, and mature marketing science routinely chains them: theory generates a hypothesis (analytical), an estimate tests it and quantifies a counterfactual (empirical), and a predictive model operationalizes the resulting policy at scale. The choice for a given study turns on three questions, asked in order.
First, what must the answer support? If the deliverable is a counterfactual—what would profits be under a price the firm has never charged, a product it has never launched, a market structure that does not yet exist—the model must recover policy-invariant primitives, which points to structural estimation (Section 30.4) disciplined by analytical theory (Section 30.3). If the deliverable is the causal effect of an intervention within the observed regime, a credibly identified reduced-form design suffices and is usually preferable for its weaker assumptions (Goldfarb, Tucker, and Wang 2022). If the deliverable is an accurate forecast or a ranking conditional on a fixed policy, a predictive model is the right tool (Section 30.5).
Second, where does identifying variation come from? A causal or structural answer requires variation in the cause that is independent of the unobserved drivers of the outcome—an instrument, a natural experiment, or a randomized trial. Absent such variation, no amount of modeling sophistication recovers an effect; the honest move is to retreat to prediction and say so. Predictive models, by contrast, need only that the future resemble the past, a much weaker and more often satisfied condition.
Third, what is the cost of a wrong assumption? Structural models deliver the most but assume the most, and a misspecified structure quietly corrupts every counterfactual built on it. Reduced-form designs assume less and fail more visibly. Predictive models assume least about mechanism but stake everything on distributional stability. The right model is the one whose fatal assumption (the last column of Table 30.1) is the most defensible in the application at hand—not the most sophisticated available.
30.7 Illustration: One Question, Three Lenses
To make the contrast concrete, consider a single question—how does a price change affect demand?—and run it through all three traditions on a common simulated market. The exercise is deliberately small and fully reproducible; its point is conceptual, not computational. We simulate a market in which firms set prices partly in response to an unobserved demand shock, creating textbook price endogeneity, and then show how each tradition treats it.
Code
set.seed(2025)
n <- 4000
# Unobserved (to the analyst) demand shifter xi; firms see it and price on it.
xi <- rnorm(n, 0, 1)
cost <- rnorm(n, 2, 0.4) # exogenous cost shifter (a valid instrument)
# Pricing rule: price rises with both cost and the latent demand shock -> endogeneity
price <- 1.5 + 0.8 * cost + 0.6 * xi + rnorm(n, 0, 0.2)
# True structural demand: utility decreasing in price, increasing in xi.
alpha_true <- -1.2 # the policy-invariant price coefficient
util <- 3 + alpha_true * price + xi
prob <- plogis(util) # logit choice probability
buy <- rbinom(n, 1, prob)
dat <- data.frame(buy, price, cost, xi)
round(cor(dat[, c("price", "cost", "xi")]), 2)
#> price cost xi
#> price 1.00 0.46 0.85
#> cost 0.46 1.00 0.01
#> xi 0.85 0.01 1.00The correlation matrix confirms the design: price is correlated with the unobserved xi, so any model that regresses purchase on price without addressing this correlation will recover a biased price effect. Now the three lenses.
Predictive lens. Fit a flexible classifier to predict buy from price and cost, and evaluate it out of sample. The model is excellent at forecasting and says nothing trustworthy about the effect of moving price.
Code
train <- 1:3000; test <- 3001:n
pred_fit <- glm(buy ~ poly(price, 2) + cost, data = dat[train, ],
family = binomial)
phat <- predict(pred_fit, newdata = dat[test, ], type = "response")
# Out-of-sample area under the ROC curve (rank accuracy), computed from scratch.
pos <- phat[dat$buy[test] == 1]; neg <- phat[dat$buy[test] == 0]
auc <- mean(outer(pos, neg, ">")) + 0.5 * mean(outer(pos, neg, "=="))
cat("Out-of-sample AUC:", round(auc, 3), "\n")
#> Out-of-sample AUC: 0.602Reduced-form (naive) lens. A logit of purchase on price recovers a coefficient that is biased toward zero, because price proxies for the omitted positive demand shock xi.
Structural / IV lens. Exploit the exogenous cost shifter as an instrument via a control function: regress the endogenous price on the instrument, take the residual, and include it in the choice model to absorb the correlation with xi (Petrin and Train 2010). The price coefficient now recovers the policy-invariant primitive and supports counterfactual repricing.
Code
first <- lm(price ~ cost, data = dat) # first stage on the instrument
dat$cfv <- resid(first) # control-function residual
struct <- glm(buy ~ price + cfv, data = dat, family = binomial)
cat("Control-function price coefficient:",
round(coef(struct)["price"], 3), " (true:", alpha_true, ")\n")
#> Control-function price coefficient: -0.993 (true: -1.2 )The three numbers tell the chapter’s whole story. The predictive model attains high out-of-sample accuracy yet its fitted price terms are not the demand slope and must not be read as such. The naive reduced-form coefficient is attenuated by endogeneity, a quantitative bias that no added flexibility would cure—only identifying variation does. The control-function estimate, using the exogenous cost instrument, recovers the true policy-invariant price coefficient and so, alone among the three, can answer the counterfactual “what happens to demand if the firm reprices.” Same data, same question, three answers—because the three models target different things and rest on different assumptions. Choosing among them is the skill this part of the book builds.
30.8 Key Takeaways
- A model is an argument about a marketing system, valued for the gap between the assumptions it imposes and the conclusions it then licenses; the traditions differ in what they hold fixed when the environment changes (Section 30.2).
- Analytical models derive equilibrium comparative statics from stated games (Nash 1950); they yield sign-definite, universally valid conclusions about the model, and are fatal to mis-specify the game (Section 30.3).
- Empirical models split on policy invariance: structural estimation recovers invariant utility and cost primitives and so supports counterfactuals (Tülin Erdem and Keane 1996; McFadden 1986), whereas reduced-form work estimates a possibly policy-variant response function and supports causal claims only under its identifying design (Goldfarb, Tucker, and Wang 2022; Frennea, Han, and Mittal 2018) (Section 30.4).
- Predictive models minimize out-of-sample risk over a flexible function class (Wedel and Kannan 2016; Athey, Catalini, and Tucker 2017); they excel at forecasting under a fixed policy but collapse under distribution shift and never, by themselves, license a causal claim (Section 30.5).
- The traditions are complements: choose by what the answer must support, where identifying variation comes from, and which fatal assumption is most defensible (Section 30.6); the same pricing question yields three different, internally valid answers depending on the lens (Section 30.7).
Akerlof, George A. 1970. “The Market for "Lemons": Quality Uncertainty and the Market Mechanism.” The Quarterly Journal of Economics 84 (3): 488. https://doi.org/10.2307/1879431.
Allenby, Greg M, Robert P Leone, and Lichung Jen. 1999. “A Dynamic Model of Purchase Timing with Application to Direct Marketing.” Journal of the American Statistical Association 94 (446): 365–74.
Ansari, Asim, Florian Stahl, Mark Heitmann, and Lucas Bremer. 2018. “Building a Social Network for Success.” Journal of Marketing Research 55 (3): 321–38. https://doi.org/10.1509/jmr.12.0417.
Athey, Susan, Christian Catalini, and Catherine Tucker. 2017. “The Digital Privacy Paradox: Small Money, Small Costs, Small Talk.” National Bureau of Economic Research.
Berry, Steven T. 1994. “Estimating Discrete-Choice Models of Product Differentiation.” The RAND Journal of Economics, 242–62.
Berry, Steven, James Levinsohn, and Ariel Pakes. 1995. “Automobile Prices in Market Equilibrium.” Econometrica 63 (4): 841. https://doi.org/10.2307/2171802.
Borkovsky, Ron N., Avi Goldfarb, Avery M. Haviv, and Sridhar Moorthy. 2017. “Measuring and Understanding Brand Value in a Dynamic Model of Brand Management.” Marketing Science 36 (4): 471–99. https://doi.org/10.1287/mksc.2016.1020.
Bradlow, Eric T, Ye Hu, and Teck-Hua Ho. 2004. “A Learning-Based Model for Imputing Missing Levels in Partial Conjoint Profiles.” Journal of Marketing Research 41 (4): 369–81.
Büschken, Joachim, and Greg M Allenby. 2016. “Sentence-Based Text Analysis for Customer Reviews.” Marketing Science 35 (6): 953–75.
Cameron, A Colin, and Pravin K Trivedi. 2005. Microeconometrics: Methods and Applications. Cambridge university press.
Chintagunta, Pradeep K., and Sudeep Haldar. 1998. “Investigating Purchase Timing Behavior in Two Related Product Categories.” Journal of Marketing Research 35 (1): 43. https://doi.org/10.2307/3151929.
Chintagunta, Pradeep, Tülin Erdem, Peter E Rossi, and Michel Wedel. 2006. “Structural Modeling in Marketing: Review and Assessment.” Marketing Science 25 (6): 604–16.
Clarke, Darral G. 1976. “Econometric Measurement of the Duration of Advertising Effect on Sales.” Journal of Marketing Research 13 (4): 345. https://doi.org/10.2307/3151017.
Dixit, Avinash. 1980. “The Role of Investment in Entry-Deterrence.” The Economic Journal 90 (357): 95. https://doi.org/10.2307/2231658.
Dubé, Jean-Pierre, Pradeep Chintagunta, Amil Petrin, Bart Bronnenberg, Ron Goettler, PB Seetharaman, K Sudhir, Raphael Thomadsen, and Ying Zhao. 2002. “Structural Applications of the Discrete Choice Model.” Marketing Letters 13: 207–20.
Dubé, Jean-Pierre, Ali Hortaçsu, and Joonhwi Joo. 2021. “Random-Coefficients Logit Demand Estimation with Zero-Valued Market Shares.” Marketing Science 40 (4): 637–60.
Erdem, Tulin. 1998. “An Empirical Analysis of Umbrella Branding.” Journal of Marketing Research 35 (3): 339. https://doi.org/10.2307/3152032.
Erdem, Tülin, and Michael P. Keane. 1996. “Decision-Making Under Uncertainty: Capturing Dynamic Brand Choice Processes in Turbulent Consumer Goods Markets.” Marketing Science 15 (1): 1–20. https://doi.org/10.1287/mksc.15.1.1.
Erdem, Tülin, Michael P Keane, and Baohong Sun. 2008. “The Impact of Advertising on Consumer Price Sensitivity in Experience Goods Markets.” Quantitative Marketing and Economics 6: 139–76.
Frennea, Carly, Kyuhong Han, and Vikas Mittal. 2018. “Value Appropriation and Firm Shareholder Value: Role of Advertising and Receivables Management.” Journal of Marketing Research, February. https://doi.org/10.1509/jmr.17.0123.
Gilbride, Timothy J., and Greg M. Allenby. 2004. “A Choice Model with Conjunctive, Disjunctive, and Compensatory Screening Rules.” Marketing Science 23 (3): 391–406. https://doi.org/10.1287/mksc.1030.0032.
Goldfarb, Avi, Catherine Tucker, and Yanwen Wang. 2022. “Conducting Research in Marketing with Quasi-Experiments.” Journal of Marketing 86 (3): 1–20. https://doi.org/10.1177/00222429221082977.
Gowrisankaran, Gautam, and Marc Rysman. 2012. “Dynamics of Consumer Demand for New Durable Goods.” Journal of Political Economy 120 (6): 1173–1219.
Greene, William H. 2003. Econometric Analysis. Pearson Education India.
Guadagni, Peter M., and John D. C. Little. 1983. “A Logit Model of Brand Choice Calibrated on Scanner Data.” Marketing Science 2 (3): 203–38. https://doi.org/10.1287/mksc.2.3.203.
Hauser, John, Gerard J. Tellis, and Abbie Griffin. 2006. “Research on Innovation: A Review and Agenda forMarketing Science.” Marketing Science 25 (6): 687–717. https://doi.org/10.1287/mksc.1050.0144.
Hotelling, Harold. 1929. “Stability in Competition.” The Economic Journal 39 (153): 41. https://doi.org/10.2307/2224214.
Iyer, Ganesh, David Soberman, and J. Miguel Villas-Boas. 2005. “The Targeting of Advertising.” Marketing Science 24 (3): 461–76. https://doi.org/10.1287/mksc.1050.0117.
Kadiyali, Vrinda, K Sudhir, and Vithala R Rao. 2001. “Structural Analysis of Competitive Behavior: New Empirical Industrial Organization Methods in Marketing.” International Journal of Research in Marketing 18 (1-2): 161–86.
Lambin, Jean-Jacques. 1970. “Optimal Allocation of Competitive Marketing Efforts: An Empirical Study.” The Journal of Business 43 (4): 468. https://doi.org/10.1086/295310.
Lilien, Gary L, Philip Kotler, K Sridhar Moorthy, et al. 1992. Marketing Models. Vol. 803. Prentice-Hall Englewood Cliffs, NJ.
Little, John D. C. 1970. “Models and Managers: The Concept of a Decision Calculus.” Management Science 16 (8): B-466-B-485. https://doi.org/10.1287/mnsc.16.8.b466.
———. 1976. “Models and Managers: The Concept of a Decision Calculus.” In Lecture Notes in Economics and Mathematical Systems, 208–9. Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-51565-1_67.
McFadden, Daniel. 1986. “The Choice Theory Approach to Market Research.” Marketing Science 5 (4): 275–97.
———. 2001. “Economic Choices.” American Economic Review 91 (3): 351–78.
Misra, Sanjog, and Harikesh S Nair. 2011. “A Structural Model of Sales-Force Compensation Dynamics: Estimation and Field Implementation.” Quantitative Marketing and Economics 9: 211–57.
Moorthy, K. Sridhar. 1985. “Using Game Theory to Model Competition.” Journal of Marketing Research 22 (3): 262. https://doi.org/10.2307/3151424.
———. 1988. “Product and Price Competition in a Duopoly.” Marketing Science 7 (2): 141–68. https://doi.org/10.1287/mksc.7.2.141.
———. 1993. “Theoretical Modeling in Marketing.” Journal of Marketing 57 (2): 92. https://doi.org/10.2307/1252029.
Narasimhan, Chakravarthi. 1988. “Competitive Promotional Strategies.” The Journal of Business 61 (4): 427. https://doi.org/10.1086/296442.
Nash, John F. 1950. “The Bargaining Problem.” Econometrica 18 (2): 155. https://doi.org/10.2307/1907266.
Nelson, Phillip. 1974. “Advertising as Information.” Journal of Political Economy 82 (4): 729–54. https://doi.org/10.1086/260231.
Netzer, Oded, Olivier Toubia, Eric T. Bradlow, Ely Dahan, Theodoros Evgeniou, Fred M. Feinberg, Eleanor M. Feit, et al. 2008. “Beyond Conjoint Analysis: Advances in Preference Measurement.” Marketing Letters 19 (3-4): 337–54. https://doi.org/10.1007/s11002-008-9046-1.
Petrin, Amil, and Kenneth Train. 2010. “A Control Function Approach to Endogeneity in Consumer Choice Models.” Journal of Marketing Research 47 (1): 3–13. https://doi.org/10.1509/jmkr.47.1.3.
Reiss, Peter C. 2011. “Structural Workshop PaperDescriptive, Structural, and Experimental Empirical Methods in Marketing Research.” Marketing Science 30 (6): 950–64. https://doi.org/10.1287/mksc.1110.0681.
Rossi, Peter E. 2014. “Invited PaperEven the Rich Can Make Themselves Poor: A Critical Examination of IV Methods in Marketing Applications.” Marketing Science 33 (5): 655–72. https://doi.org/10.1287/mksc.2014.0860.
Seetharaman, P. B., Andrew Ainslie, and Pradeep K. Chintagunta. 1999. “Investigating Household State Dependence Effects Across Categories.” Journal of Marketing Research 36 (4): 488. https://doi.org/10.2307/3152002.
Spence, Michael. 1973. “Job Market Signaling.” The Quarterly Journal of Economics 87 (3): 355. https://doi.org/10.2307/1882010.
Sun, Baohong, and Li Shibo. 2011. “Learning and Acting on Customer Information: A Simulation-Based Demonstration on Service Allocations with Offshore Centers.” Journal of Marketing Research 48 (1): 72–86.
Tirunillai, Seshadri, and Gerard J. Tellis. 2014. “Mining Marketing Meaning from Online Chatter: Strategic Brand Analysis of Big Data Using Latent Dirichlet Allocation.” Journal of Marketing Research 51 (4): 463–79. https://doi.org/10.1509/jmr.12.0106.
Toubia, Olivier, and Andrew T. Stephen. 2013. “Intrinsic Vs. Image-Related Utility in Social Media: Why Do People Contribute Content to Twitter?” Marketing Science 32 (3): 368–92. https://doi.org/10.1287/mksc.2013.0773.
Villas-Boas, J. Miguel, and Russell S. Winer. 1999. “Endogeneity in Brand Choice Models.” Management Science 45 (10): 1324–38. https://doi.org/10.1287/mnsc.45.10.1324.
Wedel, Michel, and P.K. Kannan. 2016. “Marketing Analytics for Data-Rich Environments.” Journal of Marketing 80 (6): 97–121. https://doi.org/10.1509/jm.15.0413.