RESEARCH DOSSIER

Structural Models of Differentiated Product Competition

Abstract. This research dossier examines the economic and econometric foundations of differentiated-product competition, with particular attention to demand inversion, price endogeneity, instrument validity, supply inversion, and merger counterfactuals. It derives homogeneous and heterogeneous discrete-choice demand, multiproduct Bertrand pricing, diversion, and welfare measures; compares alternative identifying information and computational methods; and considers evidence from automobile, cereal, beer, and European truck markets. A reproducible synthetic three-product example isolates the effects of ownership and marginal-cost efficiencies. The analytical synthesis establishes how the evidentiary requirements for consumer substitution, conduct, costs, and counterfactual invariance interact. The study is an interpretive and formal synthesis of existing research, not a new empirical estimation or a new named economic theory. Its empirical conclusions retain the scope of the cited studies.

Keywords: Differentiated products; industrial organization; structural econometrics; discrete choice; random coefficients; BLP; instrumental variables; market power; merger simulation; welfare; oligopoly; identification

Research object and economic primitives

Differentiated-product competition is an interaction among consumers who select among alternatives, producers who control sets of alternatives, and institutions that determine the feasible transactions. Structural analysis describes these mechanisms explicitly and relates them to observed market outcomes through restrictions that can be investigated with data. The central empirical task is to recover substitution, marginal costs, and conduct sufficiently well to evaluate a specified intervention.

A market is indexed by t; its alternatives by j in {0, 1, ..., J_t}, with 0 denoting the outside alternative. A product is a well-defined bundle of attributes and terms offered in a particular place and period. The relevant choice-maker might be an individual, household, procurement officer, or intermediary. Market size M_t counts eligible choice occasions in the period; q_jt is the number of units purchased on those occasions; s_jt = q_jt/M_t. A single-unit model requires each occasion to have one mutually exclusive choice, including abstention. Repeated purchases can be modeled as distinct occasions only when the relevant dependence and inventory dynamics are appropriately addressed.

Differentiation may be horizontal, reflecting disagreement over attributes, or vertical, reflecting shared rankings with varying valuations. A single product usually combines both. Consumer income, geography, search costs and existing ownership alter the feasible and preferred choice sets. Observed product shares record the outcome of these interacting factors, alongside prices. Price is simultaneously a consumer cost and a strategic decision variable. The firm that controls a product cares about sales of that product and, under joint pricing, sales of its other products.

The term structural identifies the maintained mapping between primitives, choices and equilibrium outcomes. The application of that term demands transparent specification of what changes and what remains invariant under intervention. A descriptive regression estimating the association between price and sales can serve an important evidentiary role, but is insufficient by itself to infer the effect of reallocating product ownership. The intervention changes firms' incentives and potentially the behavior that generated the original price variation.

The dossier addresses static single-unit demand and multiproduct oligopoly as a core benchmark. It examines departures in conduct, dynamic use, availability and bargaining through explicit scope conditions. It studies mechanisms and research designs, rather than offering a fresh econometric estimate of any national industry. The numerical example is synthetic. The international case evidence is drawn from identified published studies and their research populations.

Historical development

The intellectual transition from homogeneous-good supply and demand to product-level differentiated oligopoly required the simultaneous treatment of many substitution channels, product characteristics and unobserved demand determinants. When each cross-price term in a demand system is independently parameterized, the dimensionality grows rapidly with the product count. A characteristic-space model economizes on parameters by relating preferences across products through a shared distribution of consumer types.

Berry's 1994 discrete-choice estimation study formulated inversion of observed shares into utilities as a route to estimating demand with endogenous prices. The intellectual importance of the inversion is the emergence of an economically interpretable product-level residual. Once the observed share is connected to a mean utility, the difference between that utility and the measured components can enter instrumental-variable restrictions. The same algebra that reproduces shares for a maintained parameter vector also makes visible the price endogeneity problem.

Berry, Levinsohn, and Pakes (1995), in Automobile Prices in Market Equilibrium, linked distributions of preferences to an equilibrium model of US automobile prices and sales. The resulting approach, conventionally called BLP, organizes demand and supply around common primitives rather than estimating an unconstrained matrix of cross-price effects. Its influence derives from the ability to recover one coherent demand system and use it for economically distinct tasks, including markups and changes in product offerings.

A parallel line of work turned tractable logit substitution into a practical benchmark for merger exercises. Werden and Froeb's 1994 logit merger study is an early contribution to that line. The modern research agenda subsequently extended the available information through household data, purchase moments and richer instruments. These developments respond to distinct problems: identifying the price coefficient, identifying heterogeneous substitution, and solving the computational problem reliably. Treating these as one task obscures the consequences of improvement in any single dimension.

The history also establishes an important asymmetry. Inverting demand can reveal the residual that rationalizes observed shares conditional on a parameter vector. Recovering supply requires an additional proposition about strategic conduct. Later conduct comparisons make the distinction consequential: a demand system can remain appropriate while the assumptions governing firm incentives change. Structural research is therefore best understood as the development of a connected series of empirically defensible arguments.

Market definition, quantities, and measurement

The empirical market must correspond to an economically plausible choice occasion. A monthly metropolitan grocery category, a yearly national new-car market, and an enterprise tender market generate different demand objects. A geography should follow the actual boundaries of substitution and delivery rather than the convenience of administrative reporting. The time interval should reflect the decision process; monthly scanner transactions may include inventory accumulation, while annual durable-goods purchases include postponed replacement.

The outside share is s_0t = 1 - sum_j s_jt. Its magnitude determines the fraction of consumers whose response takes the form of leaving the modeled category. It must be supported by a population definition or an empirically defensible category choice rather than chosen to obtain plausible elasticities. A change in denominator modifies the utility normalization and may modify estimated slopes through cross-market differences. Consequently a sensitivity exercise using households, eligible buyers, or purchase occasions is part of measurement QA.

If product sales are aggregated across outlets, prices should represent the terms associated with the relevant quantities. An average transaction price is a potentially endogenous composition-weighted measure: a change in customer composition or promotion can alter the average without a uniform price change. List prices omit negotiated rebates; nominal prices conceal inflation; taxes may be included differently on demand and supply sides. Convert prices to comparable units within market and define seller net revenue separately from what buyers pay. For multipacks, distinguish item purchases, packs, and physical volume.

Products need stable identifiers linked to ownership, availability, characteristics and time. A rebranding may represent a new SKU while preserving the economic alternative. Conversely a model with unchanged identifier may receive a substantial quality upgrade. Reconstruct an ownership map at the actual decision-making level, allowing common pricing control to differ from brand ownership or corporate holdings. Document changes in distribution arrangements and tax treatment. Missing sales are not identical to zero sales, and a zero share may reflect censoring, unavailable inventory, or a genuinely rare option.

The minimum archive for an empirical study contains product-market data, units, dates, construction of prices and shares, identification of source datasets and access rights, market-size logic, ownership histories and a missing-data audit. No real microdata are used here; the distinction between desirable documentation and data actually retrieved is maintained throughout.

Logit demand and share inversion

Let consumer i in market t have utility u_ijt = x_jt' beta - alpha p_jt + xi_jt + epsilon_ijt. The parameter alpha is positive and denominated in inverse price units; xi_jt is a product-market characteristic observed by market participants but omitted from the measured attributes. Normalize outside systematic utility to zero, and assume the epsilon terms are independent type-I extreme value with scale one. Write delta_jt = x_jt' beta - alpha p_jt + xi_jt.

Conditional choice probabilities, and therefore population shares under homogeneous coefficients, are

sj=exp⁡(δj)1+∑k=1Jexp⁡(δk),s0=11+∑k=1Jexp⁡(δk).s_j = \frac{\exp(\delta_j)}{1+\sum_{k=1}^{J}\exp(\delta_k)},\qquad s_0=\frac{1}{1+\sum_{k=1}^{J}\exp(\delta_k)}.

The ratio s_j/s_0 equals exp(delta_j). Thus delta_j = log(s_j) - log(s_0). This inverse requires strictly positive shares and an outside share greater than zero. Berry (1994) situates this share inversion method within differentiated oligopoly estimation.

At a fixed positive alpha, differentiation yields the Jacobian D_jk = partial s_j / partial p_k:

Djk=−αsj(1{j=k}−sk).D_{jk}=-\alpha s_j(\mathbf 1\{j=k\}-s_k).

Own-price effects are negative; cross-price effects among available inside products are positive. The price elasticity of demand for j with respect to p_k is D_jk p_k/s_j. The own elasticity is -alpha p_j(1-s_j); the cross elasticity is alpha p_k s_k. These quantities concern percentage changes, unlike derivatives, which depend on price units.

A useful implication is proportional substitution. The ratio of shares of two unchanged alternatives is independent of the price of a third. The diversion fraction from j to k equals s_k/(1-s_j) for k different from j, including the outside good. An entrant with the same observable appeal as an incumbent does not, in this model, mechanically draw a special fraction of sales from the incumbent beyond the latter's initial share. This is the independence of irrelevant alternatives restriction. It is a substantive restriction, not merely a computational convenience, and can be tested against observed switching or independent variation in availability.

Logit is a benchmark for deriving transparent relationships and calibrating small cases. It should not be accepted solely for exactly reproducing observed shares: every positive observed share vector is representable with suitable mean utilities.

Nested and heterogeneous demand

A nested logit model groups alternatives into nonoverlapping nests g, permitting stronger substitution within nests. Let sigma be a common within-nest dependence parameter, 0 <= sigma < 1, and s_j|g the conditional share within g. Under the conventional nested logit normalization,

log⁡(sj/s0)=xj′β−αpj+ξj+σlog⁡(sj∣g).\log(s_j/s_0)=x_j'\beta-\alpha p_j+\xi_j+\sigma\log(s_{j|g}).

At sigma = 0 the model becomes simple logit. Higher sigma gives greater weight to within-nest substitution, though the actual magnitude still depends on shares and alpha. Both price and conditional share can be endogenous. Instruments must supply variation for the two endogenous terms, not just price. A nesting partition is a behavioral hypothesis; changing the partition can change merger conclusions even when each candidate model fits baseline shares.

Random-coefficients demand instead allows individual utilities to vary with observable demographics and unobserved tastes. Write u_ij = delta_j + mu_ij(theta) + epsilon_ij, with mu capturing deviations from mean tastes. For example, a household with greater need for cargo space receives a larger utility contribution from capacity. A mixed logit integrates the conditional logit probability over the distribution F of these heterogeneous tastes:

sj(δ,θ)=∫exp⁡(δj+μij(θ))1+∑kexp⁡(δk+μik(θ)) dF(i).s_j(\delta,\theta)=\int \frac{\exp(\delta_j+\mu_{ij}(\theta))}{1+\sum_k\exp(\delta_k+\mu_{ik}(\theta))}\,dF(i).

Berry, Levinsohn, and Pakes (1995) give an influential heterogeneous automobile demand application. The key economic gain is that products appealing to overlapping types can be stronger substitutes than products with equal aggregate shares but different audiences.

When alpha_i varies across consumers and is positive, the derivative is the integral of -alpha_i P_ij(1{j=k}-P_ik). This derivative depends on the joint distribution of tastes and choice probabilities. Consumers on the margin of changing products, rather than the average buyer alone, determine the relevant substitution. A normally distributed price coefficient can assign some individuals positive utility from a price increase; a researcher may instead constrain price sensitivity with a transformation, while reporting its distributional implications for welfare. Non-price coefficients can reasonably permit different signs depending on the attribute.

Neither nesting nor mixing automatically validates a demand system. In a market with sparse changes in relative prices and product composition, heterogeneous parameters may be driven by parametric curvature or distributional restrictions. Observed switching, demographic information and convincing supply-side variation can add independent information. Distributional assumptions also matter when extrapolating beyond the observed products.

Random-coefficients share inversion

To invert mixed-logit shares, fix the candidate nonlinear parameters theta and the draws or quadrature points representing F. Use the observed positive shares and initialize delta, for instance with the homogeneous-logit inverse. Iterate

δj(r+1)=δj(r)+log⁡sjobs−log⁡sj(δ(r),θ).\delta^{(r+1)}_j=\delta^{(r)}_j+\log s_j^{obs}-\log s_j(\delta^{(r)},\theta).

The standard PyBLP contraction documentation states this update and the associated supply and estimation steps. The economic meaning of the iteration is that a product whose modeled share is too low receives increased mean utility, holding the current heterogeneity parameters fixed. An outside alternative provides a utility-level normalization.

A converged inner solution does not estimate theta. Many choices of theta admit an inverse, with a different vector of residuals xi. The outer estimation problem distinguishes these candidates through moment restrictions. It follows that an exact share reproduction residual pertains to the numerical quality of the inversion, while the fit of identifying moments pertains to the economic and statistical model. Publishing the first as evidence for the second is a category error.

The inversion requires care with extremely small shares. Logarithms amplify relative error; computing exponentials without stable normalization may overflow. Hold the integration points common across outer candidate values so that the objective is not contaminated by unnecessary simulation noise. Evaluate both the utility change and the share discrepancy at termination. For finite simulated mixtures with positive weights and strictly positive probabilities, the mapping is smooth; finite precision can nevertheless induce artificial zeros. Stabilization and explicit bounds must be disclosed when they alter the calculation.

The researcher should compare different integration resolutions and record whether the recovered elasticities and marginal costs are stable. Inner tolerance must be tight relative to the outer objective and derivative calculations. Inversion can be delegated to established software, but the reproducible research record retains the software version, numerical settings, and diagnostics. A unique converged delta conditional on theta leaves identification of theta as a separate question.

Price endogeneity and identification

Firms observe demand signals that the researcher may not measure. If high unobserved quality supports both higher prices and greater sales, price correlates positively with xi. In the homogeneous model, regressing log(s_j/s_0) on price and attributes by ordinary least squares may then attenuate the negative coefficient or even reverse its sign. The direction and magnitude of bias are conditional on the covariance after accounting for controls; competitive responses or measurement errors can change it.

Let Z_D denote a vector of candidate instruments. A conditional or unconditional orthogonality assumption takes the form E[Z_D xi] = 0. Exclusion concerns the structural demand error after all included attributes and controls. Relevance concerns the ability of the instruments to shift the endogenous objects in a sufficiently informative way. Instruments must also possess adequate support; a variable shifting prices over a tiny interval supplies limited evidence for a large price intervention.

Berry and Haile's 2010 identification working paper distinguishes demand identification from the identification of costs and discrimination among conduct models under specific index, exclusion and support conditions. These results provide a reason to investigate the source of identifying variation explicitly rather than equating a specified parameter vector with an identified one.

An observational equivalence example clarifies the problem. Given positive shares and prices, choose any positive alpha and set xi_j = log(s_j/s_0) - x_j' beta + alpha p_j. Each candidate reproduces the same observed shares. Each implies a different demand Jacobian. If prices also satisfy single-product Bertrand pricing with unobserved costs, c_j = p_j - 1/[alpha(1-s_j)] rationalizes the same observed outcomes whenever the implied costs are feasible. Requiring positive costs can eliminate some candidates but need not uniquely select one. With one market alone and unrestricted demand and cost residuals, neither demand slopes nor markups are point identified.

Across markets, the proposed moments restrict the residuals and may distinguish the candidates. Local identification requires a moment Jacobian with adequate rank for the parameters after normalization. Global identification demands more than rank at a single stationary point. Near-singular directions can generate weakly identified elasticities even if the criterion has a unique reported minimum. The magnitude of these directions matters for policy outputs, which may depend heavily on poorly identified combinations of coefficients.

A research design must therefore identify the specific prices, characteristics or choice-set changes that generate informative variation, and examine whether these changes are independently determined relative to the omitted demand factors. The structural model specifies how those instruments propagate through the equilibrium; it does not confer their validity.

Instrument design and exclusion

Cost shifters, characteristics of rival and own products, and certain prices in other geographic markets are common instrument candidates. Their usefulness derives from explicit economic assumptions. An input-cost change can shift equilibrium prices while leaving utility unchanged conditional on product characteristics; a manufacturer-wide delivery shock may shift prices and availability simultaneously, breaking the simple exclusion. Local wage costs may correlate with income and demand. Exchange-rate movements can shift imported costs while also changing household purchasing power.

Product characteristics can shift market competition and markups. Sums of rivals' characteristics, for example, summarize aspects of the competitive environment. Their validity requires that the characteristics used are predetermined or independent of omitted demand shocks under the maintained conditioning information. Product entry and design targeted at anticipated local preferences threaten this argument. A firm that selects products after observing a persistent demand shock creates endogenous portfolio variation.

Gandhi and Houde's differentiation instruments study (2019, revised 2020) analyzes strong instruments for substitution parameters using exogenous variation in degrees of product differentiation. A product surrounded by close rivals has different competitive conditions from one isolated in characteristic space. Counts of nearby rivals or transformations of pairwise attributes can encode this variation. Constructing such instruments is meaningful only after the exogeneity of the underlying characteristics and availability decisions has been investigated.

Other-market prices may be correlated through shared costs. They may also be correlated through regional or national preference shocks, nationwide advertising, centralized pricing, and common promotions. Fixed effects can remove some components of such shocks but cannot guarantee that the remaining variation satisfies exclusion. Instruments based on transport costs require information about delivery geography and incidence.

An instrument audit should record the proposed causal path to price and other endogenous components, the timing relative to demand shocks, likely competing paths, conditioning variables, and the direction of variation available. In a linear benchmark, first-stage coefficients and weak-instrument diagnostics are useful; in nonlinear GMM, strength also depends on how instrument moments respond to nonlinear parameters. An overidentification test can detect some inconsistencies within a joint model but cannot establish the validity of every exclusion. Instruments that are jointly invalid may satisfy similar sample correlations.

For this reason the authorial assessment recommends separately reporting price-coefficient identification and substitution-heterogeneity identification. A strong cost shifter that moves all products together may identify category participation while offering little direct information about which products are closest substitutes. A product-composition change may reveal substitution patterns yet be unusable if its timing responds to demand. Complementary variation is valuable because these dimensions need not coincide.

Microdata and independent variation

Market shares aggregate heterogeneous decisions and therefore discard information about who chooses which products. A survey of demographics among purchasers, transaction panels, second-choice observations, or observed choice sets can help identify heterogeneity. Micro information also reveals accessibility: products unobserved in local stores cannot be treated as continuously available to all buyers.

Petrin's 2002 minivan welfare study shows how demographic information conditional on purchases can discipline demand and materially affect welfare estimates. The methodological insight is that an aggregate share can be explained by different combinations of systematic preferences and idiosyncratic utility. When household composition systematically predicts a choice, the plausible allocation of explanatory variation changes.

A micro moment is formally an expectation of an observable statistic under a specified sampling frame, such as the mean number of children among minivan buyers. Its model counterpart integrates the product choice probability and the family-size statistic, divided by the predicted probability of the conditioning event. Survey sampling weights, missing responses and measurement errors must be respected. The conditioning event in the data must coincide with the model event. A moment describing new-car buyers should not be interpreted as describing all households unless nonbuyers are represented in its denominator.

Independent price and assortment changes may provide evidence about substitution directly. Store-level randomization of prices or coupons can estimate responses within the experimental setting. Quasi-experimental cost variation may offer another design, conditional on its assumptions. Such evidence can validate local derivatives used by a structural model; nevertheless a localized short-run experiment may not identify strategic equilibrium effects over a large policy change. Spillovers across stores, anticipatory purchases and interference between experimental units need investigation.

A well-documented data combination is often more informative than collecting more observations of the same aggregate variable. Market-level data identify aggregate choices; micro moments constrain composition; experimental variation constrains causal responses; supply records inform costs or conduct. Combining them requires specifying whether they describe the same population and institutional setting. Sampling precision, structural adequacy and external validity remain distinct.

Multiproduct pricing and supply inversion

Assume a static simultaneous price-setting game, differentiable demand, interior prices, and constant marginal costs with respect to the firm's pricing decision. Let F_f denote the set of products controlled by firm f, and let m_j = p_j - c_j be the per-unit markup. Profit in market t is

πf=M∑j∈Ff(pj−cj)sj(p)−Kf,\pi_f=M\sum_{j\in F_f}(p_j-c_j)s_j(p)-K_f,

where K_f is fixed cost independent of current price. For product j in F_f, the first-order condition divided by M is

0=sj+∑k∈Ffmk∂sk∂pj.0=s_j+\sum_{k\in F_f}m_k\frac{\partial s_k}{\partial p_j}.

Define H_jk = 1 if products j and k are jointly controlled, zero otherwise, and D_kj = partial s_k/partial p_j. Then Delta_jk = -H_jk D_kj. Provided Delta is invertible,

m=Δ−1s,c=p−m.m=\Delta^{-1}s,\qquad c=p-m.

This orientation matches the multiproduct markup equations in the PyBLP documentation. The transpose matters when the demand Jacobian is asymmetric. A general model with price-specific taxes, nonconstant marginal cost, common capacity, vertical contracts or price constraints changes the relevant first-order conditions. Full cost function data are unnecessary for the conditional inversion when the benchmark applies, but marginal costs inferred this way are not directly observed costs.

Under homogeneous logit and a single-product firm, m_j=1/[alpha(1-s_j)]. For a multiproduct firm with group share S_f and a common alpha, the candidate markup for every product of that firm is 1/[alpha(1-S_f)]. Substituting this expression in the first-order condition verifies it: the sum of own-portfolio derivatives in column j equals -alpha s_j(1-S_f). Hence one common markup satisfies all the equations. This result is special to the homogeneous-logit demand and benchmark pricing structure; heterogeneous price coefficients can destroy the common-markup property.

A local first-order solution is not sufficient evidence of equilibrium. Check second-order conditions for firms' own price vectors, feasibility and sensitivity to starting values. A matrix with small singular values generates unstable inferred costs. Supply comparisons should also determine whether the dependent variable is marginal production cost, retailer incremental cost or a broader cost inclusive of transaction-specific taxes and distribution. These concepts cannot be used interchangeably.

When a cost equation is included, write c or log(c) as a function of measured cost shifters and an unobserved residual omega. Log costs require strictly positive inferred marginal costs. Negative inferred costs should trigger scrutiny of data, demand derivatives and conduct rather than silent clipping. The residual's relation to excluded demand variables supplies further empirical restrictions under the maintained specification.

Generalized method of moments

For each candidate theta, compute the share inverse and, when estimating supply, the markup and implied cost. Define demand residual xi_jt = delta_jt(theta) - x_jt' beta + alpha p_jt, for a price coefficient written with a negative utility sign. Define cost residual omega_jt = h(c_jt) - w_jt' gamma, where h is a specified transformation. The sample moment vector stacks averages of Z_D xi and Z_S omega, and may include appropriately scaled micro moments.

ϑ^=arg⁡min⁡ϑgˉ(ϑ)′Wgˉ(ϑ),\hat\vartheta=\arg\min_\vartheta \bar g(\vartheta)'W\bar g(\vartheta),

where W is positive definite on the relevant moment space. The parameter vector vartheta contains linear and nonlinear coefficients, and the criterion is defined in the sample. The GMM implementation equations in PyBLP describe how linear parameters may be concentrated out subject to the model; the price coefficient may need to remain in the nonlinear search when markup recovery depends on it.

A two-step GMM procedure commonly estimates the model under an initial weight matrix and updates it using an estimate of the moment covariance. The construction of that covariance must match the sampling structure. Products within a market share shocks and equilibrium constraints; across markets the same products or firms may share persistent shocks. Treating every product-market observation as independent is often an inappropriate assumption. Cluster definition, effective cluster count and serial dependence should be justified by data provenance and the mechanism of variation.

With local identification and suitable regularity conditions, the estimated moment Jacobian and covariance determine an asymptotic variance. In weak identification or at constrained parameter boundaries, standard approximations may fail. Resampling whole relevant clusters may be more appropriate than resampling rows, provided the resampling design reflects the underlying data process. The entire estimation and counterfactual procedure must be repeated when propagating parameter uncertainty.

A model comparison should show the role of each block of moments. Adding the supply block may improve precision under valid supply restrictions, but apparent precision can be misleading if the conduct or cost equation is wrong. The appropriate exercise compares demand-only estimates, joint estimates, independent margins where available, and alternative plausible instruments. Report objective scaling, weight matrices, parameter restrictions, and any moments excluded after inspection. This prevents an apparently minor computational convention from masking a substantive change in estimand.

Computation and numerical diagnostics

The estimation problem contains a numerical inversion within a nonlinear search, often with simulation or quadrature for heterogeneous tastes. Consequently a low reported objective may conceal errors in the inner inversion, integration, or approximate derivatives. The algorithmic objective is to solve the maintained statistical problem reliably, not to use numerical convergence as a substitute for economic identification.

Dube, Fox, and Su (2012) examine constrained optimization for demand estimation, treating equilibrium restrictions as explicit constraints rather than repeatedly solving a nested inner problem. This alternative is especially relevant when repeated solution of choice or dynamic equations is expensive. In a constrained formulation the feasibility tolerance is part of the estimator's numerical definition.

Conlon and Gortmaker's 2019 best-practices working paper reports computational experiments combining integration, instrument and optimization improvements in a common implementation. Its results are conditional on the simulated data-generating processes and correctly specified restrictions considered; they do not establish that every empirical application is well identified or has a unique numerical optimum.

A reproducible workflow first verifies the logit special case against closed-form equations. Next it verifies analytic gradients against finite differences at several feasible parameter points, with step size small enough for accuracy and large enough to avoid roundoff. It checks simulated integration at multiple settings with common nodes, and evaluates the share-contraction tolerance relative to outer derivatives. Parameter bounds should have an economic interpretation and the report should identify active constraints.

The workflow compares multiple starting values whose implied elasticities are economically distinct. For each result preserve the objective, gradient norm or constrained optimality statistic, share errors, implied marginal costs, moment residuals, and relevant curvature diagnostics. A practically indistinguishable objective across widely different merger predictions is evidence of sensitivity even when one optimum is numerically preferred. It may reflect weak identification rather than computational failure.

Counterfactual solvers require their own checks: first-order residuals, economically feasible prices and shares, local second-order conditions, and multiple starts or continuation where appropriate. A reproduction archive stores exact source data versions, scripts, package versions, random seeds and solver settings. If confidential data are unavailable publicly, release reproducible code and a synthetic example within disclosure constraints rather than describe the original industry estimates as independently replicated.

Elasticities, diversion, and pass-through

For a price vector p, the derivative D_kj measures the change in demand for k when price j changes, holding other prices fixed. The diversion ratio from inside product j to k is

dj→k=Dkj−Djj,k≠j,d_{j\to k}=\frac{D_{kj}}{-D_{jj}},\qquad k\ne j,

when D_jj is negative. If alternative shares sum to one and there is no other adjustment, the diversion fractions including the outside good sum to one. Aggregate diversion to a firm is the sum for its affected products. For multinomial logit the fraction is s_k/(1-s_j), which highlights the restrictive dependence on shares. With heterogeneous demand, it varies with overlap in marginal consumer preferences.

Diversion must be distinguished from a cross elasticity. The latter expresses a proportional change in demand for k relative to a proportional price change for j; it can be large for a small product even when the actual number of diverted buyers is small. Likewise a survey asking what buyers would choose if their first choice disappeared is informative about a discrete removal, whereas a local price diversion ratio is a derivative around current terms. Their numerical equality requires additional restrictions.

A change in costs causes equilibrium price reactions depending on the entire strategic system. Let F(p,c)=0 denote the vector of pricing first-order conditions. At a differentiable isolated equilibrium with invertible F_p, the derivative of equilibrium prices with respect to costs is

dpdc=−Fp−1Fc.\frac{dp}{dc}=-F_p^{-1}F_c.

The matrix incorporates responses across products and the curvature of demand. This is why a dollar of lost competition incentive or cost savings need not translate into a dollar of price movement. A local change can also induce rival price responses.

The accounting dimensions must be explicit. A diversion ratio is dimensionless, a margin is currency per unit, their product is currency per diverted original unit, and a price effect is currency per unit after equilibrium adjustment. A ratio divided by the original price gives a percentage-type incentive metric. Confusing these quantities creates a misleading apparent price forecast. The subsequent numerical example illustrates that distinction without introducing an additional behavioral theory.

Merger counterfactuals and efficiencies

Consider a proposed merger of firms A and B under the static differentiated Bertrand benchmark. The baseline ownership matrix has separate blocks for A and B. The counterfactual combines their product sets while holding the estimated preferences and measured product characteristics fixed. Marginal costs remain fixed unless verifiable efficiencies or other changes are introduced as an explicitly specified scenario.

At the original price vector the acquiring firm internalizes positive marginal profits on products of the acquired firm that would receive diverted demand. The incremental incentive to increase the price of j, expressed in the familiar per-lost-unit scale, is sum over k in the acquired portfolio of d_j->k times m_k. If marginal cost of j is reduced, the offset enters the pricing condition on the same unit scale, although other products' cost changes and the full equilibrium system also matter. A gross incentive metric is not an equilibrium price change.

Werden and Froeb's logit-based merger analysis (1994) provides the historical benchmark for a tractable logit calculation. For the full simulation the researcher must solve

s(p∗)−Δ(p∗,H∗)(p∗−c∗)=0,s(p^*)-\Delta(p^*,H^*)(p^*-c^*)=0,

where H* and c* are the post-merger ownership and cost assumptions. The baseline equilibrium must be reconstructed before solving counterfactuals. Deviations from baseline can be evidence of measurement or model problems even when the nonlinear solver itself converges.

Efficiency evidence requires the counterfactual marginal-cost effect, its timing, and its relationship to the merger. Fixed-cost savings increase profit but do not automatically reduce short-run marginal costs. A uniform percentage reduction in marginal cost differs from a currency-unit reduction. A reduction available without the merger belongs in both baseline and counterfactual, not exclusively in the merger scenario.

A structural merger simulation and an observed post-merger event answer different questions. The former predicts a maintained mechanism with controlled changes; the latter involves every contemporaneous economic change. To compare the two, consider input prices, entry, demand trends, promotion, new ownership practices and product mix. The discrepancy may be informative about conduct or omitted factors, but requires causal investigation. Legal conclusions depend on jurisdiction and broader competitive evidence; an equilibrium price estimate is only one input.

Welfare, product entry, and distribution

For a homogeneous-logit consumer with constant marginal utility of money alpha, the expectation of maximum utility, expressed in currency units and up to a constant that cancels in changes, is

CS=α−1log⁡(1+∑jexp⁡(aj−αpj)).CS=\alpha^{-1}\log(1+\sum_j\exp(a_j-\alpha p_j)).

The term a_j includes measured and unmeasured nonprice systematic utility. Holding the distribution of utility shocks fixed, compare the expression under the baseline and intervention. Multiplying the per-choice-occasion difference by M gives aggregate consumer surplus change under the model. For heterogeneous alpha_i, integrate the individual money-metric changes rather than divide an average utility change by an average alpha.

This calculation relies on quasilinear utility or a constant marginal utility of money over the relevant change. It can fail as a precise welfare measure when income effects are central, price sensitivity is near zero for some simulated consumers, or cardinal scale differs across choice environments. In a durable-goods model, utility includes future services, resale and replacement decisions; the static purchase-choice logsum may not capture the complete intertemporal welfare effect.

Producer surplus in the static benchmark is the sum of markups times quantities, less any relevant fixed-cost change. Tax revenue, externalities and distributional concerns are separate components of social welfare. A reduction in profits to rivals can accompany a welfare-enhancing innovation; a rise in aggregate surplus can coexist with consumer losses. These statements follow from the accounting identities and are not universal predictions for actual mergers.

Petrin (2002) uses demographic purchase information to estimate the value of minivan introduction, demonstrating the importance of heterogeneity for new-product welfare. A counterfactual product removed from the choice set must be removed from the denominator of each consumer's choice probabilities. Consumers with especially strong preferences for that product can suffer much larger losses than the market average. Reporting effects across income groups, locations and types is therefore valuable whenever the model and data identify those distributions.

When discussing entry, distinguish the immediate consumption value of additional variety from long-run changes in price, innovation, product repositioning and investment. A static choice set comparison can quantify a defined component of welfare. It should not be labeled a complete dynamic social valuation.

Reproducible numerical illustration

All numbers in this chapter are constructed for this monograph, in arbitrary currency units per choice occasion. There are three inside products, initially independently controlled, with baseline prices (5, 5, 5), inside shares (0.30, 0.20, 0.10), outside share 0.40 and alpha = 1. The homogeneous logit inverse gives mean utilities log(0.30/0.40), log(0.20/0.40), and log(0.10/0.40). Adding the original price recovers the fixed nonprice utility vector a = (4.712318, 4.306853, 3.613706), rounded here for display.

Baseline costs recovered under single-product pricing are (3.571429, 3.750000, 3.888889). The implied markups are (1.428571, 1.250000, 1.111111). Baseline own-price elasticities are -3.5, -4.0 and -4.5. Diversion from product A to B equals 0.20/0.70 = 0.285714, and diversion from B to A equals 0.30/0.80 = 0.375. Multiplying by the respective receiving-product markup gives gross incentive metrics 0.357143 and 0.535714. These are currency-unit metrics, not predicted price rises.

Let A and B merge and let C remain independent. Consider three cost cases: no reduction, a reduction of 0.25 per unit for both merged products, and a reduction of 0.50 per unit for both. Product C's cost is fixed. Solve all three pricing conditions simultaneously with the original utilities a and alpha unchanged. The following computed values are rounded; unrounded values are used in the residual and welfare calculations.

Marginal cost reduction for A and B Price A Price B Price C Share A Share B Share C Consumer surplus change per occasion
0.00 5.280272 5.458843 5.019158 0.266305 0.148503 0.115255 -0.161133
0.25 5.137664 5.316236 5.011374 0.288520 0.160891 0.109119 -0.098645
0.50 5.003492 5.182063 5.003952 0.309711 0.172708 0.103189 -0.035348

The no-efficiency scenario raises A's price by about 5.61 percent and B's by about 9.18 percent. They receive a common merged-firm markup of approximately 1.708843, consistent with the homogeneous-logit multiproduct formula. Baseline operating profit per occasion summed across firms is 0.789683; post-merger values are 0.839112, 0.938721 and 1.047126, respectively. They exclude fixed costs. Consumer surplus declines in each case, with the largest reduction in the no-efficiency case. These values are not intended to predict an actual merger.

The calculation was executed with NumPy arrays and scipy.optimize.root, starting at the baseline prices. The maximum absolute pricing first-order residual was below 5.0e-10 in each case. That verifies the computed roots to the stated tolerance, but this session did not perform a global equilibrium uniqueness or full best-response proof. The reproducibility specification below uses the exact unrounded primitives.

import numpy as np
from scipy.optimize import root
p0 = np.array([5., 5., 5.])
s0 = np.array([.3, .2, .1])
a = np.log(s0/.4) + p0
c0 = p0 - 1/(1-s0)
H = np.eye(3)
H[:2, :2] = 1

def shares(p):
    x = np.exp(a-p)
    return x/(1+x.sum())

for reduction in [0., .25, .5]:
    c = c0.copy()
    c[:2] -= reduction
    def foc(p):
        s = shares(p)
        D = -(np.diag(s)-np.outer(s,s))
        return s + (H*D.T) @ (p-c)
    r = root(foc,p0)
    assert r.success
    assert np.max(np.abs(foc(r.x))) < 1e-8
    p = r.x
    cs = np.log1p(np.exp(a-p).sum()) - np.log1p(np.exp(a-p0).sum())
    print(p, shares(p), cs, ((p-c)*shares(p)).sum())

The cases establish a transparent conditional result: in this parameterization efficiencies reduce predicted consumer losses, without fully offsetting them in the displayed range. The costs and demand parameters were calibrated rather than estimated. Consequently the exercise confirms internal algebra and the operational distinction between incentive, equilibrium response and welfare, while supplying no evidence about the frequency or magnitude of such outcomes in real markets.

Automobile markets and empirical interpretation

The US automobile research in Berry, Levinsohn, and Pakes (1995) is particularly informative because vehicles differ along many attributes, and ownership and pricing decisions span portfolios. Vehicle characteristics, market-level sales and prices provide a basis for demand and cost estimation. The original automobile equilibrium study examines a roughly twenty-year period and computes elasticities and cost parameters within its model. Its estimates do not describe all modern vehicle markets automatically.

Automobiles illustrate why the outside option is economically consequential. A household may keep an existing vehicle, buy a used one, switch transportation modes or defer purchase. A model restricted to new vehicles must represent some of these possibilities as the outside alternative or explicitly extend the choice set. Raw registration counts record completed purchases rather than the universe of eligible households. Price measurements may combine trims, incentives and financing arrangements. Fuel prices and complementary infrastructure can alter the value of technical characteristics across places and years.

The minivan application adds a second perspective on the same broad sector. Petrin's study supplements aggregate outcomes with demographic information, illustrating how a new product can provide value concentrated in particular households. The evidence demonstrates an empirical method and an application, rather than a universal valuation of product variety.

For international transfer, distinguish taste for a physical attribute from the utility generated by a local institution. Fuel economy has a different monetary implication when fuel prices and annual driving distances differ. Vehicle size interacts with parking, road networks and household composition. Electrification changes charging access and policy incentives. An international vehicle application should re-estimate or validate these components rather than transplant estimated US taste and cost parameters. The relevant comparisons concern both the common method and the distinct populations, institutions and time intervals.

Conduct and retrospective validation

Demand recovery and supply recovery differ in their dependence on the maintained conduct assumption. One-shot Nash-Bertrand competition is a useful baseline in many applications because it specifies each firm's ownership-adjusted price incentives. It may be inappropriate if contracts determine prices through bargaining, firms interact repeatedly, capacity constrains competition, or vertical arrangements affect the party that receives the markup.

Nevo's 2001 cereal market-power study analyzes margins attributable to differentiation, multiproduct pricing and possible collusion. Its empirical question highlights why a large inferred margin does not by itself prove coordination. Attractive brands and portfolio ownership may support substantial markups under independent profit maximization. Direct financial information, institutional evidence and restrictions on cost variation are valuable in evaluating alternative explanations. The cited abstract supports this distinction; this dossier does not independently reconstruct that article's estimates.

A particularly informative retrospective is Miller and Weinberg's 2017 MillerCoors joint venture study. In the studied US beer market, the authors reject an explanation confined to movement from one static Nash-Bertrand equilibrium to another. Their model-based counterfactual estimates put prices after the joint venture about 6-8 percent above the counterfactual Nash-Bertrand level; these are their conditional estimates, not general merger effects. The finding supports examining conduct shifts in addition to ownership changes.

A conduct specification can be represented by an alternative matrix of strategic internalization or by different behavioral equations, but simply allowing every conduct coefficient to vary freely can destroy identification. For any candidate conduct matrix that generates feasible marginal costs, an unrestricted cost residual can absorb the difference from the benchmark. Independent cost evidence or excluded demand shifters enter the identification argument. Retrospective price data and a plausible causal design can further distinguish mechanisms.

Observed prices of nonmerging firms are useful, since they may move because of normal strategic interaction or a wider change in competitive conduct. The model should establish the response predicted under independent rivalry before ascribing an additional change to coordination. Documentary evidence, such as contemporaneous descriptions of pricing strategies, may discipline mechanisms that are difficult to distinguish statistically. A research conclusion should clearly separate the observed event, the estimated counterfactual, and the interpretation.

International comparison and institutional variation

The structural method is portable across market economies, but its parameters and conduct assumptions are local to the studied institutions. The European heavy-truck market, examined by Ivaldi and Verboven (2005) in their European merger simulation study, illustrates an industrial application of nested-logit demand and multiproduct oligopoly. Truck purchasing may be linked to fleet operations, financing and service networks; a research design should investigate these factors rather than treating all purchases as equivalent to retail household choices. The accessible publisher excerpt confirms the application and specification, not the numerical estimates from the complete paper.

The European Commission's historical 2004 horizontal merger guidelines place competitive effects within the merger assessment framework. This citation concerns the 2004 document as a historical institutional source. The US agencies' 2023 guidance has a specific discussion of substitution, diversion, documentary evidence and simulation. A difference between these documents in terminology does not by itself establish a difference in the underlying economics or a current legal rule. This monograph does not offer a contemporary jurisdiction-by-jurisdiction legal opinion.

Cross-national comparisons require at least four forms of harmonization. The physical or service unit must be comparable; the price must reflect purchasing terms and tax treatment; the choice set and population must cover similar activities; and the organization of strategic pricing must match the model. A failure in one dimension can generate differences in inferred elasticities even when underlying preferences are similar. Harmonization may require country-specific models with common definitions rather than a pooled regression.

In less extensively measured markets, administrative transaction records, carefully designed household surveys or retailer panels may provide useful partial evidence. Coverage of informal sellers, stockouts and cash discounts can be economically important. The absence of a transaction record is not proof of the absence of a competitive alternative. Cross-country comparisons using only formal retail data risk systematically omitting substitutions available to lower-income households. A global research frame therefore distinguishes the universe of relevant consumers from the measured sample.

The case studies in this dossier provide US and European examples. They establish mechanisms and methodological comparisons, rather than numerical benchmarks for Asia, Africa or Latin America. Extending empirical coverage requires documented datasets and local institutional knowledge; geographical breadth cannot be established by attaching country names to a single fitted demand system.

Digital markets, dynamic choices, and bargaining

Digital markets can change the relation between price and demand through ranking algorithms, search and personalized offers. Zero monetary price does not necessarily mean zero economic sacrifice: time, privacy, advertising exposure and quality may enter utility. A price derivative calibrated at zero cannot be converted into a conventional elasticity using a positive price denominator; quantity and utility-based measures may be more useful. Two-sided platforms coordinate participants, and demand on one side can depend on participation and prices on the other. A single-sided ownership intervention omits these feedbacks unless they are explicitly modeled.

Dynamic durable-goods decisions have another source of dependence. The utility of buying today is compared with the value of waiting, repairing, or purchasing a later version. Expectations about future price and quality changes can shift current sales. In a static dataset this may be confounded with unobserved product quality. A dynamic extension needs a state description, transition assumptions, discounting and a model for consumers' expectations. Durable stock and resale may influence demand independently of current price.

When prices are negotiated, equilibrium price conditions derive from disagreement values and bargaining rules rather than a multiproduct posted-price first-order condition. The outside option of each bargaining participant is particularly important. A merger changes the alternatives available to buyers and sellers and may influence bargaining leverage even when product-level prices are set separately. Similarly, procurement auctions require a model of bidder participation and information, not simply an aggregate retail market share.

These departures indicate clear empirical triggers for model revision. Observe whether common prices are posted or negotiated; whether buyers can delay; whether availability or ranking is personalized; whether other market sides affect utility; and whether there are capacity or contractual constraints. If the departure is substantial, treat it as part of the structural object rather than an unstructured error term. The basic logit-Bertrand equations then remain a transparent comparison case, with the extended model justified by observed institutions.

Policy evidence and counterfactual uncertainty

The US agencies' 2023 competition evaluation guidance explicitly considers switching evidence, historical events, pricing incentives and the limited scope of simulation. A policy simulation is most useful when it is grounded in the actual choices firms and customers face. A researcher can estimate the effect of removing competition between two portfolios while holding conduct fixed, then explore alternative conduct or product-choice responses separately. Combining assumptions without labeling them makes it difficult to know what an estimated welfare loss represents.

Quantitative results should separate sampling uncertainty from sensitivity to model specification and counterfactual assumptions. Sampling uncertainty asks how repeated samples under the maintained model affect estimated objects. Numerical uncertainty concerns integration or solver error. Specification sensitivity asks how alternative plausible models of substitution, costs or conduct alter the results. Scenario uncertainty concerns the magnitude and timing of efficiencies, entry, and other intervention-dependent variables. A single standard error cannot summarize all four.

For a merger exercise, a credible report shows the baseline demand elasticities, diversion matrix, margins, and the observed source of identifying variation. It then specifies the ownership and cost changes; reports the implied equilibrium prices, shares and welfare; and provides a sensitivity range along empirically defensible assumptions. Implausible specifications can be excluded with explicit evidence. It is misleading to average unrelated scenario estimates and call the result a statistically identified expected merger effect.

Independent documentary, experimental and accounting evidence can strengthen or challenge model predictions. Survey responses and internal sales-loss reports can test the directional substitution pattern. Entry or product removal can test predicted reallocations. Observed wholesale and retail margins can help assess supply inversion, conditional on accounting definitions. Retrospective analysis can assess whether observed effects align with ex ante predictions once concurrent changes are accounted for.

A policy analyst should distinguish the positive economic conclusion from normative or legal judgment. Price and surplus changes conditional on assumptions are economic outputs. The applicable legal burden, remedy choice and treatment of efficiencies are institutional questions. A structural economic model provides evidence within that decision process; it does not constitute the legal decision.

Independent analytical synthesis

The central analytical finding of this dossier is that the credibility of a structural counterfactual depends on a chain of empirically supported mappings. These mappings connect a choice occasion to observed shares, shares to utilities, utilities to causal derivatives, derivatives and ownership to incentives, incentives and costs to equilibrium, and resulting outcomes to the chosen welfare measure. The inferential strength of the final result depends on this entire chain. No new theoretical system is proposed; the finding is a synthesis of the established methods and their documented conditions.

To see why baseline fit cannot validate the counterfactual, consider again two homogeneous-logit systems with the same observed shares and prices. For arbitrary positive alpha, choose product-specific intercepts to match the observed baseline. Under feasible Bertrand markups recover the corresponding costs. Both systems can match prices and quantities exactly. But diversion at the baseline in the simple logit model is a function of shares, while the receiving-product markup is inversely proportional to alpha. The gross merger incentive is therefore inversely proportional to alpha even though the diversion ratio remains unchanged. When prices subsequently move, the share response also depends on alpha. The observed baseline alone cannot select a unique price effect.

A second distinction concerns the quality of demand evidence versus the quality of supply evidence. Suppose product switching and price variation identify demand closely. If firms' costs are unobserved and conduct is unrestricted, alternative ownership-adjusted strategic assumptions can still rationalize baseline prices with different costs. The counterfactual then changes across conduct assumptions. Conversely reliable accounting margins may restrict supply but cannot by themselves recover flexible cross-product substitution unless sufficient additional variation is available.

These two examples explain why additional evidence should be collected at the weak link rather than merely increasing the volume of observations elsewhere. A market with rich variation in category participation may need second-choice evidence to identify within-category substitution. A market with convincing substitution data may need procurement and contract information to identify the party that sets prices. A market with clear static incentives may require a dynamic model when consumers are able to delay purchasing. These are concrete implications for research design.

The synthetic calculation makes the same point at the counterfactual stage. Conditional on one chosen alpha, demand system and conduct model, the ownership change can be evaluated transparently. Varying marginal-cost efficiencies yields different consumer losses. The difference is a substantive scenario effect rather than statistical sampling noise. The direction of the comparison in this example is robust over the displayed cases, but the magnitudes are specific to its primitives. Scientific communication must disclose that distinction.

Finally, a credible estimate should be falsifiable beyond the outcomes used to fit it. Identifying variables, observed diversion, price changes outside the estimation sample, independent margins, entry or exit, and actual institutional practices can provide distinct validation opportunities. A result becomes more informative when the supporting evidence constrains the mechanisms that matter for the intended intervention. This is the authorial contribution: a unified interpretation of the dependence among the established components and a practical account of what evidence a counterfactual requires.

Research agenda and reproducibility

A research program for applying these models should begin with a sharply defined intervention. For a prospective merger, specify the portfolios, jurisdiction, pricing stage and time horizon. For product introduction, specify the replacement or addition, availability, costs and the relevant welfare population. For a tax experiment, identify tax incidence and price measurement. The intervention determines which structural parameters and institutional mechanisms require evidence.

A transparent data pipeline preserves raw records separately from derived measures. Document the population, product identifier, market identifier, transaction quantity, purchaser price, seller net price, observed characteristics, ownership and choice availability fields. Store exact dates of price and ownership changes. Establish separate procedures for deflating monetary quantities, assigning products to markets, handling censored sales and matching survey or administrative sources.

Before a nonlinear estimation, fit and inspect a transparent benchmark, evaluate market definitions and instrument behavior, and check whether the model's assumptions correspond to the observed purchasing process. Then evaluate alternative demand specifications and instrument sets. Collect evidence for costs and conduct independently whenever possible. Nonlinear equations, software behavior and numerical convergence should be checked at the level of the counterfactual outputs rather than only the parameter vector.

The strongest future research opportunities involve validation on exogenous assortment changes, credible price interventions, independently observed cost and margin data, and cross-country records that identify institutional differences. Where research combines experiments and structural models, the experiments can inform local causal responses and the structural model can translate those responses into specified equilibrium settings. The strength of that translation depends on the invariance and support assumptions.

A completed empirical replication would require access to the original product-level data, market definitions, instruments, estimation code and the precise versions of the published papers. Such a replication is outside the present synthetic and literature-based dossier. Some cited journal sources were accessible only as indexed abstracts or author working-paper versions; the evidence register specifies those distinctions. The current manuscript preserves a reproducible computational illustration, rather than representing the published studies' estimates as newly verified empirical results.

Conclusion

Structural models of differentiated-product competition link individual choice with market shares, prices, strategic behavior and welfare. Their usefulness stems from their capacity to articulate which mechanisms an intervention changes and to solve for the resulting market outcomes. Share inversion recovers mean utilities under maintained demand parameters; valid identifying variation is required to determine those parameters. Supply inversion recovers marginal costs conditional on demand derivatives and pricing behavior. Counterfactual outcomes require both an altered economic environment and a defensible account of the mechanisms that remain invariant.

The literature and derivations demonstrate distinct roles for data about quantities, prices, demographic choice, availability, margins, and institutional behavior. A fit to observed market-level prices and quantities does not by itself validate the relevant causal derivatives or conduct. Numerical reliability and exact source attribution strengthen research, while sampling, structural and scenario uncertainty must retain their separate meanings.

The synthetic example illustrates ownership internalization and the offsetting influence of modeled marginal-cost efficiencies. Its results are conditional on a known homogeneous-logit demand and the stated simultaneous-pricing benchmark. The international cases establish the application of structural methods in particular industries and the importance of context in any transfer of empirical estimates.

The scientific value of this approach resides in the transparent connection between econometric identification, economic behavior and the proposed intervention. A research dossier can establish that connection by making its assumptions, evidence, calculations and limitations inspectable. Further empirical work can then investigate specific markets without confusing the methodological foundation with an estimate already established for those markets.

References

  • Berry, Steven T. 1994. Estimating Discrete-Choice Models of Product Differentiation. RAND Journal of Economics 25 (2): 242-262. DOI: 10.2307/2555829. Research source.

  • Berry, Steven T., and Philip A. Haile. 2010. Identification in Differentiated Products Markets Using Market Level Data. Cowles Foundation Discussion Paper 1744, updated February 2010. Related final article: Econometrica 82 (5), 2014: 1749-1797, DOI: 10.3982/ECTA9027. Research source.

  • Berry, Steven, James Levinsohn, and Ariel Pakes. 1995. Automobile Prices in Market Equilibrium. Econometrica 63 (4): 841-890. DOI: 10.2307/2171802. Research source.

  • Conlon, Christopher, and Jeff Gortmaker. 2019. Best Practices for Differentiated Products Demand Estimation with pyblp. May 7 working-paper version. Related final article: RAND Journal of Economics 51 (4), 2020: 1108-1161. DOI: 10.1111/1756-2171.12352. Research source.

  • Dube, Jean-Pierre, Jeremy T. Fox, and Che-Lin Su. 2012. Improving the Numerical Performance of Static and Dynamic Aggregate Discrete Choice Random Coefficients Demand Estimation. Econometrica 80 (5): 2231-2267. DOI: 10.3982/ECTA8585. Research source.

  • European Commission. 2004. Guidelines on the Assessment of Horizontal Mergers under the Council Regulation on the Control of Concentrations between Undertakings. Official Journal C 31, February 5, 2004: 5-18. CELEX 52004XC0205(02). Research source.

  • Gandhi, Amit, and Jean-Francois Houde. 2019. Measuring Substitution Patterns in Differentiated-Products Industries. NBER Working Paper 26375, revised November 2020. DOI: 10.3386/w26375. Research source.

  • Ivaldi, Marc, and Frank Verboven. 2005. Quantifying the Effects from Horizontal Mergers in European Competition Policy. International Journal of Industrial Organization 23 (9-10): 669-691. Research source.

  • Miller, Nathan H., and Matthew C. Weinberg. 2017. Understanding the Price Effects of the MillerCoors Joint Venture. Econometrica 85 (6): 1763-1791. DOI: 10.3982/ECTA13333. Research source.

  • Nevo, Aviv. 2001. Measuring Market Power in the Ready-to-Eat Cereal Industry. Econometrica 69 (2): 307-342. DOI: 10.1111/1468-0262.00194. Research source.

  • Petrin, Amil. 2002. Quantifying the Benefits of New Products: The Case of the Minivan. Journal of Political Economy 110 (4): 705-729. DOI: 10.1086/340779. Research source.

  • PyBLP contributors. n.d. Background. PyBLP 1.2.0 documentation. Accessed November 17, 2026. Research source.

  • U.S. Department of Justice and Federal Trade Commission. 2023. Merger Guidelines. Section 4.2, Evaluating Competition Among Firms. December 18, 2023. Research source.

  • Werden, Gregory J., and Luke M. Froeb. 1994. The Effects of Mergers in Differentiated Products Industries: Logit Demand and Merger Policy. Journal of Law, Economics, and Organization 10 (2): 407-426. DOI: 10.1093/oxfordjournals.jleo.a036857. Research source.