RESEARCH DOSSIER
Product Differentiation and Discrete Choice Demand
Abstract. This research monograph investigates differentiated-product demand through characteristic-based preferences, random utility, homogeneous and heterogeneous choice models, market-share inversion, identification, price endogeneity, variety, and market power. It derives logit probabilities, own- and cross-price effects, diversion, surplus measures, and multiproduct pricing conditions; examines how nested and mixed choice structures change substitution; and compares published applications in automobiles, cereal, and spatial gasoline markets. Reproducible hypothetical examples establish that equal aggregate shares can conceal unequal substitution, that market-size conventions alter inferred diversion, and that independent utility disturbances can attribute welfare value to product duplication. The analytical synthesis separates preference measurement, feasible choice sets, and equilibrium mechanisms. The work is a source-based methodological and comparative study with explicit limits on empirical transportability; it reports no original econometric estimates or newly named theory.
Keywords: Product differentiation; discrete choice; random utility; multinomial logit; nested logit; random coefficients; demand identification; market-share inversion; substitution; diversion ratios; consumer surplus; product variety; industrial organization; microeconomics
The Economic Object of Differentiation
Product differentiation describes an economic environment in which alternatives have distinct characteristics, perceived qualities, locations, service arrangements, or contractual conditions that affect their attractiveness to buyers. Demand is consequently distributed across alternatives that are imperfect substitutes. A change in the price of one product redistributes purchase probabilities among the competing products and the possibility of making no purchase within the category. The economic object of this dossier is that redistribution, together with its implications for the value of variety, market power, and counterfactual policy.
Horizontal differentiation arises when different consumers prefer different characteristics at comparable prices. Vertical differentiation arises when consumers agree about a quality ordering under specified conditions, though their willingness to pay for quality varies. Actual markets combine both. An electric vehicle may offer lower operating costs and limited range relative to another vehicle; a premium service may be valued more highly by all customers while only a subset can justify its price. Labels like horizontal and vertical refer to preference restrictions, not immutable properties of the physical good.
Let a market be an economically defensible population of decision makers over a stated interval, with a feasible set of alternatives indexed by j. Each alternative is defined at the level at which an actual choice is made. An automobile trim, a cereal package, a gasoline station, and a subscription contract imply different units of observation. A characteristic is a measurable attribute such as size, fuel efficiency, or travel time. Unobserved quality comprises demand-relevant attributes unresolved by the available information set: it may include brand reputation, distribution convenience, perceived durability, and local promotion. An outside option represents alternatives excluded from the inside choice set, including postponement or purchase elsewhere when appropriate.
Differentiation in a market depends on both the characteristics of products and the distribution of consumer tradeoffs. Products can be technically distinct and economically close substitutes. Conversely, nearly identical physical items can serve different customers because access, timing, bundled service, or perceived trust differs. A defensible demand analysis therefore begins with the effective choice set and population rather than a classification of products by manufacturer alone.
The dossier uses single-unit discrete choice as its organizing framework. Multiple-unit demand, baskets, repeated purchases, dynamic replacement, and platform interactions are extensions that require explicit additional assumptions. The parent discipline is Microeconomics. The present study concentrates on demand-side substitution, heterogeneity, and variety; the full strategic structure of differentiated-product competition and firm production duality are separate scientific objects.
Historical Development and Intellectual Foundations
The intellectual history of differentiation connects the study of market competition to the study of consumer characteristics and statistical choice. Hotelling's Stability in Competition (1929) is a foundational antecedent of spatial competition. Its economic importance lies in representing proximity as part of the competitive relation. A firm may be protected by the cost of reaching its rivals even when the physical good is homogeneous. Later spatial models treat a location as an abstract position in product-characteristic space as well as a point in physical geography.
Lancaster's characteristics approach (1966) placed the attributes provided by goods at the center of consumer analysis. In contemporary discrete choice models, characteristics enter utility together with prices and individual valuations. This representation allows a researcher to extrapolate cautiously from observed products to a newly introduced bundle of characteristics, a task that unrestricted separate demand equations cannot easily accomplish. The relationship is intellectual rather than a claim that Lancaster originally specified the modern structural econometric model.
McFadden's conditional logit analysis (1974) provided an econometric foundation for qualitative choice: utility-maximizing behavior can be modeled through probabilities when the analyst lacks complete information about individual valuations. The resulting probability system connects utility differences to observable choices. The probabilistic description belongs to the analyst's information set; individuals need not literally randomize their decisions.
Berry's discrete-choice models (1994) connected choice probabilities, observed product shares, and price-setting in differentiated oligopoly. In the simplest logit case, the difference of logarithmic shares yields an estimating equation for mean utility. Berry, Levinsohn, and Pakes (1995) subsequently applied a heterogeneous-consumer demand and equilibrium framework in automobile prices in market equilibrium. Their contribution opened an empirical route from market shares and product characteristics to own- and cross-price responses while addressing simultaneous pricing.
The methodological movement involved a progression in what could be measured: locations and attributes provided representations of difference; random utility provided a choice probability; market-share inversion connected product data to latent mean valuations; heterogeneous coefficients provided richer substitution; instruments and equilibrium restrictions confronted simultaneity. Each step introduced additional assumptions. The history is consequently a cumulative development of identifiable economic objects rather than a succession in which one estimator universally supersedes another.
The subsequent development emphasized micro-level information, numerical reliability, and the distinction between identifying a demand system and estimating a convenient parameterization. Berry, Levinsohn, and Pakes (2004) studied the new-car market with combined micro and macro data. Berry and Haile's (2021) Foundations of Demand Estimation provides an institutional account of the importance of identification, instruments, and data settings for the modern field.
Random Utility and the Multinomial Logit Benchmark
Consider a decision maker i with a feasible set containing J inside goods and an outside option indexed by 0. Write utility as a systematic component plus an unobserved component:
Here x denotes observed characteristics, p denotes the full decision-relevant price, and xi denotes product quality observed by market actors but omitted from measured characteristics. The parameter alpha is positive when higher prices reduce utility. The choice probability is the probability that one alternative yields utility at least as large as every feasible alternative. Ties have probability zero with continuous taste disturbances.
The independent type-I extreme-value disturbance assumption with common scale yields multinomial logit. Normalize outside systematic utility to zero and normalize the disturbance scale to one. For homogeneous coefficients, define the systematic inside utility delta. Then
The derivation uses the distribution of the maximum of independent extreme-value variables. More explicitly, integrate the density of the disturbance of the candidate good against the product of the cumulative probabilities that every other disturbance produces a lower utility. The exponential factors combine to give the denominator above. Prices have meaning relative to the scale normalization; utility is ordinal at the individual deterministic level, and its difference distribution is normalized for estimation.
The odds relation follows directly:
This is independence of irrelevant alternatives (IIA) for the homogeneous logit model. Changes in the availability or utility of a third alternative do not alter the odds between two unchanged alternatives. This strong property can be acceptable for a deliberately limited benchmark, but becomes a problem when a new option is a close replica of only one incumbent. Train's Discrete Choice Methods with Simulation (2009, second edition) surveys the relevant model families and computational approaches.
If a product disappears and all other systematic utilities remain fixed, the surviving shares are divided by one minus the removed product's initial share. This is a prediction of the particular benchmark. It does not follow from rationality alone. For example, give products A and B each a share of 0.30 and the outside option 0.40. Introduce a new product C whose systematic utility equals A's and leave the other utilities unchanged. The new denominator is 1.30 times the old one, so A and B each receive 0.230769, C receives 0.230769, and the outside option receives 0.307692. A and B lose exactly equal shares even if C is a near duplicate of A and physically unlike B. This is a transparent counterexample for the empirical adequacy of IIA, not a universal claim about consumers.
Price effects follow by differentiating the share formula. For a common positive price coefficient alpha,
The own derivative is negative and the cross derivative positive for distinct products. With the market population fixed, price elasticity equals the derivative times the price divided by the initial share. Thus the own elasticity is minus alpha times own price times one minus own share; the elasticity of j with respect to the price of k is alpha times the price of k times the share of k. These expressions are conditional derivatives holding other prices, characteristics, market size, and latent quality fixed.
For infinitesimal price increases the diversion ratio from j to k is the cross derivative for k divided by the absolute own derivative for j:
Summing across all alternatives other than j, including the outside option, gives one. The benchmark assigns diversion in proportion to competitors' initial shares. An analyst who uses that formula to evaluate a merger is importing a strong substitution restriction. The restriction should be understood before choosing the estimator.
Heterogeneity, Correlation, and Flexible Substitution
Taste heterogeneity changes both the scale and allocation of demand responses. Customers who most value a particular attribute are disproportionately represented among that product's buyers. When their preferred product becomes more expensive, the appropriate question is where those customers go, not simply which products have the greatest market shares. Aggregate substitution is governed by the joint distribution of preferences, feasible alternatives, and price sensitivities.
Nested logit groups products into nests that share some unobserved utility variation. For a common nesting parameter sigma between zero inclusive and one exclusive, let the within-group inclusive value be the log of the sum of exp(delta divided by one minus sigma). Conditional shares within a group are proportional to these exponentials. The group share is proportional to the exponential of one minus sigma times the inclusive value, with the outside treated as a singleton. The overall probability is the product of the group share and the conditional share. At sigma equal to zero this reduces to ordinary logit; larger sigma permits stronger within-group substitution. The restriction still imposes a specific correlation structure and does not make categories scientifically self-validating.
Random-coefficients logit permits beta and alpha to vary across decision makers according to a distribution conditional on observed demographics and latent taste factors. Conditional on coefficients, the logit probabilities retain their simple form; the aggregate shares integrate over the heterogeneity distribution:
The corresponding derivative is
The average of the products of conditional shares is generally different from the product of the average shares. This is precisely why aggregate share ratios no longer determine diversion. For positive individual price coefficients, own demand falls and cross demand rises in the static single-unit model; complementarity across jointly consumed goods requires a different choice structure. A normally distributed untransformed price coefficient can assign the wrong sign to part of the population. A constrained sign transformation, income interaction, or a specification of willingness-to-pay space should be evaluated rather than assumed innocuous.
McFadden and Train (2000) establish a flexible approximation result for mixed multinomial logit under regularity conditions. Such representational flexibility is distinct from empirical identification: a family that can approximate rich patterns can still be poorly pinned down with a finite dataset. A more flexible distribution imposes additional demands on observed variation, simulation accuracy, and model validation.
A multinomial probit model instead permits correlated normal errors and evaluates probabilities of multivariate inequalities. It can express substitution patterns through estimated covariance structure, with scale and location normalizations. Computational cost and identification constraints increase with the dimension. A latent-class model represents consumers as a mixture of a finite number of preference types. It is useful when group-specific information is meaningful, but fitted classes may shift with the specification and need not correspond to social categories.
A specification should be chosen in relation to the target inference. If the target is prediction of total category participation, the outside option and market-size definition may deserve more attention than fine nest structure. If the target is a particular merger, own and rival diversion among the affected products is critical. If the target is a new characteristic never previously marketed, preference extrapolation and support become central. The quality of a demand system is task-specific; estimation likelihood alone is insufficient.
Variety, Duplication, and the Value of Choice
Variety expands the feasible set when consumers gain access to alternatives previously unavailable. Under stable preferences, fixed prices for incumbent products, and voluntary choice, an expanded feasible set weakly raises maximum utility for every buyer. The statement is a choice-set comparison; it is not a complete equilibrium welfare conclusion. Entry can change prices, alter quality, impose search costs, and consume fixed resources. Some varieties appeal to a small group whose gain is economically significant despite low aggregate market share.
If income enters utility linearly with a common coefficient alpha, the logit expected maximum utility equals the log-sum plus a constant. With prices and characteristics represented in money-consistent utility units, the expected consumer surplus change from state A to B is
The outside option is included in each sum. The disturbance-scale constant cancels under the same distribution and normalization. This follows from the distribution of the maximum of independent extreme-value errors. With heterogeneous positive alpha, compute and average the individual money-metric log-sum changes. A varying marginal utility of income, income effects, or a significant change in feasible consumption calls for compensating variation determined from indirect utility rather than a mechanically rescaled log-sum.
The benchmark assigns a positive surplus effect to a formally duplicated alternative if its independent disturbance is retained. For the A/B/outside example, the addition of C raises the log-sum by log(1.30). If alpha equals 0.10 per dollar, the implied gain is 2.6236 dollars per decision maker at unchanged prices. This is a calculation in a hypothetical model. A literal relabeling of exactly the same alternative, understood as the same utility realization and feasible purchase, should produce zero gain. The independent-error representation creates the difference. Whether product duplication is economically meaningful depends on utility correlation and actual choice opportunities. Counting catalog entries as variety can therefore overstate welfare improvements.
The analysis distinguishes entry benefit, incremental production cost, and strategic price effects. Suppose that a new product X attracts customers who already purchased from the same firm. Its commercial value may be modest because of cannibalization, yet its matching benefit to customers may remain positive. Alternatively, a differentiated product can discipline competing prices and thus generate benefits for nonbuyers of the new good. Fixed entry costs and changes in producer surplus matter for total welfare, as do any externalities not reflected in consumers' utility.
Petrin's (2002) minivan welfare study illustrates the new-product problem in the U.S. vehicle market. The appropriate takeaway here is methodological: valuations of new goods require credible substitution and counterfactual comparisons. No published numerical surplus estimate is reproduced because the underlying parameter estimates and data have not been independently audited for this dossier.
Endogeneity, Instruments, and Micro-Level Information
Prices are typically decisions of firms. A desirable quality attribute omitted by the analyst can increase both demand and the price charged for the product. Consequently, a regression relating log shares to prices and observed characteristics confounds the demand effect of price with the association between prices and unobserved quality. The bias need not have a universal sign in a multi-equation market, although positive price-quality correlation often attenuates an estimated negative price sensitivity in a simple setting.
A cost shifter can serve as a price instrument when it changes price through supply without directly affecting the relevant utility conditional on included information. Shipping expense, input prices, or tax changes require careful institutional interpretation: they may also affect availability, perceived quality, or household disposable income. A rival characteristic can help predict competitive pricing but may be endogenous when firms choose characteristics strategically based on demand conditions. Firm and time fixed effects remove some common patterns; they do not automatically remove product-market-level shocks.
Gandhi and Houde (2019) investigate differentiation instruments that reflect exogenous proximity in characteristics space. Their relevance is naturally tied to the substitution patterns the researcher seeks to learn. A nearby competitor can create pricing pressure different from that exerted by a distant competitor. The authorial inference is that instrument construction should be organized around the demand derivatives required by the intended counterfactual, while the exogeneity of the underlying characteristics still requires substantive justification.
Aggregate data may reveal the popularity of each product without identifying the demographic composition of buyers or their second choices. A micro dataset can provide moments linking product attributes to individual income, household composition, travel constraints, and alternatives considered. Actual second choices or responses to exogenous assortment changes are especially informative about substitution. Stated second choices and hypothetical conjoint decisions should be distinguished from revealed market behavior. Survey design, sample selection, and survey weights affect which population is represented.
Conlon and Gortmaker's (2023) micro-data estimation extension describes incorporating individual-level evidence into a differentiated-product estimation workflow. Micro information strengthens a model when it constrains latent distributions or substitution directly; it cannot by itself validate a contaminated price instrument. Combining independent forms of variation is particularly valuable. An experiment on a price attribute, a panel recording repeated choices, aggregate shares, and a competitor's entry may each identify different features of the same system.
Causal identification must be tied to the intervention. For a local price experiment with randomized exposure, an intention-to-treat effect can be estimated even if the structural demand specification is disputed. Moving from that local effect to an equilibrium policy prediction requires assumptions about representativeness, rival response, and stability of preferences. The structural framework is useful precisely because these extrapolation assumptions can be made explicit and tested against additional evidence.
Demand, Pricing, and Market Power
The pricing consequences of differentiated demand depend on both the matrix of derivatives and ownership. Assume static simultaneous price competition, single-unit purchase opportunities, constant marginal cost with respect to output locally, and interior optimal prices. Let each firm's profit be the market size times the sum across its products of price minus marginal cost multiplied by market share. Holding rival prices fixed, the first-order condition for a price p_j set by firm f is
Define the matrix Delta with row j and column k equal to minus the derivative of share k with respect to price j when j and k share ownership, and zero otherwise. The markup vector solves Delta times markups equals shares. If Delta is nonsingular, markups equal its inverse multiplied by the share vector. This is an algebraic implication of the maintained pricing model and derivative orientation. It is not a direct measurement of marginal cost. The validity of inferred costs depends on the conduct assumption and on accurate demand derivatives.
For an isolated single-product firm under homogeneous logit, the first-order condition implies the markup is one divided by alpha times one minus its product share, with parentheses around the entire product in the denominator. For alpha equal to 0.10 per dollar and share 0.30, the implied markup is 14.2857 dollars. This is a hypothetical optimum condition, not evidence of a real markup or a claim of global optimality. Multiproduct ownership changes the formula because firms internalize sales diverted to sister products.
After a merger, diversion between the merging parties can create an incentive to raise prices in a local first-order calculation at premerger prices, even without changing demand preferences. Cost efficiencies and responses by other products matter for net equilibrium outcomes. Market concentration alone cannot determine the incentive: two firms with small shares might compete intensely for the same specialized customers. Firms with larger shares could sell to almost disjoint customer segments.
Nevo's (2001) ready-to-eat cereal study separates the estimated importance of differentiation, multiproduct ownership, and possible coordination in a specific industry. The linked author repository abstract reports that the first two components explain most of the estimated markups in that analysis. This is a conditional empirical conclusion, not a general finding that collusion is absent in concentrated industries. It illustrates why demand estimation can discipline interpretations of high margins that would otherwise be observationally ambiguous.
Recovering marginal costs from demand requires a nonsingular and numerically stable derivative system, defensible multiproduct ownership data, and the chosen conduct model. Negative inferred costs or implausible marginal margins may reveal misspecification, inaccurate transaction prices, erroneous outside shares, or omitted vertical contracts. They warrant investigation rather than automatic trimming. An independently observed cost sample can help distinguish supply misspecification from demand misspecification.
The supply analysis in this dossier is deliberately limited to the consequences of substitution and ownership. Strategic entry, capacity commitments, price discrimination equilibrium, repeated interaction, and vertical bargaining deserve their own model-dependent analyses.
Comparative Market Evidence and International Interpretation
Automobile markets offer distinct combinations of durable consumption, observable engineering characteristics, substantial price variation, and an outside option consisting of postponement, used vehicles, or alternative modes. Berry, Levinsohn, and Pakes (1995) is the central example of aggregate structural automobile demand. The empirical unit and numerical elasticity estimates in such a study belong to its actual sample and assumptions; they do not transfer automatically to a different period of fuel prices, vehicle finance, or electrification. The 2004 extension links consumer information with aggregate product data. These two studies illustrate distinct information environments even within the same broad industry.
European automobile markets provide a different institutional question. Verboven's (1996) European car market study is a published application devoted to price discrimination across national markets. Even when vehicles carry the same name, their equipment, taxes, exchange-rate exposure, warranty, dealer networks, and distribution regulations can differ. Quality-adjusted prices should be placed on a consistent tax and currency basis, and the estimation should permit country-specific outside options and demand shifters. A cross-country price gap is not itself evidence of discriminatory market power. The title and bibliographic scope of the cited study are verified; specific coefficient estimates are not imported.
Spatial differentiation extends beyond durable goods. Houde's (2012) Quebec gasoline study defines consumer location through commuting paths and estimates a spatial model using a gasoline-market panel in Quebec City. The paper compares a merger simulation with a difference-in-differences analysis. This is particularly instructive because the effective distance between alternatives depends on where consumers travel rather than solely on residential proximity. The direct journal abstract reports broad support for its demand-model assumptions from the comparison; that conclusion is specific to the studied market and research design.
The U.S. cereal study presents a frequently purchased good with relatively low transaction values. A brand-choice interpretation requires attention to package sizes, promotion, multiple-unit purchases, and store assortment. The durable-vehicle examples involve infrequent replacement and larger expenditure commitments. Gasoline is recurrent and spatially constrained. These markets should not be pooled into one universal price-elasticity statistic: the denominators, time horizons, feasible sets, and purchase units differ.
The international evidence in this dossier is a comparative review of documented applications in the United States, European national automobile markets, and Canada. It is not a representative worldwide sample. Applicability to other regions is a research question: informal retail, cash-based purchasing, differences in credit availability, unreliable product metadata, and distribution scarcity may fundamentally alter the observable choice set. A model built on complete national product-level sales may need local survey or transaction data before it can describe a decentralized marketplace. No country rankings or synthetic cross-country estimates are presented as empirical findings.
For comparison across studies, preserve the original price definition, currency basis, quantity unit, population, time window, market segmentation, specification, method of identifying price, and estimation uncertainty. An elasticity is dimensionless yet can be institutionally incomparable if it measures a different intervention or time span. Comparing long-run vehicle replacement with short-run grocery substitution on numerical magnitude alone obscures the underlying decision process.
Measurement Design, Estimation, and Numerical Reliability
A reproducible empirical study begins with a versioned data dictionary. At minimum, the product-market panel needs a stable product identifier, market identifier, period, quantity, transaction-equivalent price, observed characteristics, ownership, availability, a defensible market-size denominator, and candidate instruments. Product definitions must remain stable across changes in packaging, trim, branding, or seller. Record whether quantities represent units, households, visits, prescriptions, or subscriptions, and whether prices include taxes, delivery charges, discounts, or contract obligations.
| Field | Meaning | Primary validation |
|---|---|---|
| product and market identifiers | Choice alternative and risk set | Uniqueness, rebranding and geographic concordance |
| quantity and market size | Purchases and eligible opportunities | Same time period and unit; positive inside and outside shares |
| price | Actual or economically equivalent transaction payment | Taxes, discounts, currency and units documented |
| ownership | Decision-making firm | Effective dates and joint ownership |
| observed characteristics | Utility-relevant measured attributes | Units, missingness and availability |
| candidate instrument | Excluded shifter | Economic exclusion argument and relevance |
| micro moment | Buyer characteristics or choice statistics | Sampling weights, linkage and population |
Estimation with individual-level choices can use maximum likelihood formed from the probability of each observed choice, conditional on choice sets and covariates. Repeated choices require appropriate treatment of within-individual dependence and panel heterogeneity. Aggregate-market structural estimation instead often uses market-share inversion and GMM moments. Maximum simulated likelihood, simulated moments, and numerical integration are distinct computational procedures, even where they use a common choice representation.
The share inversion and outer optimization introduce multiple layers of numerical error. At each candidate parameter vector, record the inner residual and convergence criterion. The outer optimizer should operate on a smooth and accurately evaluated objective, with documented gradients where available. Simulation draws or quadrature rules should be held fixed across nearby objective evaluations when practicable, while sensitivity to alternative integration grids is checked separately. Multiple starting points, parameter scaling, gradient checks, convergence reports, and the condition number of the markup or moment Jacobian belong in a reproducibility record.
Dubé, Fox, and Su (2012) investigate numerical performance of demand estimation, including the behavior of nested fixed-point computation. Conlon and Gortmaker (2020) describe PyBLP estimation practices, with attention to computation and instruments. The scientific consequence is that a software success flag is evidence about one numerical task, not independent evidence of a well-identified economic structure. A rigorous report publishes convergence tolerances, integration choices, alternative specifications, and uncertainty alongside the selected parameter estimates.
A defensible output includes elasticity and diversion matrices, own- and cross-price response distributions, outside diversion, parameter estimates with appropriate uncertainty, and a transparent counterfactual solver. For finite changes, integrate the equilibrium path or recompute shares and prices at the counterfactual rather than applying a derivative mechanically to a large shock. Uncertainty in the parameters, market size, conduct assumption, and counterfactual equilibrium should be distinguished. Bootstrap procedures or asymptotic covariance estimates need to reflect the sampling structure and any clustering of shocks. Structural misspecification requires scenario analysis even when sampling standard errors are small.
Analytical Counterexamples and Sensitivity
The following examples are author-derived illustrations using hypothetical shares and taste segments. They contain no survey observations or estimated market coefficients.
Consider two alternatives A and B and an outside option. Suppose aggregate shares equal 0.30, 0.30, and 0.40. Compare a homogeneous logit model with two equally weighted consumer types. Type one has conditional shares (0.50, 0.10, 0.40); type two has (0.10, 0.50, 0.40). Their average shares exactly reproduce the homogeneous benchmark. Each positive conditional probability vector can be generated by logit systematic utilities equal to the logarithm of the option's share divided by its outside share. Hold a common individual price coefficient of 0.10 per dollar in both models.
In homogeneous logit, the magnitude of the derivative of A with respect to its own price is 0.10 times 0.30 times 0.70, or 0.021. The cross derivative of B with respect to A's price is 0.10 times 0.30 times 0.30, or 0.009. Outside diversion has derivative 0.012. For the heterogeneous population, the own derivative magnitude is 0.10 times the average of 0.50 times 0.50 and 0.10 times 0.90, or 0.017. B's cross derivative is 0.10 times the average of 0.50 times 0.10 and 0.10 times 0.50, or 0.005. The outside cross derivative remains 0.012.
| Quantity | Homogeneous | Two preference types |
|---|---|---|
| A share | 0.30 | 0.30 |
| B share | 0.30 | 0.30 |
| Outside share | 0.40 | 0.40 |
| Absolute own derivative for A | 0.021 | 0.017 |
| Cross derivative to B | 0.009 | 0.005 |
| Cross derivative to outside | 0.012 | 0.012 |
| A-to-B diversion | 0.428571 | 0.294118 |
| A-to-outside diversion | 0.571429 | 0.705882 |
Both derivative columns satisfy adding-up: each own derivative magnitude equals the sum of cross derivatives toward the other options. The same aggregate share vector supports different assessments of competition. In the illustrative merger of A and B, a calculation based on homogeneous logit would assign substantially higher lost sales to the partner than the two-type alternative. The example demonstrates observational insufficiency of a single share vector; it does not establish that heterogeneous estimation always produces smaller diversion.
A second sensitivity concerns market size. Suppose observed sales equal 300 units for each product in one period. Under a population of 1,000 decision opportunities, shares are 0.30, 0.30, and 0.40, producing the mean utility log(0.75) for each inside good. Under 1,500 opportunities, shares are 0.20, 0.20, and 0.60, giving log(1/3). Holding alpha fixed at 0.10, homogeneous-logit diversion from A to B changes from 0.30/0.70 = 0.428571 to 0.20/0.80 = 0.25. The unit sales are unchanged. This is sensitivity to the assumed participation denominator, not an estimate of the true market size.
A third case considers a new entrant's predicted share. A differentiated product designed for a small high-valuation consumer segment may achieve a substantial welfare gain within the segment while having limited average sales. If the estimation sample poorly represents that segment, the same model can substantially mispredict both penetration and matching benefits. Reporting only category-level forecast accuracy is insufficient: validation should separately compare conditional choices of populations whose tastes are intended to determine the counterfactual.
A practical sensitivity protocol varies the outside share, taste distribution, nest structure, instruments, price definition, integration rule, and cost assumptions separately before studying combinations. The output of interest is the stability of each economically meaningful prediction. If alternative specifications that fit observed shares equally well produce divergent merger effects or welfare estimates, the scientifically appropriate conclusion is a range conditional on competing assumptions or a documented need for discriminating evidence.
Original Analytical Synthesis and Boundary Findings
The central analytical finding of this monograph is that the value of differentiation is determined by the joint structure of alternatives and demand responses, while the measurement of that value depends on the information that identifies substitution. Product distance, market shares, markup, and welfare measure different aspects of the same market. None is a universal sufficient statistic for the others.
The counterexamples establish an important connection between measurement and competition. A population composition assumption changes the cross-price derivative even when product sales and total participation remain identical. A risk-set assumption changes diversion even when actual unit sales remain identical. A utility-correlation assumption changes the welfare value of an added alternative even when its physical characteristics are unchanged. These are three independent sources of counterfactual uncertainty: preference distribution, market definition, and correlation structure. They must be examined as separate specifications because correcting one does not identify the others.
This synthesis clarifies three distinct empirical tasks. First, the researcher needs to establish the economically feasible alternatives and participation denominator. Second, the researcher needs independent variation that distinguishes preference heterogeneity and substitution among those alternatives. Third, the researcher needs a conduct and resource-cost specification when answering equilibrium and social-welfare questions. The steps are logically connected without being interchangeable. A market-power estimate that passes the third step conditional on a weak second step remains sensitive to demand assumptions; a scientifically identified demand system can itself be inadequate for an equilibrium prediction if entry or capacity is central.
The most useful validation evidence depends on which uncertainty dominates. Observing choices following the temporary unavailability of a product directly probes its substitute set, conditional on the reason for unavailability. Observing responses to an exogenous price change probes local derivatives. Observing category participation before and after access changes probes the outside option. Comparing stated and actual second choices tests the connection between surveyed preferences and observed demand. Observing costs and prices independently helps test the conduct assumptions required to infer margins.
This organization is an interpretive synthesis of the established literature and the elementary calculations above. It does not propose an independent named economic theory or claim an original econometric identification theorem. In particular, the examples demonstrate logical implications of common models rather than previously unknown mathematical facts. Their contribution is to expose the different decision inputs that can generate apparently consistent but substantively divergent market diagnoses.
Differentiation also links consumer welfare and market power in a way that deserves care. Better matching can increase consumers' willingness to pay, and the same improvement can provide sellers with an advantage against close competitors. Whether buyers gain after the equilibrium price response depends on the distribution of marginal valuation, competitive proximity, and resource cost. Treating market power and product quality as separate empirical dimensions is preferable to inferring one mechanically from the other.
The analysis supports cautious scope-specific judgments. Strong substitution is a statement about a derivative or a finite counterfactual, not a label assigned by an industry taxonomy. An attractive product is a statement about utility for a population under conditions, not necessarily a statement about market dominance. Valuable variety requires a substantive change in feasible choices and consumer valuation. These distinctions preserve the explanatory role of differentiated demand across otherwise incomparable markets.
Extensions and Unresolved Research Questions
The static single-choice model is often a useful approximation, but its limits should be made visible in applications. Durable goods require attention to the value of waiting, expectations about future prices and characteristics, used-goods markets, depreciation, and household replacement cycles. A static outside option may absorb all postponement behavior, whose determinants will change under dynamic policy. Introducing a dynamic state variable and continuation value can change the interpretation of short-run elasticities.
Multiple-category choice and baskets require a model allowing goods to be jointly purchased. In the static unit-demand setting with positive price sensitivity, all cross-price derivatives are nonnegative. In actual baskets, goods can be complementary. A bundle choice specification permits the joint consumption of products but increases the dimension of the alternative space. A model of discrete-continuous demand separates the decisions to participate, select an alternative, and determine quantity, subject to explicitly stated budget and technology constraints.
Consideration sets merit particular attention in digital markets. A consumer may observe only goods surfaced by a ranking system, have a subscription restricting alternatives, or face switching costs related to past purchases. An observed zero share can reflect invisibility or infeasibility rather than low utility. Changes in search ranking alter exposure and may simultaneously affect measured popularity. A regression using observed rankings as characteristics without a design for endogenous exposure can conflate preference with information allocation.
Product design is itself an endogenous decision. Firms may choose characteristics in response to anticipated demand; a market with identical observed prices and product attributes can have different counterfactual innovation incentives depending on design constraints and fixed costs. Extrapolating a demand model into a previously unobserved region of characteristic space therefore calls for evidence about feasibility, consumer understanding, and the distribution of preferences under novelty.
Further empirical work should distinguish changes in mean utility from changes in taste dispersion, and changes in consumer valuations from changes in accessible alternatives. Particularly useful data include randomized offers, second-choice questions validated against transactions, natural changes in distribution networks, linked household-panel observations, and transparent ownership histories. International comparability requires common attribute coding, institution-specific outside choices, and transparent translation of prices and units. Full replication of the empirical studies cited here is outside this literature-based dossier; access to underlying data and independent computational replication would be additional scientific work.
Emerging computational methods can accommodate more flexible preference distributions and richer micro moments, but their credibility remains conditional on identification, reproducible code, and a defensible comparison of observed and counterfactual choice environments. The open scientific problem is not simply greater model complexity; it is the acquisition of evidence that distinguishes consequential forms of complexity.
Conclusion
Product differentiation gives demand its structure across competing alternatives. The effects of a price increase, a new product, a merger, or an alteration of market access are determined by the distribution of consumer valuations, availability, and the relevant institutional setting. Random-utility models organize these effects in a tractable framework, while heterogeneous choice models allow the observable structure of substitution to depart from aggregate-share proportionality.
The formal derivations show why choice probabilities, derivatives, diversion, and surplus must be distinguished. The illustrative calculations show why fitting aggregate market shares alone cannot establish credible counterfactual substitution or welfare. The empirical literature demonstrates the role of product data, consumer-level information, instruments, market institutions, and numerical diagnostics. Its applications remain tied to their populations and maintained assumptions.
Reliable demand research establishes the market denominator, measures prices and characteristics consistently, identifies economically meaningful price responses, tests alternative substitution structures, and separates consumer-choice implications from the conduct and cost assumptions of equilibrium analysis. Scientific confidence should follow the evidence supporting each of those operations. A differentiated market becomes economically intelligible when the analyst can explain who substitutes toward what, under which change, with what consequences for participation, prices, and welfare.
References
- Berry, S. T. (1994). Estimating discrete-choice models of product differentiation. RAND Journal of Economics, 25(2), 242–262. Source record.
- Berry, S. T., & Haile, P. A. (2014). Identification in differentiated products markets using market level data. Econometrica, 82(5), 1749–1797. Source record.
- Berry, S., & Haile, P. (2021). Foundations of demand estimation. Cowles Foundation Discussion Paper 2301 (September 2021; revised May 2022). Yale University. Source record.
- Berry, S., Levinsohn, J., & Pakes, A. (1995). Automobile prices in market equilibrium. Econometrica, 63(4), 841–890. Working-paper version: NBER Working Paper 4264 (1993). Working-paper record.
- Berry, S., Levinsohn, J., & Pakes, A. (2004). Differentiated products demand systems from a combination of micro and macro data: The new car market. Journal of Political Economy, 112(1), 68–105. Source record.
- Conlon, C., & Gortmaker, J. (2020). Best practices for differentiated products demand estimation with PyBLP. RAND Journal of Economics, 51(4), 1108–1161. Source record.
- Conlon, C., & Gortmaker, J. (2023). Incorporating micro data into differentiated products demand estimation with PyBLP. NBER Working Paper 31605. Journal version: Journal of Econometrics, 257 (2026), DOI: 10.1016/j.jeconom.2024.105926. Working-paper record.
- Dubé, J.-P., Fox, J. T., & Su, C.-L. (2012). Improving the numerical performance of static and dynamic aggregate discrete choice random coefficients demand estimation. Econometrica, 80(5), 2231–2267. Source record.
- Gandhi, A., & Houde, J.-F. (2019). Measuring substitution patterns in differentiated-products industries. NBER Working Paper 26375. Source record.
- Hotelling, H. (1929). Stability in competition. The Economic Journal, 39(153), 41–57. Source record.
- Houde, J.-F. (2012). Spatial differentiation and vertical mergers in retail markets for gasoline. American Economic Review, 102(5), 2147–2182. Source record.
- Lancaster, K. J. (1966). A new approach to consumer theory. Journal of Political Economy, 74(2), 132–157. Source record.
- McFadden, D. (1974). Conditional logit analysis of qualitative choice behavior. In P. Zarembka (Ed.), Frontiers in econometrics (pp. 105–142). Academic Press. Linked Berkeley IURD Reprint No. 101 (1973). Source record.
- McFadden, D., & Train, K. (2000). Mixed MNL models for discrete response. Journal of Applied Econometrics, 15(5), 447–470. Source record.
- Nevo, A. (2001). Measuring market power in the ready-to-eat cereal industry. Econometrica, 69(2), 307–342. Linked UC working-paper version (1999). Author repository version.
- Petrin, A. (2002). Quantifying the benefits of new products: The case of the minivan. Journal of Political Economy, 110(4), 705–729. Working-paper version: NBER Working Paper 8227 (2001). Working-paper record.
- Train, K. E. (2009). Discrete choice methods with simulation (2nd ed.). Cambridge University Press. Source record.
- Verboven, F. L. (1996). International price discrimination in the European car market. RAND Journal of Economics, 27(2), 240–268. Source record.