AI summary by TwinMind
Lecture 07: Estimation of Demand, Supply, and Market Power, Part 1
MIT OpenCourseWare
1h 6m 36s
Instructor Introduction and Course Overview
- Instructor: Toia ZZ, office E52 fourth floor, office hours Thursdays 1-2pm
- Background: PhD from NYU, worked at Columbia before MIT, was on sabbatical last year
- Research interests: decentralized markets with frictions, digital markets/antitrust, consumer financial markets
- Course goal: teach essentials of static demand models, focus on models currently used rather than historical evolution
Why Model Demand
- Consumer welfare analysis for normative questions
- Back out supply parameters (marginal cost, fixed cost) - cost data often unavailable
- Firms consider opportunity costs (e.g., cannibalization) not captured in accounting data
- Recent literature on rising markups in US economy (Yan Dela and Yan Eout paper)
- Counterfactual analysis
- Predict demand for new goods
Model Objectives and Constraints
- Handle many products without restriction on number
- Estimate with aggregate data (cannot see individual consumer purchases or link demographics)
- Allow realistic substitution patterns (conflicts with aggregate data limitation)
- Allow unobserved product characteristics (demand shocks) - things consumers/firms observe but researchers don't
- Maintain computational tractability (conflicts with many products and realistic substitution)
- These objectives often conflict with each other
Endogeneity Problem
- Prices and quantities jointly determined in equilibrium
- Price reflects unobserved information in error term
- Price correlated with error term in demand equation
- Solution: instrumental variables (supply/demand shifters)
- Problem persists and becomes more complex in sophisticated models
- Need more rather than fewer instruments for advanced models
Product Space vs Characteristic Space Approaches
Product Space Approach
- Treat each product as unique entity
- Demand system: quantities = f(all prices, all demand shocks)
- Issue: curse of dimensionality - parameters grow at squared rate with new products
- Each new product requires new row and column in coefficient matrix
- Very demanding in terms of instruments (all prices potentially endogenous)
- Cannot predict demand for new goods
Characteristic Space Approach
- Products described by characteristics (horsepower, storage, features, etc.)
- Combined with discrete choice framework
- Consumers buy one product or none
- Build from utility functions - consumers choose product maximizing utility
- Adding products doesn't necessarily increase parameters if characteristics observed
Discrete Choice Framework
- Utility function: U = observed attributes + unobserved attributes + demographics + random shifters
- Consumers choose product with highest utility
- Random part (epsilon) helps explain different consumer choices
- Market shares determined by choice regions in epsilon space
Logit Model Solution
- Assume type 1 extreme value distribution for epsilon terms
- Provides closed-form solution for choice probabilities
- Market share = exp(utility_j) / sum(exp(utility_k) for all k)
- Normalization: outside good utility = 0
- Outside good = everything else consumers could do instead
Data Requirements
- Aggregate market shares, prices, product attributes for each market
- Distributions of consumer characteristics (not linked to specific purchases)
- From census data or similar sources
- Instruments needed for endogenous prices
- Assume no sampling uncertainty in market shares
- Requires judgment on market size definition
Simple Logit Model Limitations
- Unrealistic substitution patterns
- Cross-price elasticity only function of price coefficient and market shares
- Diversion ratios only depend on market shares
- Example: Yugo (10% share) and Lexus (10% share) would have identical substitution patterns
- Problem: only epsilon terms (IID) drive different choices
- Doesn't capture that Lexus buyers systematically different from Yugo buyers
Random Coefficient Model (BLP)
- Utility: U_ijt = X_jt * beta_i + alpha_i * p_jt + xi_jt + epsilon_ijt
- Coefficients vary across consumers (beta_i, alpha_i)
- Allows different consumers to value attributes and prices differently
- Coefficients = intercept + demographics + random component
- Don't observe individual consumer characteristics, only market distributions
Model Components
- Delta_jt = average product quality (X_jt * beta_bar + xi_jt)
- mu_ijt = consumer-specific deviations from average
- Market shares require integration over consumer distribution
- Much more complex than simple logit
Why Random Coefficients Needed
- Unobserved product characteristics (xi) needed to explain market patterns
- Example: Apple higher prices/shares despite inferior observable attributes
- Creates endogeneity: firms know xi and price accordingly
- Random coefficients allow realistic substitution patterns
- People with higher income less price sensitive, prefer luxury features
- Elasticity formulas become complex integrals over consumer distributions
Consumer Welfare
- Inclusive value formula for expected welfare before epsilon draws
- Convert to dollars by dividing by price coefficient
- Integrate over consumer distribution in random coefficient model
- Issue: welfare implications from adding/removing products due to "epsilon harvesting"
- Random coefficients may ameliorate this problem through local substitution
Instrumental Variables for BLP
Simple Logit vs BLP
- Simple logit: one instrument sufficient (linear after transformation)
- BLP: need multiple instruments due to nonlinear parameters
- Supply shifter alone insufficient to identify substitution patterns
Types of Instruments
- Different markets: Presence/absence of competitors (concern: endogenous entry)
- Cost shifters: Material costs, distribution costs, tariffs, taxes (often unavailable)
- Hausman instruments: Firm's prices in other markets (concern: demand correlation across markets)
- BLP instruments: Attributes of other products in market (concern: endogenous product attributes)
- Waldfogel instruments: Zone pricing with demographic variation (concern: correlated demographics)
Estimation Procedure
- Estimate via GMM: minimize weighted sum of moment conditions
- Moments: instruments interacted with error term (xi)
- Two-loop algorithm:
- Outer loop: Search over nonlinear parameters (Theta2)
- Inner loop: Solve for deltas using contraction mapping
- Contraction mapping: iteratively update delta using difference between observed and predicted shares
- Converges to unique solution for any guess of nonlinear parameters
- Linear parameters (beta, alpha) estimated via regression once deltas known
- Many implementation details and potential failure modes
