Title: Industrial Organization
1Industrial Organization
- Product Differentiation (II)
2Differentiation Some remarks
- Why is there differentiation?
- Increases market power PgtMC
- When is there differentiation?
- Horizontal differentiation increases with income
(more developed countries have wider product
variety) - Differentiation increases with product
fungibility, e.g. - Numerous cell phones
- Energy, telecommunication services (internet,
phone calls), commodities, are difficult to
heterogenize - Demand for quality increases with income
- Income distribution of consumers varies by
region, country.
3Differentiation Some remarks
- Hotelling and vertical models
- Few (usually), symmetric firms
- Single unit demands
- Single dimension of differentiation
- In reality
- Heterogeneity in preferences along several
dimensions - Many firms, possibly with multiple brands
- Consumers may demand multiple units
4Differentiation Empirical Work
- Again, focus on market power plus
- Antitrust applications
- Definition of markets
- Mergers
- Managerial/policy applications
- Welfare analysis (e.g. product introductions)
- But
- Environment is more complex
- Demand has to capture consumers prefernces over
many products - Supply has to capture strategic interactions with
respect to many brands and firms
5Differentiation Market Power
- Extension of conjectural variations approach to a
complex environment - Each firm f maximizes profit over its portfolio
of brands - Each firm has Ff FOC, where Ff is the number of
brands in
Bertrand-Nash Competition
6Estimation Issues
- To study market power, one needs to estimate
demand - But, J products implies the computation of J2
own- and cross-price coefficients - Assume a simple linear demand equation
- With 50 brands (for example)
- 50 equations and 25,000 cross- and own- price
coefficients. - How to reduce dimensionality?
7Estimation Issues Solutions
- Solutions
- Nest products into mutually exclusive categories
and estimate coefficients in every nest - Impose restrictions in estimation (e.g. symmetry)
Cereal (25 brands)
Kids (8)
Healthy (10)
Healthy (7)
8Estimation Issues Solutions
- Solutions
- Assume a discrete choice model to project J onto
a lower dimensional space (namely characteristics)
Each consumer has an idiosyncratic
shock Assumption IID, distributed extreme value
Logit Model (McFadden, 1978)
9Logit Discrete Choice (DC) Demand
- Parsimonious many substitution patterns
recovered via few parameters - In theory one needs each consumers purchase
decision - Search for parameters that make the logit
probability as close as possible to the actual
0/1 choice - In practice this model is widely used with
aggregate data - Daily data 100 lbs of Cereal A
(price/lb1.2) 150 pounds of Cereal B
(price/lb1.5) - Assumptions
- There is serving size (e.g. 0.1 lb/day) -
consumers purchase one serving size of the brand
that gives them the highest utility - There is a fraction of consumers that consider
buying but do not (outside good) this number is
assumed
10Logit Discrete Choice (DC) Demand
- Logit model with aggregate data
- Example
- 3,000 consumers consider buying cereal
(potential 300 lbs) Note this is unobserved
and hence assumed by researcher - Observed data 100 lbs of Cereal A (1,000
servings) 150 pounds of Cereal B (1,500
servings) - Transform aggregate quantities to market shares
(sj) - Cereal A 1/3 of market potential 0.333
market share - Cereal B ½ of market potential 0.500
- Outside good 0.1667
- Now, data looks discrete choice friendly
11Logit Discrete Choice (DC) Demand
- Very parsimonious many substitution patterns
recovered via few parameters - Logit Simplest DC model
- Independence of Irrelevant Alternatives property
off-diagonal entries in a column of elasticity
matrix are equal. - Substitution patterns are driven solely by market
shares (sj).
12Logit Discrete Choice (DC) Demand
- Nested Logit
- Products grouped into mutually exclusive sets.
- Cross-elasticities across different groups are
not restricted. - IIA property remains within groups.
- Example IIA remains for cereals within kids
category
13Logit Discrete Choice (DC) Demand
- Random Coefficients Logit
- Berry, Levinsohn and Pakes (BLP)
- Also known as mixed logit (McFadden and Train)
- Most general of DC models.
- Allows taste parameters to have a distribution
- Implication flexible substitution patterns
14Other Models Continuous Choice (CC) Demand
- Also called representative consumer models
- Not parsimonious in nature. Suitable to model
broad categories of goods. Solutions - Multistage Budgeting (e.g. Hausman, Leonard and
Zona) - Demand estimated in stages
- Bottom stage has mutually exclusive sets of
products. - Problems separability structure is difficult to
test as the number of products increases, the
problem of having to estimate too many parameters
arises again.
15Other Alternatives Continuous Choice (CC) Models
- Distance Metric (DM) Method (Pinkse, Slade,
Brett) - Based on brands location in product space (need
product characteristics data) - Intuition cross-price effects are a function of
closeness in product space to reduce
dimensionality - It does not restrict the choice of CC demand
model
16Other Alternatives Continuous Choice (CC) Models
- Distance Metric (DM)
- Example
One parameter instead of n-1
17Nearest neighbors Common Boundaries
logP
logP
A?
A?
- Location and Common Boundaries of Beers in
Chicago, 4Q-1992
18Approaches to Estimation
- Disadvantages of Continuous Choice
- Dimensionality is usually larger (J2 parameters)
- Certain analyses are difficult (e.g. evaluating
the introduction of a new brand) - Disadvantages of Characteristics Space approach
- Data on characteristics may be hard to get
- Dealing with non-discrete choice goods and
complements is difficult - Computational burden
19Estimation Remarks
- Continuous Choice Models
- Several functional forms
- Convenient Linear, log-log
- Theory based Almost Ideal Demand System
trans-log. Parameters estimated have a
theoretical meaning (e.g. you can impose
homogeneity of degree 1) - Regardless of functional form, the ultimate goal
is to obtain a measure of to conduct
empirical analyses
20Application
- Competition in the US Beer Industry (Journal of
Industrial Economics, 2008)
21Differentiation Application
- Beer in the US
- gt 200 brands
- Large degree of subjective differentiation hard
to distinguish brands in blind testing - High concentration (2 brands gt40 market)
- (P-CM)/P is high gt 30
- Question Why is there a 30 mark-up?
- Product Differentiation?
- Collusion?
- Leadership models?
22Accelerated Concentration
23Data
- Three-dimensional data on prices and quantities
58 cities, 64 brands (13 brewers), 20 quarters
(1988-1992) Source IRI - Data representative of supermarket sales in
broadly defined metropolitan areas - Product characteristics
- Continuous alcohol content, container size,
city-coverage - Discrete product type (e.g. import/domestic
light/regular budget/premium/superpremium)
brewer type national coverage (regional/national)
24Market Definitions
25(No Transcript)
26Intuition
- Mark-up depends on prices, marginal cost and
demand elasticity
Assumed
Observed
Estimated from data
Unknown
27What this paper does
Estimate demand for differentiated brands
Pre-increase period Use demand estimates in FOC
of various games to recover corresponding MCs
Add 100 tax increase to pre-increase MCs to
simulate post-increase prices for each game
Compare with actual price increases
28Demand Estimation (Results)
- SAMPLE OF PRICE ELASTICITIES (MEDIAN)
29Supply
- Firms profit function
- Using the demand parameters, marginal costs (ck)
are solved from FOCs from different games - Games - Bertrand-Nash
- - Stackelberg Leadership (Budweiser /
Anheuser-Busch) - - Collusive Leadership (All brands match
Buds increase) - - 2 Collusive scenarios
30Recovered PCMs
31Predicted Actual Price Increases
- Using ck, predicted prices are solved in FOC
above under each game when Excise tax E is
introduced
32Actual Price Increases Predicted Price
Increases Bertrand-Nash
33Actual Price Increases Predicted Price
Increases Budweiser Leadership