Chapter 1 Flashcards
(102 cards)
1
Q
In order to determine a firm’s dominance in the market…
A
We need to determine the extent of its individual market power
2
Q
Can we tell from market outcomes whether firms are imposing genuine competitive constraints on one another?
A
idk!
3
Q
The structure-conduct-performance approach (Pros and cons)
A
-Regarded as old fashioned (vs game theory)
-Presumes a CAUSAL link between the structure of the market, the nature of the competition and market outcomes (p,q, profits)
4
Q
in the SCP approach, the indicators for market power are useful…
A
For structural thresholds, but they are not used as conclusive evidence of market power.
5
Q
Structural proxies for market power are usually justified by?
A
- The Cournot model.
- Good information on MC is rare, if we assume competition we can infer probability
6
Q
Derive the Lerner Index from a Cournot model
A
7
Q
Takeaways from the Cournot theory (and Lerner index)
A
- Determine the markup of the firm using the market share and elasticity.
- If the firm has a higher share (s_i), then it can increase the markup.
- If the price sensitivity is high, then the markup is likely lower.
Thus, a high share of the market is not sufficient to ensure high markups, we need to evaluate price sensitivity as well.
8
Q
What are the most used structural indicators of profitability?
A
The Herfindahl-Hirschman Index and the Concentration ratios
9
Q
The HHI, formula
A
10
Q
Why is the HHI used?
A
As a preliminary benchmark in merger controls when post-merger data is not observed.
11
Q
How do we compare HHI pre and post merger?
A
For post s_i, we assume that it will be the sum of the s_i of the merging firms.
12
Q
What are the main disadvantages of HHI?
A
needs data on volume or value of sale for ALL companies in the market.
13
Q
Concentration index formula
A
14
Q
Criticisms of the Structure Conduct Performance approach
A
- In real life, causality is not unidirectional (econometrics).
- For Bain 1951, using accounting data is not good enough to measure MC.
- Comparisons across industry are problematic
- profits, L, and s_i should be positively correlated, but we can’t affirm a causal link from concentration to profitability:
– A more efficient firm will have lower costs and higher concentration.
– If we introduce a policy to reduce concentration, production will move from efficient to inefficient firms in this case, therefore the CS might not change or be worsened.
-Other criticism: difficulty in obtaining firm’s profits.
15
Q
Alternative to SCP
A
Instead, use Game theory approach: market structures, conduct and performance emerge together as jointly determined outcomes of a model.
16
Q
Conclusion on the SCP cons from observable data:
A
The positive correlation between profits and market structure is robust across industries, but it probably has complex causes that may differ across industries.
17
Q
Since assessing profits can be difficult, what can we use instead for Lerner?
A
(AR-AVC)/AR, but only if AVC is similar to MC
This ignores fixed costs of entry, which reduces the analysis to short term comparative statistics.
Data from financial documents (accounting) is used in other assessments, but typically not in IO.
18
Q
Directly identifying the nature of competition with GT. Why?
A
Small differences in institutional or tech characteristics can lead to very different equilibrium outcomes.
19
Q
Directly identifying the nature of competition with GT. How?
A
Use observed data on cost drivers, p and q in order to infer the nature of competition.
When we have two models that may have generated the same data, we use “identification strategies” to tell them apart (features of the data).
20
Q
What does the theoretical analysis of the formal structure models of supply and demand show?
A
When a variable affects both S and D, we will only be able to separately identify the magnitude of the effect on p and q outcomes in particular circumstances. We can generally observe reduced-form effects but we will only be able to trace back the actual parameters of the demand and supply functions in particular circumstances.
We can learn about the parameters using changes in exogenous cost and demand shifters.
21
Q
Conditions for Identification of Pricing Equations
A
Traditionally, there must be a shifter of demand that does not affect supply and a shifter of supply that does not affect demand.
By including variables in the regression that are present in one of the structural equations but not in the other, we allow one of the structural functions to shift while holding the other one fixed.
22
Q
Conduct Parameters (Bresnahan (1982)). What does he show?
A
He shows the conditions under which we can use data to tell apart three classic economic models of firm conduct, namely Bertrand price competition, Cournot quantity competition, and collusion
23
Q
Conduct Parameters (Bresnahan (1982)). What is the firm’s strategy for all three cases?
A
Firms that maximize static profits do so by equating marginal revenue to marginal costs.
However, under each of the three different models, the firms’ marginal revenue functions are different. As a result, firms are predicted to respond to a change in market conditions that affect prices in a manner that is specific to each model.
24
Q
Conduct Parameters (Bresnahan (1982)). Example: perfect competition with zero fixed costs.
A
A firm’s pricing equation is simply its marginal cost curve and hence movements or rotations of demand will not affect the shape of the supply (pricing) curve since it is entirely determined by costs. In contrast, under oligopolistic or collusive conduct, the markup over costs—and hence the pricing equation—will depend on the character of the demand curve.
25
Conduct Parameters (Bresnahan (1982)). The marginal revenue by market structure model. Nested equation and parameter meaning.
This parameter was named the “conjectural variations parameter” as it reflects firm j’s conjecture on how Q changes when qj changes (BUT name debated because identification challenges)
26
Conduct Parameters (Bresnahan (1982)). The firms’ total revenue TR in each case
If we take the derivative, we find the MR, which we can then nest into the Bresnahan equation
27
Conduct Parameters (Bresnahan (1982)). Equilibrium
MR=MC
The pricing equation encompassing these three models will depend on both the quantity and the cost variables. Its parameters are determined by the parameters of the demand function, the parameters of the cost function, and the conduct parameter, lambda.
28
Which conditions allow us to identify lambda when cost information is available?
- If the MC are constant in q, then if we can estimate the demand parameter and the gamma, we can identify lambda.
-Alternatively, if we are confident of our cost data, then we could estimate a cost function, or a marginal cost function and then we could equally potentially estimate ˇ1 directly.
29
Which conditions allow us to identify lambda when cost information is NOT available? Demand Shifts
Without information about costs, the only market events that one could use for identification are changes in demand.
30
When do demand shifts arise? And demand rotations? Do they suffice to identify conduct parameters?
Demand shifts arise, for instance, because of an increase in disposable income available to consumers for consumption. Demand shifters are not generally useful for identifying the nature of conduct in the market.
Demand rotations on the other hand must be factors which affect the price sensitivity of consumers. Demand rotators will usually suffice to estimate lambda.
31
Draw the changes in price and quantity in a market following a shift in demand from D1 to D2.
The aim of the figure is to demonstrate that the demand shift provides no power to tell the two potential underlying models apart (unless we have additional information on the level of costs) even though demand shifts do successfully trace out the pricing equation for us.
32
Identifying Conduct when Cost Information Is Not Available: Demand Rotations
Because each competition strategy sets MR=MC but has different MR, the determinants of the pricing curve are different.
Each model places a differential importance on the slope of (inverse) demand for the pricing equation.
33
Draw the changes in price and quantity in a market following a rotation in demand from D2 to D3 for a price taking firm (C) vs a monopolist (M)
If a variable affects the slope of demand, then each of the three models will make very different predictions for what should happen to prices at any given marginal cost.
In the competitive case absolutely nothing should happen to markups while a monopolist will take advantage of any decrease in demand elasticity to increase prices.
34
Effects of a demand rotation on the marginal curve of oligopolistic firms
Flatter demand and marginal revenue curves will cause firms with market power to lower their prices. On the other hand, price-taking firms will keep the price unchanged since lowering the price would cause them to price below marginal cost and make losses
35
Summary of effects of demand rotations on each competition strategy
Intuitively, demand rotations allow us to identify conduct even when we have no information about costs because such changes should not cause any response in a perfectly competitive environment, there should be some response in a Cournot market and a much larger response in a fully collusive environment.
If demand becomes more elastic, prices will decrease and quantity will increase in a market with a high degree of market power. If, on the other hand, demand becomes more inelastic and consumers are less willing to adjust their quantities consumed in response to changes in prices, then prices will increase in oligopolistic or cartelized markets. Prices will remain unchanged in both scenarios if the market is perfectly competitive and firms are pricing close to their marginal costs.
36
SCP paradigm direction of effects
37
Typical SCP regression
38
What does Bain 1951 find?
Uses data on cross-section of industries and accounting data to measure MC.
He finds that the parameter related to concentration is positive, so from the SCP paradigm standpoint, concentration increases prices.
This has a Cournot motivation, market power is proportional to market shares and elasticity
39
Why is accounting data problematic?
- Physical capital is typically valued incorrectly (not opportunity costs)
- Investments that provide future profits may be treated as expenditure at the current time even though the impact is long term
- Difficult to find accurate depreciation rates.
- Firms are not typically mono-product
40
Application related to Market Structure Endogeneity
From Mad Men to Maths Men: Concentration and Buyer Power in Online Advertising, Decarolis & Rovigatti, AER, 2021
Hypothesis?
Mergers lead to higher bargaining power of intermediaries
BUT not clear what is the effect of the resulting concentration Potential “avoid sleeping with the enemy” mechanism. Advertisers might be avoiding mergers with direct rivals
41
Application related to Market Structure Endogeneity
From Mad Men to Maths Men: Concentration and Buyer Power in Online Advertising, Decarolis & Rovigatti, AER, 2021
Empirical strategy? Issue?
Effect of market concentration (HHI) on revenues (R) using instrumental variables
In markets concentration is not exogenous/randomly assigned -> OLS may be biased because of endogeneity -> IV strategy!
42
Application related to Market Structure Endogeneity
From Mad Men to Maths Men: Concentration and Buyer Power in Online Advertising, Decarolis & Rovigatti, AER, 2021
IV strategy
for HHI: Changes in ownership of ads intermediaries’ network
M&A operations between intermediaries, especially the larger ones, are unlikely to be driven by the expectation of revenue in local markets
43
We can take profits/R=PCM (price cost margin)...
=HHI/eta
industry price-cost margin or the industry profit-sales ratio are proportional to the Herfindahl concentration index and inversely proportional to the price elasticity of demand.
44
Counterfactual analysis - What for? Which steps?
To predict the effects of changes that have
not yet occurred making use of a structural model
Two steps:
1) We use our data to estimate the structural parameters of the model
2) Based on the model and the parameters estimated, we predict the responses to changes in some parameters due to the (counterfactual) policy change.
45
Conduct when cost information is available : How + drawbacks?
Direct way to estimate the Lerner index using cost data:
1) Estimate total cost function
2) Take derivative to find MC
Issue: relies on accounting data to measure the ECONOMIC marginal cost
46
NEIO General approach when conduct is known
Derive L implied by model of conduct, as a function of demand elasticities
Estimate demand model to get elasticities and the (LI)
47
How do we find the elasticity from the demand function?
- If in level form: (-) derivative of Q wrt P times P/Q
- If in log form, it's the coefficient in front of P
48
Homogeneous products supply side, what does it depend on?
- The marginal cost in market t
- Cost shifters (input prices, technology shifters)
- Variables unobserved (to the econometrician)
49
Homogeneous products model, what are the key assumptions?
There exist suitable demand shifters and cost shifters,
Shifters are exogenous
Endogenous variables: P and Q
I Firms make pricing or quantity decisions: take other market factors as
given to maximize own profits
50
Homogeneous products model, Estimating cost and demand parameters with known conduct. Where do we get the supply equation
Take the firm's profits (assumed to be identical) and find MC
Recall that MC is not observed, we need to make an assumption on the conduct to relate it to price (which is observed)
51
Homogeneous products model, Estimating cost and demand parameters with known conduct. What do we determine when there is competition?
p=MC
52
Requirements for good instrument for Q :
- corr (P , instrument) is not 0
- corr (Q , instrument) = 0
Solution: W (exogenous cost shifter: ie workers' wages)
53
Requirements for good instrument for P:
- corr (Q , instrument) is not 0
- corr (P , instrument) = 0
Solution: X (exogenous demand shifter: ie income)
54
Stages using the IV approach
1) use OLS to estimate reduced form price and quantity regressions
2) use OLS to estimate demand and supply equation, replacing endogenous dependent variables by their predicted values Qˆ and Pˆ from first stage (predicted values have ”filtered out” the correlated part)
55
Estimating the demand and supply equations: Perfect competition
56
Estimating the demand and supply equations: Cournot
Same endogeneity problem as Bertrand, use IV!
57
Estimating conduct (Intuition from Bresnahan). Cournot with identical firms. Obraining the Lerner index depending on the conduct estimation parameter
58
How can we use phi?
Currently: phi as a parameter that permits to distinguish between three different nested models. Instead of interpreted as a direct measure of market power
59
Estimating the conduct parameter when marginal costs are known
1) Substitute in the marginal cost function the parameter relating to Q_t to beta1-phi\alpha_1
2) Obtain \alpha_1 from the demand estimation
3) if MC are constant, then b1=0-> the only unknown in the new parameter is phi
Note: this equation comes from MR=MC, where MR depends on phi-> solve for P
60
If we do not know MC, how can we identify phi?
- Taking conduct as given, the parameters could be identified with
demand and cost shifters
- Unfortunately, this is not enough to identify the conduct parameter :
--We cannot disentangle phi from the other parameters
HOWEVER, we can take advantage of variables that affect the slope of the demand function: Demand rotators!
Intuition: With a rotation in the demand curve, the three models make very different predictions on what would happen.
61
Estimating conduct when cost information not available: demand rotators
The last equation being the new pricing equation
62
Estimating conduct when cost information not available: demand rotators - Challenges
-Demand rotators: Finding an exogenous shock that a↵ects sensitivity to prices and that can be tell apart from a demand shift.
- We will rarely get estimates 0 or 1 -> we will need to test the hypothesis of phi=1
- Identification concerns on identifying marginal costs using first-order
conditions (e.g., see Genovese and Mullin, 1998)
63
Summing up and clarifications for chapter 2
64
Cartel theory from Green-Porter
65
Different types of demand models
Homogeneous vs. Differentiated Products
Discrete choice (e.g, cars) vs. continuous choice (e.g., groceries)
Product space vs. characteristics space
66
Product differentiation is a main source of market power; Bertrand case
-Bertrand with homogeneous products -> Bertrand Paradox
Bertrand with heterogeneous products -> products are imperfect
substitutes for consumers -> prices > mc
With differentiated products a source of market power arises even with Bertrand competition.
67
Challenges of heterogeneous markets:
- Different prices (and therefore also, different IVs for each price!)
- Aggregating demand and consumers’ heterogeneous preferences
68
What are the issues with aggregating demand functions with differentiated products?
Effect of income on aggregate demand is different than the effect sum of all individuals income.
Aggregating demand systems generally violate Slutsky symmetry: meaning that the effect of a main product's effect on price may reduce it's consumption considerably and provoke a huuuge increase on the non main brand, while the opposite is not true (since it has a lower quantity sold to begin with)
69
Main problems in IO Applications of Demand Estimation
- Too many parameters
- Prices are endogenous variables: prices are set by the firms, who take into account unobserved product characteristics. Thus, observed prices are correlated with the error term
- Consumer heterogeneity: difficult to compute + slutsky symmetry no longer holds in general
70
What is slutsky symmetry
71
Broad approaches to address the issues arising from demand estimation
- Representative consumer models: assume a single consumer with a taste for diversity. Commonly based on the product approach
- Discrete choice models: many consumers who each buy at most one product. Commonly based on the characteristics approach -> Nested Logit + BLP!
72
Representative consumer models
Use standard demand theory without having to aggregate.
Think of how to transform the following dimensions into the behavior of a representative consumer
73
Representative consumer models: Data in Hausman, Leonard and Zona (1994) - Multi-stage budgeting approach
Divide the market in three segments with 5 brands each: e.g., {Popular, Light, Premium}
Step 1: Estimate product demand system segment by segment: 3 X (5x5)=75 parameters
Step 2: Estimate the segment demand system: 3x3 = 9 parameters. Instead of income we use overall expenditure on beer.
Step 3: Aggregate demand for beer based on price index (1 price parameter) -> “Only” 85 parameters instead of 225
Most common multi-stage budgeting model: AID System (more details below)
74
Main disadvantages with AIDS approaches
- Inflexibility: When anything changes (more consumers, introduction of new products, imperfect availability in some markets), it is di cult to modify the AID approach to account for the changes
- No corner solutions: underlying theory of individual choice assumes no corner solution (i.e. each consumer buys a positive amount of each good)
-- This can be true for broad categories of products
---Recall: AID interprets market shares as the fraction of purchases of a
representative consumer
--But can be a strong assumption for specific markets
75
Demand equation in multi-stage budgeting (AIDS)
The representative consumer allocates total expenditures in different stages: Broad category, subcategories and the individual products.
Note: prices are replaced by price indices in the category
Income is replaced by total expenditures in the category
76
Demand equation in multi-stage budgeting (AIDS)
77
homogeneity of degree ... in individual vs aggregated demand functions (AIDS)
Individual demand functions: From choice theory, these functions are homogeneous of degree 0 in price and income.
Aggregated demand functions: homogeneity is not straight forward
-The homogeneity property survives aggregation BUT when prices and ALL consumers’ incomes increase by lambda (Debreu-Mantel-Sonnenschein theorem)
-HOWEVER the theorem no longer holds if prices and aggregate income are multiplied by but individual incomes do not
e.g., prices double, income multiplies by 10 for some individuals
78
Possible price indices for AIDS
We need indices that are invariant to the choice of units of measurement (homogeneity of degree 0)
79
Hausman IV instruments for prices in AIDS
To instrument price in a given market, use prices from other markets.
We need instruments that vary at the product level (e.g., Coke Zero, Coke regular, Pepsi Light) rather than at the product category-level (e.g., cola), as in the case of homogeneous products
Thus, cost shifters (e.g., industry wages) are not su cient because they commonly affect all categories.
Recall that we use instruments because there may be endogeneity in prices/quantities (from control variables not included and captured in the error term instead)
80
Hausman IV instruments for prices in AIDS - What does the author propose?
The authors propose using prices in one city to instrument for prices in another
-Prices in other markets for the same product in the same period I
-Assumption: uncorrelated city demand shocks
e.g., Do Madrid and Barcelona have the same shocks?? When would identification break down?
81
Elasticities in AIDS model
82
Final remarks on representative consumer models
On AID systems estimation:
- Need to make assumptions on how to divide the market in segments
- Need to address endogeneity issues for many variables (e.g., different
IV for each brand) taking advantage of panel data.
- Large data requirement. The main advantage here is computational:
we can proceed in several steps, and linear estimation in each step.
Overall for representative consumer models:
- They allow one to apply flexible specifications from traditional consumer theory.
- Relatively simple models to estimate.
- They do not account for consumer heterogeneity
83
B. Discrete choice models. Model setup
Each consumer buys one product. We analyze “which car” (rather than “how many cars”)
To avoid the “too many parameters” problem ! Considers products as a bundle of characteristics
-cars: horsepower, size, fuel, economy
-drugs: pack size, active substance, format, price
-yogurt: fat content, flavor, pack size
-computer: memory, speed, screen size?
Important identification assumption: Product characteristics (other than price) are exogenous
84
Popular Discrete Choice Model: The Multinomial Logit. Virtues of the model
Individual random utility that allows for wide consumer heterogeneity
Tractability: Despite the heterogeneity, the resulting demand functions
are entirely analytical and easy to estimate
85
When is IIA reasonable
When there is low individual unobserved heterogeneity
86
Multinomial Logit general approach
87
Logit approach: 1.-Individual utility
88
Assumptions in the logit model
- All consumers are assumed to have the same valuations for observed product characteristics ad price
- the mean utility of the outside option is normalized to 0
- price of the outside option is determined exogenously
89
the P that a random consumer i buys in the logit model
coincides with the expected market share
90
What is the process of the logit model? (extreme value function)
The utility functions have a common element across individuals and an individual error econometric error term.
Said error term is modeled as iid and as extreme value function (logit distribution)
For the derivation of choice probabilities with extreme value functions, we compute the probability that an option has higher random utility that all the other options.
91
What is the process of the logit model? Market shares
From the random utility model we can compute how likely it is that individual i buys product j from:
Individuals choose the product out of the J + 1 products that maximizes utility
The probability that individual i chooses product j takes the standard logit form
Note that the choice probability depends on delta, which is the vector of mean utilities
92
Estimation problems in the logit model for discrete choices (in terms of the market share system)
Parameters enter highly non-linearly
error term enters non-linearly
93
Logit approach 5. Inverting the market share system
Equating the market share logit function to the observed market share (qj/L) gives us a system of equations... although highly nonlinear!
Berry shows that from this system we can recover mean utility
94
95
96
For 4, it would be good to mention that we need BLP instruments to be independent of a product’s specific demand shocks. This is the case if you assume that competitors can’t/don’t adjust their observed product characteristics in response to unobserved product characteristics/demand shocks of the specific product.
● Each market is a country/year combination. Therefore, your instruments should be calculated at a country/year level, instead of just a year level. Notice the same car will show up multiple times if you do it across all countries in the same year, so the instruments don’t make sense. I’m guessing this is what caused your weird results later for IV versions of the model and the questions using the IV estimated coefficients
97
Caveats of the logit model:
Key drawback: problematic implications of own- and cross-price elasticities
Own-elasticity is linearly increasing in price, which is somewhat unrealistic (e.g., we could think that people that buy more expensive products are less sensitive to prices)
ii.- Independence of Irrelevant Alternatives: The model leads to substitution patterns that do not seem plausible.
Due to iid assumption, by construction market shares do not depend on how close products are.
Cross-elasticity depends only on market shares and prices but not on similarities between goods. This is typically called the IIA property
*** Most extensions try to correct for the above
98
The multinomial logit imposes IIA by construction because market shares are determined by the utilities provided by each product
! Two (almost) identical products always have equal market shares
This makes sense BUT IIA implies that we get contraintuitive consequences when for instance an alternative is added or subtracted.
Very unrealistic! By adding a red bus that is almost identical to the green bus, I am considerably reducing the market share of “walking.”
99
IIA and how to relax it
When adding a new product, the relative probability of choosing between any of the two initial choices in the group remains unchanged.
Solution: Nested logit: we create nests of products that are closer substitutes. E.g., first I choose between walk or bus. Then I choose which type of bus.
100
Nested logit with aggregate level data
Berry 1994 model
101
Possible BLP instruments for price endogeneity of a product j produced by firm f :
IV1: Sum product characteristics of other products of the same firm
IV2: Sum product characteristics of products of all other firms
IV3 for nested logit: sum product characteristics of products of the same segment
102
