Industrial Organization

Class 2: Introduction

Yasutora Watanabe
HKUST

February 2, 2017

Yas () Class 2 February 2, 2017 1 / 31

Paper Today

Robert H. Porter, 1983. “A Study of Cartel Stability: The Joint Executive

Committee, 1880-1886,” Bell Journal of Economics, vol.14(2), pp.
301-314.
A classical paper taking structural approach in empirical IO.
The …rst paper to consider structural model of a cartelized industry.
An excellent example of how structural work should be conducted.

Yas (HKUST) Class 2 February 2, 2017 2 / 31

Theoretical Background 1: Modelling Competition

Consider a duopolistic market:

two identical …rms (the same cost function) o¤er homogeneous product
two …rms play one-shot price setting game (Bertrand competition)
p = MC is the unique (pure strategy) Nash equilibrium, which is
unrealistic in reality.
Doubtful assumptions:
Do they o¤er homogeneous product?
Are they two identical …rms (the same cost function)?
Are they playing Bertrand competition for in…nitely many times?

Yas (HKUST) Class 2 February 2, 2017 3 / 31

Theoretical Background 2: Tacit Collusion

Consider the following trigger strategy

In period 1, each …rm charges monopoly price, p m .
If both …rms have charged p m in every period preceding t, each …rm
charges p m in period t.
Otherwise, it sets its price at MC forever. ! A single deviation triggers
a halt in the cooperation.
This trigger strategy can be a subgame perfect Nash eq.(SPNE), if
the future discount rate (δ) is su¢ ciently high.
Tacit collusion: Firms behave sel…shly (maximizes its pro…t)
independently, but succeed to avoid prisoner’s dilemma and sustain a
collusion.

Yas (HKUST) Class 2 February 2, 2017 4 / 31

Theoretical Background 3: Green and Porter (1984)

One more ingredients: Imperfect price information

A …rm cannot observe the demand curve it faces (unobserved shocks)
nor rivals’price. Only the market price can be observed.
Detecting and deterring cheating is not easy, i.e., unexpected low
price might be the cause of negative demand shock or a cheat.
To sustain a cartel, …rms employ a trigger strategy or a Nash
reversion strategy.
Price wars occur after …rms observe an unexpected fall in price. Even
without a secret price cut, price wars actually occur with a certain
probability.
A predicted fall in demand will lead to an adjustment of the
cooperative price, but not a price war.

Yas (HKUST) Class 2 February 2, 2017 5 / 31

Introduction 1: Research Question

Green and Porter (1984) presented a model that explains how a cartel
can be sustained by individual pro…t maximization and price wars
occur under imperfect price information.
There are two competition regimes: tacit collusion and price wars.
Can we empirically distinguish between these two competition
regimes?
A drop in prices does not necessarily mean a price war. There might be
other reasons.
Thus, the reduced form approach would not work
We need suitable data, economic theory, and an appropriate structural
econometric model.

Yas (HKUST) Class 2 February 2, 2017 6 / 31

Introduction 2: Overview of the paper

Industry: The Joint Executive Committee – a cartel among US

rail-road …rms shipping grain from Chicago to the East Coast in the
late 19th century.
Data: Weekly time series data from 1880 to 1886 on
prices
aggregate output level (not individual …rm level)
before the anti-trust law.
Economic Model: Homogeneous product, in…nitely repeated game,
behavioral switches between cooperative and punishment periods.
Econometric model: A simultaneous equation switching regressions
model.

Yas (HKUST) Class 2 February 2, 2017 7 / 31

Industry Background: The Joint Executive Committee

Was it illegal?
The Joint Executive Committee (JEC) was a cartel which controlled
eastbound freight shipments from Chicago to the Atlantic seaboard in
the 1880s.

Yas (HKUST) Class 2 February 2, 2017 8 / 31

Industry Background: The Joint Executive Committee

Was it illegal?
The Joint Executive Committee (JEC) was a cartel which controlled
eastbound freight shipments from Chicago to the Atlantic seaboard in
the 1880s.
It was formed in April 1879 by an agreement of the rail-roads involved
in the market.

Yas (HKUST) Class 2 February 2, 2017 8 / 31

Industry Background: The Joint Executive Committee

Was it illegal?
The Joint Executive Committee (JEC) was a cartel which controlled
eastbound freight shipments from Chicago to the Atlantic seaboard in
the 1880s.
It was formed in April 1879 by an agreement of the rail-roads involved
in the market.
The …rms involved publicly acknowledged this agreement, as it
preceded the passage of the Sherman Act (1890) and the formation
of the Interstate Commerce Commission (1887).

Yas (HKUST) Class 2 February 2, 2017 8 / 31

Industry Background: The Joint Executive Committee

What was the cartel mechanism?

Each …rm individually (and secretly) sets price.

Yas (HKUST) Class 2 February 2, 2017 9 / 31

Industry Background: The Joint Executive Committee

What was the cartel mechanism?

Each …rm individually (and secretly) sets price.
Each …rm must report the quantity to the JEC and the industry total
amount transported is publicized.

Yas (HKUST) Class 2 February 2, 2017 9 / 31

Industry Background: The Joint Executive Committee

What was the cartel mechanism?

Each …rm individually (and secretly) sets price.
Each …rm must report the quantity to the JEC and the industry total
amount transported is publicized.
Firms can detect cheating, though they cannot detect who is the
cheater. This makes the cartel sustainable.

Yas (HKUST) Class 2 February 2, 2017 9 / 31

Industry Background: The Joint Executive Committee

What was the cartel mechanism?

Each …rm individually (and secretly) sets price.
Each …rm must report the quantity to the JEC and the industry total
amount transported is publicized.
Firms can detect cheating, though they cannot detect who is the
cheater. This makes the cartel sustainable.
But, the demand each …rm faces depended not only on the prices of
other …rms but also unpredictable demand ‡uctuations. Detecting a
cheat is not straightforward.

Yas (HKUST) Class 2 February 2, 2017 9 / 31

Industry Background: The Joint Executive Committee

What was the cartel mechanism?

Each …rm individually (and secretly) sets price.
Each …rm must report the quantity to the JEC and the industry total
amount transported is publicized.
Firms can detect cheating, though they cannot detect who is the
cheater. This makes the cartel sustainable.
But, the demand each …rm faces depended not only on the prices of
other …rms but also unpredictable demand ‡uctuations. Detecting a
cheat is not straightforward.
To enforce the cartel mechanism, a (variant of) trigger strategy had
been employed: If the cartel member …rms think that cheating has
occurred, they cut prices for a certain period, and then return to the
collusive price.

Yas (HKUST) Class 2 February 2, 2017 9 / 31

Industry Background: The Joint Executive Committee

Assumptions
Is the product homogeneous?

Yas (HKUST) Class 2 February 2, 2017 10 / 31

Industry Background: The Joint Executive Committee

Assumptions
Is the product homogeneous?
Yes. The transportation of grain, which does not require the speed.

Yas (HKUST) Class 2 February 2, 2017 10 / 31

Industry Background: The Joint Executive Committee

Assumptions
Is the product homogeneous?
Yes. The transportation of grain, which does not require the speed.
Most freight was for export and price should be adjusted depending on
the destinations.

Yas (HKUST) Class 2 February 2, 2017 10 / 31

Industry Background: The Joint Executive Committee

Assumptions
Is the product homogeneous?
Yes. The transportation of grain, which does not require the speed.
Most freight was for export and price should be adjusted depending on
the destinations.
No entry/exit?

Yas (HKUST) Class 2 February 2, 2017 10 / 31

Industry Background: The Joint Executive Committee

Assumptions
Is the product homogeneous?
Yes. The transportation of grain, which does not require the speed.
Most freight was for export and price should be adjusted depending on
the destinations.
No entry/exit?
During the sample period, entry occurred twice.

Yas (HKUST) Class 2 February 2, 2017 10 / 31

Industry Background: The Joint Executive Committee

Assumptions
Is the product homogeneous?
Yes. The transportation of grain, which does not require the speed.
Most freight was for export and price should be adjusted depending on
the destinations.
No entry/exit?
During the sample period, entry occurred twice.
No exit constraint.

Yas (HKUST) Class 2 February 2, 2017 10 / 31

Industry Background: The Joint Executive Committee

Demand Shifter: Great Lakes

Great Lakes is the principal competitors for the JEC.

Yas (HKUST) Class 2 February 2, 2017 11 / 31

Industry Background: The Joint Executive Committee

Demand Shifter: Great Lakes

Great Lakes is the principal competitors for the JEC.
The Great Lakes operation had a big impact on demand for the
railroad transportation and prices.

Yas (HKUST) Class 2 February 2, 2017 11 / 31

Industry Background: The Joint Executive Committee

Demand Shifter: Great Lakes

Great Lakes is the principal competitors for the JEC.
The Great Lakes operation had a big impact on demand for the
railroad transportation and prices.
Need to control for the Lake navigation season.

Yas (HKUST) Class 2 February 2, 2017 11 / 31

Industry Background: The Joint Executive Committee

Demand Shifter: Great Lakes

Great Lakes is the principal competitors for the JEC.
The Great Lakes operation had a big impact on demand for the
railroad transportation and prices.
Need to control for the Lake navigation season.
However, this is predicted demand ‡uctuations and no need for
punishment.

Yas (HKUST) Class 2 February 2, 2017 11 / 31

Econometric Model: Demand Side

The total quantity, Qt , is assumed to be a function of price, pt :

log(Qt ) = α0 + α1 log(pt ) + α2 Lt + U1,t

where
α1 is a price elasticity of demand and supposed to be negative.
Lt is a dummy variable which take one if the Great Lakes were open
to navigation.
α2 is also supposed to be negative.
fU1,1 , ..., U1,T g is a sequence of independently distributed normal
variables with zero mean and variance σ21 .

Yas (HKUST) Class 2 February 2, 2017 12 / 31

Econometric Model: Supply Side

There are N …rms with di¤erent cost functions. The cost of producing
output qit for …rm i is

Ci (qit ) = ai qitδ + Fi , for i = 1, ..., N

where
ai is a …rm-speci…c cost shift parameter
Fi is a …xed cost for …rm i
δ, the (constant) elasticity of variable costs with respect to output,
must exceed one to guarantee the existence of an equilibrium.

∂VCi qit
=δ
∂qit VCi

Yas (HKUST) Class 2 February 2, 2017 13 / 31

Econometric Model: Equilibrium Condition
Porter considers three di¤erent behavioral assumptions and claim all three
can be summarized into one equation:

θ it
pt 1+ = MCi for i = 1, ..., N
α1

Case 1: (Bertrand competition) The equilibrium condition is

pt = MCi (qit ), implying θ it = 0.
∂π it
Case 2: (Cournot competition) The equilibrium condition is ∂q it = 0.
∂Q t p it q it
Using = α1 , we have θ it =
∂p t Q t = sit . Qt
Case 3: (Joint pro…t maximization) The equilibrium condition is
MR = MCi (qit ), implying θ it = 1.
This is very clever method. The value of θ can be interpreted as
the intensity of competition from Bertrand competition to
monopoly.
Yas (HKUST) Class 2 February 2, 2017 14 / 31
Econometric Model: Market Share

Porter assumes that …rms charge equal prices in equilibrium, as the

product is homogeneous (no product di¤erentiation).
The market shares can be derived as:
1/(1 δ)
ai
sit = = si
∑ j aj
1/(1 δ)

The share only depend on ai (cost parameter) and δ (elasticity).

Market share is decreasing in ai , because δ > 1.

Yas (HKUST) Class 2 February 2, 2017 15 / 31

Econometric Model: The Supply Relationship

Only aggregate level data were available. Need to aggregate:

θ it
pt 1+ = MCi for i = 1, ..., N
α1

Porter aggregate them using weighted average. Let θ t = ∑N

i =1 sit θ it ,

N
θt
pt 1+
α1
= ∑ sit MCi .
i =1

1/(1 δ) 1 δ
Furthermore, de…ning D = δ ∑i ai , we have

θt 1
pt 1+ = DQtδ .
α1

Yas (HKUST) Class 2 February 2, 2017 16 / 31

Econometric Model: Supply Function
To estimate the model, we need an empirical structural supply function.
Rearranging the aggregate market condition:
θt 1
pt 1+ = DQtδ .
α1
we have
θt
log pt = log D + (δ 1) log Qt log 1 +
α1
Then, we can
log pt = β0 + β1 log Qt + β2 St + β3 It + U2,t
where
It is an indicator variable which equals to one if cooperative and 0
otherwise.
St is a vector of structural dummies to control for entry etc...
fU2,1 , ..., U2,T g is a sequence of independently distributed normal
variables with zero mean and variance σ22 and Cov (U1,t , U2,t ) = σ12 .
Yas (HKUST) Class 2 February 2, 2017 17 / 31
Econometric Model: Interpretation

log pt = β0 + β1 log Qt + β2 St + β3 It + U2,t

β1 measures δ 1 > 0 and is supposed to be positive.

β3 measures the average di¤erence between two regimes:
In case of the Bertrand competition, β3 approaches to 0
α1
In case of joint pro…t maximization, β3 approaches to log 1 + α1

α1
Thus, β3 should be between 0 and log 1 + α1

Yas (HKUST) Class 2 February 2, 2017 18 / 31

Estimation: Simultaneous Equation

Now we have two structural equations to estimate:

log(Qt ) = α0 + α1 log(pt ) + α2 Lt + U1,t

log pt = β0 + β1 log Qt + β2 St + β3 It + U2,t

How do we estimate this simultaneous equation model? How

should we deal with It ? Porter estimate two versions of the model:
1 Estimate the model using the observed It.

Yas (HKUST) Class 2 February 2, 2017 19 / 31

Estimation: Simultaneous Equation

Now we have two structural equations to estimate:

log(Qt ) = α0 + α1 log(pt ) + α2 Lt + U1,t

log pt = β0 + β1 log Qt + β2 St + β3 It + U2,t

How do we estimate this simultaneous equation model? How

should we deal with It ? Porter estimate two versions of the model:
1 Estimate the model using the observed It.
Porter obtained a series of data for It from a news paper and an
industry journal.

Yas (HKUST) Class 2 February 2, 2017 19 / 31

Estimation: Simultaneous Equation

Now we have two structural equations to estimate:

log(Qt ) = α0 + α1 log(pt ) + α2 Lt + U1,t

log pt = β0 + β1 log Qt + β2 St + β3 It + U2,t

How do we estimate this simultaneous equation model? How

should we deal with It ? Porter estimate two versions of the model:
1 Estimate the model using the observed It.
Porter obtained a series of data for It from a news paper and an
industry journal.
Cartel is not illegal at that point.

Yas (HKUST) Class 2 February 2, 2017 19 / 31

Estimation: Simultaneous Equation

Now we have two structural equations to estimate:

log(Qt ) = α0 + α1 log(pt ) + α2 Lt + U1,t

log pt = β0 + β1 log Qt + β2 St + β3 It + U2,t

How do we estimate this simultaneous equation model? How

should we deal with It ? Porter estimate two versions of the model:
1 Estimate the model using the observed It.
Porter obtained a series of data for It from a news paper and an
industry journal.
Cartel is not illegal at that point.
However, it turns out that this It is not so precise.

Yas (HKUST) Class 2 February 2, 2017 19 / 31

Estimation: Simultaneous Equation

Now we have two structural equations to estimate:

log(Qt ) = α0 + α1 log(pt ) + α2 Lt + U1,t

log pt = β0 + β1 log Qt + β2 St + β3 It + U2,t

How do we estimate this simultaneous equation model? How

should we deal with It ? Porter estimate two versions of the model:
1 Estimate the model using the observed It.
Porter obtained a series of data for It from a news paper and an
industry journal.
Cartel is not illegal at that point.
However, it turns out that this It is not so precise.
2 Estimate the model and It together.

Yas (HKUST) Class 2 February 2, 2017 19 / 31

Estimation: With Observed Data for It

If we have the data for It , the model is a typical simultaneous

equation model, i.e., demand and supply.
Identi…cation condition is exclusion restrictions or the rank condition.
Once the identi…cation condition is satis…ed, the model can be
consistently estimated by two stage least squares (2SLS) under usual
assumptions (see, e.g., Hayashi, pp190-192)

Yas (HKUST) Class 2 February 2, 2017 20 / 31

Estimation: Without Data for It

Assume that It follows a simple binomial (or Bernoulli) distribution:

1 with prob λ (cooperation)

It =
0 with prob 1 λ (price wars)

This model is a simultaneous equation switching regression model

How to estimate this model? OLS or 2SLS will not work because of
nonlinearity.
We specify all the distribution for error terms - we can use maximum
likelihood!

Yas (HKUST) Class 2 February 2, 2017 21 / 31

Estimation: Introducing Matrix Notation

log(Qt ) = α0 + α1 log(pt ) + α2 Lt + U1,t

log pt = β0 + β1 log Qt + β2 St + β3 It + U2,t

Now, we can summarize two equations by writing

Byt = ΓXt + ∆It + Ut

where 2 3
1
log Qt U1,t
yt = , Xt = 4 L t 5 , U t = ,
log pt U2,t
St
and the coe¢ cients vectors:
1 α1 0 α0 α2 0
B= ,∆ = ,Γ = ,
β1 1 β3 β0 0 β2

Yas (HKUST) Class 2 February 2, 2017 22 / 31

Estimation: Introducing Matrix Notation

Ut is identically and independently distributed as N (0, Σ) where

σ21 σ12
Σ=
σ12 σ22

Remember the standard normal PDF:

1/2 z2
φ(z ) = (2π ) exp
2
or
1/2 1
φ(z ) = (2π ) k /2
jΣj exp (x µ)0 Σ 1
(x µ)
2
for multivariate normal distribution.

Yas (HKUST) Class 2 February 2, 2017 23 / 31

Estimation: Likelihood with It

The PDF of yt conditional on It is:

kB k
h (yt jIt ) = jΣj 1/2
2π
1
exp (Byt ΓXt ∆It )0 Σ 1
(Byt ΓXt ∆It )
2

Then, the likelihood function is

T
L(∆, B, Γ, Σ) = ∏ h(yt jIt )
t =1

Notice that we can estimate the model by this ML problem if the

sequence is available. However, 2SLS is much simpler.

Yas (HKUST) Class 2 February 2, 2017 24 / 31

Estimation: Likelihood without It

The PDF of yt conditional on It is:

kB k 1/2
h (yt jIt ) = jΣj
2π
λ exp 1
2 (Byt ΓXt ∆It )0 Σ 1 (Byt ΓXt ∆It )
+(1 λ) exp 1
2 (Byt ΓXt )0 Σ 1 (Byt ΓXt )

Then, the likelihood function is

T
L(∆, B, Γ, Σ) = ∏ h(yt jIt )
t =1

Yas (HKUST) Class 2 February 2, 2017 25 / 31

Data: List of variables

GR: Grain rate, in dollars per 100lbs. (pt )

TQG: Total quantity of grain shipped, in tons. Qt
LAKES: Dummy variable for Great Lakes operation period.
PO: Cheating dummy observed in data.
PN: Estimated cheating dummy.
DM1: Dummy re‡ecting entry by the Grand Trunk Railway.
DM2: Dummy re‡ecting an addition to NY Central.
DM3: Dummy re‡ecting entry by the Chicago and Atlantic.
DM4: Dummy re‡ecting departure of the Chicago and Atlantic from
the JEC.
Other: Monthly dummies are also used to adjust for the seasonality

Yas (HKUST) Class 2 February 2, 2017 26 / 31

Result

Yas (HKUST) Class 2 February 2, 2017 27 / 31

Result

Yas (HKUST) Class 2 February 2, 2017 28 / 31

Result

Generally speaking, all variables have expected sign and are

signi…cantly di¤erent from zero.
ML produces much better …t than 2SLS.
Price is signi…cantly higher in cooperative periods.
The relationship
θt
β3 = 1+
α
enables us to recover θ t , and computed value of θ t is about
0.336. This number is close to the Cournot competition case.

Yas (HKUST) Class 2 February 2, 2017 29 / 31

Conclusion

The author con…rmed the statistical evidence that switches actually

occurred and were statistically signi…cant.
The estimates show signi…cantly lower prices and lower revenues
during the Nash revision (or punishment) periods.
As predicted by the Green and Porter (1984)
The cartel price is substantially higher than Bertrand but not as high as
the level that maximizes the joint pro…t.
PO and PN are not related to the opening/closing of Great Lakes
(expected demand shock).
Frequency of revisions increases with the number of participants in the
market.

Yas (HKUST) Class 2 February 2, 2017 30 / 31

Comments

The paper o¤ers an understanding about market power, which cannot

be observed directly.

Yas (HKUST) Class 2 February 2, 2017 31 / 31

Comments

The paper o¤ers an understanding about market power, which cannot

be observed directly.
The data used here is quite simple, (p,Q), and some demand and
supply shifters.

Yas (HKUST) Class 2 February 2, 2017 31 / 31

Comments

The paper o¤ers an understanding about market power, which cannot

be observed directly.
The data used here is quite simple, (p,Q), and some demand and
supply shifters.
Structural estimation allows us to go very far!! Much further than
descriptive work.

Yas (HKUST) Class 2 February 2, 2017 31 / 31

Comments

The paper o¤ers an understanding about market power, which cannot

be observed directly.
The data used here is quite simple, (p,Q), and some demand and
supply shifters.
Structural estimation allows us to go very far!! Much further than
descriptive work.
Note that you do not have to model everything - indeed Porter has a
lot of non-structural components.

Yas (HKUST) Class 2 February 2, 2017 31 / 31

Comments

The paper o¤ers an understanding about market power, which cannot

be observed directly.
The data used here is quite simple, (p,Q), and some demand and
supply shifters.
Structural estimation allows us to go very far!! Much further than
descriptive work.
Note that you do not have to model everything - indeed Porter has a
lot of non-structural components.
At the same time, do you think the results are reliable? The results
are heavily rely on the “structures” – the economic, functional form
and stochastic assumptions

Yas (HKUST) Class 2 February 2, 2017 31 / 31

Comments

The paper o¤ers an understanding about market power, which cannot

be observed directly.
The data used here is quite simple, (p,Q), and some demand and
supply shifters.
Structural estimation allows us to go very far!! Much further than
descriptive work.
Note that you do not have to model everything - indeed Porter has a
lot of non-structural components.
At the same time, do you think the results are reliable? The results
are heavily rely on the “structures” – the economic, functional form
and stochastic assumptions
Each assumption must be carefully examined and justi…ed.

Yas (HKUST) Class 2 February 2, 2017 31 / 31

Comments

The paper o¤ers an understanding about market power, which cannot

be observed directly.
The data used here is quite simple, (p,Q), and some demand and
supply shifters.
Structural estimation allows us to go very far!! Much further than
descriptive work.
Note that you do not have to model everything - indeed Porter has a
lot of non-structural components.
At the same time, do you think the results are reliable? The results
are heavily rely on the “structures” – the economic, functional form
and stochastic assumptions
Each assumption must be carefully examined and justi…ed.
But these are common problems any empirical research that attempts
to go beyond simple correlations.

Yas (HKUST) Class 2 February 2, 2017 31 / 31