Chapter 17

EMPIRICAL STUDIES OF INDUSTRIES WITH

MARKET POWER
TIMOTHY F. BRESNAHAN*
Stanford University

Contents
1. Introduction 1012
2. Oligopoly theory and the measurement of market power 1014
2.1. Notation, cost, and demand 1014
2.2. Alternative treatments of firm conduct and of 0 1019
3. How the data identify market power 1031
3.1. Comparative statics in demand 1032
3.2. Comparative statics in cost 1034
3.3. Estimation of marginal cost more directly 1039
3.4. Supply shocks 1041
3.5. Comparative statics in industry structure 1042
3.6. On the identification of market power 1045
4. Market power in product-differentiated industries 1045
4.1. "Market definition", policy analysis, and product differentiation 1049
4.2. Product differentiation in the LR 1050
5. What has been learned about market power? 1051
6. The future: The sources of market power 1053
6.1. Predation 1054
6.2. Entry 1054
6.3. Final remark 1055
References 1055

*I owe a tremendous debt to the editors of this handbook for their encouragement and comments.
Their contribution is so large that it would be diminished if I named any of the other very helpful
friends and colleagues who contributed ideas and discussion. Though Richard E. Quandt's name does
not appear in the reference list, his influence on the field shows in dozens of papers. The remaining
errors are mine.

Handbook of Industrial Organization, Volume II, Edited by R. Schmalensee and R.D. Willig
Elsevier Science Publishers B.V., 1989
1012 T.F. Bresnahan

1. Introduction

This chapter treats econometric studies of market power in single markets and in
groups of related markets. The recent increase in the number of such studies and
substantial advances in the methods for carrying them out constitute a dramatic
shift in the focus of empirical work in the industrial organization (IO) field. The
new literature treated here is based largely on time series data from single
industries, or on data from closely related markets. It has taken a markedly
different view of what can be observed, and how economic quantities are to be
measured, than earlier work did. At some risk of oversimplifying a growing and
varied literature, I summarize the new approach as having these central ideas:
Firms' price-cost margins are not taken to be observables; economic marginal
cost (MC) cannot be directly or straightforwardly observed. The analyst infers
MC from firm behavior, uses differences between closely related markets to
trace the effects of changes in MC, or comes to a quantification of market
power without measuring cost at all.
Individual industries are taken to have important idiosyncracies. It is likely that
institutional detail at the industry level will affect firms' conduct, and even
more likely that it will affect the analyst's measurement strategy. Thus, practi-
tioners in this literature are skeptical of using the comparative statics of
variations across industries or markets as revealing anything, except when the
markets are closely related.
Firm and industry conduct are viewed as unknown parameters to be estimated.
The behavioral equations by which firms set price and quantity will be
estimated, and parameters of those equations can be directly linked to analyti-
cal notions of firm and industry conduct.
As a result, the nature of the inference of market power is made clear, since the
set of alternative hypotheses which are considered is explicit. The alternative
hypothesis of no strategic interaction, typically a perfectly competitive hypothe-
sis, is clearly articulated and is one of the alternatives among which the data
can choose.

This "new empirical industrial organization" (NEIO) is clearly somewhat

different than the previously dominant empirical method in the field, the struc-
ture-conduct-performance paradigm (SCPP).
For the quarter century following the pioneering work of Bain (1951), the focus
of the SCPP was the cross-section study of many industries. Industry and firm
profits were predicted from various structural measures. The NEIO is partly
motivated by dissatisfactions over three maintained hypotheses in the SCPP: (i)
Ch. 17: Industries with Market Power 1013

economic price-cost margins (performance) could be directly observed in

accounting data, (ii) cross-section variation in industry structure could be cap-
tured by a small number of observable measures, and (iii) empirical work should
be aimed at estimating the reduced-form relationship between structure and
performance. 1 Furthermore, the SCPP has been caught in a kind of gridlock over
the question of whether high accounting profits are to be interpreted as a sign of
good or of bad performance. 2 The SCPP did, however, introduce something into
the field of tremendous value: systematic statistical evidence. The NEIO is an
attempt to continue the use of such evidence while returning to the study of
single (or related) industries. On its more optimistic days, the NEIO therefore
sees itself as taking the best from the two great empirical IO traditions: SCPP
and industry case studies.
A typical NEIO paper is first and foremost an econometric model of an
industry. Thus the new literature has been able to draw closely on economic
theory to guide specification and inference in the empirical models. Quite a bit of
this chapter will treat method: it will attempt to provide a review of the manner
in which theory has been applied, and the way in which economic inferences have
been drawn from the empirical work. Much of the work in the literature has been
focused on one set of issues, those surrounding price and quantity determination
in oligopoly. The major subtopics covered include the formation and enforcement
of tacitly collusive arrangements, the nature of noncooperative oligopoly interac-
tion in the world, the degree of single-firm market power under product differ-
entiation, and the size and determinants of the industry price-cost margin. The
next two sections take up the question of measuring market power in con-
centrated single product industries; Section 2 reviews the various empirical
models of monopoly power and of oligopoly interaction, and Section 3 covers the
theoretical and empirical arguments for why it is monopoly power that is being
measured. Section 4 takes up the question of measuring market power in the
presence of product differentiation. These three sections form the bulk of the
chapter, as the material they cover forms the bulk of the literature. The chapter
has two conclusions. The first is a review of what the NEIO has learned about the
extent of market power, the second a biased view of where the literature should
go to learn more about the sources of market power.

lThis is the extreme "structuralist" view associated primarily with Bain. [See Scherer (1980, ch. 1)
for the label.] Other positions which take more of a view that conduct is sometimes observable are not
importantly different for my purposes. The SCPP and NEIO remain very distinct on how perfor-
mance and conduct are to be measured in any view.
2See Demsetz (1974) and Schmalensee, Chapter 16 in this Handbook. The conventional story is
that high profit measures poor performance, i.e. measures the Lerner index ( ~ ) . Demsetz' alternative
interpretation was that high profits measure good performance, i.e. low costs, an argument he
buttressed by the observation that much of the profit in concentrated industries goes to large firms.
These firms might therefore be large because of cost advantages. Bain (1951) had already provided a
"poor performance" interpretation of this observation, however.
1014 T.E Bresnahan

2. Oligopoly theory and the measurement of market power

Many of the advances in methodology for measuring market power can be seen
most clearly in a stylized econometric model of oligopoly interaction in a
single-product industry. The central inference in the stylized model is about firm
and industry conduct: the goal is the estimation of parameters measuring the
degree of competition. In parallel to laying out the stylized model, I will follow a
single specific treatment, namely Porter's (1984) study of strategic interaction
among nineteenth-century railroads. This organization is slightly repetitive, but
permits the simultaneous treatment of two topics: the relationship of the em-
pirical inferences to oligopoly theory, and their relationship to the data.
The stylized model has three sets of unknown parameters: costs, demand, and
firm conduct. The observable variables that are endogenous to the industry
equilibrium include industry price and each firm's quantity (sometimes only
industry quantity) in time series; price-cost margins are not taken to be directly
observable. This focus was reflected early in the rhetoric of the literature: in his
title, Posse (1970) cast the econometrician's problem as "Estimating Cost Func-
tion Parameters Without Using Cost Data". The observables are also taken to
include variables that shift cost and demand functions. 3 Oligopoly theory is used
to specify the equations of the model to be estimated. In this section, the use of
theory to specify the model will be emphasized over inference. Inferences about
market power will be identified only through refutable implications of the theory
contained in the comparative statics or comparative dynamics of oligopoly
equilibrium. Another paper title is symptomatic of the form of departure from
tradition: Panzar and Rosse (1977a) "Structure, Conduct and Comparative
Statics". The question of exactly how the comparative statics identify the conduct
parameters is sufficiently important to deserve discussion in a separate section.
The stylized model's specification of demand functions and of cost functions
tends to follow fairly standard applied econometrics treatments, so I will be terse
in describing them. Specification of conduct parameters is more novel, less
standardized, and less well understood, so I will treat it at somewhat greater
length.

2.1. Notation, cost, and demand

The dependent variables of the stylized model are market price It, and each
firm's quantity Qir Throughout, i will index firms and t will index observations,

3A few studies have also included variables which can be interpreted as directly shifting firm
conduct: changes in regulation, entry (or the number of firms), mergers, and so on. These kinds of
variation is much more prevalent in the studies of closely related markets than in single-industry time
series.
Ch. 17: Industries with Market Power 1015
normally taken to be in time series. Since we are treating the single-product case,
Qt = ZiQit is well defined. For clarity, it is also useful to assume that the demand
function for the product contains no intertemporal linkages from durability,
habits, learning or other sources. It is convenient to write the demand function in
inverse form:

Pt = D ( Q t , Yt, 8, edt), (1)

where Yt are all variables shifting demand, ~ are unknown parameters of the
demand function. The demand equation econometric error terms edt are written
as entering in a potentially nonlinear way and not necessarily treated as a scalar.
In Porter's (1984) study of an 1880s railroad cartel, Qt is grain shipped by rail
from Chicago to the East Coast, measured in tons. 4 The time index t refers to a
week between the first week of 1880 and the sixteenth week of 1886. The price
data, Pt, are based on a weekly poll taken by the cartel of its members; given the
possibility of secret price cutting, P, is probably to be interpreted as if it were a
weighted average of list prices. The demand function takes the constant elasticity
form: 5

log Pt = 8o + 8110g Q, + 82Lt + Edt' (1')

where Lt, the only demand-side exogenous variable, is a dummy variable equal to
1 if the Great Lakes were open to navigation. (Shipment by water is a seasonal
competitor to the railroads.)
Returning to the stylized model, the treatment of costs similarly takes a
familiar form. For reasons of later convenience, we start from the total cost
function:

c,, = c ( o . , w,,, z., r, (2)

where W/t is the vector of factor prices paid by firm i at observation t, Zit are
other variables that shift cost, F are unknown parameters, and ec, are economet-
ric error terms, treated as in (1). The distinction between factor prices and other
cost shifters is maintained because many important developments in the litera-
ture concern the comparative statics of market equilibrium in W. I have put an i
subscript on Z and W, since in some applications the comparative statics of
equilibrium in the costs of a single firm or subgroup of firms are emphasized.

4porter provides an argument for why this industry should be treated as a single-productone on
pp. 302-303. He considers the aggregationof all grain, the dropping of all nongrain shipments, and
the dropping of all westbound shipments.
5I shall mention a consistent notation throughout,rather than adopting the notation of individual
papers. (1') is Porter's (1) in inverse form.
1016 T.F. Bresnahan
More commonly, however, Z and W will not have an i subscript, since they will
be measured at the industry level. In applications where the cost function being
treated is a short-run cost function, Z will include the ( S R ) fixed factors. The
definition of marginal cost follows from (2):

M C = CI(Q,,, Wit, Z,t , , ec.), (3)

where the nonlinearity of cost in the econometric error has been exploited for the
first time. C1(. ) is written as random, just as C(.). The error term in (3) is
(harmlessly) written the same as (2), though obviously an additive error in the
cost function will not appear in MC.
Outside the perfectly competitive model, firms do not have supply curves. 6
Instead, price- or quantity-setting conduct follows more general supply relations:

Pt = C~(Q,, Wit, Z , , F, ec, ) - D I ( Q t , Yt, ~, edt)QitOi,. (4)

Since P - DIQ is monopoly M R , (4) has the interpretation of M C = "perceived"

M R for oligopoly models. The parameters 0 index the competitiveness of
oligopoly conduct. As Oit, a positive unknown parameter, moves farther from 0,
the conduct of firm i moves farther from that of a perfect competitor. At first
glance, (4) appears to contain MC, even though I have said the approach
assumes that M C is unobservable. The contradiction is not real, however, only
apparent. Marginal cost, C1(. ), in (4) appears as a function of unknown parame-
ters, F, not as an accounting datum. Only after F has been somehow estimated
can M C be calculated; most of the methods described below draw inferences
about and 0 in the same econometric step.
The next section will take up the conduct interpretation of 0 at some length.
Obviously letting 0 vary both by firm and observation results in an overpara-
meterized model. It is written here in such generality to permit nesting all of the
known theories of oligopoly. Any empirical study will put some structure on the
way 0 varies across time and firms.
In Porter's railroad study, the constant-elasticity demand assumption implies
that (4) can be written with the explicit Oit terms substituted out. Furthermore,
(4') is not estimated for the individual firms, but rather aggregated to industry-
wide data:

log(Pt(1 + ~10t)) = M C = Fo + FllOg Qt. (4')

It is assumed that firms have heterogeneous marginal costs arising from a
log-linear cost function, but that there are no exogenous shifts in the level of

6By a supply curve, I mean a solutionfor Q as a functionof P of the equation P = MC(Q).

Ch. 17: Industries with Market Power 1017
costs over time. Thus, M C in (4') is interpreted as the marginal cost of the
average firm in the industry at time t. The overparameterization of 0it is solved
by aggregation: the interpretation of 0t is as the average of the conduct parame-
ters of the firms in the industry. Since there is some entry and some acquisition
activity during the sample period, the average firm's marginal cost can be
expected to shift over time: this is captured by a series of structural dummies St,
which enter as ( F 2, St). 7 Similarly, these changes in industry structure may have
changed conduct: this is captured in the same dummies. The lack of other cost
shifters (factor prices) or demand shifters (the price of lake shipping) is dealt with
not by new data but by a close reading of the estimation results in light of the
likely omitted variable biases.
To Porter, a crucial question is whether Ot varies over time because of changes
in industry conduct. High 0t periods are to be interpreted as successful cartel
cooperation, low 0t times as price wars or similar breakdowns in cooperation.
Furthermore, it is assumed that the probability of successful collusion is 1 = yr.
After a transformation, the supply relation ultimately estimated takes the form:

Supply relation Probability

log et = Fo + aa + Fllog Q: + (F2, St) + ec,t vr (4")
log Pt = Fo + aa + ab + Fllg Ot + (F2, St) + ecit 1 - ~r

where aa is a transformation of the conduct parameter in periods of successful

collusion, and ab measures the change in conduct when collusion breaks down. 8
Of course, it is clear that a a cannot be separately estimated from F0 on the basis
of estimating these equations, but there is considerable interest in estimating a b,
the percentage amount by which a breakdown in the cartel changes prices. As the
form of (4") suggests, Porter estimates the system (1', 4") by "switching equa-
tions" methods, in which the probability vr as well as the regular parameters are
estimated from the data.
In general, the endogenous variables of the stylized model are Pt and Q,.
Many empirical studies, like Porter's, use only industry-wide data and thus have
endogenous variables Pt and Qt- In either event, the core econometric methods
are those for simultaneous equations. Some empirical studies proceed by estimat-
ing (1) and (4) directly as structural equations: this is attractive, as the supply
parameters 0 are of primary interest. Other studies, however, may lack data on
price or quantity or on firm-specific quantity. In the first case, some reduced form

7The four structural dummies capture the entry by the Grand Trunk railway and the entry by the
Chicago and Atlantic, an addition to the New York Central and the departure of the Chicago and
Atlantic from the Cartel. The ( , ) notation is the inner product.
8Let 0, be able to take on one of two values, 0 a and 0 b. Then a a = - l o g ( 1 + O a 6 1 ) and
cta nc otb = - l o g ( 1 + 0b ~t).
1018 T.F. Bresnahan
will be estimated. In the second, aggregate data will be used. For both purposes,
it is useful to briefly define a few related functions. If (1) and (4) are solved
simultaneously for all firms, they yield the reduced forms for price and each
firm's quantity:

P, = e * ( w , , z,, Y,, (5)

Q , = Q * ( W t, z , , Y,, 12, e,), (5')

where 12 = ($, F, 0) is the vector of all parameters, et is the vector of all

structural error terms, Wt is the superset of all the W/t, 9 and Z t and Yt are
similarly defined. Similarly, we have the reduced form equation for industry
quantity, Q*(.), and so on.
Discussion of the results of these analyses in general must be delayed until
after the discussion of the interpretation of the parameters 0. Porter's results,
however, will be discussed briefly here. The value of a b estimated in (4") implies
that price was raised about 40 percent by successful collusion in the industry. The
implicit estimates of 0 are consistent with collusive behavior approximately as
anticompetitive as that implied by the Cournot model. 1 The implicit fraction of
the time that the cartel broke down was 28 percent; both the amount and timing
of cartel breakdown activity differed somewhat from the patterns detected by
earlier scholars and by contemporaneous sources.
The advantages of the modelling technique embodied in (1) and (4) seem to me
to be threefold. The first is that the econometric approach is structural. Each
parameter has an economic interpretation, and substantial departures of the
estimated parameters from expected values can serve as clues to difficulties with
or shortcomings of the analysis. For example, Porter estimates a negative/'1; he
provides an argument for why M C might be downward-sloping, an argument
that is more convincing in the context of railroads than it might be in many
others. The second advantage of the approach I take up in the next section: if the
interpretation of 0 is correct, the relationship of the estimates of conduct to
theory will be clear. A third advantage is in the section after that: given the
structural nature of the econometrics, the reason why the data identify the
conduct parameters can be made clear.

9If there are only market factor prices, Wt = Wit without any loss of information. If all firms buy
factors at different prices, then W~= (WI,, W2..... ).
lThe assertion about Cournot depends on the (unidentified) inference that noncooperative
behavior is approximatelyprice-taking.
Ch. 17: Industries with Market Power 1019
2.2. Alternative treatments of firm conduct and of 0

Supply relations are more general than supply equations because they permit the
possibility of nonprice taking conduct, captured in the strategic interaction
parameters (0). Clearly the form in which this nonprice taking conduct is
modelled will be central to the inferences about market power drawn in any
particular study. The approach covered in Subsection 2.2.1 takes the specification
of 0 directly from a theory or group of theories. The logical extension of that
approach, testing a small number of distinct theories of oligopoly interaction, is
covered in Subsection 2.2.2. Another group of papers, covered in Subsection
2.2.3, has had a looser connection to oligopoly theory. A typical paper in it
reports its inferences as "estimating oligopoly conjectural variations". There has
been enormous confusion about the interpretation of this work, however, prim-
arily because of a language gap. To resolve the confusion, the phrase "conjectural
variations" has to be understood in two ways: it means something different in the
theoretical literature than the object which has been estimated in the empirical
papers. A brief subsection, 2.2.4, discusses the interpretation of estimates when
only industry-wide data are used, another area of some confusion. Throughout,
individual papers will be used to illustrate how the analytical ideas are actually
carried out in the data.

2.2.1. Supply relations from single, specific theories

The first approach to writing ( 4 ) a n d specifying 0 uses explicitly theoretical

language. Important examples include Rosse (1970), in which the supply relation
is derived from the theory of monopoly. Bresnahan (1981) has product differenti-
ation with Bertrand pricing. Porter (1984) uses the Green and Porter (1984)
version of Stigler's (1964) theory of collusive oligopoly over time to specify (4);
since that theory implies that both collusive and price-war periods will be seen in
the data, 0 varies over time in the empirical model. After the specification of
models based on a single oligopoly theory, I discuss the specification of alterna-
tive models based upon a small number of different solution concepts (Bertrand,
collusion, etc.) each of which leads to a different version of (4), as in Bresnahan
(1980). Geroski (1983) applies this approach to the coffee industry data of Gollop
and Roberts (1979), specifying theories in which conduct (and therefore 0) varies
across firms (such as Stackelberg leader/follower models). All of the studies
which take this approach to specifying the supply relation impose dramatically
more structure on the way 0 enters (4) than the way I have written it, the
structure implied by the theory or theories used. More generally this approach
focuses on the estimation of specific models of strategic conduct. The flavor of
this analysis can be gotten by considering a few examples. The first of these is
1020 T.F. Bresnahan
monopoly and the Cournot model, or (in a higher or at least higherfaulting
theoretical language) one-shot Nash equilibrium with quantities as the strategic
choice variables. This takes the form Oit = 1 for all i, t:

Pt = C l ( a i t , Wit, Zit, F , Ecit) -- Ol(Qt , Yt, 8, e d t ) a i t . (6)

When there are several firms in the industry, (6) provides an estimating equation
for each firm. When there is only one firm in the industry, Cournot behavior
(among others) is the same as monopoly behavior. Thus, 0 = 1 for monopoly,
and (6) holds.
Rosse's (1970) study of American newspapers estimated an equation like (6)
simultaneously with one like (1). In his work, Qit is taken to be a three-vector:
column inches of advertising, Qa; circulation, Qt; and column inches of "edi-
torial" (nonadvertising) material, Q~. There are two prices associated with these
variables: the price of circulation is measured as the average price per subscrip-
tion copy, and the price of advertising is the average price charged per inch. u It
is reasonable to expect that the amount of circulation affects advertisers' de-
mand; and that the amount of editorial material and the amount of advertising
affect subscribers' demand for the newspaper itself. Using superscripts on prices
for advertising and for circulation, the demand equation (1) takes the form:

et a = D a ( Q t , Qt, Yt, a, Eat ) (advertising demand),

(1')
Pt~ = D ( Q t , Qt, Qt, Yt, 8, ecr) (circulation demand).

The treatment of M R needs also to reflect this interdependence. The appropriate

definition of M R is the derivative of the firms' entire revenue with respect to a
single product's quantity. For example, the form of (6) for circulation is:

Pt = MCC - QtD~(Q~, Q~, Q~,... ) - Q~D~(Qt, Q t , . . . ), (6 ~)

where the last term on the right is the unusual M R term: the advertising-price
effect of higher circulation. Simultaneous estimation of cost function and supply
relation exploits the cross-equation restrictions between demand and M R . If
demand can be estimated, and M R thereby inferred, an estimate of monopoly
M C can be obtained from M C = M R . This permits Rosse to take a very
sophisticated view of the M C function. For example, his treatment permits
transitory shocks to the sizes of firms in his sample to drive a wedge between

llClearly, this three-product approach to the newspaper involves some aggregationof products,
such as the different sizes of advertisementthat can be purchased. Also, in the interest of uniform
notation, I have suppressed some of the details of Rosse's treatment, and have written the demand
equations in inverse form.
 Ch. 17: Industries with Market Power 1021
LRMC and S R M C . Obviously, the degree to which such an analysis is convinc-
ing depends on the quality of the demand estimates and on the reliability of the
M R inference; a linear demand specification (like Rosse's) may fit well, yet
provide poor estimates of M R . As (6 c) suggests, the ability to support this
inference against criticism turns on a sophisticated use of the institutional detail
of the industry, in this case newspapers.
The other static, noncooperative, symmetric oligopoly model to receive atten-
tion is that of Bertrand (one-shot Nash equilibrium with prices as the strategic
variables). Since this model is not interestingly distinct from perfect competition
in the case of single product industries with fiat marginal costs, the greatest
attention has been focused on the product-differentiated case. Bresnahan (1981)
estimates such a model on 1977 and 1978 cross-sections of automobiles by type.
The demand system comes from a spatial treatment of the demand for automo-
biles by type. Automobiles of different types are assumed to lie in a one-dimen-
sional space, and consumers are assumed to be distributed according to a
one-dimensional parameter describing differences in their demands. 12 The own-
price and cross-price elasticities are determined by how close products are in this
space. Letting X i be the quality of good i, the demand for a typical good is given
by

oi = 80((ej- t , ) / ( x + - x , ) - ( t , - x.)) + ..,, (7)

where products h, i, j are adjacent in quality space. Here the price of automobile
i is the manufacturer to dealer list price, which is identical to the transactions
price for the vast majority of sales. Quantity is model-year production, and the
unit of observation (i) is now not the firm but the product. Quality is taken to be
ex ante unknown, but to be a hedonic function of observable characteristics:

X i = f(Z, 81) q- $'xi' (8)

where Z is a vector of the typical hedonic characteristics: length, weight,
horsepower, etc. Under Bertrand competition, the equivalent of (6) is given by

Q, = ( P~ - MC~) 8o(1/( X / - Xi) - 1/(X~ - Xh) )

+di+PjSo/( Xj - Xi) + d,hPhSo/( Xi -- Xh) + (6")

where dij is a dummy for whether the firm that produces product i also produces
product j. As in the Rosse treatment, the definition of M R is taken on a

12The demand system is that of "vertical product differentiation"; see Prescott and Visscher
(1977).
1022 T.F. Bresnahan
whole-firm basis. In this cross-section work, marginal cost is a function only of
product type Z, not of factor prices. It is clear from (6") that the closeness of
competitive products is the key determinant of market power in such a model.
Bresnahan (1981) reports two related findings: the large price-cost margins
appear to be on the larger vehicles, and the larger vehicles appear to be much
farther apart in product-quality space, i3 The obvious problem with this kind of
modelling is the highly restrictive assumptions made about the form of the
demand system. Bresnahan (1980, 1981), like Rosse (1970), attempts to deal with
these primarily by analysis of the residuals; the devices include introducing firm
dummies, adding exi to (7), and so on.
A second class of models of some importance have separate leaders and
followers. Of these, the first is the Stackelberg leader model. I adopt the
notational convention that there are I firms in the industry, so i = 1 . . . . . I. The
Stackelberg model writes (6) for firms 2,... I, but for the "leader," firm 1 writes:

Pt = Ci(Qit, Wit, Zlt, F , ~clt) -- DI(Qt, Yt, 8, edt)Qlt(l q- 0S) , (7)

where 0s is obtained by first solving (6) simultaneously for all firms except firm 1;
this yields functions Qi(Ol .... ). Then 0s is the derivative of the sum of these
with respect t o Qlr
A closely related model, the dominant firm model, has been estimated by
Suslow (1986). In her treatment, there is a fringe of firms that are price-takers,
the producers of "scrap" or recycled aluminum. The dominant firm is Alcoa, and
the sample refers to the interwar period in which Alcoa had a monopoly on the
production of new aluminum. In a general dominant firm model, the fringe firms
have supply relations which are supply curves. Without loss of generality, the
supply curve of the entire fringe is determined by its collective MC:

Pt = Cl(Qf,, wet, Zft, F, ecft), (10)

Qf, = s ( Pt, we,, zf,, F, ecft), (a0')

where Qft = ~"~i=2,I Oit, and W and Z are taken to be common to all firms in the
fringe: the function in (10') is the inverse of that in (10). Suslow (1986) permits
the dominant firm and the fringe to sell somewhat differentiated products. For
clarity I will assume they sell the same product. In her paper, Qft is the physical
quantity of "secondary" aluminum recycled, and eft the recycled price. The
demand shifters Zrt include an estimate of the stock of aluminum available for
recycling. The leader in a dominant firm model is taken to be sophisticated, and

i3This confirms the view of many industry observers, based on accounting profits data. It received
further confirmation from a study of auto dealers' (as well as manufacturers') prices in Bresnahan and
Reiss (1985).
 Ch. 17: Industries with Market Power 1023

to set:

e, = c1(.) + Dg.)O ,(a + s o ) , (11)

S D = DI(.)SI(.)/(1- DI(.)Sx(.)). (11')

In Suslow's paper, Alcoa's M C is shifted by cost variables, such as the firm's

accounting average variable cost, which dearly do not enter the supply function
of the fringe. Price is measured by realized average revenue, quantity by sales in
physical urtits. 14 She finds that Alcoa enjoyed considerable monopoly power,
despite the competition of the fringe, since the quantity S D is estimated to be
small. This approach has some of the same interpretational difficulties as the
earlier one: M R still needs to be carefully selected. There is also the need to
establish or test that the selected firm is in fact the leader; Suslow accomplishes
this by arguing from the industry structure of new aluminum (monopoly) and of
recycling (many firms).
Leaving static noncooperative theories, leads us to dynamic theories. The
problem of repeated oligopoly interaction has received a great deal of theoretical
attention, since it is reasonable to presume that long-run considerations might
reduce the competitiveness of oligopoly conduct. Oligopolists might go along
with a collusive arrangement, even though deviations are profitable, if they
recognize that deviations will lead to a general breakout of competition. A
pathbreaking model is that of Stigler (1964): if it is purely this self-interest which
holds collusive arrangements together, they should be expected to sometimes
break down as the parties grow suspicious of each others' motives and behavior.
In an uncertain environment, firms will not always know whether secret price-
cutting has occurred, and this will lead to some price wars even if there is no
actual secret price-cutting. Thus, one should expect cartels to break up and to
reform: data on a cartelized or tacitly collusive industry should show both
periods of successful cooperation and periods of outright competition. Em-
pirically, of course, this will show up as time-varying 0.
Several recent formalizations have put considerably more structure on the
theory, and on the pattern of industry equilibrium over time. The first of these is
Green and Porter (1984) (GP). In their theory, price wars actually break out
because of shocks that firms cannot observe. When profit drops in the SR, firms
cannot tell with certainty whether this is because of chiselling on the cartel
agreement or because the aggregate situation has worsened. Bad enough shocks
will trigger price wars. GP, and a later paper by Porter (1983) treat the problem
of "design" of a cartel arrangement. In their theory, firms decide (collectively)

14Suslow's treatment of the marginal cost of a monopolistic that can produce to store as inventory
is somewhat shortchanged by this discussion.
1024 T.F. Bresnahan

how high to raise prices in collusive periods, how trigger happy to be, and how
long price wars should be. The degree of competition in a price war is taken as
exogenous (it is assumed to take the Cournot form). The theory predicts that
there will be alternating periods of price war and of successful collusion. The
length of the price wars and the size of a shock needed to trigger a price war are
picked to maximize industry profit. The intuition of these and of other Stigler-
esque theories comes from incentive economics. Why should collusive firms not
raise price all the way to the monopoly level? If they did, it would give too much
of an incentive to deviate from the cartel arrangement. Why should price wars
last a while before the cartel is reformed? Otherwise, the possibility of a price war
could not deter any opportunistic behavior.
Abreu et al. (1986) find cartel designs that are even more profitable for firms
than the GP ones, by permitting more complex arrangements among firms. These
designs still have alternating periods of successful collusion and of price wars, but
now there are "triggers" both for beginning a price war and for ending one.
Thus, the length of price wars is random. Furthermore, the amount of competi-
tion in a price war is endogenous to the model: it, too, is optimized to maximize
the returns to the colluders. Rotemberg and Saloner (1986) take a somewhat
different tack. In their model, the environment in which firms operate shifts over
time. As a result, the optimal cartel price shifts as well. Suppose (as in the
analysis they provide) that current demand is not a predictor of future demand.
Then in boom periods, the gain from defecting from a cartel is unusually large at
any given price. Therefore the cartel must set an unusually low price to reduce
the incentive to defect. The reverse line of argument holds in demand busts.
These various theories have in common the idea that in an imperfectly
informed world, "successfully" collusive industries will have periods of cartel
pricing and periods of competition. 15 In general, they imply models with Oit not
necessarily equal to 0i~. The theories differ somewhat in the expected time-series
behavior of these two regimes, as the exact equations determining passage from
one regime to the other vary between theories. Green and Porter theories, for
example, seem to suggest that 0 changes from the collusive to the competitive
value when there is an unanticipated shock to demand, and that returns to
collusion will follow with a fixed lag. Abreu et al. have 0 following a time-series
process driven by demand shocks as well, but the process is Markov. The
Rotemberg and Saloner theory suggests endogenous strategic variation in con-
duct within the collusive regime. It is easy to imagine other theories of success or
failure in tacit collusion which predict different patterns; taking all of these
theories at once would lead to even more complex potential time-series behavior
for 0.

l~Rotemberg and Saloner do not write their theory in this way, because they assume that there is
no imperfect information. This is clearly an assumption of convenience (irrelevant to the point they
are making) rather than a central feature of the model.
Ch. 17: Industries with Market Power 1025
I have already discussed the details of Porter's (1983) approach to time-varying
0, and the discussion of the previous paragraph begins the discussion of the
relationship of that approach to theory. For clarity, let me reprint (4") here. The
supply relation estimated in Porter (1983) [see also Lee and Porter (1984)] is

Supply relation Probability

log Pt = Fo + aa + FllOg Qt + (F2, St) q- ecit ~r (4")
log Pt = Fo + aa + ab + Fllg Qt + (1"2, St) + geit 1 - - ~7

By estimating only a single probability of colluding or competing, ~r, Porter puts

less structure on the problem than the theories suggest. Estimation of ~r in (4")
can tell us how frequently cartels break down, but cannot tell us what predicts
breakdowns and reformations- the area in which the theories disagree. At a
minimum, one would like rr to be state dependent. Alternatively, whether
changes in markup from regime b to regime a appear statistically to be the same
event as shocks to demand could be investigated. An initial attempt to investigate
the time-series behavior of the regimes is contained in Porter (1985). Rotemberg
and Saloner (1986) provide a somewhat less structural analysis of the same
question, using Porter's reported regime shifts as well as the time-varying pattern
of automobile industry competition and collusion reported in Bresnahan (1987). I
see little in these investigations that differentiates among the various theories; the
time-series behavior of conduct is harder to estimate than the average level of
conduct.
It will be surprising (to me, anyway) if further similar investigations succeed in
strongly differentiating among the theories. Investigations of the question, Do
there seem to be price wars?, can essentially take advantage of the data in all
periods. The investigator is trying to find out that there are two distinct regimes.
On the other hand, an investigation of the question, What sets price wars off and
what stops them?, necessarily will have a much smaller sample size. Instead of
having the number of observations equal to the number of sample periods, it has
the number of observations equal to the number of price wars. This low
assessment of the success probability of further testing is not particularly trou-
bling (to me, anyway) since the various theories are identical for practical
purposes.
A final remark on this subsection goes better here than anywhere else.
Industrial organization economists have frequently felt that their field was
data-starved, or at least starved of appropriate data. The studies reviewed here
show this to be false. The dissertation of work of Rosse, Bresnahan and Suslow
each involved collection of data from industry sources in the public domain, with
no reliance on government sources for endogenous variables. Porter's work, too,
is based on industry sources.
1026 T.F. Bresnahan

2.2.2. Supply relations from a small set of theories

The papers treated in the previous subsection took a single (in the last case,
complex!) theory as a starting point for the specification of the supply relation. A
closely related development has been attempts to estimate the supply relations of
a small number of different theories and to test among them. This is the approach
of Bresnahan (1980, 1987). The data used are again cross-sections of automobiles,
this time from the mid-1950s. The demand equations are exactly as in (7). But the
supply relations for joint monopoly pricing as well as those for Bertrand pricing
are estimated. These differ from (6") only in that dij = 1 whether the neighbor-
ing products are produced by the same firm or not. (MR for a joint monopoly of
all firms is the derivative of industry revenue with respect to quantity.) The
estimates, otherwise much like those described above, show collusive behavior in
some years, but competitive (Bertrand) behavior in 1955. This provides a stra-
tegic explanation of part of the large expansion of auto production in that model
year. In related work, Geroski (1983) and Roberts (1983) put the structure
implied by a small number of leader-follower type theories on the data for coffee
roasting firms in the United States, finding that the smaller firms in the industry
are price-taking followers. The leading finns do not appear to joint profit
maximize (even given the constraint implied by the fringe's supply curve) but do
behave less competitively than Cournot firms.
Since each of the theories reviewed up to now in this section is associated with
different values of the parameters in 0, one might decide to treat 0 as a
continuous-valued parameter and estimate it. This approach is the one discussed
in the next section. It risks the possibility that values of 0 which are "in
between" existing theories will be estimated, but that is hardly a disaster. The
distinction between the continuous-valued 0 and the distinct 0's from several
theories is purely econometric. The researcher who has estimated 0 from a
continuum will test theories by nested methods. The other researcher will use
non-nested tests to distinguish among the few theories entertained, as I did in the
work described in the previous paragraph) 6

2.2.3. Supply equations in "'conjectural variations" language

The second approach to specification of (4) uses "conjectural variations" (CV)

language and treats conduct as a continuous-valued parameter to be estimated.
In this language, the parameters describing firms' conduct are not written in the
form of 0 in (4). Instead, the parameters are described in terms of firms'
conjectural variations, that is, their "expectations" about the reaction of other
firms to an increase in quantity. These parameters are typically allowed to take

16See Geroski, Phlips and Ulph (1985) for a different position on this issue.
Ch. 17: Industries with Market Power 1027

on any values in a broad range. An important early paper is Iwata (1974), whose
title "Measurement of Conjectural Variations in Oligopoly" is illustrative of the
thrust of the literature. 17 He saw the question as inferring where, in the contin-
uum between perfect competition and monopoly, the Japanese plate glass indus-
try lay. Another important early paper was Gollop and Roberts (1979), which
permitted conjectures to vary across firms. Later work by Spiller and Favaro
(1984) and Gelfand and Spiller (1987) on banking continued this emphasis on
heterogeneity of firm conduct.
In the CV language, the empirical supply relationship is written in the form:

e,= c1(-)- DI(.)O,,(1 + (12)

Here Qjt is the vector of all other firms' quantities, and the dependence of cost
and demand on exogenous variables and parameters can be temporarily sup-
pressed. Note first that the only difference between (4) and (12) is that the term
Oit in (4) has been replaced by the term 1 + r(-). This does suggest some practical
differences between the papers discussed in the previous subsection and the CV
papers. The CV papers tend to permit all values of ri, not just those associated
with particular theories. There is dearly nothing fundamental about this: as
discussed above, one could easily treat 0 in (4) as lying in a continuum. Second,
the CV papers have tended to emphasize the relationship between firm size and
conduct: hence the explicit dependence of ri(. ) on quantities in (12). There is
also an implicit dependence: different values of are often permitted firms in
different size classes.
As a matter of logic, there is absolutely no difference between (4) and (12) in
general, since the identity Oil- 1 = ri(Q,, Qjt, z , , ~) implies that the two
specifications can nest the same models. (Nothing in the previous subsection
implied that Oil needed to be a constant, though many theoretical models have
constant Oil.) Therefore I will use 0 to mean 1 + r, and vice versa according to
convenience, in the rest of this chapter. I cannot overemphasize this point: there
is no difference between saying what the "conjectural variation" is and saying
what theory of oligopoly holds in the data. Misunderstandings of the phrase
"conjectural variations" to mean something other than it does mean in the
empirical papers has been rife, however.
There are some cases where no misunderstandings arise: for example, the
Coumot model is labelled "zero conjectural variations". It is usually innocent to
think of Cournot firms as ones that "expect" other firms' quantities to be

17Iwata (1974) differs from many of the papers surveyed here in that it assumes that accounting
data reveal PCMs. Its role in using the data to try to draw inferences about conduct was very
influential, however.
1028 T.F. Bresnahan

constants. Similarly, r(.) of - 1 is perfect competition, and it is completely

correct to say of a competitive firm that it "expects" price to be a constant.
The linguistic difficulties arise in other cases because allusion to underlying
theoretical models is typically made in the same "expectations" or "conjectures"
language. Suppose we think of (12) as the derivative of a single-firm profit
function:

maxD
Qi
(
. j
~.,Qj,(Qi) . . . . )
Qi- C(Q, .... ) (13)

Equation (13) has every other firm's quantity as a function of Q~. Then we read
1 + r~ in (12) by

1 + ri ( . ) = d Q i / d Q , + ~_, d Q f f d Q , , (14)
J

where the sum over firms j is j ~ i. Some minor variations in language occur,
but the typical understanding is that the d Q j / d Q ~ terms measure the way firm i
"expects" firm j to "react" to an increase in quantity. 18
It is when the estimated "conjectures" are ones which lead to prices close to
the collusive level that the simple "expectations" interpretation is suspect. The
point can be seen under the assumption that all firms have the same cost curves
and "conjectures". Let there be I firms in the industry. Suppose we get the
conjectures associated with the collusive level of output, (1 + ri) = 1:

P + C1(. ) - D I ( . ) I Q i (15)
for each firm, since that particular value is the solution to the problem "maximize
the profits of all ! firms". How, then, are we to interpret 1 + ri = 1? Taken
literally, the coefficient says that the firm picks its output to maximize industry
profits because it "expects" the other firms in the industry to match its output:
ri > 0 is an expectation of matching behavior, and the ri of (15) imply an
expectation of proportional matching: d Q j / d Q i = 1.
In a great many treatments of oligopoly as a repeated game, firms produce
output in most periods according to (15), but the reason they do is that they
expect deviations from that level of quantity to lead to a general breakdown in
restraint. (The exact form this would take varies among the theories: see Subsec-

18Some papers describe estimates of (11) as "estimating conjectural variations", other as "estimat-
ing firms' first-order conditions". If (11) is solved for Qi it is "estimating firms' reaction functions".
Adding up the first-order conditions and interpreting the result in fight of Cournot theory even leads
to the language "estimating the equivalent number of firms". Obviously, there are no important
distinctions between these languages.
Ch. 1~ Industnes with Market Power 1029

tion 2.2.1 above, and see Shapiro, Chapter 6 in this Handbook, for a much fuller
treatment.) Thus, the matching behavior is unobserved; firms expect that if they
deviate from the collusive arrangement, others will too. This expectation deters
them from departing from their share of the collusive output.
The crucial distinction here is between (i) what firms believe will happen if they
deviate from the tacitly collusive arrangements and (ii) what firms do as a result
of those expectations. In the "conjectural variations" language for how supply
relations are specified, it is clearly (ii) that is estimated. Thus, the estimated
parameters tell us about price- and quantity-setting behavior; if the estimated
"conjectures" are constant over time, and if breakdowns in the collusive arrange-
ments are infrequent, we can safely interpret the parameters as measuring the
average collusiveness of conduct. 19 The "conjectures" do n o t tell us what will
happen if a firm autonomously increases output (and thereby departs from the
cartel agreement), the normal sense in which theoretical papers would use
"conjectural variations".
A second set of interpretive questions arises when the variation across firms in
r i is modelled by estimating tli in ri(Qit, Qjt, zit, ~), as in Gollop and Roberts
(1979), Spiller and Favaro (1984), and Gelfand and Spiller (1987). A speculative
interpretation is that the dependence of the ri(. ) on own quantity tells us
something about "mutual forbearance". The notion here is that one can read the
derivative of ri(-) with respect to own quantity as revealing something about how
competition would change if firms were to deviate from agreed production. This
seems to trip over the distinction just raised between what the conjectures enforce
and what they are. I would therefore not use the "mutual forbearance" language.
The use of different strategic parameters for firms of different sizes suggests
that the CV models may provide a strategic explanation of the size distribution of
firms, since their endogenous variables include the quantity produced or market
share of each firm. By permitting the firms to have different conduct, such models
permit ex ante identical firms to be of different sizes in equilibrium. For example,
the three papers just listed all permit different conjectural variations for different
size classes of firms. When economies of scale are permitted (as in these papers)
this can provide information about the details of large-firm-small-firm interac-
tion. It was this which motivated Geroski (1983) and Roberts (1983) to test
specific theories in this context: they were particularly interested in questions

19All of these theories of going along with a collusive arrangement suggest that it is not necessarily
the fully collusive outcome that will arise. Somewhat smaller r~'s than in (13), and therefore somewhat
larger equilibrium output, can also arise (when worse information makes collusive arrangements
harder to enforce, for example.) In these circumstances, firms' "expectations" about what would
happen if they deviated from planned output might be exactly the same, but their production levels
would be different. That is to say, there is no information about firms' expectations contained in the
estimates of the r/(.). The ri(. ) tell us how close to a completely collusive outcome the expectations
induce. The only thing they tell us about expectations is that the expectations are sufficient to deter
departure from the normal arrangement.
1030 T.F. Bresnahan

like: "Are large firms, taken as a group, leaders and small firms followers?" I
think this very interesting line of research is still incomplete, since alternative
explanations of the size distribution of firms have yet to be introduced. In
particular, it would seem important to let different firms have different cost
functions. Then it would be possible to test the alternative hypothesis that the
size distribution is driven by relative costs despite identical conduct parameters. 2

2.2.4. Work with aggregate industry data and aggregate conduct

In many circumstances, the lack of single-firm data will prohibit estimation of

supply relations for individual firms. Instead aggregate industry data must be
used. One approach is to simply rewrite (4) in aggregate form:

Pt = CI(Q,t, Wit, Z., 1', e~.) - DI(Q,, Yt,6, edt)QitO,, (4)

becomes

Pt = Cl( O,, Wt, Zt, F, ect ) - Dl( Ot , Yt, 6, edt)OtO r (16)

This approach is familiar from as different works as Appelbaum (1979) and

Porter (1984). Under the null hypothesis of no market power, 0 = 0, the
interpretation of Ca as industry marginal cost is clear. When there is market
power, however, different firms will almost certainly have different marginal costs
in equilibrium. Analysis like that of Cowling and Waterson (1976) shows that
noncooperative oligopoly will tend to have variation in price-cost margins across
firms unless they have identical, constant MC. In these circumstances, a stable
industry marginal cost curve need not exist, and interpretation of (16) may be
clouded.
There are clearly some circumstances in which industry-wide marginal costs
are equal to each firm's MC; consider the example of a cartel that succeeds in
rationalizing production. Generally, however, (16) will need to be interpreted as
some sort of average. In this context, an error in the interpretation of (16) has
crept into the literature: the assertion that Oit = Ojt is an implication of theory. 21
This is clearly incorrect, as there is nothing in the logic of oligopoly theory to
force all firms to have the same conduct. It is better to follow Cowling and
Waterson (1976) and interpret the aggregate 0t in (16) as industry average
conduct, and ( P - C1)/P as the industry average markup.

2This is not a trivial task. The existing specifications of a common cost function with scale
economies do permit heterogeneity in the level of MC across firms. The interesting hypothesis turns
on whether firms of different sizes would have similar MC at the same output, a tricky measurement
problem.
21Appelbaum (1982) made this argument; it has been picked up by Lopez (1984).
Ch. 17: Industries with Market Power 1031
2.2.5. Final thoughts on 0 and ri

Both the work closely based on economic theory and the conjectural variations
work has overwhelmingly cast its (logical) tests of theories of strategic interaction
as (statistical) hypothesis tests about 0. These studies tend to state their problem
as one of measuring conduct or strategic interaction rather than as of measuring
market power. Thus, they focus on hypothesis tests about 0 or 1-,.. In this
connection, an important observation was made by Appelbaum (1979): setting
the entire vector 0 to zero in (4) or (equivalently) setting all of the ri to - 1
imposes the restriction of price-taking conduct. Thus all approaches to specifying
(4), even those which do not use explicitly theoretical language, can be thought of
as "Testing Price-Taking Behavior", Appelbaun,'s title. This would not be
particularly interesting, except that the particular alternative hypothesis against
which price-taking has been rejected is one with market power.
The next section treats the question of what constitutes an adequately rich
specification of cost and demand so as to permit a reasonably convincing case
that a strategic interaction hypothesis is in fact being tested. The section will
show that the hypothesis of market power is in fact identified on reasonable data.
This is an important step: if it were merely true that perfect competition were
rejected, and that no positive indicia of power over price were found, the
observation that the results might be a statistical artifact would be compelling.
For now, however, let me point out an extremely important advance implicit in
this approach. The alternative hypothesis includes price-taking behavior: when it
is rejected, it is rejected against specifications based on theories in which firms
succeed in raising prices above M C . Only econometric problems, not fundamen-
tal problems of interpretation, cloud this inference about what has been de-
termined empirically.

3. How the data identify market power

An advantage of the use of structural econometric models and explicit theories of

industry equilibrium is that the class of models the data are allowed to treat is
made explicit. Thus, the class of alternatives within which the inference of
monopoly power has been drawn can be dearly stated. This in turn limits the
number of alternative explanations which can be reasonably advanced. More
importantly, however, this procedure permits an explicit answer to a central
question: Why is the economic inference of monopoly power identified by the
data? What implication of the theory of perfect competition has been found to be
false when market power is measured by these methods? These questions do not
arise for SCPP methods, of course, since SCPP takes price-cost margins to be
observable. In the NEIO, PCMs are to be estimated, and it is therefore of
1032 T.E Bresnahan

immediate interest what observable feature of the data, and what natural experi-
ments, reveal them to the analyst. To date, there are four new classes of
identification arguments: (i) comparative statics in demand, (ii) comparative
statics in cost, (iii) supply shocks, and (iv) econometric estimation of MC. This
section takes up these arguments in turn. There is a fifth area, the comparative
statics in industry structure, which is familiar in its logic; price is predicted
as a function of the number of firms or of other concentration measures. This
area is, I believe, awaiting its identification arguments, for reasons I lay out in
Subsection 3.5.

3.1. Comparative statics in demand

A natural empirical procedure is to write out a system consisting of (1) and of

one (4) for each firm, under one of the parameterizations of 0 from the previous
section. Some appropriate econometric method yields estimates of 8, F and
especially O, under the assumption that they are all separately identified. Thus,
the same data, and inferences based on the comparative statics of equilibrium,
provide estimates of cost, demand and conduct parameters. This natural proce-
dure, however, should make clear why it is that it works. What idiosyncratic
implication of perfect competition has been rejected, what idiosyncratic implica-
tion of market power or oligopoly interaction has been observed in the data? The
first approach to this question was to ask whether the comparative statics of
monopoly, oligopoly, and perfect competition models are logically distinct, and if
so how.
The encouraging answer to this question can be seen in a very simple model, in
which only an aggregate supply relation is estimated and in which 0 is taken to
be a constant over time. 22 To further simplify, assume that the econometric
errors enter both demand and supply in an additive way, and that the slope of
the demand curve does not depend on Qt- Then we write (1) and (4) as:

Pt = D ( Q t , Yt, 8 ) + edt, (17)

Pt = CI(Qt, Wt, Zt, F) + ect - DI(Yt, 8)Qt#. (18)

Obviously, 3 can be estimated: (17) has only quantity as an included endogenous

variable, and instruments are available from the cost function. Call any estimate
which has been obtained 3*. Then one could calculate the "datum" D,.* =

22The analysis of this section is based on Bresnahan (1982) and Lau (1982), which make the
argument presented here in a more precise way.
Ch. 17: Industries with Market Power 1033

DI(Yt, 8"), and consider estimation of the equation: 23

Pt = C I ( Q t , ~ t , Zt, r ) + i~ct - Di*QtO. (19)

This has two endogenous variables: Qt, which occurs in cost, and D~*Qt, the
variable whose coefficient is 0. When can 0 be estimated? Econometrically, two
conditions must hold. First, the two endogenous variables must not be perfectly
correlated. The definition of Di* makes clear that they will in fact be perfectly
correlated unless D1, the slope of the demand curve, depends on Yr Second,
instruments must be avail/tble for both endogenous variables. This will obviously
be the case if Yt is a two-vector, with one element of Yt entering D1, the other
not. More generally, Lau (1982) has shown that a sufficient condition for
identification is that the inverse demand function D(.) cannot be written in a
way such that Yt is separable from Qt; since Qt is a scalar, this clearly requires
that Yt be a two- (or more) vector.
The economics of this argument can be stated very simply. The comparative
statics of models with market power have a particular role for changes in the
slope of the demand curve. Suppose that the exogenous variables entering
demand can (in principle) perform a particular natural experiment: they can
rotate the demand curve around a given point, say the industry equilibrium point.
Under perfect competition, this will have no effect: supply and demand intersect
at the same point before and after the rotation. Under any oligopoly or mo-
nopoly theory, however, changes in the elasticity of demand will shift the
perceived marginal revenue of firms. Equilibrium price and quantity will respond.
Thus, the comparative statics of models with monopoly power do have idiosyn-
cratic predictions, and the market comparative statics of perfect competition are
distinct from those of monopoly.
The papers that have relied on this identification principle are those that have
had a good natural experiment shifting the demand equations in an appropriate
way. In Just and Chern (1980), it is the buying side which has the market power:
a concentrated manufacturing industry buys tomatoes from atomistic farmers.
The crucial exogenous variable was a change in harvesting technology, one which
they argued changed the elasticity of supply. 24 In Bresnahan (1981, 1987) the
firms possibly having market power are sellers of automobiles: demand elastici-
ties depend on how close to one another firms' products are in a product space.
More generally, we might think about the most attractive applications of this
identification argument. The two elements of Yt might be something measuring
the size of the economy, such as national income, and a second variable

23This is describing an econometric procedure so ugly that no one would ever undertake it: it does
however, show that and why more powerful techniques can identify.
24just and C h e m treat the case of oligopsony, so it is the supply elasticity which is shifted by the
technical change. This has no effect on the logic of the argument.
1034 T.F. Bresnahan
measuring the price of a substitute or substitutes. 25 Use of this method of
identification obviously turns on the ability to estimate the demand elasticities in
a reliable way. The analyst will need to answer such criticisms in any particular
case with standard econometric techniques for investigation of the robustness of
results. Many observers have noted that alternative interpretations of the Just
and Chern technology shift are available. The dependence of the Bresnahan
automobile results on the exact ordering of the products in quality space is
frequently pointed out. Equation (8), above, introduces an unobserved error into
product quality, thereby relaxing the assumption that the ordering of products
can be determined solely on the basis of observable proxies for quality.
A further refinement of this line of reasoning is available if one is prepared to
assume that marginal cost is homogeneous of degree one in observed factor
prices. 26 Then the 0 in (17) or the more general (4) is clearly identified. The
analyst interprets the coefficients of Ot that are interacted with cost shifters to be
part of M C , and those that are not to be part of the perceived marginal revenue
term with coefficient O. (Since M C is homogeneous of degree one in factor prices,
Ot c a n n o t enter M C in a way such that it is not interacted with one or more
elements of W.) Implicitly, this is how early papers' like Appelbaum (1979, 1982)
obtained identification. The homogeneity property is guaranteed if M C is ob-
tained by differentiation of a total cost function, possibly one of the forms
(translog, generalized Leontief, etc.) commonly used in factor demand system
estimation. Then all coefficients in the supply relation which are not functions of
factor prices are interpreted as indicators of market power. This line of argument
obviously leans very hard on the assumption that the functional form of M C is
correct and that all of the true marginal prices of the inputs can be observed. The
true marginal price of capital is a potential problem for such studies. The way for
such a study to rebut alternative interpretations of the results is to explore the
robustness of the results to alternative treatments of MC: alternative functional
forms, alternative treatments of the quasi-fixity of capital and labor, nonaccount-
ing definitions of the cost of capital, etc.

3.2. Comparative statics in cost

An alternative comparative statics analysis is that of Panzar and Rosse (1977a,

1977b, 1987) (PR). PR propose two separate ideas: first, that an appropriate
method for analysis is estimation of the reduced form, with particular attention
given to the coefficients of factor prices Wt. The second idea is that reduced-form

25Clearly, this idea leans on an older line of thought, especially in connection with cartelization in
the markets for primary commodities. See Scherer (1980, pp. 229ff).
26See the discussion of duality and cost below, and in Panzar and Rosse (1987).
Ch. 17: Industries with Market Power 1035
revenue equations are likely to be estimable in many circumstances, since revenue
is likely observable even where price and quantity are not. In light of this, let us
begin with the reduced-form revenue equation, called R*(.). This is the total
revenue for a single firm. R* is equal to equilibrium quantity (which depends on
cost, demand and conduct) times equilibrium price (which has the same determi-
nants). The observable shifters of cost and demand - Z, W, and Y - all enter this
function. In the case of monopoly, solve the single firm's (1) and (4) for the price
and quantity as a function of exogenous variables, parameters, and error terms,
and then continue by calculating revenue. This yields a reduced-form revenue
function of the form:

R i , = R*(W,.,, Zit , Y,, params, e,). (20)

Equation (20) will be written in the same form when there is more than one firm
in the market, and (1) and (4) have been simultaneously been solved for several
firms. It will depend on the exogenous variables for all firms, of course. Let
R w(W~t, Z i , Yt, params, ~,) be the vector of derivatives with respect to all inputs,
and let (,) be the inner product. The PR statistic is

H R = (Wit, R w ( . ) ) / R * ( . ) , (21)

the sum of the elasticities of the reduced-form revenues with respect to all factor
prices. 27 The PR statistic requires little data on endogenous variables in the
system, although it does need all of the variables which shift demand or cost. The
analyst proceeds by estimating the reduced-form revenue equation, R* including
all available information on W, Z and Y. Then H R is calculated. 28 A particular
advantage of estimating only a revenue equation is that no quality correction
need be made to define a true price for the industry. The product may be better
in some markets, so that its price per pound overstates its true price there. Yet
this tricky data problem does not affect the reduced-form revenue equation. More
generally, H n can obviously be calculated whenever the structural system (1), (4)
has been estimated. But R*(.) can also be estimated in many circumstances when
the structural equations, especially the supply relations, cannot.
The PR statistic has a clear economic interpretation in several cases. First,
suppose that the market studied is a monopoly. Then H R < 0. A very general
proof is available in PR: the intuition, however, can be seen here, for the case of

27Existing applications have either parameterized the reduced form so that this statistic is a
constant or have reported estimates near the center of the sample in some sense.
2SThe idea that W are the only exogenousvariables needed to estimate HR has been somewhat
oversold. In some circumstances, the estimating equation for revenueis misspecifiedwhen only cost
variables are included as exogenousvariables. As the results discussed below imply, it is appropriate
to omit demand shifters only when the hypothesis being tested is perfect competition. The test for
monopoly requires a revenuefunction with all exogenousvariables.
1036 T.F. Bresnahan

"elementary" monopoly. By elementary monopoly I mean simply the case in

which a single firm picks only a single quantity. I will suppress the econometric
errors and the parameters. Let R(Q, Y) be monopolist's revenue function in the
usual sense, so that RQ is MR. 29 Then the monopolist solves:

R * ( W , Z, Y ) = m a x R ( a , Y ) - c ( a , w , z )
Q

= QD(Q, Y ) - c ( a , W, z ) . (22)

Let R~' be the derivative of equilibrium revenue with respect to the kth factor
price, and W k be the kth factor price. A comparative statics analysis of (22)
implies:

n R = ZR~Wk/R = (RQ)2(RQQ - CI1)-l/R ~_~ O. (23)

k

Thus, the statistic H R is signed for the elementary monopolist. The intuition of
the simple result is straightforward. The H R statistic gives the percentage change
in equilibrium revenues that would follow from a 1 percent increase in all of the
firm's factor prices. A 1 percent increase in all factor prices must lead to a 1
percent upward shift in MC. Thus H n reveals the percentage change in equi-
librium revenue that would follow from a 1 percent change in cost. Elementary
monopoly theory tells us that a monopolist's optimal revenue will always fall
when costs rise: otherwise, the monopolist's quantity was too large before the
cost rise.
PR show, in a powerful result, that this finding generalizes to the case of a
monopolist that has many choice variables, including both the case where the
variables are the outputs of many products and the case where the variables
include variables such as quality, advertising, etc.
Even this straightforward implication of monopoly theory has important uses.
Suppose we have a sample of "monopolists" that face competition from sellers of
other related products. A natural question is whether they are in fact monopo-
lists, or whether the competition from other firms means that they should be
treated as in a larger, more competitive market. The PR statistic speaks directly
to this question; if they are monopolists, H R should be negative. Unfortunately,
it is not necessarily true that H n has to be positive if the firms are not
monopolists. PR show that in some specific models of oligopoly and of monopo-
listic competition, H n must be positive. Thus, it is appropriate to see H n as a
statistic which has some ability to discriminate among alternative competitive
hypotheses. There can, however, easily be a false finding of monopoly, since
H R < 0 can occur for reasons other than monopoly.

29R(Q, Y) is distinct from R*(.), the reduced-form equation for revenue.

Ch. 17: Industries with Market Power 1037

A second economic hypothesis that can be cast as a test on H R comes when the
markets studied are in LR perfectly competitive equilibrium in the strong sense:
free entry has driven out inefficient firms, and remaining firms produce at the
bottom of their LR A C curves. Then H R = 1. Let MA C be minimum average
cost, and QMA C be the quantity which minimizes A C. A proportional shift in
all factor prices will raise MA C by the same proportion without changing
QMAC. The estimation of H R proceeds using data on the revenue for single
firms. At the firm level, revenue will shift proportionately to cost in LR equi-
librium. At the industry level, revenue will expand less than proportionately to
cost, as the increase in price will lead to lower quantity demanded: in this LR
theory, the supply adjustment comes through entry and exit. Furthermore, the
use of single-firm data is warranted, since the only determinants of price and of
firm revenue in LR equilibrium are MA C and QMAC. 3 On the same argument,
R*(.) should not be a function of demand variables in a test of this hypothesis.
Thus, the reduced-form revenue equation, estimated on firm data, has two
distinct testable restrictions under the hypothesis of LR perfect competition.
The LR flavor of the comparative statics analysis in the PR analysis, both of
monopoly and of competition, is reflected in the existing applications of the PR
statistic, which are on cross-section data in similar local markets. Panzar and
Rosse (1977b) treat the case of newspaper firms in local media markets. An
observation is a newspaper, with its revenue as the dependent variable (of course,
the majority of revenue is from advertising). If newspapers are monopolies, it is
because they do not face intense competition from other media. They are able to
reject the hypothesis that newspapers are monopolies even when they are the only
newspaper in the market: the interpretation goes to the importance of competi-
tion from other media. Shaffer (1982) applies the PR ana!ysis to a cross-section of
banking firms in New York State in 1979, finding that the hypothesis of
monopoly as well as the hypothesis of LR perfect competition could be rejected.
How convincing these studies are depends on two areas: whether estimates on the
cross-section of local areas reveals differences in LR equilibrium, and whether all
of the variables which shift cost have been identified and correctly entered. The
first of these points has been thought through: see Rosse (1970) on the "perma-
nent plant hypothesis". At a minimum, it is clearly important to treat the cases of
markets with rising demand separately from those with falling demand. The
second point is very similar to one discussed in the previous section.
Recent work by Sullivan (1985) and Ashenfelter and Sullivan (1987) has
extended the PR comparative statics in W idea to circumstances where variables
other than revenue are observable: the results are based in the comparative

3Thus, it is appropriate to use only cost shifter exogenous variables in a test of LR perfect
competition, and the reduced-form equations are not misspecified under the null if all demand
variables are omitted.
1038 T.F. Bresnahan
statics of market price and quantity in factor prices. As a result of the additional
observables, they can treat the oligopoly estimation problem of attempting to
draw inferences about conduct. It will be most convenient to write the supply
relation for a typical oligopolist in the conjectural variations form:

Pt -~ Cl(Ot, ~t, Zt, F, ~c,)--[- DI(')Qi(1 + ri(.)). (24)

In Sullivan's treatment, only market-wide data on price and quantity, and

exogenous variables are available. Solving (24) for all firms simultaneously with
the demand curve will yield reduced-form equations for P and Q, call these P*
and Q*. Following PR, estimation of these reduced-form equations could yield
HQ and Hp, the elasticities of Q and of P with respect to marginal cost, possibly
measured by a comparative statics exercise involving a proportional increase in
all factor prices. Assume that all firms have common marginal cost, and let
L~'= ( P - C1)/P be the Lerner index. In our notation, Sullivan shows:

)-~(1 + ri) -1 >__ - n p / ( n Q , . ~ ) >_ - n p / n Q , (25)

i

where the last inequality follows because ~ must be less than or equal to unity
with non-negative MC.
The left-hand inequality in (25) relates one unobservable quantity to another,
because only H l, and H e are estimated by the technique: no estimates of
marginal costs or of the price-cost margin are formed. However, under assump-
tions that M C is no less than zero, the right-hand inequality of (25) does imply a
bound o n the competitiveness of conduct. The statistic on the far right can be
estimated. The larger is the statistic on the far left, the closer is conduct to
competition. Thus, (25) can permit rejection of the hypothesis of successful
collusion, though not of competition.
In his empirical work, Sullivan (1985) uses a cross-section (states of the United
States) time series (years) on the cigarette industry. The crucial exogenous
variable is state taxes, which clearly proxy for MC; all other exogenous variables
are captured in an ANOVA procedure. The analysis obtains a slightly fighter
bound than (25) by assuming costs are at least as large as taxes (paid by the
seller), and is able to reject the hypothesis that cigarette prices are set as if by a
cartel. Ashenfelter and Sullivan (1987) take a nonparametric approach to the
same data, using year-to-year changes in tax rates and in the endogenous
variables in the same state to estimate Hp and H a. In thinking about this
approach, it is clear that its main potential problems in application are similar to
those of PR: Has it been established that the variables which shift MC are not
acting as proxies for any other variables? Is the quantitative relationship between
these variables and MC certain? The use of tax data is obviously particularly
Ch. 17: Industries with Market Power 1039
strong on the second point. I suspect that the first point will usually turn on a
detailed argument from the institutional detail of the particular industry at hand,
from econometric investigations of robustness, and from ancillary data.
Why is it that the comparative statics in cost can only lead to inequality
restrictions on oligopoly conduct, while comparative statics in demand variables
provide an estimate of the degree of oligopoly power? [Compare (25) and (18).]
The answer follows directly from the nature of the econometric exercise in each
case. Consider the two-equation system determining industry price and quantity:
there is a demand equation, and a supply relation. The conduct parameters we
are particularly interested in are in the supply relation. When demand is shifted
by some exogenous variable, it tends to trace out the supply relation, which is
after all what we are trying to estimate. The statistic based on the comparative
statics in cost could very easily identify the demand equation. They can only cast
indirect light on parameters in the supply relation.

3. 3. Estimation of marginal cost more directly

The two methods discussed in the previous subsections have in common that they
treat the comparative statics of the industry or market equilibrium in isolation.
Price and quantity are the only endogenous observables. To the extent that
price-cost margins are estimated, the inference is based on the supply behavior
of firms. I now turn to an alternative approach, which attempts to econometri-
cally estimate M C from cost data or from factor demand data. This approach
uses the methods of cost and factor demand function estimation using flexible
functional forms. It relies heavily on the economic theory of cost as dual to
production. 31
The pioneering work in this area was done by Gollop and Roberts (1979) and
Appelbaum (1979, 1982). Their approaches start from the total cost function,
C(-) [see (2)]. To the observables of the stylized model they add quantity
demanded of factors of production: typically broken down only into labor,
capital and materials (sometimes energy is separate from other materials inputs).
I label the demand for a particular factor of production xk, that for all factors
taken together as X. The key to the approach is to note that M C is the derivative
of C(.) with respect to quantity, C1(.) and that (using standard duality results)
the factor demand equations are the derivatives of C(.) with respect to factor
prices. 32 Then the approach estimates the demand equation, the supply relation,

31Obviously,this approach and the onesdescribedin Subsection3.1 are complementsrather than

substitutes.
32This followsfrom Shephard'slemma: see Diewert(1971).
1040 T.F. Bresnahan
and appends to that system the factor demand equations:

?, = D(Q,, Y,, 8, (26)

Pt ~- c l ( o i t , Wit, Zi,, F, Ecit) - DI(Qt, Yt, 8, edt)QitOit , (27)

Xt = Cw(Q,t, Wit, Z., F, ec,,). (28)

Clearly, appending equations (28) to the system offers at least the possibility of
substantial increases in the precision with which MC can be estimated, since
there will be cross-equation restrictions between the factor demand equations and
the supply relations; the cost parameters F appear in both. It is reasonable to
expect these restrictions to be quite powerful. Since C(-) is necessarily homoge-
neous of degree 1 in W, its derivatives Cw(. ) will (taken together) depend on all
of the parameters of C(.). Thus, all of the parameters in MC also appear in the
factor demand equations.
Clearly, the important questions about the utility of this technique in practice
turn on the success in estimating MC. Questions of the appropriate functional
form for C(.) can probably be addressed by trying several alternatives, or by
using prior information about the industry at hand to specify the technology. To
the extent that (28) includes a demand equation for capital, users of this
approach must face the problem of valuing the capital assets of the firm: capital
needs to be decomposed into the price of capital services and their quantity.
Thus, the body of criticisms of t-he SCPP which centered on the accounting
treatment of capital will likely reappear as criticisms of the cost function
approach. Furthermore, if all factors are treated as SR variable in (28), the
price-cost margins will need to be interpreted as price relative to LRMC.
Another approach to using factor demand information has recently been
introduced by Hall (1986). He starts from the attractive notion that MC could be
directly observable by the conceptual experiment of changing quantity produced,
holding everything else constant, and measuring by how much expenditures on
inputs changed. As the empirical analog of this, Hall works with data on the rates
of change of output and of the labor input. One way to think of this is that
average incremental cost is revealed by the data: the discrete changes in outputs
that occur between sample periods lead to discrete changes in inputs, and the
resulting empirical AIC is taken to be the estimate of MC. The second notion in
Hall is that changes in the labor input alone can reveal MC. Under the
assumption of (LR) constant returns, the wage rate times the change in labor
demand divided by labor's share in cost should be AIC. To date, this approach
has been largely implemented on aggregate data. 33

3SHall only attempts to estimate Z#; Shapiro (1987) extends the same logic to estimate 0 as well.
Ch. 17: Industries with Market Power 1041
This approach has dearly closely related to the previous one, in much the same
way that index number approaches to cost and productivity are related to
econometric cost and production functions, and therefore shares many of the
same advantages and disadvantages. Some of the problems of interpretation have
been overcome: the use of the labor demand only helps somewhat with the
problem of capital valuation, though labor's share in cost still needs to be
calculated. The index-number flavor adds another potential difficulty: if MC is
not flat, AIC can be a poor proxy for it. Since the MC curve of interest is
SRMC, it is unlikely to be fiat in applications.
Although I said earlier that the methods described in this section were a
complement to, rather than a substitute for, other methods, empirical practice
done not yet reflect this. There are two regularities in scholarly practice to note.
First, none of the papers cited in this subsection uses industry detail to provide a
defense of its maintained hypothesis. Second, all of the papers in the literature
can be divided into two classes: those cited in this subsection, which argue
identification argument only from the restrictions between MC and factor
demands, and all other papers described in this chapter, none of which tried to
use factor demands to get better estimates of MC.

3.4. Supply shocks

I argued above that a core implication of modern theories of cartels, as well as an

ancient empirical assertion about them, is that their conduct is not constant over
time. I would now like to return to Porter's (1983) [Lee and Porter (1984)]
switching regressions method for determining this. The question will be: What is
it in the data that identifies the inference that cartels break up and reform? I will
write (4") slightly differently. Under the assumption that there are two kinds of
conduct, which I think of as "price wars" and "collusion", there are two supply
relations that hold in the data:

P = C1( ) - - DI(Yt, 8)OOr, during price wars, (29)

P = C1( ) - DI(Yt, 8)QOc, during periods of collusion. (30)

Porter completes the model by specifying the (constant) probability ~r that (29)
holds vs. (30). In price-war periods, prices and quantities are determined by the
intersection of (29) and of the demand curve (1), while in periods of successful
collusion, these are determined by (30) and (1). The analyst does not know
whether there are in fact these two different regimes in the data: that inference is
to be drawn from the pattern of prices, quantities and exogenous variables.
A natural question to ask is why this inference can in fact be drawn. I believe
that the inference comes from a particular shape of the joint distribution of P
1042 T.F. Bresnahan
and Q conditional on Z, W and Y. Let us hold all of those exogenous variables
fixed at some arbitrary levels. Let the P and Q which solve (29) and (1) be called
Pr and Qr, and those which solve (30) and (1) be called Pc and Qc- These are
random variables, since all of (29), (30), and (1) have econometric error terms.
But the two different random variables have two different centers of distribution.
In the r-regime, the mean of price will be lower and the mean of quantity will be
higher: the regimes differ only in that the r-regime has a lower supply relation
than the c-regime. If 0c is much larger than Or, i.e. of collusion is successful at all,
we should expect these two means to be far apart.
Now consider the entire distribution of P and Q conditional on the exogenous
variables. It has two local modes: one each at (Pr, Qr) and (Pc, Qc). Empirical
techniques for dealing with bimodal distributions, of which the switching regres-
sions method is a leading example, will be able to detect the presence of the two
modes. Thus, the Stigler-esque theories do have an idiosyncratic implication
about the shape of the distribution of prices and of quantities. This line of
inference departs somewhat from those described above, where the emphasis was
on the comparative statics in observable exogenous variables. Here, the variable
describing which regime the industry is in was taken to be unobservable. The
implications of the theory for the data were drawn by making a simple assump-
tion about that unobservable: that it took on two distinct values. Thus, the nature
of the inference here comes from identifying a specific component in the error
term: a component that enters the system as a supply shock.
The main potential difficulty with this inference is this: some unobserved shock
other than changes in conduct may be moving in the supply equation. Since the
inference is based on an error component, there is nothing in the procedure itself
to guarantee that the conduct interpretation is the right one. If there are
uncaptured changes in factor prices (recall Porter has no factor-price data) or
shocks to technology, these could shift the supply relation as well. The size of the
effect in the railroad data - the shocks lead to changes in both price and quantity
on the order or 50 p e r c e n t - makes it appear likely that the conduct interpreta-
tion is the right one, particularly in light of the extensive contemporaneous
discussion of cartel adherence. 34

3.5. Comparatioe statics in industry structure

The methods for identification of market power described in the previous four
subsections yield estimates of the degree of power over price of a particular

34The alternative approachis to decide a priori on the sample split, and then attempt to separately
estimate 0c and Or, as in Bresnahan (1987). This has the substantial disadvantage of requiring prior
information, but the advantage of being able to more directly assess whether the apparent shocks to
supply are due to changes in conduct.
Ch. 17: Industries with Market Power 1043

industry in its particular setting. The industry's structure, in the SCPP sense, is a
given of the analysis. Substantial time-series changes in industry structure are
rare events; thus, the single-industry case study method only rarely, and only on
some bodies of data, permits the question of how changes in market structure
affect conduct and performance. Methods based on cross-sections of similar
markets have also cast some new light on the relationship between industry
structure and market power. Unlike earlier mainstream work, which used
accounting profit as the dependent variable, many recent studies use price or a
price index as the dependent variable. The goal of the investigation is to see how
concentration affects prices. Let me briefly outline the work in this area before
discussing its interpretation.
The cross-section study of similar markets has been focused on businesses that
are geographically local. The dependent variable is price, either a price index or
one of the prices of a multi-product firm. The estimating equation is typically a
reduced form for price. The industries studied include banking, for which
Rhoades (1982) lists dozens of studies, retail food [Cotterill (1986), Lamm
(1981)], gasoline suppliers [Marvel (1978)], airline city-pair markets [Graham,
Kaplan and Sibley (1983)], cement [Koller and Weiss (1986)] and no doubt
others. These studies typically take concentration to be exogenous. 35 The equa-
tion they estimate is therefore close to the reduced-form equation for price,
departing from it only in that quantity (as transformed into a concentration
measure) is included as exogenous. These studies confirm the existence of a
relationship between price and concentration, which is at least suggestive of
market power increasing with concentration. An interesting variant uses time-
series changes in industry structure: see Barton'and Sherman (1984) on the effects
of a merger in the microfiche film industry on prices and profits.
Most of these studies offer the interpretation that the empirically estimated
relationship can be interpreted to cast light on the prediction of almost all
theories of oligopoly that higher concentration causes higher price-cost margins
by changing conduct. I have seen no careful defense of this interpretation, and I
am troubled by it; I offer a series of interpretational difficulties here not because I
believe they are true but because they have not yet been rebutted.
If markets are less concentrated when they are larger, and more firms will
"fit", then what relationship are we seeing in the data? Take the stark case of free
entry as soon as entrants' profits are positive. We interpret the relationship in the
data as being about concentration and P - MC, yet there is another equation in
the model: P = A C for entrants. In the larger markets, firms are also larger, have

3SAn important exception is Graham et al. (1983) which tests for the exogeneity of its concentra-
tion measures. Exogeneity is not rejected, even though the coefficients of the concentration measures
change substantially when they are treated as endogenous. This suggests that exogeneity cannot be
rejected because the test has little power, rather than because the assumption is substantively
innocuous.
1044 T.E Bresnahan
lower (average) costs, but everywhere firms break even. This reinterpretation is
not necessarily a hostile one, but the welfare economics are somewhat different:
price and concentration are related in a way that has no obvious bad effects, and
does not imply entry barriers.
A somewhat different endogeneity problem arises within industries with the
same number of firms. If the firms are selling the same products, then the more
concentrated industry likely has more heterogeneity in costs. Greater heterogene-
ity in costs might interact with conduct in a way that increases prices or it might
not. 36
Even given the exogeneity of concentration, if firms are heterogeneous in their
cost functions, markets with more firms allow more statistical "draws" on the
lowest-cost firm. One should expect, on average, that the lowest-cost firm out of
five has lower costs than the lowest out of three. If the lowest-cost firm is
particularly important in the determination of price, as in some competitive
models as well as in some oligopoly models, then this purely statistical effect will
result in lower P in less concentrated markets. Since estimates that link price to
concentration are necessarily on market-wide data, we are in the world of
Subsection 2.2.4. The crucial equation is:
P = Average[MCl + D I ( . ) Q Average[0], (17')
where the notation Average[. ] means share-weighted average taken over firms in
the market. In general, we do not know whether it is Average[MC] or Average[0]
or both that is lower in the less concentrated markets. It is the latter interpreta-
tion most authors have in mind. Furthermore, there can be links between these
two: a firm with a substantial cost advantage may have less competitive conduct
than it would facing more equal competition.
That last point can be fleshed out with some observations of the way actual
heterogeneity has been sometimes explicitly measured. Consider the Graham
et al. (1983) finding that not only concentration affects price in airline city-pair
markets, but also that who the competitors are matters. Markets in which one or
some competitors are new entrants into the airline business overall have lower
prices than other markets, all else equal. It is extremely likely (see Graham et al.,
section 5) that these entrants have lower MC. It is also possible that their
presence changes the conditions of competition. Which is it? The analysis of the
paper cannot say. Exactly the same question applies to the interpretation of the
industry structure dummies used in Porter (1983) [see (4") above].
These questions of interpretation are not unanswerable; the previous four
subsections discussed explicitly methods of telling MC from 0. The questions are,
however, unanswered.

36Spiller and Favaro (1984) escape the problem of endogeneityin their time-series study of
Uruguayan banking. Their sample period includes a change in the regulation of entry into their
industry. They find that freer entry shifts the supplyrelationdownward.
Ch.. 17: Industries with Market Power 1045

3.6. On the identification of market power

This section has reviewed a large body of method, all developed in the recent
past, for empirical investigation of the hypothesis of market power. Several
distinct lines of argument have been advanced, each of which relies on a distinct
implication of market power for identification. 37 This variety reflects the variety
in the data available in different industry studies. In any particular industry, the
available information and institutional detail allows different kinds of analysis
and different defenses of different analyses. We can therefore expect some
continuing variation in desired method. It also reflects the great many implica-
tions of the comparative statics of equilibrium in markets with market power
which are not found in competitive markets.

4. Market power in product-differentiated industries

Product differentiation raises two kinds of empirical questions, loosely divisible

into the SR and LR. In the SR, the stock of products offered by firms and the
attributes of these products is fixed. In the SR the measurement question of
interest is how much monopoly power firms have because of existing product
differentiation. This is (at least) a two-part question. First, as firms' products
grow more distinct, each firm's profits will depend less on the policies of other
firms. This first part of the question can be adequately answered by investigation
of the elasticities of demand, including the cross-elasticities. Second, as products
grow more distinct, each firm will respond less to competitive moves by rivals.
Understanding this part of the question requires an empirical model of competi-
tive interaction. The measurement of this SR market power has seen tremendous
progress in recent times, and that forms the subject of this section.
The market power conceptual issues associated with the SR product differenti-
ation questions are not all that distinct from those in the single-product case. The
measurement problems are more severe, however. There are more demand
parameters to be estimated: even under the assumption of constant elasticity (or
slope) and symmetry, an N-product industry has N own-price elasticities, N
income elasticities, and (N - 1)N/2 cross-price elasticities. There is almost no
industry for which the position that there are more than 100 products is
untenable: without putting more structure on the problem, the analyst could need
to estimate literally thousands of elasticities. On the cost side, the fact that firms
produce multiple products suggests the existence of economies of scope: the cost

37At least one false identification argument has been proposed, as well. Koutsoyiannis (1982)
argues that sales maximization by oligopolists can be empirically distinguished from entry-deterring
behavior and from static profit-maximizing behavior. His model assumes monopoly rather than
oligopoly: see his equation (26).
1046 T.F. Bresnahan

function may have some new complexities as well. The use of prior information
to guide the specification of the model becomes crucial in such circumstances.
Fortunately, in many contexts, prior information will be available from sources
like industry trade journals, marketing studies, and so on. As a result of the
industry specificity of this prior information, there is considerable variation
across industries in the way one would like to proceed with the analysis. 38
Far and away the most common technique for apparently product-differenti-
ated industries is to assume that the products in the industry are basically fairly
close substitutes, use an index of several products' prices as the observable price,
and proceed. 39 This procedure is not inherently wrong. It may, however, result in
the attribution of market power to noncompetitive conduct when in fact the
source of the market power is differentiated products.
When the analyst wishes to study the product-differentiation issues directly,
some procedure to reduce the complexity of the analysis from its full size must be
employed. There is some experience with three general forms: modelling the
product choice part of the demand system, aggregating similar products until
there are only a few left in the system, and estimating only a few functions of the
parameters of interest.
Tools for the product choice elements of demand have been a major topic of
econometric theory and practice in recent times. The work of McFadden (1982)
and others on discrete choice has provided a framework for modelling individu-
a/s' choices of products. These techniques, such as nested logit models, are
clearly appropriate in circumstances where there is prior information about
groupings of products, such as when industry sources emphasize the existence of
distinct product segments within which competition is much more direct than
without. A parallel literature, in the theory of "spatial" product differentiation,
has concentrated on the relationship between heterogeneity in consumers' tastes
and the demand curves facing differentiated oligopolists. It is more appropriate
to industries in which there are no clear segment boundaries, i.e. where the fact
that products A and B are both important parts of the competitive environment
of C need not imply that A is an important part of B's environment. The spatial
models thus emphasize the localization of competition as a way to reduce the
number of demand parameters, while the discrete choice models emphasize
grouping. 4 Both modelling approaches treat product quality similarly. Not

3SSomeother approaches have been attempted as well. Haining (1984) uses the spatial autocorrelao
tion of prices of retail gasoline stations to attempt to infer something about the pattern of interaction
among them. I could see not relationship between his statistical hypotheses- "pure competition" is
the name of one and "supply and demand" is another- and any economichypothesis.
39See,for example, Gollop and Roberts (1979), Roberts (1983), and Gerosld(1983) on roast coffee;
Appelbaum (1982) on tobacco and textiles; Sullivan (1985) and Ashenfelter and Sullivan (1987) on
cigarettes.
4In the limit, models like multinomiallogit (without any nests) have the entire industry forming
the market segment. Then competitionis completelysymmetric.
Ch. 17: Industries with Market Power 1047

surprisingly, there has been considerable interest in these models in the market-
ing field; Schmalensee and Thisse (1986) provide an overview of both the relevant
economics and marketing literatures.
Empirical examples of this approach can be found in Bresnahan (1981, 1987),
which use a spatial model of the demand for automobiles by type as the demand
system. The flavor of this approach is that explicit functional form assumptions
are made about the distribution of demands across individuals. These distribu-
tions, in turn, determine the form of the aggregate demand system. As in
econometric work in discrete choice, typical distributional assumptions lead to
empirical models with many fewer parameters than the unstructured approach
described above. The degree to which such an analysis is convincing turns
critically on the quality of the information used to specify the demand system.
The best procedure for this is undoubtedly a close reading of the industry trade
journals and of typical marketing practice.
Nonetheless, any approach which begins with a highly structured demand
system naturally raises questions about the appropriateness of the particular
structure. The distinction between localized competition and more systematic or
segmented models is particularly important in this regard. Schmalensee (1985)
devises test statistics for competitive localization that uses only the measurable
movements of endogenous variables, the prices and quantities of particular
brands. If exogenous shocks to the system are either market wide (i.e. shift the
demand or supply of all products together) or are product-specific, then the
extent to which particular products' prices and quantities tend to move together
are an indicator of competitive localization. In an application to the ready-to-eat
breakfast cereal industry, Schmalensee is able to decisively reject the symmetric
model in which all products compete equally. The particular pattern of localiza-
tion implied by his estimates leads him to doubt the (covariance) restrictions that
identify the degree of localization, however.
The approach of aggregating products until there are only a few elasticities to
be estimated was taken up by Gelfand and Spiller (1987), Suslow (1986), and
Slade (1987). Gelfand and Spiller use data on banking firms competing to make
loans of a great many different types. They aggregate the loans until only two
types are left, and investigate the demand elasticities in the resulting two-by-two
system. An important advance in their work is a model of interrelated oligopoly
in the multiple markets, as firms' profits in each market are affected by strategies
of other firms in both markets, or even possibly strategies that link the two
markets. 41 Presumably such effects can only be studied when the number of
markets has been reduced to a reasonably small level. Slade's (1987) treatment
of gasoline station "majors" and "independents" is quite similar. The work of

41Gelfand and Spiller cast this intermarket interdependence in CV form: each firm has "conjec-
tures" about how other firms will "react" in each of the two markets.
1048 T.F. Bresnahan
Suslow (1986) aggregates outputs into two: all of those produced by the domi-
nant firm (Alcoa in the aluminum industry before the Second World War) and
those produced by the fringe. The dominant firm's M R is a function of the
degree of substitutability between its product and the product sold by the fringe,
as well as by the usual determinants, the market demand elasticity and the fringe
supply elasticity.
A third approach to the problem of multiple products has been taken by Baker
and Bresnahan (1983, 1985). In their approach, the problem of estimating all of
the cross-elasticities of demand is avoided by estimating only the interesting
summary statistics of the demand for that product. To estimate the market power
associated with a particular product, it is unnecessary to estimate all of the effects
of all of the other products' prices in the market. Instead, only the total effect of
competition from other products as a brake on the pricing power of the firm
owning a particular product is of interest. Consider the seller of product 1, facing
the demand system: 42

P1 = D ( Q , , P2 . . . . . IN, Y, ~), (31)

where P2 . . . . , PN are the prices of products of competitors and econometric

errors are suppressed for convenience.
The measurement problem is that the demand parameter vector 8 can be very
long, containing the cross-elasticity of demand with each of the 2 , . . . , N prod-
ucts. Baker-Bresnahan substitute out 1'2 . . . . , PN in (31) by solving the supply
and demand equations for each of them. Suppose that there are ( N - 1) more
equations like (31), one for each Of the other products. Also, there is a supply
relation for each of products 2 . . . . . N:

Pi = Cl(Qi, Wi, zi, I'i) - DI(Q, Y, 8)QiOi, i = 2 . . . . , N, (4 d)

where Q is the vector of all firms' (and equivalently, all products') quantities. The
2 * ( N - 1) equations (31), (4 d) can be solved for the prices and quantities of
products 2. . . . . N. Call the solution for P:

Pi = eT ( Ql, Wu, zN, ru, oN, r, ~),

where the dependence on Q1 arises because only products 2,..., N have been
solved out, and the subscript I refers to the superset of all the subscripts i. The
equation to be estimated is:

P1 = D(Q1, P f f ( ' ) . . . . , P ~ ( . ) , Y, 8)

= DR(Q1, Wn, ZN, IN, ON, Y, t~), (32)

42This is slightly unfamiliar notation. It wouldbe more familiar to write quantity for this product
as a function of the prices of all products: the functionpresented is simply the inverseof that.
Ch. 17: Industries with Market Power 1049
the residual demand curve for product 1. There are two immediate observations
here. First, an enormous amount of information has been lost here by substitut-
ing out the prices of all the other products; it will be impossible to estimate all of
the separate elements of 8 from (32), much less all of the other parameters in it.
But this is of little importance. The elasticity of D R with respect to QI tells us
how much power the firm has over product l's price, taking into account the
adjustment of all other firms' prices and quantities.
A somewhat similar example may clarify the technique. Suppose that (4 d)
takes the form P i - - M C i ( W i ) for all the other firms: they are price-takers.
Equation (32) then predicts the price of product 1 as a function of its quantity,
and variables shifting the costs of all other products. If firm 1 has no market
power in this product, the prices (and in this example, the costs) of other
products will determine its price. In the no-market-power case, the elasticity of
D R with respect to Q1 will be zero.
Thus, the Baker-Bresnahan approach estimates the demand elasticity facing
the single firm or product, taking into account the competitive reaction of all
other firms in the market. This demand elasticity summarizes the market power
of the firm: knowing it is insufficient to determine the sources of that market
power. 43
The relationship between the Gelfand-Spiller or Suslow approach and the
Baker-Bresnahan approach is this: in the first approach, all of the elasticities of
demand, supply, and competitive interaction are estimated. From them the
market power of any particular firm could be calculated. Of course, for practical
implementation this approach requires specifying a relatively small number of
different products. The second approach works when there are a large number of
products, but does not yield estimates of all of the elasticities, only of the
summary statistics relating to each firm's market power.

4.1. "'Market definition", policy analysis, and product differentiation

Some of the techniques for assessing market power have been applied to the
problem of "market definition" in antitrust analysis. In antitrust applications, it
is frequently of some importance to determine whether a group of firms would
have any market power if they chose to act in concert, or in other contexts
whether a single firm in fact has any market power. 44 The latter question can be

44These questions are well-posed, even where the usual method of answering them, defining a
"relevant market" and calculatingmarket shares in it, is senseless.
43A11of this presumes that the residual demand curve can be estimated. The condition for that is
that firm l's costs have movedindependentlyof all other firms' costs. An obvious appficationof the
technique, therefore, is in the international context. One would ask how steep the demand curve
facing producers in a single country was; the natural experimentfor estimating that quantity would
be good if, for example, exchangerate movementshad movedreal relativecosts in differentcountries.
1050 ZF. Bresnahan
directly answered by the Baker-Bresnahan technique. Scheffman and Spiller
(1987) extend the Baker-Bresnahan technique to estimate the elasticity of de-
mand facing a group of firms, thereby providing an answer to the former
question. Baker and Bresnahan (1985) ask how much steeper the demand curve
facing two firms would be post-merger than the pre-merger level. If two firms sell
products that are very close substitutes, then each likely provides an important
part of the competitive brake on the other, unless there are several other firms
providing similar products as well. The increase in the steepness of the residual
demand curve measures this effect.
Methods based on the Panzar-Rosse statistic have also been u s e d in this
connection. 45 Panzar and Rosse (1987) give a new interpretation to their mo-
nopoly test which is directly relevant here. Suppose a particular firm (firms) has
been studied by PR methods. A rejection of "monopoly" for this firm (group of
firms) implies that it (they) cannot be treated as acting in isolation. Other firms
must be interacting with the firm (firms) at hand.
In a slightly different context, Schmalensee and Golub (1983) examine the
spatial product differentiation of firms in the electricity market. They use models
of the demand for electricity, the costs of transmission, and of competition to
assess the likely impact of deregulation.

4.2. Product differentiation in the L R

In the LR, firms can add products, change their attributes (either physically or in
consumers' perceptions) or new firms can enter with new products or imitations.
This is a very complex area, full of hypotheses. Strategic interaction effects of
many kinds are possible: preemption by establishment of a reputation for
product superiority, preemption by filling out the product space, coordination of
investment in distinct product types so as to reduce competition, and so on.
Essentially nothing empirical is known about any of these hypotheses. Further-
more, the welfare implications of SR market power in a product differentiated
industry are not transparent. In the Chamberlinian tangency of monopolistic
competition, every firm has market power in the sense of this section. Yet that
does not establish that there is any inefficiency, once the need to cover the fixed
costs of product design are taken into account. The crucial issue here is an
adequate empirical treatment of the supply curve of new firms and of new,
different products. Empirical work on this area is likely to be forthcoming soon,
but little exists now. 46

45See Shaffer(1983) and Slade (1983).

46Many of the relevant analytical issues to support empirical work can be found in Panzar and
Rosse (1981). Bresnahan and Reiss (1986) estimate an equation for the entry of a second product
differentiated firm into a monopoly.
Ch. 17: Industries with Market Power 1051

5. What has been learned about market power?

However useful its methodological contributions, the industry case-study nature

of much work in the NEIO has raised questions of interpretation. How general
are the results? What do these studies, taken together, reveal about market power
in the economy as a whole? What have we learned about the conditions under
which market power tends to arise? What is known about the easily measured
correlates of market power, such as concentration? In short, a single industry case
study cannot paint a broad picture; it can only reveal the nature of industry
conduct and performance in the industry studied. The original idea of the SCPP
was that empirical research would estimate a function mapping structural char-
acteristics into measures of conduct (where that is possible) and of performance
(more commonly). Empirical knowledge of this map is obviously valuable. It
contains information about the sources as well as the location of market power. It
could be used to guide policy in those areas, such as merger policy, where it could
influence structure rather than conduct directly. An industry case study, whether
done by the methods of the 1930s or the 1980s, can hope to reveal at most one
point on the function. The integration of different case studies to give a unified
picture of the whole map is an obviously attractive prospect. It can only be
partially carried out now, even though the empirical papers described in this
chapter have treated well over a dozen industries. In Table 17.1 I reproduce the
estimated price cost margins (Aa) from several different NEIO studies reviewed

Table 17.1
Summary of existing empirical work

Author Industry .LP

Lopez (1984) Food processing 0.504

Roberts (1984) Coffee roasting 0.055/0.025 a
Appelbaum (1982) Rubber 0.049 c
Appelbaum (1982) Textile 0.072 c
Appelbaum (1982) Electrical machinery 0.198 c
Appelbaum (1982) Tobacco 0.648 c
Porter (1983) Railroads 0.40 b
Slade (1987) Retail gasoline 0.10
Bresnahan (1981) Automobiles (1970s) 0.1/0.34 d
Suslow (1986) Aluminum (interwar) 0.59
Spiller-Favaro (1984) Banks "before ''e 0.88/0.21 f
Spiller-Favaro (1984) Banks "after ''e 0.40/0.16 f

Largest and second largest firm, respectively.

bWhen cartel was succeeding: 0 in reversionary periods.
c At sample midpoint.
d Varies by type of car; larger in standard, luxury segment.
e Uruguayan banks before and after entry deregulation.
f Large firms/small firms (see their table 2).
1052 T.F. Bresnahan

here. 47 Different scholars will undoubtedly differ on the extent to which it offers
answers to the questions of the last paragraph. A few preliminary conclusions,
however, are available. These are cast in somewhat guarded language primarily
because of the limited coverage of the available studies.

Conclusion A

There is a great deal of market power, in the sense of price-cost margins, in some
concentrated industries.

The conclusion seems almost forced by the last column of Table 17.1. Several
studies have found substantial power over price. Available data do not permit a
systematic assignment of concentration indexes to the industries listed in Table
17.1, since they are not drawn from the economic censuses; several are based on
the primary data gathering of different scholars. A glance down the list of
industries, however, is sufficient to demonstrate that they are overwhelmingly
drawn from the highly concentrated end of the industrial spectrum.
Finding A, I think, cannot be controversial, particularly with its qualification
that it refers to "some" concentrated industries. The finding would be less
controversial without the "some", and I think this is right. There are at least two
reasons to suspect the generality of the findings in the papers reviewed in this
chapter. First, authors who invent methods for the detection of market power are
likely to first apply them in industries where they expect to find it. Thus, the
existing studies have largely treated quite concentrated industries, industries
where there were known or suspected cartels, industries where a solid old-style
case study suggested anticompetitive conduct, and so on. The field is now ripe for
revisionism! Or at least for continued expansion of the set of industries in which
conduct and performance are well measured.
Second, the list of industries studied to date is special in another sense. Since
the data are often drawn from trade journals, regulatory bodies, court proceed-
ings or similar sources, the industries covered may be unrepresentative. Consider
the industries with excellent trade journals; they are ones in which information
about what competitors are doing is quite good. Thus the repeated finding of
successful collusive arrangements might reflect the particular information struc-
ture of these industries. Similarly, Suslow's ability to mine the trial transcript in

47Afew papers were left out of the table because their estimates were not given in such a form as to
permit calculation of the Lemer index. (These papers heavily emphasize conduct over performance, of
course.) Panzar-Rosse methods and Sullivan methods are excluded because they do not provide an
estimate of the Lemer index. Instead, they provide an estimate of conduct parameters or a test of
certain hypotheses about conduct. (Comments on a draft of this chapter suggest that this is not well
understood: see Subsection 3.2, above.) Baker-Bresnahan methods are excluded because the demand
elasticity estimates they provide correspond exactly to the Lemer index only in certain circumstances.
(See Section 4.)
Ch. 17; Industries with Market Power 1053

Alcoa for data arose because there was reason to suspect that firm had substan-
tial market power. More new data sources are needed!

Conclusion B

One significant cause of high price-cost margins is anticompetitive conduct.

The studies under review distinguish between conduct, in the sense of firms'
behavioral rules for price-fixing, and performance, in the price-cost margin
sense. It is not the case that, systematically, we see tiny departures of conduct
from price-taking plus very steep demand curves leading to large departures of
performance from the competitive standard. Instead, some of the studies appear
to be finding conduct well toward the collusive end of the spectrum. For example,
Porter (1983) and Bresnahan (1987) both find explicitly collusive behavior. I
should emphasize that conduct is not uniform. Roberts (1984), for example, finds
that most of the firms in his industry should be classified in an (essentially)
price-taking fringe. The largest firms have much less competitive conduct, but
have not succeeded in raising prices to the profit maximizing level given the
fringe's behavior. The variety of conduct across industries, as well as the variety
of performance, suggest the importance of the continued study of market power
as a phenomenon.

Conclusion C

Only a very little has been learned from the new methods about the relationship
between market power and industrial structure.

There are two points here, one implicit in the discussion below Conclusion A.
Table 17.1 is drawn mostly from the highly concentrated end of the industry
spectrum. We therefore have new information about the map from structure to
conduct and performance over only a very limited range. The second point is that
the causes of market power have not been addressed by very many of these
studies. One presumes, for example, that long-surviving market power is an
indicator of some failure of the entry process to discipline conduct. Yet entry has
hardly been discussed in the papers. That leads, naturally enough, to my final
section.

6. The future: The sources of market power

This should properly be a short section: although the NEIO has had a great deal
to say about measuring market power, it has had very little, as yet, to say about
1054 T.F. Bresnahan

the causes of market power. In particular, the topics of entry, predation, entry
deterrence, and strategic competition in the LR generally have not yet been
extensively taken up in empirical work with explicit theoretical foundations.
These topics, then, remain primarily for the future. There are a few scattered
contributions.

6.1. Predation

Bums (1986) casts considerable light on the possibility and profitability of

predatory pricing in a study of the tobacco trust. He finds that acquisitions of
competitors made by the trust became cheaper after predatory incidents. Since
predation has long been believed not to be an equilibrium phenomenon, this is a
useful and important contribution. It relies on nineteenth-century data, perhaps
the unique data available for its method. The tobacco trust often thought it was
predating; in that era before antitrust law prevented such self-revelation, the trust
left a solid documentary record of when the predatory incidents took place.
Investigations of the trust's behavior by the Federal government and the courts
yielded a rich data source for Bums' study. More circumspect modem firms will
be less well documented. We have not yet taken the methodological step of
discovering the empirical implications of predatory conduct when acts of preda-
tion must be inferred, not observed.

6.2. Entry

The problem of entry has received some useful methodological contributions

have been made in working papers. Panzar and Rosse (1977b) work out an
empirical model of LR entry, treating the number of firms in the industry as a
continuous variable. They argue that the comparative statics of monopolistically
competitive industry equilibrium have distinctive, testable implications. The key
to their argument is the addition of a third set of equations to (1), demand, and
(4), pricing. The new equations take the form:

P = LRAC(W, Y) (33)

for the marginal firm. The application of this model, or similar ones, in monopo-
listic competition contexts is obviously an attractive prospect.
Bresnahan and Reiss (1986b) take up the problem of econometric models of
entry with an integer number of firms. They provide an empirical application to
monopolies and duopolies on a sample of automobile dealers in small, isolated
towns. They model the entry decision of each of the first two firms into a market,
Ch. 17: Industries with Market Power 1055

and exploit the comparative statics of the level of firm profits in the size of the
market to draw inferences about monopoly and duopoly conduct and entry
behavior. Like Panzar and Rosse, they argue strongly that the right kind of
sample for the study of entry is a cross-section of closely related markets. The
particular characteristics of those markets in which more firms have entered will
reveal the determinants of firm profitability.
Clearly much more work is needed on the determinants of, and the effects of,
entry. In Subsection 3.4, above, I discussed the existing literature on the effects of
entry and concentration on price in cross-sections of related markets. A careful
working out of the analytics of the number of firms in an industry, their sizes,
and so on, is a crucial step in the successful analysis of the effects of entry. This is
one area in which we do not lack for data; many studies treating entry as
exogenous have already been carried out.

6.3. Final remark

In stating the need for further study so strongly, I do not mean to suggest that
the accomplishments to date are small. By departing from the tradition of
treating performance as observable in accounting cost data, the N E I O has
provided a new form of evidence that there is substantial market power in the
economy, a form of evidence that is not susceptible to the standard criticisms of
earlier approaches. Furthermore, the individual studies of particular industries
are specific and detailed enough that alternative explanations of the findings can
be rebutted. The current state of affairs is quite encouraging: we know that there
is market power out there, and need to know a lot more about exactly where. We
know essentially nothing about the causes, or even the systematic predictors, of
market power, but have come a long way in working out how to measure them.

References

Abreu, D., Pearce, D. and Stacchetti, E. (1986) 'Optimal cartel equilibria with imperfect monitoring',
Journal of Economic Theory, 39:251-269.
Anderson, J. (1984) 'Identification of interactive behavior in air service markets: 1973-76', Journal of
Industrial Economics, 32:489-507.
Anderson, J. and Kraus, M. (1984) 'An econometric model of regulated airline flight rivalry',
Research in the Economics of Transportation, 2.
Appelbaum, E. (1979) 'Testing price taking behavior', Journal of Econometrics, 9:283-294.
Appelbaum, E. (1982) 'The estimation of the degree of oligopoly power', Journal of Econometrics,
19:287-299.
Ashenfelter, O. and Sullivan, D. (1987) 'Nonparametric tests of market structure: An application to
the cigarette industry', Journal of Industrial Economics, forthcoming.
Bain, J.S. (1951) 'Relation of profit rate to industry concentration: American manufacturing,
1936-40', Quarterly Journal of Economics, 65:293-324.
1056 T.F. Bresnahan
Baker, J. and Bresnahan, T. (1983) 'Estimating the elasticity of demand facing a single firm: Evidence
for three brewing firms', mimeo.
Baker, J. and Bresnahan, T. (1985) 'The gains from merger and collusion in product differentiated
industries', Journal of Industrial Economics, 33:427-444.
Barton, D. and Sherman, R. (1984) 'The price and profit effects of horizontal merger: A case study',
Journal of Industrial Economics, 33:165-177.
Bresnahan, T.F. (1980) 'Three essays in the American automobile oligopoly', dissertation, Princeton
University.
Bresnahan, T.F. (1981) 'Departures from marginal-cost pricing in the American automobile industry:
Estimates for 1977-1978', Journal of Econometrics, 11:201-227.
Bresnahan, T.F. (1982) 'The oligopoly solution concept is identified', Economics Letters, 10:87-92.
Bresnahan, T.F. (1987) 'Competition and collusion in the American automobile oligopoly: The 1955
price war', Journal of Industrial Economics, forthcoming.
Bresnahan, T.F. and Reiss, P. (1985) 'Dealer and manufacturer margins', Rand Journal of Economics,
16:253-268.
Bresnahan, T.F. and Reiss, P. (1986) 'Entry in monopoly markets', mimeo.
Bulow, J. and Pfleiderer, P. (1983) 'A note on the effect of cost changes on prices', Journal of Political
Economy, 91:182-185.
Bums, M. (1986) 'Predatory pricing and acquisition cost of competitors', Journal of Political
Economy, 94:286-296.
Chamberlin, E.H. (1933) The theory of monopolistic competition. Mass.: Harvard University Press.
Cotterill (1986) 'Market power in the retail food industry: Evidence from Vermont', Review of
Economics and Statistics, 68:379-386.
Cowling, K. and Waterson, M. (1976) 'Price cost margins and market structure', Economica,
43:267-274.
Demsetz, H. (1974) 'Two systems of belief about monopoly', in: H.J. Goldschmld, H.M. Mann and
J.F. Weston, eds., Industrial concentration: The new learning. Mass.: Little, Brown.
Diewert, E. (1971) 'Application of Shephard's duality theorem', Journal of Political Economy,
79:481-507.
Gelfand, M.J. and Spiller, P. (1987) 'Entry barriers and multi-product oligopoly', International
Journal of Industrial Organization, forthcoming.
Geroski, P. (1983) 'The empirical analysis of conjectural variations in oligopoly', mimeo.
Geroski, P., Phlips, L. and Ulph, A. (1985) 'Oligopoly, competition and welfare: Some recent
developments', Journal of Industrial Economics, 33:369-386.
Gilbert, R. (1984) 'Bank market structure and competition - a survey', Journal of Money, Credit and
Banking, 16:617-645.
Gollop, F. and Roberts, M. (1979) 'Firm interdependence in oligopolistic markets', Journal of
Econometrics, 10:313-331.
Graham, D.R., Kaplan, D.P. and Sibley, D.S. (1983) 'Efficiency and competition in the airline
industry', Bell Journal of Economics, 14:118-138.
Green, E.J. and Porter, R.H. (1984) 'Noncooperative collusion under imperfect price information',
Econometri ca , 52:87-100.
Haining, R. (1984) 'Testing a spatial interacting-markets hypothesis', Review of Economics and
Statistics, 66:576-583.
Hall, R. (1986) 'The relationship between price and cost in U.S. industry', mimeo.
Iwata, G. (1974) 'Measurement of conjectural variations in oligopoly', Econometrica, 42:947-966.
Just, R. and Chem, W. (1980) 'Tomatoes, technology and oligopsony', Bell Journal of Economics,
11: 584-602.
Koller, R. and Weiss, L. (1986) 'Price levels and seller concentration: The case of Portland Cement',
mimeo.
Koutsoyiannis, A. (1982) 'Goals of oligopolistic firms: An empirical test of competing hypotheses',
Southern Economic Journal, 51:540-567.
Lamm, R. (1981) 'Prices and concentration in the food retailing industry', Journal of Industrial
Economics, 30:67-78.
Lau, L. (1982) 'On identifying the degree of competitiveness from industry price and output data',
Economics Letters, 10:93-99.
Ch. 17: Industries with Market Power 1057

Lee, L.F. and Porter, R.H. (1984) 'Switching regression models with imperfect sample separation
information- with an application on cartel stability', Econometrica, 52:391-418.
Lerner, A. (1934) 'The concept of monopoly and the measurement of monopoly power', Review of
Economic Studies, 11:157-175.
Lopez, R.E. (1984) 'Measuring oligopoly power and production responses of the Canadian food
processing industry', Journal of Agricultural Economics, 35:219-230.
Marvel, H. (1978) 'Competition and price levels in the retail gasoline market', Review of Economics
and Statistics, 60:252-258.
McFadden, D. (1982) 'Econometric models of probabilistic choice', in: C. Manski and D. McFadden~
eds., Structural analysis of discrete data with econometric application. Cambridge, Mass.: MIT.
Panzar, J. and Rosse, J. (1977a) 'Structure, conduct and comparative statics', mimeo.
Panzar, J. and Rose, J. (197To) 'Chamberlin vs. Robinson: An empirical test for monoptly rents',
mimeo.
Panzar, J. and Rosse, J. (1987) 'Testing for 'monopoly' equilibrium', Journal of Industrial Economics,
forthcoming.
Porter, R.H. (1983) 'A study of cartel stability: The joint executive committee, 1880-1886', Bell
Journal of Economics, 14:301-314.
Porter, R.H. (1984) 'Optimal cartel trigger price strategies', 29:313-338.
Porter, R.H. (1985) 'On the incidence and duration of price wars', Journal of Industrial Economics,
33:415-426.
Prescott, E.C. and Vissclier, M. (1977) 'Sequential location among firms with foresight', Bell Journal
of Economics, 8:378-393.
Roberts, J. (1984) 'Testing oligopolistic behavior', International Journal of Industrial Organization,
2:367-383.
Roberts, M. (1983) 'Testing oligopolistic behavior: An application of the variable profit function',
International Journal of Industrial Organization, 2:367- 383.
Rhoades, S. (1982) 'Structure-performance studies in banking: An updated summary and evaluation',
FRB staff studies no. 119, see also Federal Reserve Bulletin.
Rosse, J.N. (1970) 'Estimating cost function parameters without using cost data illustrated methodol-
ogy', Econometrica, 38:256-275.
Rotemberg, J. and Saloner, G. (1986) 'A supergame-theoretic model of business cycles and price wars
during booms', American Economic Review, 76:390-407.
Scheffman, D. and Spiller, P. (1987) 'Geographic market definition under the DOJ guidelines',
Journal of Law and Economics, forthcoming.
Scherer, F.M. (1980) Industrial market structure and economicperformance, 2nd edn. Chicago: Rand
McNally.
Schmalensee, R. (1985) 'Econometric diagnosis of competitive localization', InternationalJournal of
Industrial Organization, 3:57-70.
Schmalensee, R. and Golub, B. (1983) 'Estimating effective concentration in deregulated wholesale
electricity markets', Rand Journal of Economics, 5:12-26.
Schmalensee, R. and Tlfisse, J.F. (1986) 'Perceptual maps and the optimal location of new products',
Center of Operations Research and Econometrics, core reprint no. 8620.
Shaffer, S. (1982) 'A nonstructural test for competition in financial markets', in: Proceedings of a
conference in bank structure and competition. Chicago: Federal Reserve Bank of Chicago.
Shapiro, M. (1987) 'How (monopolistically) competitive is U.S. industry', mimeo.
Slade, M. (1987) 'Conjectures, firm characteristics and market structure: An analysis of Vancouver's
gasoline-price wars', mimeo.
Spiller, P. and Favaro, E. (1984) 'The effects of entry regulation on oligopolistic interaction: The
Uruguayan banking sector', Rand Journal of Economics, 15:244-254.
Stigler, G.J. (1964) 'A theory of oligopoly', Journal of Political Economy, 72:44-61.
Sullivan, D. (1985) 'Testing hypotheses about firm behavior in the cigarette industry', Journal of
Political Economy, 93:586-598.
Sumner, D. (1981) 'Measurement of monopoly behavior: An application to the cigarette industry',
Journal of Political Economy, 89:1010-1019.
Suslow, V. (1986) 'Estimating monopoly behavior with competitive recycling: An application to
Alcoa', Rand Journal of Economics, 17:389-403.