I
^^
HD28
.M414
no.
3?57-
96
The International Center for Research on the
Management of Technology
Market Segmentation and the Sources of
Rents from Innovation: Personal
Computers in the late 1980's
Timothy
F.
Scott
Bresnahan
Stem
**
Manuel Trajenberg
***
WP # 154-96
August 1996
Sloan
*
*
WP # 3927
Stanford University and NfBER
* MIT Sloan School of NBER
*** Tel
©
Aviv University and
NBER
1996 Massachusetts Institute of Technology
Sloan School of Management
Massachusetts Institute of Technology
38 Memorial Drive, E56-390
02139
Cambridge,
MA
li),--^
Wwl
FEB g^
-^^37
Market Segmentation and the Sources of Rents from Innovation:
Personal Computers in the late 1980's
By Timothy F.
Scott Stem**, and
Bresnahan*,
Manuel Trajtenberg*
August, 1996
NBER
NBER
Tel Aviv University and NBER
Stanford University and
MIT
Sloan School and
We would like to thank seminar participants at Harvard, Princeton, UCLA, Stanford, MIT, the
Justice Department, the NBER 10 and Productivity Workshops, Econometrics in Tel Aviv 1995,
and especially the Conference
criticisms.
honor of Richard Quandt's retirement for helpful comments and
Cristian Santesteban and Michele Sullivan provided able research assistance. We are
in
grateful to the Alfred P. Sloan Foundation, to the
Technology and Economic Growth Program
(Stanford), to the International Center for Research in the
Management of Technology (MIT),
and to the Stanford Computer Industry Project for support.
ABSTRACT
Market Segmentation and the Sources of Rents from Innovation:
Personal Computers in the late 1980's
By Timothy
Bresnahan,
F.
Scott Stem, and Manuel Trajtenberg
This paper evaluates the sources of transitory market power
computers (PCs) during the
PC
to entry into the
manufacturers.
We
late 1980s.
Our
analysis
is
in the
market for personal
motivated by the coexistence of low barriers
industry and high rates of innovative investment by a small
number of PC
attempt to understand these competing phenomena by measuring the role that
of product differentiation (PDs) may have played in segmenting this dynamic
reflects some notion of product similarity and offers a potential source of market
different principles
market. Each
PD
segmentation,
in that
products which share similar characteristics
products which belong to separate groups. Our
first
PD
may be
closer substitutes than
measures the substitutability between
Frontier and Non-Frontier products, where the technological
fi-ontier is
defined to include those
computers which incorporate the Intel 80386 chip. The second PD measures the differentiation
between those products sold by brand-name firms (IBM, Compaq, etc..) and those products sold by
the non-branded fiinge. Our modeling strategy is flexible enough so that neither PD is assumed, a
priori, to be
of product
more important than
differentiation,
the other. Expanding
we measure
upon recent advances
in the
measurement
the separate roles that technical advance (incorporating the
(e.g., by IBM) played in contributing to transitory market power. In so doing,
paper attempts to account precisely for the market origins of innovative rents in the PC industry.
80386) and branding
this
Our
principal finding
is that,
during the late 1980s, the
PC
market was indeed highly
segmented along both the Branded (B versus MB) and Frontier (F versus NF) dimensions. The
effects of competitive events (e.g., entry, imitation, price cuts) in any one cluster were confined
mostly to that particular cluster, with little repercussion on other clusters. For example, less than 5%
of the market share achieved by a hypothetical clone entrant would be market-stealing fi-om other
clusters. In addition, the product differentiation advantages of B and F were qualitatively different.
The principal benefit associated with F was limited to the isolation it provided fi^om the large number
of
NF
competitors.
segmenting
B
In contrast, brand
products fi-om
transitory market
innovation in the
power
PC
NB
names
competition.
(arising fi^om
shifted out the product
These
demand curve
results help explain, in a precise
as well as
way,
how
market segmentation) shaped the underlying incentives for
industry during the mid to late 1980s.
INTRODUCTION
I.
Product innovations contribute to economic progress by expanding the range of choices
available to
consumers and by improving the performance dimensions of existing products. However,
the incentives to innovate
do not stem
directly
from the social value of new or improved products
but from the private returns that innovators manage to capture in the marketplace. In order to be
an effective inducement, these potential rents must be secured from the competitive threats posed by
pre-existing products, from imitators, and, at least for
some
As Arrow (1962) pointed
technological frontier even further.
from products extending the
time,
out, iimovators have to enjoy
degree of transitory market power in order to pay for their innovations. In this paper,
some of the
critical
measurement issues
transitory market
power
computers, or just
PCS)
in
we examine
by evaluating the sources of
that arise in this context,
the market for microcomputers
some
(commonly
referred to as personal
in the late 1980s.
In differentiated, fast changing product markets, innovative firms might be able to take
advantage of several different sources of transitory market power.
incorporate novel features,
Second, firms
making
protective umbrella of a pre-existing brand-name reputation over
premium;
this branding,
Of
in the
new good.
marketplace by extending the
new product
introductions.
To
take
innovators might exploit a widespread willingness to pay a brand
alternatively, the innovator
latest offerings
new products might
existing products only an imperfect substitute for the
may try to slow down the pace of rent-dissipation
advantage of
First,
might concentrate on serving high-margin niche markets for the
from "high-quality" branded producers (Teece, 1988).
course, the classical justification for patents, trademarks, and other forms of formal
intellectual property protection
is
precisely their ability to provide temporary
order to prompt innovation and creativity; yet, these forms of protection
a minor role in ensuring innovative rents in the
PC industry during the
seem
monopoly power
in
to have played only
late 1980s.'
Consequently,
we
focus here on the role that pushing the technological frontier (in this case meaning the incorporation
of the 80386 microprocessor) and relying on a brand name reputation (such as that of IBM) played
'
Patents did not substantially delay the entry of imitators as the industry
an open technical standard. Thus, our industry
(1987)
in the
modest
role played
by
is like
is
organized around
the majority of those studied by Levin et
patents in ensuring appropriability.
al.
in
ensuring transitory market power to innovative
Our
starting point
is
PC
firms.
the fact that markets for differentiated products often exhibit
some form
of segmentation or clustering according to a small number of "principles of differentiation" (PDs).
Each
PD defines a distinct notion of product
similarity according to the
presence or absence of some
key product characteristic and offers a potential source of market segmentation. Thus, products that
belong to the same cluster (defined by these PDs) would be closer substitutes to each other than to
products belonging to other
product
clusters.
We
associated with a strong brand
is
or not the product
is
at the cutting
consider here two such principles:
whether or not the
name (Branded, B, versus Non-Branded, NB) and whether
edge of the technological
fi-ontier
(Frontier, F, versus
Non-Frontier, NF). Both contemporary industry sources as well as retrospective analysis support the
view that the
late
1980s market for personal computers indeed exhibited the 4-way clustering implied
by these two PDs:
{B, F}, {B, NF}, (NB, F}, and {NB, NF}.^
estimate, the instruments that
estimates
all
we
The model of demand
use for that purpose, and the issues that
we
draw
stem fi-om and build upon these two principles of differentiation.
that allows us to evaluate the substitution patterns within
fi-om
we
address with the
We devote the bulk of the paper to the estimation of a discrete choice model
PCS
that
and extend recent methodological advances
and across these PDs.
in the empirical
of demand for
To do
so,
we
study of product differentiated
markets (Bresnahan 1981, 1987; Trajtenberg, 1990; Berry, 1994; Berry, Levinsohn, and Pakes,
1995). Following Berry (1994),
we motivate the fiinrtional form
for the
demand system by modeling
a discrete choice maximization problem faced by a set of heterogeneous consumers.
for the potential correlation
between price and unobserved quality by
We then account
utilizing a set
of instruments
based on the attributes of competing products using a model of competition. The specific model that
we
put forward, the "Principles of Differentiation (PD) Generalized Extreme Value (GEV),"
particular application of the
GEV
class
of models suggested by McFadden (1978). The
is
a
PD GEV
allows us to treat different potential sources of segmentation symmetrically, thus overcoming the
hierarchical structure (and
Finally,
we
concomitant limitations) of the familiar Multi-Level Nested Logit model.
take advantage of our priors about the prevailing principles of differentiation and use
As a matter of semantics, we shall refer to each of these groupings as clusters
PCS belonging to the {NB, NF} cluster are commonly referred to as "clones").
^
(notice that
instmmental variables that
reflect "local" (i.e.,
within-PD) competitive conditions, thus exploiting an
additional source of variation in the context of our short panel dataset.
Our main
findings are,
1980s, along both the
F and
first,
that the
B dimensions.
the
one
as entry, imitation, or price cuts) in any
with
PC
market was indeed highly segmented during the
Consequently, the effects of competitive events (such
cluster
were confined mostly to
that particular cluster,
repercussion on the others. For example, our simulations imply that less than
little
5%
market share achieved by a hypothetical clone entrant would be market-stealing fi^om other
The second
finding
that the
is
The advantage
different.
product differentiation advantages of B and
contrast,
The paper
qualitatively quite
provided from the large number of
it
NF
competitors.
was
By
is
non-branded PCS, but also shifted out the product demand curve a great deal.
organized as follows.
in the late 1980s, including
scheme proposed.
clustering
attributes for each
Principle
clusters.
bestowing a brand name reputation on a new product not only provided some protection
fi-om competition fi-om
market
F were
of the
associated with positioning a product at the technological fi^ontier
limited to the relative isolation that
late
PC
in
1987 and 1988.
in
GEV model,
some
discuss
some key
features of the
PC
then describe the data, consisting of market shares, prices and
the estimation.
robustness, and perform
n we
sources of buyer heterogeneity and the appropriateness of the
product sold
of Differentiation
of instruments used
We
In Section
review
We
its
In Sections
IV and V, we develop
the
implementation, and present and discuss the set
present the main results in Section VI, assess their
simulations that highlight the extent of segmentation.
Section
Vn
concludes.
THE PC MARKET IN THE LATE
n.
1980s^
Technical change in personal computing has been extremely fast and sustained; moreover, the
nature of competition
among PC
1987-88 catches the technology
^
at
firms has
changed over time as
in swift transition
well.
wath the introduction of a
We limit our analysis IBM PCS and IBM-compatible PCS.
the time,
we
Our snapshot of the market
new
microprocessor, the
Following industry convention
exclude both older, 8-bit computers (Apple H, CP/M, etc.) and incompatible
architectures such as Apple Macintosh. While Macintosh and IBM-compatible machines
converge
overiap
in
in capabilities
over time (particularly during the early 1990s), there was
among commercial customers
during the period under study.
little
would
important
here with the Frontier. The role played by brand-name reputation
we associate
80386, which
highlighted, a role that
would evolve and mutate
substantial brand capital changed.
and required substantial
only the most technically capable
handful of other
making
PC
PC makers would
effective use
also
and types of firms with
At the time when the 80386 was introduced
design and production of microcomputers
technical difficulties,
later on, as the identities
is
(in late 1986), the
of the new chip posed considerable
R&D efforts fi-om PC
makers.*
By
the end of 1986,
A
firms had succeeded in introducing an 80386-based system.
see their development efforts succeed during our sample period,
whereas most other firms took even longer to come out with marketable 386 designs.
A key question posed in this paper is the extent to which PCS that belonged to the Frontier
(or were marketed by a Brand) were insulated fi^om competition by Non-Frontier (Non-Branded)
products.
The answer
buyers
their valuation
in
to this question crucially depends
of Frontier and/or Branded
on the degree of heterogeneity among
status.
In the presence
of substantial
heterogeneity. Frontier (and Brand) can be protected fi-om the competition of Non-Frontier (and
Non-Brand), while
same time coexisting with these products,
at the
neither group eliminating the
other.
A
close-up look at the transition fi^om 286- to
386-based
PCS
suggests that there were
indeed pronounced differences across buyers in their assessment of the prospective benefits of 386-
based PCS.
From an engineering
perspective, the
new systems
yielded substantial improvements in
the ability to run software due to the increased speed of the microprocessor and an
model
for the
main memory.
However, some users (and
analysts)
saw 386-based systems just
PCS. Others, already rurming or planning
faster versions of older
improved address
to run
as
more advanced and
demanding software (such as the new Microsoft/IBM operating environment, Windows), saw them
as the only sufficient platform that
*
By
the mid-1990s,
design. This
is
it
would
has become
satisfy their
much
computing needs within the
usefijl life
easier to incorporate frontier technology into
because a much higher share of PC innovation has become embedded
of the
PC
in the
components themselves (such as an Intel microprocessor) rather than in the integration among
components (see Henderson and Clark (1990) for a discussion between component-based versus
integration-based innovation). In the era under study in this paper, the opposite
equipment manufacturers had to develop specialized
skills
was
true.
PC
and make substantial expenditures
order to market a product on the technological frontier (Steflfens, 1994).
in
current purchase. These conflicting opinions,
led to significant differences
among buyers
echoed
in their willingness to
much more expensive 386-based PCS.'
also
steeply sloped
demand curve for the
Frontier
and Non-Frontier systems. Our model
to which Frontier
PCS
PCS were insulated
differed not only in terms
in lively debates in the trade press,
will
We
pay for the much more capable but
would expect such heterogeneity
PCS, as well as
in
poor
substitutability
indication that the product
name
in
the
PC
was
reliable
of
of their technical features, but also
in
terms of the brand-name
and support from
quality service
buyers valued
an
selling firms, as well as
and practical for business applications. IBM, the leading brand
market for many years, was not just the inventor of the dominant design in
PCS
but
and technical excellence from making and marketing
we
treat
AT&T
and Hewlett Packard as branded based on,
technological leadership in industry segments close to
business-oriented
PC
we
The
Compaq. The
largest advertiser in the
fourth firm
treat as branded,
PC
establishing a reputation for innovation in
IBM) and by
prior to
1994).
PCS; and
(/;)
their
(/)
their stated
market and
commitment
to the
market segment, backed up by substantial advertising and marketing
expenditures.*
market
PCS
and the only specialized
at the time,
(such as
its
Compaq
PC
firm so treated,
built its
introduction of the
brand name by
first
386-based
PC
Within a few years after our sample, several other firms
strategy;
built substantial
brand reputations
however, except for Compaq, none had been successfiil prior to
O'Malley (1989) summarized both sides of the debate: "While a 386 or 386SX system often
a wiser purchase than a 286, those suggesting the demise of the 286 misunderstand the scope
and diversity of PC applications."
*
than
is
providing reliable service and support for these frontier products (Steffen,
by imitating Compaq's
is
PC
computers to business customers. Other large electronics firms also had preexisting reputations;
these,
'
between Frontier
fi^om Non-Frontier competition at that time.
also enjoyed a reputation for service, support,
larger
to result in a
measure these features of demand, and hence the extent
reputation that the manufacturers bestowed on them. During our sample period,
Brandedness as an indicator of high
presumably
Other large computer companies sold PCS, but primarily to their existing customers rather
a mass market. We treat these, including Burroughs, Honeywell and Wang, as Non-
in
Branded.
1988.^
As with
Frontier status, the issue
of the degree of insulation from competition
firms enjoyed vis a vis non-branded ones,
Some
and product
buyers, however,
presumably
will
reliability
may have
branded
hinges largely on the extent to which buyers
heterogeneously value the attributes signaled by brandedness.
service, support,
that
presumably
is
A branded company's reputation for
valued by a// buyers (and
we
will
measure
that).
particularly high valuations for these firm-related attributes, and
be willing to pay an even higher premium for Branded products. In
heterogeneity was widely reported for
PC
purchases
in
fact, this sort
our sample period.* Buyers
organizations (e.g. those in the Fortune 500) tended to favor branded
PCS
of
in large
over those offered by
clone makers. Other, usually smaller, organizations valued brandedness significantly less tending to
buy the cheaper clone computers.
Compaq
In our primary specification,
as a separate branded group, and seek to measure
how
we
treat
IBM, AT&T, HP and
insulated their
PCS were from
competition from clones.'
In principle, not just Frontier
and Brand but other aspects of PCS, technical and otherwise,
could have been the source of market segmentation and hence of temporary monopoly power.
However, IBM's
PC
historic decision to
firms at the time of our sample to
class
have an open architecture for
be
strictly
PCS
(and the decision of the other
compatible), allows us to eliminate outright a whole
of potential sources of market segmentation. IBM's open architecture meant that buyers could
"mix and match" a wide range of peripherals and add-in cards; therefore, price
PCS due to
^
Dell
differentials
between
such components could be arbitraged away.'" Given the existence of a competitive
Computer would be
retail
the next entrant into the Branded segment through the
development of a reputation for excellent service and support
in
conjunction with substantial
advertising.
*
SeeRyan(1987), Jenkins
'
In related specifications,
branded
firms.
(1 987),
we
also
and Kolodziej( 1988).
examine whether
IBM was fiirther separated
We also experiment with the boundary of the branded
variables, such as distribution channels (store/mail order)
from other
segment. Other marketing
may have been
important but do not
receive special treatment in our models.
RAM
Some PCS are limited in the amount of
they can maximally hold. Up to this limit,
however, buyers can typically populate the computer with more of less
according to their
'"
RAM
market for unbundled components, variations
of the hard
disk, could not
in attributes
make one PC a poor
PC
for each
We therefore treat these
substitute for another.
characteristics as hedonic attributes, in the sense that
demand
such as the amount of RAM, or the size
we
allow them to
shift
but not to affect the substitution patterns between products."
Having focused here on the sources of transitory market power,
sample period
in perspective,
and the demand for desktop computing are being
The PC
processes.
it
important to put our
is
sketch the forces that brought rapid change to the
transformed time and again the patterns of competition in
PCS
the relative level of
it.
shifted continuously
invention-possibilities curve in
industries
advance
memory
PCS
lie
These
relentless
era, a
still,
Shortly after our period, a
newer and
faster
we
much
number of firms improved marketing operations,
The
losing in our sample period.'^
budget and the software they seek to run. Sellers typically picked these
bind on customers early in the
An
arbitrage
intermediate case
it,
since speed
is
is
life
number of
new frontier.
In the early 1990s, the principal suppliers of
IBM was
standards that
larger
especially delivery and customer
came
"
out the
seek to measure.
microprocessors heralded a
microprocessors and operating systems, Intel and Microsoft respectively,
PC
shifts
word processors, and spreadsheets
changes mean that both sources of segmentation that
support, thereby approaching Branded status.
control of
introduction
pushing out the demand for more technically advanced PCS.
companies sold 386-based PCS. Later
same
The
and disk drives frequently
like operating systems,
Frontier and Brand, were shifting rapidly.
In the
other fast-paced,
themselves. There are also a variety of technically progressive
downstream. Software products
rapidly,
chips,
and
by powerful dynamic
sector itself is technically progressive. Immediately upstream
microprocessors,
faster
industry,
Both the supply of technically advanced
innovative industries such as integrated circuits, mass storage devices, and so on.
of ever
PC
to hold
current
much of the
PC
industry
limits so that they did not
of the machine.
microprocessor speed, expressed
in
MHZ.
Buyers could not
"designed into" the microprocessor and controller card. The
of several different processor speeds with most models somewhat mitigated this effect
most cases, buyers could find their ideal point nearly met by an offered variation. We
examine this issue in a variant of our base specification.
availability
since, in
'^
During and
after
standard setting via the
our sample period,
IBM attempted to regain control of PC
market
OS/2 operating system and the introduction of a novel (and
proprietary)
capable
structure, with multiple technicaUy
There
in a
measurement of transitory market power
in
is
power
that
we
seek
fi-om the
the private incentives to innovate arise
moment of transition. However,
which
discounted (expected) value of the rents
transitory market
much of the
doubt that these changes destroyed
is Uttle
measure
to
branded firms competing, emerged fi-om these evems.
accme
in
each period;
in this sense, the period-specific
investment
a key step towards understanding innovative
dynamic technology-intensive markets.
THE DATA
ffl.
Estimating a product-level differentiated
prices,
demand system
and characteristics of each make/model
The
also needed.
dataset that
we
in
requires data
on
quantities sold,
each year. Estimates of total market size are
gathered for the present study consists of
all
PC
make/ models
1988.
attributes could be obtained for 1987 and
(boxes) for which accurate data on price and
set
Identifying a comprehensive, weU-defined
block.
The
gathering accurate data for each and every box,
we
relevant market presence (see Table
scope of the dataset to those vendors with a
1).
Because of the
restricted the
difficulties
resulting sample consists
Quantities
encountered
in
and 137 for 1988.'^
of 121 make/models (observations) for 1987,
were obtained from a database provided by
1987 and 1988.
covering the years
technology (IT) equipment (including
in
of distinct make/models constituted a major stumbling
the U.S.'' Using
calculate the total
COMTEC's
COMTEC
PCS) by a
COMTEC Market Analysis
reports in great detail purchases of information
large,
weighted sample of business establishments
sample data to
weighting scheme, one can extrapolate from the
number of PCS
sold for each make/model. Since
COMTEC
purchased PCS, one can use the
establishments in the U.S. whether or not they
PC
Services,
design (Micro-Chamiel Architecture).
Ultimately,
IBM's
samples business
same weighting
strategy to regain standard setting
leadership failed.
'^
See Appendix
A for a detailed description of the data gathering methodology.
Appendix
analysis.
provides precise definitions of all variables used in the
•*
The
COMTEC
establishments.
Note
"Wave
6" dataset that
we
use here
that the sample excludes households.
is
based on a sample of over 6,000
B
scheme to estimate the
total potential
market size for desktop business PCS. For our purposes
take the potential market to be the total number of oflBce-based employees (39 million.)
that
we
estimate includes an "outside good" (that
is,
price and characteristics data
come from
from a wide variety of additional sources.
We
the
focused on a small number of
RAM provided
speed (MHZ), and the standard hard disk provided (DISK,
in
above, our
principles of differentiation in the
analysis
PC
technological frontier and whether or not
with a
dummy
variable
microprocessor, and
mentioned
in
Section
(FRONTIER)
to those with
11,
we
relies
market
it
at the time:
identified
variable
BRANDED
(expressed in KB), the
whether or not the
was produced by a branded
of
to
1
firm.
PCS
8088, and older). As
name was
AT&T,
to products produced
the
at
incorporating the 80386
four vendors for which a brand
1
PC was
We capture the former
other microprocessors (the 286,
assigns
critical technical
MB).''
important determinant of demand and substitution patterns: IBM,
Compaq.'* The
as well as
on the possible existence of two
heavily
that assigns the value
all
million.
GML Microcomputer Guide,
characteristics including the microprocessor, the standard
As mentioned
The model
the option of not buying), and therefore
market shares are computed as unit sales of each make/model divided by 39
Our
we
a potentially
Hewlett-Packard and
by these four firms and
to
products produced by other vendors.
Tables 2-4 provide a
Though
the
two samples
usefijl portrait
are only a
of the
PC
industry in each of the years under study.
year apart, there are substantial differences in the average
price/performance of PCS sold. Whereas the average price changed only slightly between the
years ($3,225 to $3,205), the average hard disk size (DISK), speed
increased substantially. For example, the average standard
all
688 KB. Indeed, the average
two
years.
" These
The
rate
level
(MHZ), and memory (RAM)
RAM
is
accentuated
if the
(B&W or Color),
KB
10% between
We also collected
whether the machine was portable, model
vintage, etc.
'*
Other configurations were attempted; the qualitative
inclusion or exclusion of branded firms at the margin.
to
the
observations are weighted by the
are the characteristics actually used in the estimation.
information on the monitor type
increased from 611
of each of these characteristics rose over
of technical advance
two
results are not affected
by the
quantities sold.
The
rate
For example, the weighted average standard
of advance shows also
10% of
which grew from
BRANDED
marketed
price
in
in the
the
share of 386-based
sample
1987 to
in
RAM rises
PCS
15%
in 1988.
characteristics are estimated to
The
characteristics (Table 3).
By
to
716 KB.
FRONTIER),
contrast, the
number of
of the number of make/models
both years. The advance in the price/performance ratio
on the observed product
the average for
(i.e.,
30%
averaging
products stayed nearly constant,
KB
from 584
is
highlighted in a regression of
coefiBcients for each
be positive, while the 1988 coeflBcient
is
of the technical
substantially lower than for
1987. While a substantial amount of price variation exists in each year, our sample reflects the rapidly
falling real prices for
inception
(Bemdt and
Finally,
B/NF
cluster.
statistic
in the
NB/NF
data according to our
from
this table
cluster, at least
is
PDs (Branded and
NB/F
market segment since
its
Frontier) in Table 4. Perhaps
of
that while a relatively high share
60%
of the boxes sold
Moreover, the average price for boxes
cluster through
this
Griliches, 1993).
we "slice" the
the most striking
concentrated
computing power which has characterized
rises as
in
all
products are
each year are drawn from the
one moves from the "clone"
(or B/NF), with the highest average prices in the
B/F
cluster.
NB/NF
While
this
preliminary cut of the data suggests the presence of substantial returns to membership in either the
F or B group, one cannot
identify the precise
(through the benefit of a large demand curve,
Instead, our
way
or, alternatively,
more systematic evaluation of the returns
the estimation of the differentiated product
which these returns manifested themselves
in
to
a steep one) from this simple table.
membership
demand system,
in the
B
or
F group
the model to which
will follow
we now turn.
MODELING THE DEMAND FOR PCS
IV.
The
rest
of this paper
is
devoted to estimating a model of PC demand which allows us to
evaluate precisely the origins of transitory market
market during 1987-88.
We
different competitive positions
power and
measure the demand for
PCS
of PCS which incorporated
isolation
at the
from competition
in the
PC
product level and explore the
different attributes.
'^
Our model, based
We exploit and extend recent methodological advances which have expanded the range of
economic models which can be estimated with product-level data (Bresnahan, 1981, 1987;
'^
Trajtenberg, 1990; Berry, 1994; Berry, Levinsohn, and Pakes, 1995).
10
on the
familiar
random
utility
model, aggregates from individual buyer heterogeneity to product level
PC
market
reality, this
products and of products no longer being sold.
Our model
demand. Besides corresponding to
permits straightforward treatment of new
is
novel in that
it
permits the existence of
non-overlapping principles of differentiation (like brandedness and frontiers status) without calling
for difficult numerical integration.
A.
THE RANDOM UTILITY APPROACH TO DEMAND ESTIMATION
Our model of PC demand
developed
motivated by the theory of random
is
(GEV)
"Generalized Extreme Value"
in the
class
utility
(Quando, 1956), as
of models put forward by McFadden
Our
(1978) and following Berry's recent proposal for the estimation of such models (Berry, 1994).
point of departure
choosing
is
among J,+l
a discrete choice random
alternatives
(J,
utility
model. Each buyer
marketed products
is
assumed to maximize by
year t and the option of no purchase 0=0)),
in
as follows,
v„
j.":":,}
where
Vjj is
the value of product
characteristics, Pj
is
x/p
to buyer
j
the price and
=
^j is
I.
ap,
.
For product],
6.
=
x/p
+ ap.
+
riy is
product j from the average valuation in the population
which
is
a realization of the
but a given buyers'
draw
of
similar product characteristics.
characteristics X.
r\,j
of parameters to be estimated
random variable
Xj
is
(I)
n,
a vector of observed product
distribution
dimensional)
in this discrete
of
rj
CDF
choice
11
parts.
Let
(2)
^.
the difference between buyer I's value for
(6j).
t|.
two
Each buyer receives a draw of the
The draws
need not be independent of
The
We write this (J+1
.
the level of unobserved product quality.
be the mean valuation by buyers for product j. Thus,
r\^
5^
(1994), decomposing Vy into
Our treatment follows Berry
vector,
*
J,+l
are independent across buyers
ri^
if
products
j
and k share
depends on parameters p and the product
as F(ti; X, p).
demand system.
Thus, {p, a, p}
is
the vector
Since the
work of Quandt
to valuation, here
x], is
choice model, there
is
(1956),
we have known that the
distribution
a key determinant of the shape of demand.
in fact a precise relationship
buyer level and the pattern of cross-product
between the
elasticities at the
of the random shocks
In the context of our discrete
distribution
of r|
aggregate market
at the individual
level.
The modeler
picks a family of distributions F(r|; X, p). Particular values of the parameter vector p correspond
across products sharing
common
to patterns
of dependence
Our model
contains a parameter pp, for example, which permits dependence across the idiosyncratic
in
r]
For certain values of
shocks to Frontier products.
product
if k
»
(T]ij
0)
is
also likely to strongly prefer
and j share Frontier
Now
product
j
a rate a) and
in the distribution
other Frontier products.
competition will
in
r|
[tii^
|
riy
of particular product
some portion of those consumers whom had
of r|, a
relatively high share
More generally,
positive
»
»0]
0]
the value of
chosen j are
Because of the
of these second-best alternatives
dependence
in the distribution
of r| among
The parameters p measure
across products with similar products characteristics
become more "segmented,"
j:
initially
their second-best alternatives.
products makes those products closer substitutes.'*
As dependence
Frontier products (E
in the price
induced to switch to what had been previously
dependence
all
buyer strongly preferring any Frontier
status.)
consider the impact of an increase
falls (at
pp, a
(or similar) product characteristics.
will
be
similar
that dependence.
becomes
stronger,
the sense that the impact of competitive events (such
in
as a price decrease or entry) will be confined mostly to those products with similar characteristics.
In other words, in this fi^amework a
model of which
classes
of products are close substitutes
model of dependence among the elements oft) related to product
is
a
characteristics.
A PRINCIPLES OF DIFFERENTIATION GEV MODEL OF PC DEMAND
B.
Of course,
may be
'*
a separate question exists as to
how to identify those product
characteristics
which
important for understanding diflferentiation and market segmentation. Fortunately, economic
This type of dependence across observations can be contrasted with the case where
across choices (as in the multinomial logit).
introduction of a
new product
will
Under the assumption of independence, the
effect on all products in
have the same (proportional)
market, regardless of characteristics. This implication of the
iid assumption
of the "Red Bus/Blue Bus" problem emphasized by McFadden (1973).
12
is
t^ is iid
the
just a restatement
theory, in conjunction with an understanding of the particular market under study, provides
considerable guidance.
(such as the
Like many other consumer products, there exists several features of the
amount of RAM or the
away
price
of the hard drive) for which there
In the presence
for that feature (or component).
arbitrage
size
exists a separate
PC
market
of competitive component markets, buyers
will
premiums associated with the incorporation of these components into particular
products. In other words, to the extent that consumers can "repackage" products along a specific
dimension, each consumer will equate their marginal utility for this dimension to
sales market.
Accordingly, there will be no ex-post heterogeneity
valuations for this dimension. Thus, in choosing
X,
p),
one should incorporate
how to
price in the post-
among consumers
in their marginal
parameterize the distribution function
among products which share
potential correlation
its
characteristics
F(r|;
which
are not "repackageable."
Guided by
repackageable
this mle,
~ Branded
cannot be marketed
technical expertise
in
we
focus on two product characteristics which
the post-sales market.
was required
As
to incorporate the
80386 microprocessor
memory
of McFadden (1978), which
model.
PC. In
options
(e.g.,
we
McFadden
contrast,
RAM and
(MHZ).
need to parameterize
F(r|; p)
can be incorporated symmetrically into the distribution function. To do
(NML)
into a
the presence of non-mutually exclusive product groupings
computationally straightforward specification,
1
under study, considerable
well, during the period
and, to a lesser extent, the speed of their microprocessor
To accommodate
argue are not
and Frontier. The existence of a Brand name reputation, by construction,
buyers could entertain several post-sales options to reconfigure their
DISK)
we
states a significant generalization
so,
we
(i.e.,
PDs)
so that several
rely
PDs
upon Theorem
of the nested multinomial
introduces the Generalized Extreme Value
in a
logit
(GEV) model through
a
constructive proof demonstrating that a wide variety of dependencies (and thus patterns of
substitution) are consistent with
random
utility
maximization:
13
Proposition 1 (Adapted from Theorem
rJ+1
R
-.
>
is
a non-negative,
I,
McFadden
(1978)): If G:
homogeneous of degree one function
satisfying certain restrictions," then
„-n,j^
-nio
F(Tiio.-->nn) = exp-G(e "'"......e
is
(3)
the cumulative distribution function of a multivariate extreme value
distribution,
and
e*'
ore*"
G(e
e*'^
»,.... e
is
the corresponding equation determining the market share of product
j,
where Gj
McFadden (1978)
purpose specification
is
the partial derivative of j with respect to
suggests that Proposition
tool:
1
cross-product substitution matrix.
model which incorporates
In order to
make
Our model
can serve as a straightforward and general
1
provides a method for parameterizing the
will utilize this
theorem to parameterize a random
/Jow-we5/ec/ cross-product correlations in
t].^"
more understandable, consider
the logic of the theorem
associated with a traditional specification, the one-level
among products according
e*^.
by specifying the function G(-), the modeler chooses the pattern of
dependency of r|yS across products. Thus, Proposition
utility
'•')
NML. To
to whether or not the product
is at
the G(.) function
incorporate potential correlation
the technological frontier (F versus
NF), the flmction G(-) takes the form.
" The
limit
must alternate
of G(-) as any argument goes to
in sign
and
first partials
«>
must be equal to +~. Mixed
partials
of G(-)
must be non-negative (McFadden, 1978).
Most previous research has implemented some version of the NML (a special case of the
structure in Theorem 1). We are aware of only one other empirical study (Small, 1987) which
utilizes McFadden (1978) to parameterize a non-nested structure.
^°
14
«\
G(o')
Pf
=
«j/pF
Pf
(5)
IjeF
VjeNF
;
where the parameter pp parameterizes the degree of substitutability among products with the same
Frontier status (F or
NF)
relative to the substitution
consistent with (1), the distributional parameter pp
dependence across products which share F or
to
1,
(5) simplifies to the
element oft]
is
G function
(at least)
we
two
becomes
stronger.
As pp goes
to 0, the
Conversely, as pp goes
(i.e.,
are interested in parameterizing the potential correlation
distinct
functions, as follows:
(
=
the unit interval.
associated with a simple multinomial logit model
dimensions (F/NF and B/NB).
most straightforward way, we compose G(-)
G(e»)
status
lie in
To be
products.^*
each
independent).^
In our application,
products along
NF
must
among non-matching
ap
s)*^
to be the weighted
To accommodate
sum of two
one-level
among
this in the
NML G(-)
contributes to a term which measures
(F or
NF) and in
a term which measures
(B or NB). Note that
model
is
its
if
pp
=
substitution with products
its
the model
1
is
which share technological
status
which share marketing
status
substitution with products
a nested logit by Branded status only, and
if
pg
=
the
1
nested by Frontier status only.
The implications of (6)
for substitution in the
market can be seen more clearly
in
terms of the
market share equation for each product. Letting B(j) and F(j) denote the groups to which product
j
belongs,
—
gVPp
E
^F
S^/pp
,.VPb
-^-— E
Pf
. a3
e-^^^^
«i/Pb
gVPB \h
VkeBO)
(7)
keF(j)
^
keB(i)
J
Notice that
which
is
changes
G{^')
this share
equation
is
composed of two
proportional to the one-level
in the prices
NML.
a relatively strong impact on the market share of product
share just one
PD
for B, the numerator of
Consider a product located
or attributes of products located also in the
neither in B(j) nor in F(j) will
F and one
terms, one for
j,
same
in {B(j), F(j)}; then,
(B(j), F(j)} will
cluster
whereas changes
in
products which are
have a lesser impact. Clearly, the impact of changes
with product j will be somewhere
in
have
in
between. Each of these effects
products that
is
stronger as
the relevant p goes to zero.
While our proposed model,
of
(6), is quite simple.
Proposition
different clustering structures, including a larger
interactions between
them,
differentiation simply requires
etc.
number of
Incorporating additional
adding suitable terms to
1
(6).^'*
can accommodate a wide range
"principles
of
differentiation,"
dimensions of potential product
In this sense, the
McFadden Theorem
provides a useful and underutilized modeling tool with potentially wide applicability.
We have estimated several variant models (only a subset of which are presented in the
Results Section). We have experimented wath additional clustering by MHZ (as in Small ,1987),
^*
an
"IBM"
cluster, as well as others.
16
THE PD GEV VERSUS ALTERNATIVE MODELS OF SUBSTITUTION
C.
Given our goal to measure the degree of segmentation afforded by
(B/NB and F/NF),
which make the most
several issues arise
(at least)
two
PD's
distinct
traditional (and computationally
straightforward) parameterizations of the joint distribution oft] inappropriate for our problem. First,
because
we want
to distinguish
between several
different sources
of product
differentiation,
we
cannot employ a simple unidimensional vertical product differentiation (VPD) specification
(Bresnahan, 1981, 1987; Berry, Camall and
distributional
quality.
As
parameter measures variation
well, a nested multinomial logit
Spiller,
1996; Stavins, 1996), in which a single
among consumers
(NML) model
in their preferences for overall
product
(Trajtenberg, 1990; Goldberg, 1995; Stem,
1996), in which several mutually exclusive dimensions of product differentiation can be incorporated,
is
also inappropriate.
For example,
if
we were to
top level of the nesting determined by the
potential correlation in
r|
B/NB
specify (1) to be a two-level nested logit with the
be assuming away
among
(and the resulting implications for market-level substitution)
products which share Frontier status but do not match along the
VPD and NML
we would
distinction (Figure 1),
B/NB
models can parameterize our principal hypotheses
dimension.
Thus, neither the
~ the
separate existence
directly
of segmentation along two overiapping dimensions (F/NF and B/NB).
In order to assess the relative merits
to the two-level
NML (see Figure
1).
richly parameterized than a two-level
the
PD GEV
accommodates
It is
of the
PD GEV,
it
is
worth comparing
in
important to note that our proposed model (6)
NML.
Instead, the difference
substitution in a
way
that
we
substitutes to each other than to other products, the
models
more
is
argue
is
more appropriate
B
is
that
for our
and F status are closer
differ in their treatment
the partially matching products (note that the groups in the
detail
no more
between the two models
application. In particular, while in both models products sharing both
among
it
PD GEV
of substitution
are not mutually
exclusive).
Consider the two-level
Pf determines
this sense, the
how much
NML
is
NML (with B on top) in the upper panel of Figure
"closer" products sharing both
B
1:
the parameter
and F are than those sharing only B. In
parameterizing segmentation along
more than one dimension. However,
lowering the price of a {B, F} product has the same effect on market shares in the partially matching
17
as in the completely unmatching
{NB, F} category
permits only one of the
two
partially
{NB, NF} category."
In this sense, the
NML
Moreover, reversing
matching clusters to be close substitutes.
the order of nesting just reverses the problem.^*
These properties can be
stated also in terms
of cross-elasticities, where e^y
the average
is
cross-price elasticity of a product in cluster y with respect to a change in the price of a product in
cluster X."
According to the two-level nested
^bf.bf ^
NML with B/NB on top (as in Figure
1),
^bnf,bf
^ ^nbf,bf " ^nbnf,bf
(8)
^nbf.bf
^ ^bnf.bf ^ ^nbnf.bf
(9)
Reversing the order of nesting renders,
^bf.bf
Having
between
to choose
equalities
seems
likely to
^
and (9)
(8)
is
an uncomfortable specification choice. Neither of the
hold in the present context, since this case, as
many
others,
is
not naturally
nested.
The
PD GEV
overcomes
this limitation
by permitting both
a parallel fashion (see the second panel in Figure
^'
principles
As we have
of differentiation,
is
As
all
" For
of dependence among
this exercise,
it
is
in
^^^
^nbf.bf
picture,
it
in
<
(^^)
^bnCbf
a completely symmetric way.
also has the
Of
same impact on the
possible to add a further parameter, say
p,.,
and a
PC's versus the outside good.
well, a four cluster one-level
possibility
drawn the
not fiindamental as
third level nest that has
'
and F, are thus treated
written the model and
outside good. This
^*
B
of "local" to work
1), i.e.,
^bCbf ^ V^nbtbP ^bnf.bF " ^nbntbf
The two
definitions
assume
NML (F/B, F/NB, NF/B, NF/NB) also excludes the
partially
all
matching products.
products have equal market shares and prices. The
elasticities
vary with price and market share monotonically. This assumption just simplifies the exposition.
18
course, this specification
related products to
the nested
logit,
it
is still
be better
substitutes.
only
is
no more
for our specific application
closely
richly parameterized than
which
is
to identify the impact
PDs.
V.
IMPLEMENTING THE PD GEV MODEL
A.
THE ECONOMETRIC PROCEDURE
We
two parameters permitting more
Thus, while the model
more appropriate
is
arising fi-om separate
restrictive, as there are
outline here the procedure
by which the parameters of the
PD GEV
model can be
consistently estimated fi"om a product-level dataset such as the one described in Section in.
Essentially,
we
follow the suggestion of Berry (1994) and construct a
of expectations of the
of unobserved product
level
instruments (discussed below).
express
the
J
it
products can be written
on the exogenous
quality, ^, conditional
In order to implement this procedure
as a fiinction of data and parameters to be estimated.
demand system. Recall
GMM estimator fi^om the set
we
need to isolate ^ and
We are able to
do so by
above that the market share of product
fi^om eq. (7)
j
"inverting"
market with
in a
as.
gVPr
Sj=7;7-i^
''
G(e«)
[V
"^
Y,
e*'"^
UfO)
^^
j
U^)
(">
I
keB(j)
keFO")
Notice that every element of
e*^*^
6, the (J+1)
vector of average product quality, enters the market share
equation of each product. Thus, with a complete cross-section of products, the market share
equations form a system of J equations with (J+1)
for the implicit function 6(s, p). Setting
s
We use this functional
unknowns
(the vector 6).
6„ the mean value of the outside good, equal to
=
s(6(s, p),p)
relationship to estimate the parameter vector {a
use our assumption that
6j is
We invert this system
also the average level
19
0,
we
solve
(12)
&,
p}
We now
of utility associated with product j,
b.
Rearranging (13)
in
terms of
E„
=
X;^ ^ap.
and substituting
.
^
in 6(s, p),
(13)
we
can define the sample version of the
disturbance for our estimator as,
6/s,p) - (X/p . ap.)
=
I.
(14)
Since E(^)=0, and under the assumption that Z, the matrix of instruments,
standard
GMM estimator
is
is
exogenous, then the
defined as,
L
-
5(ZQ 'Z%
(15)
p.a,P
Q
where
is
the standard weighting matrix. Thus, the estimation
techniques, and
is
of the model requires only standard
computationally straightforward.^*
INSTRUMENTS
B.
The
GMM estimator in (15) requires an instrumental variables (IV) vector, Z, with rank at
least as large as the dimensionality
vector
is
composed of the
current application
it
is
derivatives of the disturbance
(E,)
(p, a, p}.
PD GEV
As suggested above (and
in
Of course,
the efiBcient
FV
with respect to the parameters. In the
quite difficult to construct the optimal Z, so
for the construction of Z for
^*
of the parameter vector,
we
propose a simpler strategy
models.
BLP
(1995)), a two-step procedure
is
appropriate. First, 6
is
solved for as a fijnction of the observed market shares and a "guess" for p (Eq. (12)). This
procedure involves a well-behaved numerical nonlinear procedure. With 6(s, p) in hand, the
remaining (mean-value) parameters (a,P) can be estimated with a linear instrumental variables
estimator. Because the market share function can be expressed in closed form (see (11)), there
no need for integration of the market share
fiinction
comes
relative to the simulated
at a substantial
implemented
available
in
BLP
computation savings
on each
iteration
(1995). Additional information (including
by public access
ftp archive.
20
is
of the estimation. This
method of moments estimator
GAUSS
estimation programs) are
Our
strategy relies
In the spirit of the recent literature,
characteristics, X.
functions of
X which make plausible instruments.
Xj, the
This section
on the econometric exogeneity of the
^'
we
entire
matrix of product
use a model of supply to point out
A row of our proposed Z(X) consists of
observed product characteristics of product j
counts and means of
X for products sharing a cluster with product
counts and means of
X for products sold by the firm oflfering product
counts and means of
X for products sharing both cluster and seller.
first visits
the general econometrics issues briefly.
j
j
We then explain how our IV strategy
follows fi-om an assumption of equilibrium pricing behavior by sellers in an industry with
PD GEV
demand.
We
face
diflScult to justify
two challenges
an assumption that price
is
product quality, meaning that a higher
E(P^)>0.'*' Similar arguments
^j,
so that the
difficult,
first
mean
independent of ^;
framework
last
is
E,
term
side,
it
is
has the interpretation of unobserved
should lead suppliers to set a higher
E„
that the prices
term as well as the
the econometric
With market power on the supply
constructing Z.
in
p.
As a
result,
and share of other products are correlated v^th
in the
RHS
of (14)
calls for FV.
To make
this
more
inherently nonlinear.'' In the face of such nonlinearity (and
without additional assumptions), the efficient FV vector will contain elements which are themselves
functions of the parameters to be estimated, leading to substantial computational difiBculties
(Chamberiain, 1987).
In response to the difficulties,
observed product characteristics
^'
their
in
we
year
propose to
t,
is
utilize
exogenous.
our assumption that X, the matrix of
The immediate
implication of this
of BLP (1995), who impose some of the economic restrictions of
model on Z(X). Unlike BLP, we offer no proof that our Z is the first under term in a series
Our approach
is
most
like
approximation to the optimal FV.
'"
CT
This problem was emphasized by Trajtenberg (1989, 1990),
who found that the demand
scanners was estimated to he positively sloped in the context of an
exogenous
the culprit for the positive bias.
''_i
dp
=
NML model with
prices. Further analysis revealed that omitted quality, correlated
Our
solution follows Berry (1994).
—dp
L^EZ isahighly nonlinear function of p.
21
for
with price, was indeed
assumption
that
is
we
are able to include
Xj
we are looking for easy-to-calculate variables which are likely to
In general,
coeflBcients
is
^^'^^
)
(
supply."
but are independent of
highlight this source
of a multiproduct firm
selling several products,
^.
A
natural source
of instruments, suppose
of a non-cooperative Bertrand-Nash
to the solution
j
To
price-setting
which
we
*"
variable
PD-specific indices to reflect the
which
group
that firms
game. Consider the
G
shifts the
RHS
is
will
in
substitutes
a particular group increases, the
will shift in
is
exogenous
for each product.
for observation j
j
(in
and become
flatter in (16).
First,
we
Table
in
we
on a group-specific
demand curve
Both
basis.
associated
effects should impact
number of products
in
exploit our assumption that the group-specific entry
5,
we
implement
which includes the number of products
sums over
this idea
in F(j)
by calculating an instrument vector
and
B(j), the
groups to which product
the characteristics of other products in F(j) and B(j).
also calculate "Ownership" instruments (labelled so in Table 5) which exploit the
economics of multiproduct pricing as
reflected in the
RHS
of (16). For example, as a particular firm
'^
In our base model, which has two distributional parameters, each row of Z must have
minimum of (K+3) elements (where K is the number of observed product characteristics).
"
specify
an econometric sense) to create a "local" index of the number of substitutes
As shown
belongs, as well as
We
for product
be correlated with the impact of the segmentation parameters, a and p, on prices and
the implicit fianction 6(s,p). Essentially,
process
FOC
(16)
the observed price and the product's within-group market share. Thus, the
each group
behave according
Q
a candidate instalment.'*
number and strength of close
For example, as the number of products
in that
of instruments for demand
represent by the set 0(j) (for ownership).
1
with each product
be correlated with
c(q^.w.)^.E(p,-c,)-^^
Pj =
Any exogenous
number of
^^
excluded instruments needed to dim{a,p}.
price and/or
directly in Z. This immediately reduces the
This approach was
(1981, 1987),
in
critical to
which model
the identification of the (simpler)
a
demand models of Bresnahan
simplicity allowed for the analytical solution
of equilibrium prices
and market shares.
^*
As in many other industries, it is diflBcult to obtain observable product-specific cost shifters
PCS; consequently, we do not attempt identification via that route. However, this strategy can
sometimes be used appropriately (Bresnahan and Baker, 1986; Hausman, 1994).
for
22
sells
a larger share of the number of total products available,
it
will
charge a higher price on each
product (as compared to the prices which would be observed under a more unconcentrated market
structure). Similarly, the
sum of the
characteristics
positively correlated with price, since that
two
analytical
ownership and
of the products of the firm (other than j)
would mean higher
in the literature
because some groups,
(BLP
differentiation to construct the
like Frontier,
different instruments than
(1995)). In particular,
saw much more
we
error.
It is usefiil
economic
We include
while
returns, this
dummies
for both
B
is
entry than others. This entry, which
in
successful
we assume
the short panel considered in the
to note that our exogeneity assumption
at first blush. In particular,
its
are using our assumption about the
excluded IVS. This strategy
across groups means that our proposed instruments vary even
present application.
our immediate
by changing competition. The different rates of entry
to be exogenous, shifts prices and market shares
determined by
these
cluster.
group structure of product
seems
we blend
Finally,
be
arguments and use the characteristics and count of products which share both
By using PD-specific instruments, we use somewhat
predecessors
% in eq.(16).''
will
we believe that
may
not be as strong as
entry into specific clusters
is
it
endogenously
does not mean that our instruments are correlated v^th the
and F
in
observable product quality; which means that
E,
is
merely the deviation of the products' quality fi^om the cluster means.
been predicated on the assumption that firms engage
Finally, the discussion so far has
Bertrand competition. However,
if
firms play a different type of non-cooperative
game
in
(e.g.
Coumot), then the precise form of (16) would change, but our instruments would be unaffected, since
the logic of ownership-based and group-based instruments remains the same. Outright collusion
though would make a
More
difference, since
generally, the point
is
ownership-based instruments would be longer be relevant.
that the instruments should reflect the underlying supply
conditions that prevail in the market, that
is,
the type of price setting behavior of firms, and the nature
of the heterogeneity of consumers as manifested
in the
groups. For the case at hand,
firms used individualistic pricing strategies; the pace of technical change in
intensity
^^
we
closely follow
BLP
(1995).
23
we assume that
PCS
and the ensuing
we
will explore the
of competition seems to have prevented price collusion. However,
In this argument
and demand
importance of this assumption by removing the ownership-based and PD-based instruments from
Z
the course of evaluating the robustness of our results.
in
RESULTS
VI.
We turn now to the presentation and analysis of results from the estimation of the PD GEV
for a variety
of specifications.
We also
PD GEV.
models and contrast them with the
model,
we
present estimates of alternative
Two-Level
NML and VPD
After exploring the robustness and limitations of the
meaning of the estimates by drawing hypothetical demand functions for
illustrate the
Branded and Frontier PCS and perform a
highlight the high degree
set
of counterfactuals involving hypothetical entry to
of market segmentation
that these estimates imply.
PRINCIPAL FINDINGS
A.
We first preview our main findings (drawn from Model
1
in
Table 6)
in
a concise fashion to
help frame the discussion of the results:
<
(i)
(Pp
smaller than
B
,
«1. Both of the
Pb)
1),
substitution parameters are estimated to be small (significantly
suggesting that there
is
indeed a high degree of market segmentation along both the
and F dimensions: Non-Branded products are poor substitutes for Branded products and
Non-Frontier products are poor substitutes for Frontier products. The parameter pg
quite precisely in
most
specifications, while the estimate
of Pp
is
much
is
estimated
less precise. Nevertheless,
the estimates for both distributional parameters are quite stable across specifications.
(//)
The product
differentiation advantages
frontier is limited to the insulation
it
of B and F are
provides from
NF
different.
The advantage of being
competition (pF«l). Perhaps surprisingly,
incorporating Frontier technology did not necessarily shift the product
is
small, usually negative,
out the product
demand curve out
(Pfrontier
and always imprecisely estimated). In contrast, having a brand name
demand curve (Pbrand ^
competition (pg <
at the
0)
a«^ provides B products with some
1).
24
shifts
protection from
NB
(/77)
As
expected, the price coefiScient, a,
is
sensitive to
whether or not the model accounts for price
endogeneity, and to the type of instruments used. Models that ignore the correlation of price with
unobserved quality
exploit the full set
(iv)
result in
smaller (in absolute value) estimates of a; models which
much
of proposed instruments severely reduce the precision of
The choice of model
that are highly sensitive to the order
PD GEV model,
of nesting. In
substitution parameters are reversed in the
a vis the one with the F/NF nest on top. Thus, the
alternative, at least in those cases
the
Two-Level
particular, the relative sizes
Two-Level
this estimate.
be important for analyzing the market under
specification turns out indeed to
study. In fact, the leading alternative to the
NML,
yields estimates
and significance of the
NML model with the B/NB nest on top, vis
PD GEV model offers an attractive methodological
where the nesting order
is
not well defined a
priori.
MAIN SPECIFICATION, ROBUSTNESS AND IDENTIFICATION
B.
We
which
present in Table 6 our base specification (Model
we vary the
elements of
two
1),
X (the set of product characteristics),
alternative specifications in
and a fourth one where
an additional PD. The parameters estimated are the distributional coefficients pp and
coeflBcient,
a various Ps
,
a sub-set of
DISK,
(a constant,
RAM,
In fact, even though pp
and/or Pp
=
1
•
The
is
and coefficients for
and MHZ), as well as p^Hz
As previewed above,
1
Models
positive,
1
and
and
FRONTIER, BRANDED, IBM,
and for
Model
4.
all
these specifications.
estimated very imprecisely, one can easily reject the hypotheses that pg
Models
significant while the
in fact negative in three
add
a price
i"
results for the price coefficient,
2, nil in
we
Pb,
the estimates of pp and Pe are quite small in
negative and quite stable across specifications; however,
in
do not
3
and
its
are encouraging:
level
out of four specifications).
The
estimated to be
BRANDED
bit (fair
coefficient is large,
cannot be distinguished fi-om zero (and
is
IBM coefficient is positive and gives (at the
50% incremental
boost to
IBM products above and beyond
other Branded products. Finally, our technological control variables (DISK,
25
it is
of significance varies quite a
four models, the
4). In all
FRONTIER coefficient
point estimate in the base specification) a
a,
=
RAM,
and
MHZ)
have
the expected sign, though only the
It
and
coefficient
significant (in
is
Model
1).^*
should be noted that the coefficients that determine the level and slope associated with
which are of particular
F,
DISK
interest here, are relatively stable across the four specifications.
Moreover, the point estimates suggest
differentiation
B
that there
was a
F versus
B
advantage afforded along the
the
qualitative difference in the product
dimension.
In fact,
we
are able to reject
the hypothesis:
Pf=
Ho:
at the
at the
95%
level),
fi^om having a brand
F
parameter (p^o^z
^
coeflBcients
name, being
in
at the Frontier
-99).^^
little
we compare
of the Two-Level
5,
substitutability
reject
advantages arising
we
include a
MHZ substitution
evidence that there existed segmentation along the
MHZ
~
!)•
The most
is
on
top.
PD GEV
troubling feature of Models 5 and 6
In particular, with
of products across branches
PD
that governs the split
PC
is
NF
p
at that level,
B/NB on top,
at
pg
is
the
«
Pf,
a given level of the tree
and hence a high degree of
between branches. According to
branded or not (pg
However, Model 6 implies the opposite,
confers insulation from competition from
with those from two alternative
Recall that a small
the crucial distinction resides in whether a
precisely estimated).
cannot
and F
NML model.
market segmentation according to the
Model
we
B
the estimates from the
while the reverse occurs when F/NF
poor
(e.g.,
high
proved to be beneficial more in the sense of insulating
high sensitivity ofthe estimates to the order of nesting.
signifies
~ 4.61). While the
In contrast to the substantial segmentation found along both the
dimension (we cannot reject H^. p^nz
specifications
2)
relative to the
is that,
pushing out the demand cune. In Model 4
dimensions, there seems to be
In Table 7,
5.22, X^o.so,
do not allow us to draw firm conclusions
a cautious interpretation of this result
from competition than
'^
Pb
90% level in our main specification (Wald test statistic =
standard errors of the
Ho
Pf=
Pb>
that
is,
is
both very small and
that having Frontier status
products, whereas Branded status does not provide
We tried also combinations of these attributes, but the extremely high collinearity between
them, as well as between them on the one hand and
PRICE and FRONTIER on
the other,
rendered highly imprecise estimates.
"
Recall that specifying this model
to (6), each of which
is
is
easy.
All that
is
required
is
the addition of several terms
a summation over sets of products which share a similar
26
MHZ rating.
additional protection. Obviously, there
models, contradictory as they are,
the
PD GEV
is
no way of telling within
this
the "right" one. In contrast to these
is
context which of the two
awkward modeling
choices,
confronts the issue of relative importance directly by treating the two principles of
differentiation symmetrically.
we
In Table 8
vary the instrument set in order to examine the robustness of our findings to
Model 7
alternative identification assumptions.
dependence on unobserved
absolute value), providing
quality.
The
Model
exogenous, ignoring
their
presumed
result is that the price coeflBcient shrinks dramatically (in
weak evidence that
the overidentifying restrictions). In
treats prices as
we eliminate the
8,
we cannot reject
endogenous (though
prices are indeed
ownership-based instruments, that
is,
those that stem from the optimizing behavior of multi product firms. Similarly, Model 9 removes
PD-based instruments,
that
those associated with the variance in competition across groups. Each
is,
of these exercises dramatically reduces the measured precision of the important parameters of the
model. The implication
is
quite clear: to identify our model,
on a behavioral assumption about supply and a
of these assumptions are necessary
We ran several
which appeared
in
other
structural
we
constructed instruments that relied
one about the market. As
no
GEV specifications,
accounting for
serial correlation (across
much
specification in
that these data
by Model
suggests
1
we can
which
can
Pm)-^* While there
reject the qualitative results described above, the inclusion
tell (at least
of most estimates, implying
to
accommodate
for finer distinctions: the basic
is
that
A look
is
only
Tables 2 and 4 anew
and overwhelming
in the (B,
NF)
cluster.
By contrast,
(clone) product sold less than 5,000 units/year. Thus, finer distinctions
'*
at
fact that
IBM accounted for 60% of the market (about
per year), and the bulk of those sales were
out to the stark contrast between
that there
of
through these models). In other words, the results implied
cannot be made more precise by additional modeling.
why we cannot hope
would have
products
both years), arbitrary heteroskedasticity, and additional potential substitution
additional parameters drastically reduces the precision
so
turns out, both
order for us to provide evidence for our main findings.
in
patterns (e.g., the inclusion of four separate substitution parameters, Pe, Pnb, Pf^
exists
it
1
any model
million unit sales
the average
between non-IBM
NB/NF
PCS
lose
IBM and the rest of the pack.
These sections are not reported
in tables but are available
directory has been set up for public access.
27
from the authors.
An ftp
archive
C.
A GLIMPSE AT DEMAND CURVES FOR BRANDED AND FRONTIER PRODUCTS
We noted above that the estimates of Pbranded 's positive,
PiBM
'S
also positive,
and that PpRomTER
we
as to the meaning of these results,
demand
construct hypothetical
hypothetical 1987
demand
's
and highly
significant, that
negative and insignificant. In order to gain fiirther intuition
moment
take for a
the point estimates
at face
IBM, Non-Frontier PC,
curves, one associated with a typical
One point on each demand curve
value and
Figure 2 displays two
of PCS.
fiinctions for different types
with a typical Non-Branded, Frontier PC.
large,
is
the other
the actual average price
and quantities for these two typical products. The other marked point on the demand curve for the
IBM/NF PC
corresponds to the price at which the mean buyer's valuation of an
just equal that of a "clone" product, that
calculate the price that renders 6^8, f
~
is,
the price at which
very early stage
in the difiusion
Likewise,
6nb,nf-
it
clear that the
main advantage of Frontier
status,
of the 386, seems to have been a steep demand curve rather
than a large one (recall the argument in section 2 about the type of users that chose 386-based
back
then).
The converse was
true for
IBM/Branded
demand
characterizing sellers' behavior, these
we
^nb.nf-
The demand curves thus constructed make
at that
6^^^^ =
IBM/NF PC would
While
status.
distinctions
we
PCS
have not estimated equations
would explain why
IBM
chose to take the
bulk of its product differentiation advantage as higher quantities rather than as higher prices (recall
IBM
the negative coefficient
on
manufacturers of Frontier
PCS took the
large
demand curve and
in the
reverse course.
the steep one, were both at
2)),
whereas the pioneer
The two mechanisms of appropriability,
work
in the industry,
but within
dififerent
the
PDs.
COUNTERFACTUALS
D.
In order to shed further light
counterfactual exercises. In each
and, using the estimates from
would
the
hedonic pricing equation (Table
induce.
same
The main goal
cluster, vis
we
Model
is
to
1,
on the meaning of our
introduce a hypothetical
compute the changes
compare the
a vis the impact on sales
effects
and
present in Table 9 four
new product
into a particular cluster
in sales
in other clusters.
pg.
28
of other products that such entry
of entry on sales of competing brands within
degree of insulation from competition enjoyed by products
substitution parameters p^
we
findings,
We can thus provide a sense of the
in different clusters, as
captured by the
we
In each case,
introduce a hypothetical product with a value of 6
mean value
attractiveness) set equal to the
means
fixed (which
in particular that
recompute the implied maricet shares for
all
expected
of all other products
for that cluster.'' Holding the 6s
we do
its
(i.e.
not allow for any pricing responses to this entry),
products, including the outside good.
We then calculate
the difference in sales for each product before and after the hypothetical introduction. Finally,
over
the sales
in
which an average F/B product
of this hypothetical
from within the F/B
had not purchased
segmentation
PC would
total
is
introduced (with 6
851
1
units,
PCS
statistic
The remainder comes from
before.
and
its
the outside good, that
In the last part of the table, 9(E),
(linearized) standard error.
is
is,
predicts that
"market-stealing"
drawing 1700
from individuals who
simple market-stealing
the percentage change in
is
of the entry experiment. This negative
larger (in absolute value) as the entrant
product's. This positive
statistics
The model
we show a
The numerator
segmentation becomes more important. The denominator
These
-5.2).
of which 5270 (61%)
unit sales of existing products within the cluster as a result
number becomes
=
cluster. Market-stealing across clusters is significantly smaller,
additional units (20%).
PC
we sum
products in each cluster to evaluate the competitive effect of entry across clusters. Consider
all
Table 9 (A),
all
we
is
becomes more important and
as
the percentage change in unit sales of
number increases with the importance of the product introduced.
confirm the importance of intracluster market stealing implied by the high level of
measured segmentation.
While the overwhelming
product comes from /«/racluster
the estimation.
a vis F/B)
partially
is
The
is trivial
first result is
(16
examining the
/'n/ercluster effects highlights the results
on the completely non-matching
which matches along the
303 units stolen from the F/NB
cluster
B
in
in the
among
the
dimension (NF/B, which loses 1350 units versus
cluster) the effect in
terms of percentages
which matches along the F dimension (F/NB, which
For example,
vis
terms of unit quantities
loses over
1
is
much
NB/NF
1988 category,
\y^= -5.44,
cluster in 1988.
29
larger for the
percent of its unit sales,
whereas NF/B's 1350 units represent just one tenth of one percent of total units
''
of
(NF/NB
Intercluster market-stealing occurs almost exclusively
matching clusters (NF/B and F/NfB). Moreover, while the effect
larger for the cluster
cluster
effects,
that the effect
units).
most of the market share of the hypothetical new
result is that
sold).
the average level of 8 in that
A
similar story (a large /w/racluster effect, and /w/ercluster effects confined to partially
matching clusters)
is
hypothetical low-end clone
the
For example, a
told in each of the other hypothetical presented in Table 9.
(NF/NB)
sells
Branded group (the B/F and B/NF
over 2,500
clusters);
units,
of which only 3 1 are market-stealing from
an additional 75 are drawn from the
F/NB
cluster.
According to our counterfactuals and contrary to some widely held perceptions of this industry, these
clones posed
little
CONCLUDING REMARKS
Vn.
For the
economy;
fly in
competitive threat to Branded and Frontier competitors.
last
at the
1
same
5 years, the
time,
is
it
PC
industry has been one of the
one of the most competitive. Such conjunction of forces seems to
the face of the intuition suggested
by Arrow about the necessity of monopoly power
to induce, and -ndeed pay for, costly and risky innovation.
where quality-adjusted
ever lower and
new
prices
most innovative sectors of the
fall at
Where do
rents
come from
in
in
order
a world
a rate of 25% per year, low-cost imitators keep driving prices
distribution channels eradicate the advantages
of existing brand
Motivated by a desire to identify the market origins of innovative rents,
this
capital?
paper develops
a simple discrete choice model that uses just aggregate data and can easily incorporate prior
knowledge of the
structure of the industry.
We
propose the Principles of Differentiation
drawing primarily from McFadden (1978) and Berry (1994).
multinomial logit models, ours
is
of substitution to conform to a
In applying the
market exhibited two
not hierarchical; therefore,
specific order
it
Compared
of
of what follows
differentiation, according to
relies heavily
cross-elasticities
of nesting.
regarded as having a brand name, and to whether or not the product
Much
to Multi-Level nested
does not constrain the
model to the case of PCS during 1987 and 1988,
principles
GEV,
on the above
we begin by noting
that the
whether or not the firm could be
was
at the technological frontier.
testable hypothesis: the
model
is
structured
accordingly (there are two key substitution parameters associated with these PDs); an important
subset of the instruments used are predicated on the assumption that these
slicing the
fiirther
demand
PDs
are a
good way of
market so as to capture the different degrees of competition faced by products
computations using the estimates are designed so as to compare
fiinctions facing products in different clusters,
30
in
salient properties
them;
of the
and to gauge inter-cluster market-stealing
effects.
The hypothesized PDs
even
if
the market as a whole
insulation
from competition
frontier can generate rents
also oflFer a suitable
is
highly competitive, market segmentation
in certain clusters.
more mundane
market was by no means an
clusters.
demand
These
may provide
(temporary)
Developing a brand name and/or innovating
B
or
F products can have
either large
demand curves or be poor
all
encompassing phenomenon, but that
name conferred a
whereas being early on
results refer just to
it
was
large advantage in the sense of shifting out the
at the technological frontier did not.
1987 and 1988 and, given the rapid pace of change
of the
PC
in the industry,
sector before and after. Yet
interpret the evolution
one inference stands out as eminently
plausible: contrary to widely-held perceptions, the
IBM as the paramount leader in this industry seems to have been caused
demise of
not by relentless competitive
pressures from clones, but rather by the erosion of IBM's quasi-monopoly stand within
is
in
some
own
specific,
cases also positioning themselves at the frontier.
two-way grouping postulated here
an appropriate representation of the
PC
market
is
just a starting point for
at the time.
Other
industries,
what we believe
and perhaps even the
sector itself at a different time, would require different ways of structuring and grouping the
PC
competing products and firms. Indeed,
all,
its
by Compaq, then by other entrepreneurial firms that invested heavily on developing a
brand name, and
The
PC
largely localized within
one should be wary of using them to
cluster, first
at the
competitors. Indeed, our findings indicate that competition in the
Further, having a brand
function,
raised above:
by increasing buyers' willingness to pay; depending on the strength and
the heterogeneity of this increase,
substitutes for
framework to address the puzzles
consist just of an
amorphous
it
seems that product differentiated markets very
collection
they appear to be structured according to a
a corresponding
way
menu of strategies opened
of n different products produced by
few
principles
to firms.
m
of differentiation, which
rarely, if at
firms. Rather,
relate in turn to
A systematic study of those principles,
of the
they differ across markets, and of their dynamics seems to be a promising and exciting line of
research in empirical Industrial Organization.
31
BIBLIOGRAPHY
Arrow, K. (1962). "Economic Welfare and the Allocation of Resources for Invention." The Rate and
Direction of Inventive Activity, ed. R. Nelson, Princeton Univ. Press, Princeton (NT), pp. 609-625.
Bemdt, E. and Z. Griliches (1993). "Price Indexes for Microcomputers:
Price Measurements and Their Uses, ed. Murray Foss, et
Chicago
An
Exploratory Study,"
University of Chicago Press,
al,
(IL), pp. 63-93.
Berry, S. (1994). "Estimating Discrete-Choice
Models of Product
Differentiation,"
RAND Journal
of Economics, 25(2): 242-262.
Berry,
S.,
M. Camall and
P. Spiller (1995), "Airline
Hubs: Cost and Demand," mimeo, Yale
University.
Berry, S.,
J.
Levinsohn, and A. Pakes (1995). "Automobile Prices in Market Equilibrium,"
Econometrica, 63(4): 841-90.
Bresnahan, T. (1981). "Departures from Marginal-Cost Pricing
in the
American Automobile Industry:
Estimates for 1977- \91%,'' Journal of Econometrics," 17: 201-227.
Bresnahan, T. (1987). "Competition and Collusion in the American Automobile Industry:
The 1955
Price V^dLT," Journal
of Industrial Economics, 35
Bresnahan, T. and
Baker (1988). "Estimating the Residual Demand Curve Facing a Single
J.
(4):
International Journal of Industrial Organization,
Cardell, N.S. (1991). "Variance
6:
457-82.
Components Structures
Washington State University Working Paper.
Chamberlain,
G.
Restrictions," Journal
"Buyer's Guide:
"Asymptotic Efficiency
in
Extreme Value and Logistic
for the
Distributions,"
(1987).
firm,"
283-300.
Estimation with
Conditional
Moment
of Econometrics, pp. 305-334.
Microcomputer Hardware," ComputerWorld, September
26, 1984.
COMTEC Wave 6, COMTEC Market Analysis Sen/ices,
"Complete Computer Product Book," DataSources, (1986; 1987).
GML Microcomputer Guide, GML Corporation,
Goldberg, P. (1995). "Imperfect Competition
Lexington (MA), various years.
in International
Automobile Industry," £co/iomemcfl, 63(4): 891-951.
32
Markets: the case of the
US
Hausman, J. (1996). "Valuation of New Goods Under Perfect and Imperfect Competition," The
Economics of New Goods, ed. T. Bresnahan and R.J. Gordon, University of Chicago Press,
Chicago (IL).
'
Hendel,
I.
(1994).
"Estimating
Computerization Returns,"
Multiple-Discrete
NBER Technical
Choice
Models:
An
Application
To
Working Paper #168, Cambridge (MA).
Henderson, R. and K. Clark (1990). "Architectural Innovation: The Reconfiguration of Existing
Product Technologies and the Failure of Established Firms," Administrative Science Quarterly,
35(1): 9-30.
Jenkins, A.
(1987). "In the Market for Micros," Computerworld, 2(32A):
Kolodzicj, S. (1988). "The
Levin, R., et
al (1987).
23-24.
Growth of PC Clones," Computerworld, 20(45A): 35-36.
"Appropriating the Returns from Industrial Research and Development,"
Brookings papers on Economic Activity, pp. 783-831.
Lua, K.T. (1989). "Relative Performance Measurement of 80386, 80286 and 8088 Personal
Computers," Microprocessing
&
Microprogramming.
2(2): 85-95.
McFadden, D. (1973). "Conditional Logit Analysis of Qualitative Choice Behaviour," Frontiers
in Econometrics, ed. P. Zarembka, Academic Press, New York (NY).
McFadden, D. (1978). "Modelling the Choice of Residential Lxxation," in Spatial Interaction
Theory and Planning Models, ed. A. Karlqvist, et al., North-Holland Publishing, Amsterdam
(NE).
McFadden, D. (1981). "Econometric Models of Probabilistic Choice," Structural Analysis of
Discrete Data with Econometric Applications, ed. C. Manski and D. McFadden, MIT Press,
Cambridge (MA).
PC Tech Journal,
various issues.
"Annual Hardware Review,"
PC World,
June 1984.
O'Malley, C. (1988). "Winning the Battle of Systems Overkill," Personal Computing, 12(12):
99-106.
Quando, R. (1956). "A Probabilistic Theory of Consumer Behavior," Quarterly Journal of
Economics, 70: 507-36.
Ryan, A. (1987). "Buyers Skirt Clone Service Woes," Computerworld, 21(50): 35, 41.
33
Small, K. (1987).
"A Discrete Choice Model
for
Ordered Alternatives," Econometrica,
55(2): 409-424.
Stavins,
J.
(1995).
"Estimating
Demand
Elasticities in a Differentiated
Product Industry: The
Personal Computer Market," mimeo. Federal Reserve Bank of Boston, Boston (MA).
Steffens,
J.
(1994).
NewGames:
Strategic Competition in the
PC Revolution.
Pergamon
Press,
Oxford (GB).
Stem,
S.
(1996). "Product
Demand
in
Pharmaceutical Markets," mimeo,
Teece, D. (1988). "Technological Change and the Nature of the Firm,"
Economic Theory,
Trajtenberg,
M.
ed. G.
(1990).
Dosi
et
al,
Pinter Publishers,
MIT
Sloan School.
Technical Change
Economic Analysis of Product Innovation, Harvard University
Cambridge (MA).
34
and
New York (NY).
Press,
Appendix A
Dataset Methodology and Sources
PC DATASET CONSTRUCTION METHODOLOGY
We arrived at the sample of make/models used in the analysis by a process of elimination
I.
First, we restricted ourselves to products sold by the top 25
eliminated
makes for which the number of units sold was less
and
1),
than 300 (approximately equivalent to sales of $1,000,000). This eliminates a substantial number
according to the following
criteria.
vendors of PCs (see Table
of very small
IBM clones,
but each of these small vendors accounted for a very small fraction of
the business market at the time. In fact, over
90%
of the
1988 were marketed by the top 25 vendors. Secondly,
on
Intel
number of PCs sold
in
microprocessors (the laPX 80x86 family), including "close" variation from
The
etc.
total
principal alternative microprocessor (Motorola's
68000
family)
computers marketed primarily by Apple.' These computers were only
Intel-based machines at this time.^ Finally,
we
1987 and
we constrained the sample to PCs based
AMD, NEC,
was incorporated
into
distant substitutes to
we could
eliminated makes/models for which
not
obtain reliable data on prices and/or characteristics.
DATA ON PRICES AND CHARACTERISTICS
n.
Unfortunately there
is
no
which provides consistent and
of marketed PCs. Thus, we had to
single data source
reliable
rely on a
of sources, and invested a great deal of effort in trying to establish consistency and
comparability. The primary data source for this purpose was the GML Microcomputer Guide, a
quarterly publication which provides list prices and technical characteristics of all PC's marketed.
The GML data was complemented and confirmed with an extensive set of additional primary and
secondary data sources. These included 1987 and 1988 back issue catalogues and company
histories (provided directly by vendors) and a number of contemporary surveys in
ComputerWorld, Data Sources, PC World, and PC Technical Journal. Finally, a complementary
dataset with prices and characteristics of marketed PCs was graciously provided by Zvi Griliches
information about the price and characteristics
variety
and Ernie Bemdt (see Bemdt and Griliches (1993)). In order to be included
complete and matched
or in
able
set
of prices and characteristics needed to be
in the
sample, a
identified in the
two or more of the additional sources. For many, if not most, of the
to identify at least two mutually consistent sources.'
GML data
observations,
we were
There were also a small number of PCs (mostly concentrated at the very low or very high
end of the quality dimension) which were based on other microprocessors (HP workstations.
Commodore PETs, etc.). We excluded them as well from the sample.
'
^
Graphics and marketing people tended to purchase Apple technology, whereas accounting,
manufacturing, and sales management tended to buy
Apple as an additional Branded firm
in the
qualitative difference to the estimates.
context of the
'
year.
IBM.
We experimented with including
context of nested logit models, but this
made
little
However, we have not conducted such experiments
in the
PD GEV model.
For a small number of observations, we observed sales in both years but price in only one
When this occurred, we imposed a price depreciation rate of 10% between the two years.
35
APPENDIX B
VARIABLE DEFINITIONS
QUANTITY
Number of PC j
QUAN^,
COMTEC
MARKET SIZEj
Total
Sold
in
Year
t
as Estimated
by
Market Analysis Services
Number of Non-Proprietary
Workers; Includes Workers
Business Office
Who Choose
Not
to
Purchase
MARKET SHAREj
,
Market Share of Product j
in
Year
t
QUAN / 39,000,000
CHARACTERISTICS
FRONTIER^,
Dummy
Equal
to
1
if
Product j Incorporates 80386
1
if
Product j
Microprocessor
BRANDED:,
Dummy
Equal
AT&T, HP,
PRICEjj
or
to
is
Marketed by IBM,
Compaq
List Price of Product
j
in
Microcomputer Guide or
Year
in
t
as Reported in
Two
or
More
GML
Additional
Sources
RAMj,
Standard
Year
MHZ:,
t,
Random Access Memory (RAM)
for Product
Kilobytes (KB)
Standard Microprocessor Speed Provided, in Year
Product j;
DISK,,
j; in
Provided, in
in
t,
MHZ
Standard Size of Hard Disk Provided, in Year
Product j; in Megabytes (MB)
36
t,
for
for
nCUREl
Nested Multinomial Logit and
PD GEV
Buy?
F|B
B,F
B,NF
NB,F
NB,NF
None
none
NB,F
B,F
F <cr^
Pf
j|F/NF;B/NB
^
//V
B,NF
37
/V
NB,NF
TABLE
1
SUMMARY OF TOP 25 VENDORS
FUUVf
TABLE 2
SUMMARY STATISTICS BY YEAR
TABLES
HEDONIC PRICING EQUATION
1987 & 1988 PC MARKET
(Standard Errors
DEPENDENT VAR =
PRICE
in Parentheses)
TABLE 4
SUMMARY STATISTICS
BY CLUSTER BY YEAR
YEAR =1987
Table 5
Definitions of Instruments
(Used In Model (1))*
Principles of Differentiation
F(j,t) =
52
(other products sold in year
1
that share
F or NF
Disk^
5]
B(j,t) s (other
1
products sold
E
year
in
t
that share
Disk,^
Branded^
B
or
NB
51
status with product j}
Frontier^
B(j,t)
B(j.t)
BG.t)
status with product j}
FO.t)
FG.t)
FO.t)
52
52
t
Ownership
0(j,t) = (other
E
ocj.t)
1
products sold
E
oG.t)
in
year
t
Disk,
by
seller
of product j}
TABLE 6
PD GEV MODEL
1987 & 1988 PC MARKET
1
TABLE 7
PD GEV VERSUS TWO-LEVEL NML MODELS
1987 & 1988 PC MARKET
TABLES
PD GEV MODEL
ALTERNATIVE TREATMENTS OF THE INSTRUMENTS VECTOR
1987 & 1988 PC MARKET
PARAMETERS
Table 9 (A, B)
Etemand Impact of Entry of
One
Additional F/B Product
Affected Category
Change
in
Unit Sales
Percent Change
This Product
+8569
Other F/B Products
All
All
NF/B
Products
-5390
-1350
-.0265
-.0012
F/NB
AU NF/NB
Products
Products
-303
-16
-.0107
-.000041
Outside
Good
-1512
-.000041
One Additional NF/NB Product
Outside good
-806.00
-.0000022
All
NF/B
All
F/B Products
Products
-26
-5
-.000024
All
F/NB
Other
Products
NF/NB
Products
-1590
-75
-.0000026
-.0027
This Product
+2502
47
Table 9 (C, D)
Demand Impact of Entry of
One
Additional
F/NB Product
Affected Category
Change
in
Unit Sales
Percent Change
Outside good
-463
-.0000124
All F/B Products
Other
All
NF/B
Products
-196
-14
-.0010
-.0000124
F/NB
Products
All
NF/NB
Products
-51
-854
-.00182
-.0022
This Product
+1578
One
Additional
NF/B Product
This Product
+ 7940
All
All
Outside
F/B Products
Other
NF/B
Products
-207
-5212
-.0010
-.0047
F/NB
Products
All
NF/NB
Products
-2
-29
-.000067
-.000
Good
-2490
-.000067
48
Table 9 (E)
Market-Stealing Segmentation
AQfH
^tb
'Cb
Product Introduced in
/
AQ,PC
fpc
oo
fr
•) v>
->
:>
007
Date Due
Lib-26-67
MIT LIBRARIES
3 9080 01428 2054
MHaMBMMMiaHiil