Study on E-commerce and Market Structure

Description
This paper examines the effect of the advent and diffusion of e-commerce on supply-side industry structure. We specify a general industry model involving consumers with differing search costs buying products from heterogeneous producers. We interpret e-commerce as a reduction in consumers' search costs. We show how it reallocates market shares from high-cost to low-cost producers.

E-commerce and the Market Structure of Retail
Industries
1
SHORT FORM: ‘E-commerce and Market Structure’
Maris Goldmanis, University of Chicago
Ali Horta¸ csu, University of Chicago and NBER
Chad Syverson, University of Chicago Booth School of Business and NBER
¨
Onsel Emre, Putnam Investments
March 25, 2009
1
We thank Tim Bresnahan, Francine Lafontaine, Michael Rauh, and participants at multiple
seminars for valuable suggestions. We gratefully acknowledge ?nancial support for this research
from the NET Institute (www.NETinst.org) and the NSF (award no. SES-0242031). Goldmanis
and Horta¸ csu can be contacted at the Department of Economics, University of Chicago, 1126 E.
59th Street, Chicago, IL 60637. Syverson can be contacted at the University of Chicago Booth
School of Business, 5807 S. Woodlawn Ave., Chicago, IL 60637.
Abstract
This paper examines the e?ect of the advent and di?usion of e-commerce on supply-side
industry structure. We specify a general industry model involving consumers with di?ering
search costs buying products from heterogeneous producers. We interpret e-commerce as a
reduction in consumers’ search costs. We show how it reallocates market shares from high-
cost to low-cost producers. We test the model using US data for three industries: travel
agencies, bookstores, and new auto dealers. Each industry exhibits the market share shifts
predicted by the model, but the mechanisms vary, ranging from aggregate factors in the
travel industry to local-market factors in the other two industries.
This paper explores how the advent and di?usion of e-commerce impacts the structure
of retail and similar industries. While there is a burgeoning literature studying how e-
commerce has a?ected prices and price dispersion (Brynjolfsson and Smith, 2000; Clay et
al., 2001; Scott Morton et al., 2001; Brown and Goolsbee, 2002; Baye et al., 2004), much less
work has looked at how the di?usion of the Internet has in?uenced the number or type of
producers that operate in an industry. That is, questions of which businesses most bene?t
and most su?er (perhaps to the point of having to cease operations) from the new consumer-
matching and distribution systems that e-commerce brings have received little attention.
Conventional wisdom suggests that such e?ects can be large and diverse in impact; the rapid
growth of Orbitz, Travelocity, and Expedia at the expense of local travel agencies is one oft-
cited example. Yet we do not yet know quantitatively just how large this particular e?ect
has been or whether similar mechanisms operate across di?erent industries. This paper seeks
to begin to address these issues.
It is almost certain that more than just equilibrium prices are a?ected when e-commerce
spreads in an industry. Market shares are very likely to change; given the reduction in con-
sumer search costs that e-commerce can bring, any ?rm’s price advantage will be multiplied
in terms of market-share gains. Higher cross-price elasticities imply di?erential impacts on
industry ?rms depending on whether they have a cost advantage or disadvantage relative
to their competitors. It is also quite likely that these market share changes can be drastic
enough to lead some ?rms to exit from the market entirely. On the other hand, lower search
costs could also induce new entry into the industry. Presumably, though, these entrants
may di?er on average from industry incumbents because e-commerce has raised the return
to being e?cient (or, alternatively, to being able to produce high-quality goods). In such
ways, e-commerce can have important entry and exit consequences as well.
Our investigative approach combines theoretical and empirical analyses. We ?rst model
equilibrium in an industry comprised of heterogeneous ?rms selling to a set of consumers who
di?er in their search costs. Heterogeneity across ?rms arises from di?erences in underlying
1
abilities like production costs or output quality. We embody them as di?ering marginal
costs for the sake of concreteness, though it is easy to modify the model to allow variation in
product quality levels instead. Industry consumers search sequentially when deciding from
whom to buy. Firms set prices given consumers’ optimal search behaviour as well as their
own and their rivals’ production costs. Firms that cannot cover their ?xed costs exit the
industry. Initial entry into the industry is governed by an entry cost.
We interpret the advent and di?usion of e-commerce as a leftward shift in the consumer
search cost distribution. We use our model to show how e-commerce activity impacts equi-
librium market structure. The model o?ers predictions about not just equilibrium prices,
but also market shares, the number of producers, and the producer type (marginal cost)
distribution.
Consistently with previous literature, the model predicts a decline in equilibrium average
price levels and price dispersion. The more novel implications of our work, however, regard
what happens to the equilibrium distribution of ?rm types. Here the model predicts that
the introduction of e-commerce into an industry should result in the shrinking and some-
times exit of low-type (i.e., high-cost) ?rms, a shift in market share to high-type (low-cost)
?rms, and with some additional assumptions about the ?rm type and consumer search cost
distributions, a drop in the number of producers as well.
We test the model using US County Business Patterns (CBP) data from 1994-2003. CBP
data contain, at the detailed industry level, the total number of establishments (stores) as
well as their size distribution. While we cannot measure producer types directly, we can
use size as a proxy; hence shifts in the size distribution are informative about heterogeneous
e?ects of e-commerce within an industry. The panel nature of the data allows us to focus on
changes in the distribution over time within local markets, removing possibly confounding
di?erences in technology or demand across markets. We identify local di?erences in the
impact of e-commerce (i.e., the size of the shift in the local search cost distribution) using
consumer-level survey data to measure the fraction of the local population who report buying
2
goods and services online.
We focus the empirical tests on three industries perceived to have been considerably im-
pacted by e-commerce: travel agencies, bookstores, and new auto dealers. We ?nd support
for the predictions of the theoretical model. Growth in consumers’ use of the Internet for
purchases is linked to declines in the number of small (and presumably low-type) estab-
lishments, but has either no signi?cant impact or even positive impact on growth in the
industries’ numbers of large establishments. Interestingly, while the industries experience
similar patterns in market share shifts, the speci?c mechanisms linking declining search
costs to the shifts di?ered across the industries. The shifts in the travel agency industry re-
?ected aggregate changes driven largely by airlines cutting agent commissions as consumers
increasingly shifted to online ticket sources. In bookstores and new car dealers, on the other
hand, the evidence suggests that the decline in small retail outlets re?ect market-speci?c
impacts of Internet di?usion.
We present the general industry model in the next section and explore its predictions for
how shifts in search costs impact equilibrium in an industry with heterogeneous producers.
The third section discusses the data used in the empirical analysis. This is followed by a
presentation and discussion of the empirical results. A short discussion concludes.
1. Model
Our model combines elements of two distinct theoretical literatures. One is the set of search
models with consumers that have heterogeneous search costs. Examples include Carlson and
McAfee (1983), Rob (1985), Benabou (1993), and Horta¸ csu and Syverson (2004). Our con-
ceptual approach of treating the di?usion of e-commerce technologies as shifting consumers’
search costs (perhaps disparately for di?erent consumers) is the obvious motivation for draw-
ing on this previous work. The second literature involves industry equilibrium models that
feature heterogeneous producers and endogenous selection into production, like Hopenhayn
(1992), Melitz (2003), Syverson (2004), and Asplund and Nocke (2006). Endogenising the
3
set of equilibrium producers is important to meet our goal of assessing how e-commerce
might di?erentially impact industry producers by type, including determining which types
enter and exit when search costs change.
1.1. Set-up
There is a continuum of ?rms selling a homogeneous good for consumption by a continuum
of consumers. All consumers have perfectly inelastic unit demand for the good being sold,
but are heterogeneous in their search costs s ? R
+
. The total mass of consumers is ?xed
and normalised to one. The probability distribution of consumer search costs is given by
cdf Q having a continuously di?erentiable pdf q. It is assumed that 0 is the greatest lower
bound of the support of q and that Q(0) = q(0) = 0. Like in Benabou (1993), ?rms are
also heterogeneous, di?ering in their marginal costs of production c ? R
+
, which are their
private information. The total mass of all operating ?rms is L. Unlike Benabou, we let the
mass of ?rms be determined endogenously, through a zero-pro?t condition (see Section 1.4).
The timing of decisions by ?rms and consumers is as follows. At the beginning of the
period, potential ?rms consider entering the industry. If a ?rm decides to enter, it pays
the sunk cost of entry, ? and learns its own marginal cost c, which is drawn i.i.d. from a
publicly known probability distribution with cdf ? and pdf ?, whose support lies in [0, 1].
Next, ?rms decide whether to stay in the industry or not. Those that choose to stay then
decide how much to charge and produce. Production requires a ?xed cost of operation ?,
which is identical in all ?rms. This cost can be avoided if the ?rm chooses to stay out of the
market.
1
1
We could have eliminated the ?xed cost of operation from the model, but in that case, those ?rms that
otherwise exit the market would stay in the market by charging prices equal to their marginal costs. Thus
having a ?xed cost in the model leads to the sensible implication that only ?rms that make positive pro?ts
stay in the market.
4
1.2. Consumers’ Problem
We make the standard assumption that consumers know the price distribution, F (with
density f), but must engage in costly search to learn the price charged by any particular
?rm. Consumers’ search is undirected and sequential; they visit stores one-by-one to learn
their price and after every visit compare the bene?t and cost of continued search. If the
expected price reduction from visiting another store is greater than the marginal (search)
cost s, the consumer continues to search; otherwise, she buys the product at the lowest
price in hand. Thus, as in McCall (1970), the optimal stopping rule is characterised by a
reservation price where a consumer continues to search as long as she ?nds a price greater
than some reservation price ?(s), where ?(s) is given by:
s =
_
?(s)
0
[?(s) ?p]f(p) dp. (1)
As seen in the equation, the reservation price is such that, if the price in hand is ?(s), the
marginal cost of search s equals the expected bene?t from continuing search. (The integral
on the right-hand side is the expected reduction in price from another search, accounting
for the option value of discarding higher price draws.) It also implies that a consumer
with zero search cost always buys from the ?rm with the lowest price. We convert this
optimality condition into an equivalent but slightly less intuitive form (albeit easier to work
with analytically) by integrating (1) by parts. This yields:
s =
_
?(s)
0
F(p) dp. (2)
Di?erentiating this with respect to s yields 1 = F(?(s))?
?
(s), which shows that ?(s) is strictly
increasing in s, and hence invertible on its range. The inverse is given by
?
?1
(r) =
_
r
0
F(p) dp.
5
1.3. Sellers’ Problem
We assume that ?rms do not know the marginal costs and hence the prices set by their
rivals, but instead know the marginal-cost distribution ?. Further, ?rms do not know the
search cost of any individual consumer, but they do know the distribution Q of search costs.
Taking as given the distributions of search costs and marginal costs, each ?rm determines
its optimal price based on the demand it faces, characterised by the reservation price rule
?(s) implied by (1).
Let us now consider the optimisation program of a ?rm with marginal cost draw c that
chooses to stay in the industry. We ?rst determine market share as a function of the price
p charged by the ?rm: x(p).
2
The optimal search rule implies that only consumers with
reservation prices ?(s) above p will buy from the ?rm. Take one such consumer with reser-
vation price r. Recalling that the price distribution in the market is given by the cdf F and
that the total mass of operating ?rms is L, the mass of ?rms charging a price less than r is
LF(r). The assumption of undirected search implies that this particular consumer is equally
likely to buy from any one of these ?rms. That is, the probability that she will buy from a
particular ?rm charging price p is 1/(LF(r)). Integrating over all such potential customers
of this ?rm yields an expression for market share:
x(p) =
_
?
p
g(r)
LF(r)
dr, (3)
where g(r) is the pdf of the reservation price. We can use (2) to write the corresponding cdf
as
G(r) = Q[?
?1
(r)] = Q
__
r
0
F(p) dp
_
. (4)
2
We use the market share interchangeably with the quantity because there is no outside good, each
consumer demands one unit of the good, and the total mass of consumers equals one.
6
Taking the derivative of G(r) with respect to r, we ?nd g(r) as
g(r) = q
_
?
?1
(r)
¸
F(r). (5)
We use the reservation price distribution to simplify the integral for market share. Inserting
(5) into (3) gives
x(p) =
1
L
_
?
p
q
_
?
?1
(r)
¸
dr. (6)
This equation is a standard (residual) demand curve: a ?rm faces demand determined by
its own price as well as its competitors’ prices. Here, these prices are embodied in the
distribution F(p). Note that demand is downward sloping, since
x
?
(p) = ?
1
L
q
_
?
?1
(p)
¸
< 0.
The pro?t function of a ?rm with marginal cost c choosing to stay in the industry can
be expressed as the solution to the ?rm’s optimisation program:
?(c) = max
p
(p ?c)x(p) ??. (7)
The values of p that maximise this equation for given values of c will de?ne the equilibrium
pricing function p(c). The ?rst-order condition for an optimum requires that, for all c,
[p(c) ?c]x
?
[p(c)] + x[p(c)] = 0, (8)
while the second-order condition for a maximum at this point stipulates that
[p(c) ?c]x
??
[p(c)] + 2x
?
[p(c)] < 0. (9)
7
1.4. Industry Equilibrium
Let p(·) and x(·) be, respectively, the pricing and residual demand functions in equilibrium.
Note that this implies that p(·) is optimal for each ?rm, given x(·), and therefore the ?rst and
second order conditions for individual optimality, (8) and (9), must hold at each point. The
downward-sloping demand then yields three important properties of the industry equilibrium.
Property 1. The equilibrium pricing function p(c) is increasing with marginal cost:
p
?
(c) > 0 (?c).
Proof. Applying the Implicit Function Theorem to the ?rst-order condition (8) yields
p
?
(c) =
x
?
(p)
[p(c) ?c]x
??
[p(c)] + 2x
?
[p(c)]
> 0,
since demand slopes downward and the denominator is negative by the second-order condi-
tion.
Property 2. The demand function x(p(c)) is decreasing with marginal cost:
dx
dc
p(c) < 0
(?c).
Proof.
dx
dc
p(c) = x
?
[p(c)]p
?
(c) < 0
by downward-sloping demand and Property 1.
Property 3. The pro?t function is decreasing with marginal cost: ?
?
(c) < 0 (?c).
Proof. Applying the Envelope Theorem to (7) yields ?
?
(c) = ?x[p(c)] < 0.
Note that Property 3 implies that the ?rms’ decision rule for staying in the industry or
leaving is characterised by a cut-o? value: there exists a threshold ¯ c > 0 such that ?rms
stay in the industry if and only if their marginal cost is c ? ¯ c (we assume here that the exit
8
decision is non-trivial, that is, some ?rms do exit and some produce). The threshold value
is given by
0 = ?(¯ c) = [p(¯ c) ? ¯ c]x[p(¯ c)] ??. (10)
The initial stage involves ex-ante identical potential entrants deciding whether or not to
commence operations. We assume that there is unlimited entry into the industry: ?rms keep
entering until the expected value of post-entry pro?ts equals the sunk entry cost. That is,
? =
_
¯ c
0
?(c)?(c) dc =
_
¯ c
0
[p(c) ?c]x[p(c)]?(c) dc ??(¯ c)?. (11)
Note that this entry condition implies ex-ante zero pro?ts and ex-post positive pro?ts.
Finally, note that Property 1 implies that prices will be distributed with support [p, ¯ p],
where p = p(0) and ¯ p = p(¯ c), with the cdf (for q ? [p, ¯ p]) given by
F(q) = Pr{p(c) ? q | ?(c) ? 0} =
Pr{c ? p
?1
(q) &c ? ¯ c}
Pr {c ? ¯ c}
=
?[p
?1
(q)]
?(¯ c)
. (12)
Note that F(q) = 0 for q < p and F(q) = 1 for q > ¯ p. We are now ready to de?ne the
equilibrium in this industry.
Definition 1. A search equilibrium is a set {? : R
+
? R
+
, p : R
+
? R
+
, x : R
+
?
R
+
, F : R
+
? [0, 1], ¯ c > 0} satisfying equations (2), (6), (8), (10), (11), and (12), along
with inequality (9).
2. Comparative Statics
Our goal is to determine the e?ect of a decrease in search costs on the search equilibrium.
In particular, we are interested in how shifts in search costs a?ect the equilibrium price
distribution F, the operating cut-o? cost ¯ c and the total mass of ?rms L. To this end,
consider a family of search cost distributions Q(· | t), where higher t corresponds to higher
9
search costs in the sense of the monotone likelihood ratio property (MLRP).
3
First, let us consider the function p(c, F, t), which gives the best-response price for a
?rm with marginal cost c when the price distribution of all operating ?rms is F and the
search costs are Q(· | t). Examining the ?rm’s ?rst-order condition and applying the MLRP
condition, we obtain our ?rst comparative statics result.
Proposition 1. The best-response pricing function p(c, F, t) is increasing in t.
Proof. See Appendix B.
Thus, the optimal price charged by each ?rm is increasing in the search costs, holding
?xed other ?rms’ pricing and entry/exit decisions (which a?ect F). However, this by itself
does not guarantee that the equilibrium prices will increase with search costs. Therefore, we
must look for conditions on the search cost distribution that will guarantee that the equilibria
will move in the same direction as the individual response functions. To this end, we must
?rst make precise the notion of increasing price distributions. Following Rauh (forthcoming),
we adopt the following partial order ? on the set of distribution functions with support in
(0, ?): F ? F
?
i? F ?rst-order stochastically dominates F
?
(i.e., F(p) ? F
?
(p) for all p > 0).
We now ask for conditions on q that will ensure that the equilibrium distribution F will be
increasing in t (with respect to the partial order ?).
As explained in Appendix A, a natural su?cient condition for the equilibrium distri-
bution to be increasing in search costs is that the market be supermodular in the sense of
Rauh (forthcoming). Verifying this condition is not trivial in our model, however, since our
setting di?ers substantially from Rauh’s model due to the endogenous entry/exit decisions
of ?rms. Therefore, for the rest of our analysis, we will restrict our attention to the case
when the search cost distribution is uniform, where we can characterise equilibria explicitly.
Although we are able to obtain exact results only in the uniform search cost case, numerical
simulations show that the comparative statics under other search cost distributions (such as
3
That is, for each s
1
> s
0
, the ratio q(s
1
| t)/q(s
0
| t) is increasing in t.
10
the exponential distribution) tend to be very similar to those obtained under the uniform
distribution (see Appendix C).
2.1. Uniform Search Costs
Following the discussion in Appendix A, we focus on uniform search cost distributions:
Assumption 1. The search cost distribution is uniform on [0, a] for a > 0.
With this formulation, a decrease in search costs can be identi?ed with a decrease in the
parameter a. The marginal cost distribution, on the other hand, is allowed to take a very
general form, subject only to the weak condition of log-concave cdf, which is satis?ed by most
commonly used distributions (such as uniform, normal, log-normal, gamma, exponential,
Pareto, and others; see Bagnoli and Bergstrom (2005)):
Assumption 2. The cdf of the marginal cost distribution is log-concave, i.e., ?(c)/?(c)
is decreasing in c for all c.
Given Assumption 1, the demand function (6) for any p ? p simpli?es to
x(p) =
1
L
_
?
p
1
a
I
{?
?1
(r)?[0,a]}
dr =
1
aL
_
?
p
I
{r?[?(0),?(a)]}
dr =
1
aL
[?(a) ?p]. (13)
The second equality follows because ? is increasing. The ?nal equality holds because it is not
optimal for any ?rm to charge less than ?(0), so that p ? ?(0). Note that x
?
(p) = ?1/(aL) <
0 and x
??
(p) = 0, so that the second-order condition (9) holds. Plugging (13) into (8), the
?rst-order condition becomes
p(c) =
1
2
[?(a) + c], (14)
so that the demand and pro?t functions reduce to
x(c) =
1
2aL
[?(a) ?c] and (15)
?(c) =
1
4aL
[?(a) ?c]
2
, (16)
11
and the operating threshold equation (10) yields
¯ c = ?(a) ?2
?
aL?. (17)
The upper and lower limits of the support of the equilibrium price distribution are therefore
p = p(0) = ?(a)/2 and ¯ p = p(¯ c) = ? ?
?
aL?.
We now see that a search equilibrium is fully determined by two parameters, ¯ ? ? ?(a) > 0
and L, satisfying equations (2) (for s = a), (11), (12), and (14) through (17). Plugging (14)
and (12) into (2) for s = a yields
a =
1
?(¯ ? ?2
?
aL?)
_
¯ ??
?
aL?
¯ ?/2
?(2p ? ¯ ?) dp +
_
¯ ?
¯ ??
?
aL?
1 dp
=
1
2?(¯ ? ?2
?
aL?)
_
¯ ??2
?
aL?
0
?(c) dc +
?
aL?.
Finally, we insert equations (14) through (17) into the entry condition (11), reducing the
conditions for a search equilibrium to the following system of two equations in ¯ ? and L:
?(¯ ?, L; a) ?
1
2?(¯ ? ?2
?
aL?)
_
¯ ??2
?
aL?
0
?(c) dc +
?
aL? ?a = 0; (18)
?(¯ ?, L; a) ?
1
4aL
_
¯ ??2
?
aL?
0
(¯ ? ?c)
2
?(c) dc ??(¯ ? ?2
?
aL?)? = ?. (19)
Manipulating the ?rst equation (details in Appendix B) shows that either the mass of ?rms
or the reservation threshold of the consumer with the highest search costs (or both) must
increase as the search cost distribution shifts right.
Lemma 1. At least one of the quantities ¯ ? and L must be increasing in a:
?L
?a
? 0 ?
?¯ ?
?a
> 0.
The proof, which is provided in Appendix B, amounts to showing that if both L and ¯ ?
12
were non-increasing in a, the left-hand side of (18) would be decreasing in a, which would
violate the identity. The logic of this result is straightforward: a decrease in search costs (a
lower a), if not accompanied by a decrease in search opportunities (a lower L), will result in
increased marginal bene?t of continued search, which will cause searchers to become more
selective, thus decreasing ¯ ?.
In a similar manner, equation (19) implies that if the mass of ?rms decreases as the search
cost distribution shifts left, the reservation threshold of the consumer with the highest search
costs must also decrease.
Lemma 2. If L is increasing in a, so is ¯ ?:
?L
?a
> 0 ?
?¯ ?
?a
> 0.
The proof, shown in Appendix B, consists of demonstrating that the contrary statement
would cause the left-hand side of (19) to be decreasing in a, violating that identity. Whereas
Lemma 1 relied on the consumer side, Lemma 2 relies on the producer side: the intuition is
that an increase in competition (higher L) must be accompanied by a compensating increase
in searchers’ reservation prices (thus increasing ?rms’ expected pro?ts per transaction) in
order for average pro?ts to stay constant.
The results of Lemma 1 and 2 imply that ¯ ? must be strictly increasing in a:
?¯ ?
?a
> 0.
Together with the pricing equation (14), this gives us our ?rst key result:
Proposition 2. When search costs decrease, the price p(c) charged by a ?rm with
marginal cost c decreases for any operating ?rm.
Our next objective is to determine the e?ect of a change in a on the operating cut-o?
value ¯ c and on the level of concentration in the market. It will be convenient to ?rst de?ne
13
the quantity
?(a) ?
1
aL(a)
,
where we write L(a) to emphasise its dependence on a. Note that this can be interpreted as
the per-?rm density of consumers with a given level of search costs, since the total number
of ?rms is L(a) and the density of consumers with any level of search cost s is simply 1/a.
It is easy to see that ?(a) is decreasing in a:
Lemma 3. The per-?rm density of consumers with any given level of search costs is
decreasing in a: ?
?
(a) < 0.
The proof of this result is straightforward (see Appendix B): since ¯ ? is increasing with
a, ? needs to decrease with a in order to preserve equality in (19).
The pro?t function of a ?rm with marginal cost c now becomes:
?(c; a) =
1
4aL(a)
[ ¯ ?(a) ?c]
2
=
1
4
?(a)[ ¯ ?(a) ?c]
2
, (20)
where we have written L(a) and ¯ ?(a) to emphasise the dependence of these parameters on
a. Taking the derivative of this expression with respect to a and applying Lemma 3 (details
in Appendix B), we can now easily make our next observation: if an increase of search costs
hurts any currently operating ?rm, it must also hurt all ?rms with lower search costs:
Lemma 4. If there exists c
0
? ¯ c(a) such that ?
a
(c
0
; a) ? 0, then ?
a
(c; a) < 0 for all
c < c
0
.
The intuition for this result is again quite simple. The only negative e?ect on a ?rm of
increasing a and thus increasing ¯ ?(a) is that the ?rm now has to share its current customer
base with more higher-cost ?rms. This e?ect becomes larger and larger, as the marginal cost
of the ?rm decreases. (Note, for example, that the ?rm with marginal cost ¯ c(a) was already
sharing all of its consumers with all operating ?rms, so that the only additional sharing
14
comes from the additional ?rms that were not operating before, whereas the zero-cost ?rm
now needs to share each of its customers with more of the ?rms that were operating before.)
It now becomes clear that the pro?t of the ?rm at the current marginal cost cut-o?
level ¯ c(a) must decrease as search costs decrease. If this were not the case, the pro?ts of
all currently operating ?rms would increase, which would result in an overall increase of
ex-ante expected pro?ts. This would violate the entry condition (19), which states that the
ex-ante expected pro?ts must remain constant at ?. Since the pro?t function (for each a)
is strictly decreasing in c, the fact that the pro?t of the current cut-o?-level ?rm falls below
the operating threshold ? implies that the new cut-o? level will be lower than the current
level. Formalising these arguments (Appendix B), we obtain our second key result:
Proposition 3. When search costs decrease, so does the cut-o? marginal cost, ¯ c.
Proposition 3 has the immediate empirically testable implication that some of the ?rms
with the highest marginal costs of production will exit the industry in response to a decrease
in consumers’ search costs.
Propositions 2 and 3 together yield two more testable implications: both the prices
charged in equilibrium and the marginal costs of operating ?rms will decrease, as search
costs decrease (formal details in Appendix B):
Corollary 1. When search costs decrease, the distributions of equilibrium prices and
marginal costs of operating ?rms shift to the left in the sense of ?rst-order stochastic domi-
nance.
Thus, search cost decreases lead to increased e?ciency of operating ?rms and to lower
prices for consumers. As Proposition 3 shows, this increased e?ciency is due to the fact that
the lowering of consumer search costs diminishes the pro?ts of ine?cient (high-marginal
cost) ?rms, causing some of these ?rms to exit the industry. It is easy to see, however, that
the more e?cient ?rms will actually bene?t from a search cost reduction. If a decrease in
¯ c (and thus a reduced likelihood of staying in the market) were accompanied by decreased
15
pro?ts of all operating ?rms, the ex-ante expected pro?ts would decrease, violating the entry
condition that says that those are constant and equal to the cost of entry.
Corollary 2. A decrease in search costs causes the pro?ts of ?rms with su?ciently low
marginal costs to increase: for each a, there exists ˆ c(a) < ¯ c(a) such that ?
a
(c; a) < 0 for all
c < ˆ c(a).
Similar reasoning leads to the conclusion that the total market share of low-cost ?rms
should increase in response to decreasing consumer search costs, as the share of high-cost
?rms decreases. To state this formally, let us denote the total market share of all ?rms with
marginal cost in (c, c + dc) (for in?nitesimal dc) by X(c; a) dc. Then, for each c,
X(c, a) dc = Lx(c; a)?(c) dc.
Applying similar arguments to those we used for determining the change in pro?ts, we can
readily obtain the following result (see Appendix B for details):
Corollary 3. A decrease in search costs causes the total market share of all ?rms with
su?ciently low marginal costs to increase: for each a, there exists ˆ c(a) < ¯ c(a) such that
X
a
(c; a) < 0 for all c < ˆ c(a).
The results of Proposition 3 and Corollaries 2 and 3 establish the main empirical hy-
pothesis of our model: search cost declines driven by the advent and di?usion of e-commerce
have di?ering e?ects across businesses in an industry. Low-type (high-cost) sellers are hurt,
sometimes to the point of being forced to exit. Higher types (low-cost sellers), however,
actually gain from the shift: the market share of low-cost ?rms grows, resulting in increas-
ing concentration of the market. Finally, it appears to be impossible to sign the change in
the total mass of ?rms analytically, but numerical simulations with a variety of marginal
cost distributions suggest that the mass of ?rms may decrease when search costs decrease
(Appendix C).
16
3. Data
Our empirical analysis uses data from two primary sources: industry employment and es-
tablishment counts from the US Census Bureau’s County Business Patterns (CBP), and US
consumers’ online purchasing behaviour from Forrester Research Technographics surveys.
We brie?y describe these data sets here, as well as discuss our market de?nition.
3.1. County Business Patterns
Annual County Business Patterns data contain, by detailed industry, the number of es-
tablishments in each US county. Establishments are unique geographic locations where
economic activity takes place (i.e., o?ces in the travel agency industry, storefronts in the
bookstore industry, and car lots in the auto dealerships industry). A ?rm can own one or
more establishments.
4
Both the total number of establishments and establishment counts
by employment range are included in the data.
5
In cases where disclosure of con?dential
information is not an issue, total industry employment and payroll in the county are also
reported. However, these are often missing in the industries we study, particularly in smaller
counties served by only a handful of ?rms. We can, however, impute total employment by
multiplying the establishment counts in an employment range category by an estimate of
the average number of employees per establishment in the category. We use the simple av-
erage of the categories’ endpoints for this estimate. While imputations invariably introduce
measurement error, we are reassured by the fact that the correlation between imputed and
actual reported employment for those counties where the latter is available is quite high.
4
While it would be very interesting to study the issues at hand in the context of within- and across-?rm
shifts, there is unfortunately no way to identify ?rms in the CBP data. ‘Firms’ in the model above can be
interpreted here as distinct operations (o?ces, storefronts, or lots) in an industry. While it is possible that
common ownership may a?ect individual establishments’ reactions to the shift to e-commerce, we think that
the model’s basic implications about the relative impacts on low- versus high-type producers continue to
hold to a large extent even within multiple-establishment ?rms. For example, all else equal, a ?rm seeking
to reduce its size will tend to close its low-type operations ?rst.
5
The reported ranges are: 1–4 employees, 5–9, 10–19, 20–49, 50–99, 100–249, 250–499, 500–999, and over
1000 employees. Since very large establishments are relatively uncommon in the industries we study here,
we aggregate the largest categories into a single category.
17
Further, most of the empirical work below focuses on establishment counts, which we never
have to impute.
We use data spanning 1994 to 2003, which surrounds the period when the advent of
browser software began the Internet’s di?usion into the broader population. It is also the
time span for which CBP data are available with the level of industry detail necessary for
our purposes here. We focus on three industries: travel agencies (SIC 4724/NAICS 561510),
bookstores (SIC 5942, NAICS 451211), and new auto dealers (SIC 5510/NAICS 441110).
While a major change in the industry classi?cation scheme occurred in 1997 (from the SIC
system to the NAICS taxonomy), these industries’ boundaries remained una?ected, so values
before and after the change are comparable.
3.2. Household Internet Use
The data on households’ e-commerce activity comes from Forrester Research, a market re-
search company with a program focusing on consumers’ technology use. Its annual Techno-
graphics survey is designed to be nationally representative and includes the responses of
roughly 55,000 people living in the continental US.
6
We have access to the 2003 and 2004 surveys. Survey responses re?ect behaviour in the
year previous to the title year, because the survey is typically administered from prior-year
December through title-year January. For example, when the 2004 survey asks respondents
about their behaviour over the past year, the answers re?ect actions taken in 2003.
While the survey is primarily cross-sectional, conveniently for us there is a retrospective
question asking when the respondent ‘start[ed] purchasing products or services online.’ The
respondent can choose one of several time ranges: ‘less than 1 year ago’, ‘1 year to less than
2 years ago’, and so on up to ‘8 years ago or more’. We construct from these responses the
fraction of market consumers that had started purchasing products or services online for
each year from 1994 through 2003.
7
6
See Goolsbee (2000) for additional details about the survey.
7
We used the 2003 survey to compute the fraction of online shoppers in 1994 and 1995, and the 2004
18
3.3. Market De?nition
We de?ne markets using the US Bureau of Economic Analysis’ Component Economic Areas
(CEAs). CEAs are collections of counties usually, but not always, centred on Metropolitan
Statistical Areas (MSAs). Counties are selected for inclusion in a given CEA based upon
their MSA status, commuting ?ows, and newspaper circulation patterns, subject to the
condition that each CEA’s counties are contiguous. CEA boundaries need not coincide with
state boundaries. The selection criteria ensure that counties in a given CEA are economically
intertwined. The roughly 3200 US counties are grouped this way into 348 markets that are
mutually exclusive and exhaustive of the land mass of the United States. Since our Internet
use data excludes Alaska and Hawaii, our empirical analysis uses data for the 345 CEAs in
the continental US.
8
Using CEAs o?ers a compromise between con?icting requirements of the analysis. The
most constraining observation is that, with an Internet use sample of 55,000, using smaller
market areas (like counties) would result in many markets having very thin samples. We
use the county indicator in the Technographics survey to aggregate the respondents to the
CEA level. This reduces the sampling error involved, though of course with the trade-o?
of losing some variation in market structures. Further, counties may in some cases be too
small to accurately capture market areas in the industries we investigate. This is especially
true in more rural areas, where cross-county commerce in travel agency, book sales, and auto
purchases is likely to be commonplace. CEAs should be large enough to envelop businesses’
survey to compute the fractions from 1996 to 2003. The use of two surveys was necessary because the ‘8 years
ago or more’ responses in the 2004 survey correspond to any purchases occurring before 1996, not necessarily
those in 1995 exclusively. We do see 1995 purchase patterns, however, in the 2003 survey (through the ‘7
years to less than 8 years ago’ responses). We are still left with online activity in 1994 being measured with
‘8 years ago or more’ responses from the earlier survey. However, given the small fractions of respondents
reporting buying products online in 1995 (see below), as well as the fact that the Internet’s commercial
structure at that time was quite embryonic, it is unlikely that many of the purchases attributed to 1994
actually occurred before that year. The use of two separate surveys over the observation period does not
seem to have created spurious increases in reported online purchases. There is no discernible trend break
between 1995 and 1996, the surveys’ point of contact.
8
See US Bureau of Economic Analysis (1995) for more detailed information about creation of CEAs and
the super-regions that they comprise, Economic Areas.
19
catchment areas in most cases.
9
To give an idea of the size of markets in our data, Table 1 presents summary statistics
of within-CEA establishment counts in our industries. In order to highlight across-market
di?erences, we ?rst take the within-market average establishment counts over our sample pe-
riod, and then report quantiles of the cross-sectional distribution of these averages. The table
shows quantiles for the total number of establishments as well as for each of the employment
size categories. We note, however, that our empirical speci?cations below include market
?xed e?ects, so that the estimated relationships between market structure and consumers’
online shopping behaviour re?ect within-market variation over time.
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
TABLE 1 ABOUT HERE
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
4. Empirical Tests
We seek to test the model’s implications regarding how a shift in the consumer search
cost distribution impacts industry market structure, particularly with regard to the relative
fortunes of high- and low-type businesses. Our focus, as mentioned previously, is on industries
where a shift in consumer activity to e-commerce channels has been cited as having a noted
impact on industry businesses. While these industries are in many ways suitable for our
analysis, they are not perfect matches to the stylised industry in the model. We do ?nd it
entirely plausible, as the model assumes, that there are signi?cant and persistent di?erences
in producers’ types in these industries. The most relevant type dimension in these industries
is, it seems to us, the per-dollar cost to industry businesses of delivering a bundle of goods
and services at a given quality level.
An important dimension where reality and the model depart, however, is with respect to
9
Since our consumer e-commerce use measure is built from responses of a ?xed set of consumers to a
retrospective question, we must also assume that any across-CEA population movements over our sample
period are unrelated to local growth in e-commerce infrastructure.
20
horizontal product di?erentiation. We do not model this above, but it almost surely exists
to some extent in each industry we study. Horizontal product di?erentiation may dampen
the quantitative impact of the substitutability-enhancing (via reduced search cost) features
of e-commerce. (Researchers have noted e?orts along these lines among booksellers; see Clay
et al. (2002), for example.) To the extent that any changes did occur, our estimates o?er
guidance as to the magnitude of e-commerce’s impact net of product di?erentiation shifts.
Another potential point of departure between our model and our analysis is that in two
of the industries, travel agencies and bookstores, the di?usion of the Internet has allowed
the entry of online-only retailers. As in Latcovich and Smith (2001), these businesses have
di?erent cost structures than traditional ‘brick-and-mortar’ retailers, in that they may have
higher ?xed costs, but lower marginal costs. Moreover, such Internet-only retailers arguably
provide a di?erent bundle of goods and services, in that customers cannot inspect the good
?rst-hand and must wait for it to be shipped. By assuming uniform ?xed costs and homo-
geneous products, our model does not explicitly account for the creation of Internet-only
retailers, focusing rather on how brick-and-mortar retailer demand might change in response
to a reduction in consumer search costs brought about by the Internet. An advantage of in-
vestigating new car dealers, however, is that regulations prevent similar ‘online-only’ entrants
in this industry, making it a close match to our theoretical model.
Yet another dimension we do not model is the endogeneity of certain ?xed costs, such as
advertising, which can lead to industry dominance patterns, as in Sutton (1991). Latcovich
and Smith (2001) document high level of advertising expenditure among online booksellers.
If consumers are not fully informed about the quality of their retail service, and if advertising
can signal vertical characteristics such as reliability, security, and ease of use, ?rms advertise
heavily to increase consumers’ willingness to pay. Just as with search costs, horizontal or
vertical di?erentiation decreases consumers’ abilities to substitute across industry producers.
21
4.1. Travel Agencies
Much has been made of the demise of the travel agent as consumers shifted their travel
purchases to e-commerce sites like travel search engines (e.g., Orbitz or Expedia) or to travel
service providers themselves (especially by buying tickets directly from airlines’ websites).
Aggregate statistics leave little doubt that the di?usion of the Internet coincided with
considerable establishment exit in the travel agency industry. Figure 1 plots two time series:
the total number of industry establishments, and the fraction of Technographics survey
respondents reporting that they had ?rst purchased products or services online by a given
year. The number of travel agency establishments was fairly steady, slightly rising in fact,
until 1997, at which time it began to fall substantially. The number of establishments
in the industry dropped by over 35% between 1997 and 2003. As can be seen, this exit
coincided with a post-1997 acceleration in the fraction of surveyed consumers reporting
online purchases.
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
FIGURE 1 ABOUT HERE
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
This broad exit pattern was concentrated among the industry’s smaller operations. Panel A
of Table 2 contains establishment counts by establishment size category (size is measured by
number of employees).
10
Over the sample period, establishment counts fell in the four small-
est employment categories, those including businesses with fewer than 50 employees. The
drop was especially precipitous among establishments with fewer than 10 employees. At the
same time, though, the number of establishments with 50 or more employees actually rose.
The number of operations with 100 or more employees grew 70%. The vicious shake-out at
the low end was therefore accompanied by growth among the largest industry businesses.
11
10
The US aggregate numbers in Table 2 include a few establishments not in the market-level data we use
below, since the aggregate numbers include establishments in Alaska and Hawaii as well as those not placed
into a speci?c county within a state (this latter group is referred to as ‘statewide’ establishments in the
CBP).
11
The CBP data does not allow one to track individual establishments through time. It is therefore
22
These patterns are consistent with those predicted by the model. A decline in search costs,
made possible through the di?usion of the Internet and the advent and improvement of
travel-shopping websites, shifted equilibrium production to the larger, higher-type produc-
ers in the industry. Indeed, some of these high-type producers may host the very portals
that led to the decline of their smaller competitors.
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
TABLE 2 ABOUT HERE
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
To show the connection more formally, we regress the (logged) number of industry em-
ployees and establishments in a CEA market-year observation on the fraction of people in the
market who reported making purchases online by that year. Because Internet use di?used
sooner into certain markets with high demand for travel services (e.g., New York and San
Francisco), but for reasons likely unrelated to its use for purchasing those services, there
is an underlying positive correlation across markets in the number of travel agencies and
the fraction of consumers using the Internet. If we did not control for these di?erences, we
would spuriously conclude that greater Internet use led to increases in travel agency numbers.
We therefore include CEA ?xed e?ects in this and all of our empirical speci?cations. The
estimates thus re?ect the relationship between changes in online purchase frequencies and
industry activity within CEA markets. We also control for employment across all industries
in the market-year (also taken from the CBP data) to account for the in?uence of overall
market growth or decline of the industry.
The results, reported in Panel B of Table 2, re?ect the aggregate patterns above. Higher
fractions of consumers buying goods and services online are associated with declines in the
conceptually possible that even a growing industry could exhibit net establishment losses at lower employment
ranges due to formerly small businesses growing into larger size categories. However, this scenario would
imply that the total number of establishments in the industry remained roughly unchanged. This is clearly
not the case here. One possibility that cannot be ruled out, however, is that many small establishments were
merged into larger ones. This would shrink establishment counts both at the low end of the distribution and
in total. To the extent mergers played a role, though, we show shortly that the employment growth among
large establishments did not fully make up for employment losses among the industry’s small operators.
23
numbers of industry employees and establishments in the market. The estimated impact
of consumers’ e-commerce activity is quite substantial for the smallest establishments. For
example, a 15 percentage point increase in the fraction of consumers making purchases online
– a one standard deviation change – corresponds to a 13% (21%) drop in establishments with
1–4 employees (5–9 employees). Notice, however, that this negative impact lessens as one
works up the establishment size distribution. Indeed, it eventually becomes insigni?cant
with positive point estimates for establishments with 50–99 employees and those with 100
employees or more.
12
Greater e-commerce activity among consumers is therefore associated with losses among
the smallest industry producers, but a positive in?uence on the largest producers. Despite
the inclusion of market ?xed e?ects, however, the test above does not answer the question
of whether the market structure impact of the shift to e-commerce acts locally or instead
more broadly. It could be that the many within-market changes re?ect aggregate shifts, and
while the overall increase in Internet purchasing behaviour shifts industry market shares in
the direction predicted by the model, there is no sense in which this impact is noticeably
stronger in markets that saw larger increases in consumers’ Internet use than in those that
experienced smaller gains. To answer the question of the geographic scope of e-commerce’s
impact in the industry, we add a set of year dummies to the regression. This removes
the impact of aggregate shifts in Internet use, leaving only the idiosyncratic within-market
12
The di?erent sample sizes across establishment size categories re?ect the fact that not all market-
year observations have a positive number of establishments in a particular category. The small number
of large establishments in the industry makes the sample for the largest size categories particularly small
and may in part explain the imprecise results in these cases. To explore this issue further, we estimated
an alternative speci?cation for the 50–99 and 100+ employee size categories where, rather than using the
logged number of establishments as the dependent variable, we used a dummy equal to one if there was at
least one establishment in the size category in a market-year and zero otherwise. (The numbers in Table 1
indicate most of the observations where this dummy equalled one correspond to the presence of only one
establishment.) In this case, all market-year observations can be included in the sample. This alternative
speci?cation also indicated a positive correlation between consumers using online commerce channels and
growth among large establishments, but in this case the relationship was statistically stronger (signi?cant
at the 10% level for establishments with 50–99 employees and at the 5% level for those with more than
100.) The results in the ?rst numerical column indicate that any employment gains in the larger size classes
are swamped by employment losses due to the exit of smaller operations. Overall market employment, not
shown here, enters positively and signi?cantly in most of the speci?cations, as one might expect.
24
variation in the growth of online purchasing patterns and establishment counts to identify
the coe?cient. In essence, this regression tests if markets that had unusually high increases
in Internet use also saw larger-than-average declines in small-establishment counts.
13
The regression results (with year dummy coe?cients not reported for parsimony) are
in Table 2, Panel C. In this case all coe?cients on the measure of consumers’ e-commerce
activity are statistically insigni?cant. There is no measurable market-speci?c in?uence of
online purchases on local travel agencies. This indicates, very interestingly, that the shifts
in industry market structure seen above, while coincident with consumers’ increasing use of
online sites to conduct their travel purchases, did not arise from a set of coordinated market
structure shifts in speci?c markets that produced the observed patterns once aggregated up.
Instead, the in?uence of Internet use on market structure in the industry is a completely
aggregate phenomenon.
A consideration of the speci?c way e-commerce impacted this industry o?ers a likely
explanation for this result. As Internet purchases of airline tickets became more common
over our observation period, airlines incrementally decreased the commissions they paid to
travel agents. The ?rst, modest commission cut (imposing a $50 cap per domestic ticket,
which given the standard 10% rate at the time meant it was only binding for tickets above
$500) occurred in 1995.
14
This ended up being only the ?rst cut of a series, however. By 2002,
major carriers had ceased paying commissions altogether. Since airline tickets accounted for
an estimated 58% of travel agencies’ revenues in 1996, these commission declines resulted in
a serious income loss for the industry (some lost commissions were replaced by fees charged
13
Speci?cally, the coe?cient on the fraction of consumers in the market shopping online is identi?ed
from the correlation between two values: a market’s growth rate in the number of industry establishments
relative to the average across all markets in that year, and that same market’s change in the fraction
of consumers reporting shopping online relative to the across-market average. That is, the coe?cient is
negative if markets with larger-than-average declines in establishment counts saw higher-than average growth
in Internet purchases.
14
The facts on travel agent commissions discussed in this paragraph are from a 2002 report by the National
Commission to Ensure Consumer Information and Choice in the Airline Industry (NCECICAI). The creation
of the NCECICAI was a provision of the Aviation Investment and Reform Act for the 21st Century. The
commission’s congressionally mandated mission was to study the travel agent industry and, more generally,
the airline services information available to consumers.
25
directly to the consumer, though these did not cover the losses). Small operations, having
high ?xed costs relative to their sales volume, found pro?tability increasingly di?cult to
obtain and began to exit, as seen in the data. Importantly, however, airlines cut commissions
across-the-board nationwide – presumably in response to perceived changes in consumers’
aggregate ticket purchasing patterns – rather than market-by-market. We are aware of no
instances where airlines selectively reduced commissions more in those particular markets
where online purchases were growing fastest. This would explain why the connection between
Internet use and market structure changes is starkly evident in aggregate changes over time
but not so across markets within a period. It is also consistent with the fact that any growth
among the largest establishments was uncorrelated with local Internet use, because many
of these establishments plausibly tapped into the new (and national) Internet market, and
drew their business growth largely from customers outside their local area.
One potential concern in the regressions with year ?xed e?ects is that the results might
be driven by changes in markets’ socio-economic composition over time, rather than by
a change of online shopping habits. Unfortunately, it is not possible to fully control for
market demographics, because detailed demographic information is not available on a yearly
basis. However, we were able to control for one key demographic: age. We found that
the age variable did on occasion enter the regressions signi?cantly, but that there was little
noticeable change in the key coe?cient estimates on the fraction of online purchasers in a
market, con?rming that the observed correlation between establishment size distribution and
the market’s fraction of online consumers cannot be explained away by a shift of consumer
age distribution.
4.2. Bookstores
Another line of business that has by many accounts in the popular press been a?ected by
the di?usion of Internet commerce is the retail bookstore industry. Several booksellers have
blamed their demise in large part on the competitive demands of e-commerce (Herman, 2001;
26
Weisman, 2004; Melo, 2005). The process through which this competitive e?ect would take
place is again that which is highlighted in our model: e-commerce induced reductions in
consumers’ search costs shift market share across the industry type distribution.
We investigate this possibility by repeating the empirical analyses above, this time us-
ing CBP data for the bookstores (SIC 5942/NAICS 451211) industry. We begin with the
industry-wide establishment counts shown in Panel A of Table 3. They re?ect similar pat-
terns to those seen with the travel agency aggregates: declines in establishments in the
smaller employment size categories with coincident expansion in the larger categories. For
instance, while the number of bookstores with fewer than 20 employees fell by over one-
fourth during the sample, those with more than 20 employees more than doubled. This
growth was particularly pronounced among the 50–99 employee size category. So we again
see the pattern of market share shifts from small (low-type) operations to large (high-type)
ones.
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
TABLE 3 ABOUT HERE
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
Again the question arises of whether these e?ects re?ect aggregate impacts or instead
coincide with local Internet commerce patterns. No obvious analogy exists in the bookstores
industry to the airlines’ commission reductions and their impact on travel agencies. Therefore
one might expect the impact of the Internet here to be more concentrated within particular
markets. If this is the case, the overall shift from smaller to larger bookstores noted above
re?ects aggregated changes that occurred market-by-market.
We investigate this issue by estimating the above speci?cation that includes year ?xed
e?ects, this time using bookstores CBP data. The results are reported in Panel B of Table 3.
Again we have suppressed the estimated year e?ects and the coe?cients on overall market
employment.
In contrast to the market structure shifts in the travel agency industry, there is more
27
evidence that local market e?ects matter in bookstores. Markets seeing faster growth in
local consumers making online purchases had greater declines in bookstore employment
and the total number of bookstores, with establishment exit being driven by losses among
operations having fewer than 20 employees. This increased exit was statistically signi?cant,
except for establishments with fewer than ?ve employees.
There is weaker evidence, on the other hand, that local online purchasing behaviour im-
pacted the growth seen among larger booksellers. None of the e-commerce activity (‘fraction
online’) coe?cients for the three largest size categories, while re?ecting the positive co-
movement between online shopping and the numbers of larger bookstores, are statistically
signi?cant. This is likely due to the fact that the industry classi?cation system includes an
industry separate from bookstores, ‘Electronic Shopping and Mail-Order Houses’ (NAICS
45411), into which the largest online booksellers are classi?ed.
15
The expansion seen in
large bookstores may instead re?ect the ascendance of the new-format large-store chains like
Barnes and Noble and Borders. Their growth is not strongly correlated with local online
shopping habits because, while these sellers have extensive online operations (Barnes and
Noble has its own website and Borders has teamed with Amazon), their online operations
have industrial classi?cations that are separate in the CBP data from their brick-and-mortar
locations.
4.3. New Auto Dealers
The last industry in which we investigate the impact of e-commerce on di?erent producer
types is new auto dealers. The new auto dealers industry has special appeal as a forum
for testing our model. Speci?cally, franchise law restrictions make it extremely di?cult to
15
Note that online airline ticket sales operations are not included in this industry. According to the US
Census Bureau, businesses in NAICS 45411 sold $4.16 billion of books and magazines in 2003, $2.14 billion
of which was exchanged via ‘e-commerce’ channels (these are de?ned as transactions over open networks
like the Internet or proprietary networks running systems like Electronic Data Interchange). These book
and magazine sales accounted for 3.2% and 5.3% of the industry’s total and e-commerce product sales,
respectively. See US Census Bureau (2005) for details.
28
operate Internet-only sales channels.
16
This means that e-commerce in this industry functions in a way that almost exactly
matches how we embody it in the model; that is, it is purely a demand-side device that lowers
consumers’ costs of gathering product information. The essential technology of production
and delivery in the industry is unchanged, even among any new producers that might enter
the market after the change in search costs. There are no issues of retailers selling the
industry’s product but not being counted in the industry’s CBP data (as with Amazon or
WalMart with regard to the bookstores industry, for instance), and consumers cannot use
e-commerce channels to bypass retailers altogether and buy from manufacturers directly (as
is the case for many airline ticket purchases). Thus new car dealers allow us to see quite
directly how reductions in search costs impact an essentially isolated set of retailers whose
basic distribution technology is not impacted by e-commerce. The only change they face is
in how easy it is for consumers to ?nd out about their products or to be matched to low-price
dealerships.
There is anecdotal evidence that e-commerce channels have increased the number of
dealers from whom they obtain quotes before purchasing (Gartner, 2004). Furthermore,
Scott Morton, Zettelmeyer, and Silva-Risso examine in a set of papers the relationship
between buyers’ use of e-commerce channels and the (lower) ?nal prices they pay for their
cars.
17
However, we are unaware of any attempt to formally analyse what this reduction in
search costs leads to in terms of the market structure of auto dealerships. The model o?ers
guidance as to the likely mechanism and its impact; namely, that declining search costs led to
shrinking and exit among the low-type dealers and shifted market share to the highest-type
operations.
16
See, for example, Katz and Payne (2000) and Scott Morton et al. (2001). Auto manufacturers are
prohibited from selling their cars directly. Even online buying services like autobytel.com and carsdirect.com
do not sell their own inventories of cars to their customers. Instead, they act as referral services, matching
customers to their a?liated physical dealers.
17
Scott Morton, Zettelmeyer, and Silva-Risso (2001) show that car shoppers using Autobytel.com to get
free quotes from dealers in their market end up paying lower prices. Zettelmeyer, Morton, and Silva-Risso
(2005) provide evidence that the lower prices obtained by consumers utilising online resources is not solely
due to a selection bias in which hard-bargaining or low-search-cost customers choose to use the Internet.
29
Panel A of Table 4 shows changes in the number of new auto dealer establishments by
size over the sample period. Unlike for travel agencies and bookstores, the total number of
establishments did not decline. In fact, the number rose slightly. Some of this gain came
from growth in the number of establishments with less than ten employees. It is not clear
what types of operations these are, particularly those with 1–4 employees, which is quite
small even for ‘standard’ dealerships in isolated rural settings.
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
TABLE 4 ABOUT HERE
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
Excepting these, however, the remainder of the establishment size types exhibit the
patterns seen before. (Note also from Panel C of Table 1 that, unlike travel agencies and
bookstores, the bulk of the industry’s establishments are not concentrated in the smallest
employment categories. Over two-thirds have between 10 and 100 employees.) There were
drops in the number of dealerships in the 10–19 and 20–49 employee categories – just under
20% in the former case and 10% in the latter – but growth in the number of larger-sized
establishments. Moreover, the growth rate in the establishment counts increases with the
size category. (Because the number of auto dealers with 100–249 employees is so much
larger than in the travel agency and bookstores industry, we have included this as a separate
category in our analysis here and aggregated establishments with 250 or more employees
together.)
The regression results are shown in Panel B of Table 4. Again only the speci?cation
with year ?xed e?ects is shown. The changes in the aggregate establishment counts just
discussed are in fact related to local consumers’ use of the Internet to make purchases. While
the coe?cients on fraction online are insigni?cant for the three smallest establishment size
categories, there is a signi?cant negative impact of local online purchasing on the number of
dealerships with 20–49 employees in the market. Its economic size strikes us as non-trivial;
a one-standard-deviation increase in online shopping corresponds to a 3.5% drop in the
30
establishment count. At the same time, increasing e-commerce activity drives growth in the
number of local dealerships with 50 or more employees. The result is not signi?cant for the
largest establishment size category (though the coe?cient is large and positive), likely due to
the small number of market-years in our sample with very large dealerships. If we estimate
as above an alternative speci?cation using an indicator for market-years where at least one
such establishment exists, however, we do ?nd a signi?cant positive impact of e-commerce
on the presence of very large auto dealers.
Interestingly, the ?rst two columns of Panel B suggest that e-commerce has had an
expansionary e?ect on the industry overall. Markets where Internet purchasing grew more
than average saw higher increases in auto dealer establishment counts and employment (recall
that this is controlling for overall employment changes in the market). This is of course
opposite to what is seen in the travel agency and bookstores industries. This likely re?ects
the fact discussed above that e-commerce did not facilitate growth in new auto sales through
channels external to the industry, unlike what happened for travel agents and booksellers.
Thus here, industry producers overall were able to bene?t from market expansions driven by
reductions in consumers’ search costs, rather than losing part or all of the expanded market
to sellers outside the industry.
5. Conclusions
This paper has investigated the equilibrium market structure changes spurred by the in-
troduction of e-commerce tools that reduce consumers’ search costs. We speci?ed a gen-
eral industry model involving consumers with di?ering search costs buying products from
heterogeneous-type producers. Solving for the equilibrium in the general case, we showed
how shifts in the consumer search cost distribution impact equilibrium prices and market
shares. Speci?cally, downward shifts in search costs lead to lower prices and shift market
share from low-type producers to the industry’s high-type businesses.
While there is an empirical literature investigating the advent and di?usion of e-commerce
31
on prices, little has been done regarding the market structure impacts – speci?cally, the
shifts in market share from low- to high-type businesses that our model predicts. We test
these predictions in three industries for which the introduction of e-commerce has arguably
decreased consumers’ search costs considerably: travel agencies, bookstores, and new auto
dealers.
We found evidence of the market share shifts predicted by the model. As consumers’
use of the Internet to make purchases rose, smaller establishments (where size re?ects ?rm
‘type’) declined in number and larger establishments became more dominant.
Interestingly, while the nature of the market share reallocations were similar in the in-
dustries, the speci?c mechanisms through which the declining search costs created them
were di?erent. For travel agencies, the shifts re?ected aggregate changes, common across
markets, driven in large part by airlines’ reductions in agent commissions in response to
consumers’ increasing use of online sources to buy tickets. This is evidenced by the fact
that once these aggregate changes in Internet purchasing patterns were controlled for, there
was no indication that the magnitude of the market share changes were any larger (smaller)
in markets experiencing idiosyncratically high (low) growth in consumers’ online purchases.
For bookstores and auto dealers, on the other hand, there was evidence that more exit oc-
curred among smaller stores in those markets where Internet use grew fastest. This suggests
that the industry-wide declines in small bookstores and auto dealers re?ect market-speci?c
impacts.
Appendix A: Rationale for Uniform Search Costs
Since we know (from Proposition 1) that the best-response price function p(c, F, t) is increas-
ing in t for all F and c, a natural su?cient condition for the equilibrium price function p
?
and equilibrium distribution F
?
to be increasing in t is that the market be supermodular in
the sense of Rauh (forthcoming). That is, we look for conditions under which each ?rm’s
pro?t function has increasing di?erences in own price p and market price distribution F. If
32
this is the case, the best-response function p(c, F, t) is increasing in F. Since p(c, F, t) is also
increasing in t, this implies that the equilibrium price function p
?
(c, t) is increasing in t for all
c. Intuitively, the increase of t has two e?ects on the price function of a given ?rm: ?rst, the
direct e?ect, holding the price distribution constant (which is positive by Proposition 1) and,
second, the indirect e?ect due to the increase in the prices of other ?rms (which is positive
by supermodularity). While the condition of supermodularity is not strictly necessary for
the equilibrium price function to be increasing in t, intuition suggests that an unambiguous
general comparative statics result for p
?
with respect to t is unlikely to obtain when the e?ect
of the market price on the best-response individual pricing function is ambiguous, which will
typically be the case when the market exhibits no complementarities.
We now turn to Rauh (forthcoming) for a general insight into the class of search cost
distributions that result in a supermodular search market. Before we do so, we must note,
however, that the class of models considered in Rauh (forthcoming) di?ers from ours in two
fundamental ways: ?rst, our model contains an endogenous probability of ?rms quitting the
market after realising that their marginal cost exceeds the threshold ¯ c, which is absent from
Rauh’s model, and, second, the total mass of ?rms L is endogenously determined in our
model, but is ?xed at an exogenous level in Rauh’s. These additional features make our
model more complicated than Rauh’s, yet the same basic forces are at play in determining
the interaction between own price p and market price F. In particular, the nature of the
added complexity suggests that the cases for which supermodularity can be guaranteed in
our model is a subset of those that yield supermodularity in Rauh’s model.
Unfortunately, it turns out that this observation forces us to restrict our attention to
the uniform distribution of search costs. The reason for this conclusion is as follows. First,
note that if the search cost density is increasing sharply over its range, a rightward shift of
the distribution (i.e., an increase of search costs) decreases the advantage of low-cost ?rms,
because they now need to share more of their customer base with more high-cost ?rms.
Thus, in his Proposition 3, Rauh shows that the search cost density must not be increasing
33
too sharply, lest an increase in the search costs should actually decrease the demand faced
by low-cost ?rms. Second, note that if the search cost density is decreasing sharply over its
range, an increase in search costs increases the mass of marginal consumers (i.e., consumers
indi?erent between buying and not buying) at a range of prices, consequently decreasing
incentives to raise prices. This observation is at the heart of Rauh’s Proposition 5, which
gives an upper bound on the rate of decrease in the search cost density. Propositions 3
and 5 together show that supermodularity cannot be guaranteed unless the search cost
is neither increasing nor decreasing too sharply. To quote Rauh (p. 15): ‘The uniform
distribution therefore represents the canonical example of complementarities.’ Furthermore,
the bounds that Rauh’s Propositions 3 and 5 place on the absolute value of the slope of the
search cost density reduce to zero when there is no set price cap, beyond which consumers
always have zero demand. Thus, with in?nitely inelastic unit demand, as in our model,
complementarity in the baseline model of Rauh (forthcoming) can be ensured only if the
search cost distribution is uniform. This leads us to restrict our attention to the uniform
distribution.
Appendix B: Proofs
Proof of Proposition 1
Denoting p
?
? p(c, F, t) and rewriting the FOC for the best-response price (8), we obtain
0 = ?(p
?
, F, t) ? (p
?
?c) +
x(p
?
, F, t)
x
p
(p
?
, F, t)
.
Di?erentiating with respect to p
?
yields (omitting the arguments for visual clarity):
?
p
= 1 +
(x
p
)
2
?xx
pp
(x
p
)
2
=
1
x
p
_
2x
p
?
x
x
p
x
pp
_
(8)
=
1
x
p
.¸¸.
0 i? ?
t
< 0.
Now, we can write
x(p
?
, F, t)
x
p
(p
?
, F, t)
=
_
?
p
?
q
__
r
0
F(u) du
¸
¸
t
¸
dr
?q
_
_
p
?
0
F(u) du
¸
¸
¸ t
_ = ?
_
?
p
?
q
__
r
0
F(u) du
¸
¸
t
¸
q
_
_
p
?
0
F(u) du
¸
¸
¸ t
_ dr.
Since
_
r
0
F(u) du >
_
p
?
0
F(u) du for all r > p
?
, the MLRP implies that the integrand is
increasing in t for each r. Consequently, the entire expression is decreasing in t, and thus
also ?
t
< 0. But this implies
?p
?
?t
> 0, as noted above.
Proof of Lemma 1
Implicitly di?erentiating the two identities that de?ne a search equilibrium, (18) and (19),
yields a system of equations for
?¯ ?
?a
and
?L
?a
:
?
a
+ ?
¯ ?
?¯ ?
?a
+ ?
L
?L
?a
= 0; (21)
?
a
+ ?
¯ ?
?¯ ?
?a
+ ?
L
?L
?a
= 0, (22)
where the subscripts indicate partial derivatives. Denoting
I ?
?(¯ ? ?2
?
aL?)
_
¯ ??2
?
aL?
0
?(c) dc
2?(¯ ? ?2
?
aL?)
2
=
?(¯ c)
_
¯ c
0
?(c) dc
2?(¯ c)
2
, (23)
the derivatives of the reservation price condition (18) are
?
a
=
_
L?
a
I ?1; ?
¯ ?
=
1
2
?I; ?
L
=
_
a?
L
I > 0. (24)
35
Now, by Assumption 2, ?(c)/?(c) > ?(¯ c)/?(¯ c) for all c < ¯ c, so that
_
¯ c
0
?(c) dc <
_
¯ c
0
?(c)
?(¯ c)
?(¯ c)
dc =
?(¯ c)
2
?(¯ c)
.
Plugging this into (23) yields I < 1/2, which immediately implies that ?
¯ ?
> 0. Notice also
that, since the integral in (18) is positive, (18) implies that
?
aL? < a, from which it follows
that
_
L?
a
< 1 and therefore that ?
a
< I ?1 < 0. Consequently, equation (21) implies that
at least one of ¯ ? and L must be increasing in a:
?L
?a
? 0 ?
?¯ ?
?a
> 0. (25)
If both L and ¯ ? were non-increasing in a, the left-hand side of (21) would be negative, which
would contradict (21).
Proof of Lemma 2
The derivatives in (22) are
?
a
= ?
1
4a
2
L
_
¯ ??2
?
aL?
0
(¯ ? ?c)
2
?(c) dc < 0; (26)
?
¯ ?
=
1
2aL
_
¯ ??2
?
aL?
0
(¯ ? ?c)?(c) dc > 0; (27)
?
L
= ?
1
4aL
2
_
¯ ??2
?
aL?
0
(¯ ? ?c)
2
?(c) dc =
a
L
?
a
< 0. (28)
Thus, (22) implies that if L is increasing in a, then ¯ ? must also be strictly increasing in a:
?L
?a
> 0 ?
?¯ ?
?a
> 0. (29)
If this were not true, the left-hand side of (22) would be negative, which would contra-
dict (22).
36
Proof of Lemma 3
From (22) and (28),
?¯ ?
?a
= ?
1
?
¯ ?
_
?
a
+ ?
L
?L
?a
_
= ?
?
a
?
¯ ?
_
1 +
a
L
?L
?a
_
. (30)
Since the
?¯ ?
?a
> 0, ?
a
< 0, and ?
¯ ?
> 0, this implies that
?L
?a
> ?
L
a
, (31)
that is, aL is increasing in a, and thus ? = 1/(aL) is decreasing in a.
Proof of Lemma 4
Taking the derivative of the pro?t function with respect to a, we obtain
?
a
(¯ c(a), a) =
1
4
[ ¯ ?(a) ?c]
2
?
?
(a) +
1
2
?(a)[ ¯ ?(a) ?c]
?¯ ?
?a
(a)
=
1
4
[ ¯ ?(a) ?c]
_
??
?
(a)c + ¯ ?(a)?
?
(a) + 2?(a)
?¯ ?
?a
(a)
_
.
Thus, for any c ? ¯ c < ¯ ?(a), the sign of ?
a
(¯ c(a), a) equals the sign of the rightmost term
above. Since ?
?
(a) < 0 by Lemma 3, this term is increasing in c. It follows that if the term
is negative for c
0
, it is, a fortiori, negative for all c < c
0
.
Proof of Proposition 3
We begin by showing that the pro?t of the ?rm at the current marginal cost cut-o? level
¯ c(a) must decrease as search costs decrease. First, rewrite the entry condition (19) as
_
¯ c(a)
0
?(c; a)?(c) dc ??[¯ c(a)]? = a.
Fully di?erentiating this with respect to a and noting that ?(¯ c(a); a) = ?, yields
_
¯ c(a)
0
?
a
(c; a)?(c) dc = 0. (32)
37
Together with Lemma 4, this implies that ?
a
(¯ c(a); a) > 0: otherwise, the integrand in (32)
would be everywhere negative (by Lemma 4), which would contradict (32).
It is now obvious that the marginal cost threshold ¯ c(a) decreases as a decreases. Let a
change from a
1
to a
2
< a
1
. Then
?[¯ c(a
2
), a
2
] = ? = ?[¯ c(a
1
), a
1
] > ?[¯ c(a
1
), a
2
],
where both equalities follow from the de?nition of ¯ c(a), and the inequality follows from
?
a
(¯ c(a); a) > 0. Since ?
c
(c, a) < 0 for all a and c by Property 3, this implies that ¯ c(a
1
) >
¯ c(a
2
), as desired.
Proof of Corollary 1
Let a change from a
0
to a
1
< a
0
. Let us index all corresponding quantities and functions
by 0 and 1, respectively. By Proposition 3, ¯ c
1
< ¯ c
0
. The cdf of the marginal-cost distribution
of operating ?rms is given by
˜
?(c) = ?(c)/?(¯ c). Since c
1
< c
0
, it immediately follows that
˜
?
1
(c) >
˜
?
0
(c) for all c.
Next, observe that p
1
(0) < p
0
(0) (by Proposition 2) and that p
1
(¯ c
1
) < p
1
(¯ c
0
) < p
0
(¯ c
0
)
(the ?rst inequality by Property 2 and ¯ c
1
< ¯ c
0
; the second by Proposition 2). Thus, the
support of the equilibrium price distribution shifts down, p
1
< p
0
and ¯ p
1
< ¯ p
0
. Consequently,
F
1
(p) ? F
0
(p) on the complement of [p
0
, ¯ p
1
], since F
0
(p) = 0 for p < p
0
and F
1
(p) = 1 for
p > ¯ p
1
.
Finally, by Proposition 2, p
1
(c) < p
0
(c) for all c ? [0, c
1
], so that p
?1
1
(r) > p
?1
0
(r) for all
r ? [p
0
, ¯ p
1
]. Since ¯ c
1
< ¯ c
0
, it follows from the de?nition of F ((12)) that F
1
(p) > F
0
(p) for
all p ? [p
0
, p
1
].
Proof of Corollary 2
Recall the equation from the proof of Proposition 3:
_
¯ c(a)
0
?
a
(c; a)?(c) dc = 0.
38
This would be violated if ?
a
(c; a) > 0 for all c < ¯ c. Thus, there exists ˆ c < ¯ c such that
?
a
(ˆ c; a) ? 0. But then, by Lemma 4, ?
a
(c; a) < 0 for all c < ˆ c.
Proof of Corollary 3
The total market share of all operating ?rms equals one: 1 =
_
¯ c(a)
0
X(c, a) dc. Di?erenti-
ating this with respect to a yields
0 = ¯ c
?
(a)X[¯ c(a), a] +
_
¯ c(a)
0
X
a
(c, a) dc.
Since c
?
(a) > 0 by Proposition 3, this implies that
_
¯ c(a)
0
X
a
(c, a) < 0. In particular, there
exists ˆ c < ¯ c such that X
a
(ˆ c, a) < 0.
By de?nition,
X(c, a) = Lx(c; a)?(c) =
1
2a
[ ¯ ?(a) ?c]?(c).
Thus,
X
a
(c, a) =
_
?
1
2a
2
[ ¯ ?(a) ?c] +
1
2a
¯ ?
?
(a)
_
?(c).
The sign of this expression equals the sign of the expression in parentheses, which is clearly
increasing in c. Thus, X
a
(ˆ c; a) < 0 implies that X
a
(c; a) < 0 for all c < ˆ c.
Appendix C: Numerical Simulations
Closed-form solutions for equilibrium components such as the price distribution, the marginal
cost cut-o?, and the mass of ?rms do not exist even in the case when the search cost
distribution is uniform. When search costs are not uniformly distributed, algebraic means
are even less successful: not only are there no closed form solutions, but also, as explained in
Appendix A, it is in general very hard even to derive comparative statics results. We therefore
turn to numerical simulations in this section. The goal is twofold: ?rst, to illustrate the
known comparative statics results for the uniform distribution, and, second, to determine
whether similar results can be obtained for another class of distributions. These latter
39
investigations show that comparative results analogous to those from the uniform search
cost distribution case do obtain when search costs follow an exponential distribution.
Since the equilibrium is straightforwardly de?ned by a system of equations ((2), (6),
(8), (10), (11), and (12)), there is no need for an ad-hoc numerical algorithm. We simply
discretise the search cost, marginal cost, and price spaces and solve the resulting system
of non-linear equations using the mathematical modelling language AMPL with the solvers
SNOPT and MINOS.
Uniform Search Cost Distribution
Let the search cost distribution be uniform on [0, a]. The results derived in the theoretical
section then show that the marginal cost threshold ¯ c should be increasing in a and that
equilibrium price distributions should shift to the right as a increases. Consequently, the
expected price µ
p
=
_
p(¯ c)
0
pf(p) dp should also be increasing in a. The theoretical analysis
remains silent about the direction of change in the mass of ?rms L. Using three di?erent
distributions for the marginal cost distribution, we con?rm the theoretical results for ¯ c, F,
and µ
p
. Furthermore, for all of the cases studied we also observe that the mass of ?rms, L,
increases in a.
The changes of ¯ c, µ
p
, and L with respect to a are shown in Figure 2. Equilibrium price
distributions F for three di?erent levels of a are shown in Figure 3.
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
FIGURES 2 AND 3 ABOUT HERE
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
Exponential Search Cost Distribution
Let the search cost distribution have an exponential distribution with parameter ? = ?a > 0.
Then, higher a corresponds to higher search costs (in the sense of MLRP). Using three
di?erent distributions for the marginal cost distribution, we ?nd that the local comparative
40
statics are analogous to those obtained for uniform search cost distributions. In particular,
¯ c, F, µ
p
, and L are all increasing in a.
The changes of ¯ c, µ
p
, and L with respect to a are shown in Figure 4. Equilibrium price
distributions F for three di?erent levels of a are shown in Figure 5.
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
FIGURES 4 AND 5 ABOUT HERE
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
References
Asplund, M. and Nocke, V. (2006). ‘Firm turnover in imperfectly competitive markets’,
Review of Economic Studies, 73(2), pp. 295–327.
Bagnoli, M. and Bergstrom, T. (2005). ‘Log-concave probability and its applications’, Eco-
nomic Theory, 26(2), pp. 445–469.
Baye, M. R., Morgan, J. and Scholten, P. (2004). ‘Price dispersion in the small and in the
large: evidence from an Internet price comparison site’, Journal of Industrial Economics,
52(4), pp. 463–496.
Benabou, R. (1993). ‘Search market equilibrium, bilateral heterogeneity, and repeat pur-
chases’, Journal of Economic Theory, 60(1), pp. 140–158.
Brown, J. R. and Goolsbee, A. (2002). ‘Does the Internet make markets more competitive?
Evidence from the life insurance industry’, Journal of Political Economy, 110(3), pp.
481–507.
Brynjolfsson, E. and Smith, M. (2000). ‘Frictionless commerce? A comparison of Internet
and conventional retailers’, Management Science, 46(4), pp. 563–585.
Carlson, J. A. and McAfee, R. P. (1983). ‘Discrete equilibrium price dispersion’, Journal of
Political Economy, 91(3), pp. 480–493.
Clay, K., Krishnan, R. and Wol?, E. (2001). ‘Prices and price dispersion on the Web:
evidence from the online book industry’, Journal of Industrial Economics, 49(4), pp.
41
521–539.
Clay, K., Krishnan, R., Wol?, E. and Fernandes, D. (2002). ‘Retail strategies on the Web:
price and non-price competition in the online book industry’, Journal of Industrial Eco-
nomics, 50(3), pp. 351–367.
Diamond, P. (1971). ‘A model of price adjustment’, Journal of Economic Theory, 3(2), pp.
156–168.
Gartner, J. (2004). ‘Car dealers feel net e?ect’, Wired, June 11.
Goolsbee, A. (2000). ‘In a world without borders: the impact of taxes on Internet commerce’,
Quarterly Journal of Economics, 115(2), pp. 561–576.
Herman, E. (2001). ‘Independent New York City bookstore reaches ?nal chapter’, New York
Daily News, December 21.
Hopenhayn, H. A. (1992). ‘Entry, exit, and ?rm dynamics in long run equilibrium’, Econo-
metrica, 60(5), pp. 1127–1150.
Horta¸ csu, A. and Syverson, C. (2004). ‘Product di?erentiation, search costs, and competition
in the mutual fund industry: a case study of S&P 500 index funds’, Quarterly Journal
of Economics, 119(2), pp. 403–56.
Katz, D. and Payne, H. (2000). ‘Tra?c jam: auto dealers use government to build Internet
roadblocks’, Reason, July.
Latcovich, S. and Smith, H. (2001). ‘Pricing, sunk costs, and market structure online:
evidence from book retailing’, Oxford Review of Economic Policy, 17(2), pp. 217–234.
McCall, J. J. (1970). ‘Economics of information and job search’, Quarterly Journal of Eco-
nomics, 84(1), pp. 113–126.
Melitz, M. J. (2003). ‘The impact of trade on intra-industry reallocations and aggregate
industry productivity’, Econometrica, 71(6), pp. 1695–1725.
Melo, F. (2005). ‘Bookstore no more’, St. Paul Pioneer Press, August 26.
National Commission to Ensure Consumer Information and Choice in the Airline Indus-
try. (2002). Upheaval in Travel Distribution: Impact on Consumers and Travel Agents,
42
Report to Congress and the President.
Rauh, M. T. (2007). ‘Nonstandard foundations of equilibrium search models’, Journal of
Economic Theory, 132(1), pp. 518–529.
Rauh, M. T. (forthcoming). ‘Strategic complementarities and search market equilibrium’,
Games and Economic Behavior.
Rob, R. (1985). ‘Equilibrium price distributions’, Review of Economic Studies, 52(3), pp.
487–504.
Scott Morton, F., Zettelmeyer, F. and Silva-Risso, J. (2001). ‘Internet car retailing’, Journal
of Industrial Economics, 49(4), pp. 501–519.
Stigler, G. J. (1961). ‘The economics of information’, Journal of Political Economy, 69(3),
pp. 213–225.
Sutton, J. (1961). Sunk Costs and Market Structure, Cambridge, MA: MIT Press.
Syverson, C. (2004). ‘Market structure and productivity: a concrete example’, Journal of
Political Economy, 112(6), pp. 1181–1222.
US Census Bureau. (2005). 2003 E-commerce Multi-sector Report, available at http://
www.census.gov/eos/www/ebusiness614.htm.
Weisman, R. (2004). ‘Final chapter for WordsWorth Books on Boston’s Harvard Square is
Saturday’, Boston Globe, October 27.
Zettelmeyer, F., Scott Morton, F. and Silva-Risso, J. (2005). ‘How the Internet lowers prices:
evidence from matched survey and auto transaction data’, NBER Working Paper 11515.
43
Fig. 1: Fraction of Consumers Purchasing Online and Total Number of Travel Agencies: 1994–2003
0
5
10
15
20
25
30
35
40
45
50
1994 1996 1998 2000 2002
18K
20K
22K
24K
26K
28K
30K
P
e
r
c
e
n
t
a
g
e
p
u
r
c
h
a
s
i
n
g
o
n
l
i
n
e
T
o
t
a
l
e
s
t
a
b
l
i
s
h
m
e
n
t
c
o
u
n
t
Year
Percentage purchasing online
Total establishment count
Fig. 2: Comparative Statics with Respect to Search Cost Changes When Search Costs Are Uniform
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.5 1.0 1.5 2.0
C
u
t
-
o
?
M
C
,
¯c
Search costs, a
0
5
10
15
20
25
0.5 1.0 1.5 2.0
M
a
s
s
o
f
?
r
m
s
,
L
Search costs, a
0.5
1.0
1.5
2.0
2.5
0.5 1.0 1.5 2.0
E
x
p
e
c
t
e
d
p
r
i
c
e
,
E
[
p
]
Search costs, a
Uniform (0,1) MC
Exponential (1) MC
Beta (2,1) MC
1
Fig. 3: Equilibrium Price Distributions for Three Levels of Uniform Search Costs under Three
Types of Marginal Cost Distribution
0.0
0.2
0.4
0.6
0.8
1.0
0.5 1.0 1.5 2.0 2.5 3.0
MC distribution: Uniform (0,1)
F
(
p
)
p
0.0
0.2
0.4
0.6
0.8
1.0
0.5 1.0 1.5 2.0 2.5 3.0
MC distribution: Exponential (1)
F
(
p
)
p
0.0
0.2
0.4
0.6
0.8
1.0
0.5 1.0 1.5 2.0 2.5 3.0
MC distribution: Beta (2,1)
F
(
p
)
p
a = 0.50
a = 1.25
a = 2.00
2
Fig. 4: Comparative Statics with Respect to Search Cost Changes When Search Costs are Expo-
nential
0.3
0.4
0.5
0.6
0.7
0.8
-1.5 -1.0 -0.5
C
u
t
-
o
?
M
C
,
¯c
Search costs, a
0
5
10
15
20
25
-1.5 -1.0 -0.5
M
a
s
s
o
f
?
r
m
s
,
L
Search costs, a
0.5
1.0
1.5
2.0
2.5
-1.5 -1.0 -0.5
E
x
p
e
c
t
e
d
p
r
i
c
e
,
E
[
p
]
Search costs, a
Uniform (0,1) MC
Exponential (1) MC
Beta (2,1) MC
3
Fig. 5: Equilibrium Price Distributions for Three Levels of Exponential Search Costs under Three
Types of Marginal Cost Distribution
0.0
0.2
0.4
0.6
0.8
1.0
0.5 1.0 1.5 2.0 2.5 3.0
MC distribution: Uniform (0,1)
F
(
p
)
p
0.0
0.2
0.4
0.6
0.8
1.0
0.5 1.0 1.5 2.0 2.5 3.0
MC distribution: Exponential (1)
F
(
p
)
p
0.0
0.2
0.4
0.6
0.8
1.0
0.5 1.0 1.5 2.0 2.5 3.0
MC distribution: Beta (2,1)
F
(
p
)
p
a = ?1.5
a = ?1.0
a = ?0.5
4
Table 1: Cross-Sectional Comparison of CEA Markets: Average Establishment Counts
A. Travel Agencies
Mean 25%ile Median 75%ile
Total establishments 74.3 10.8 22.4 58.3
Estabs. with 1–4 employees 48.4 6.6 13.6 35.7
Estabs. with 5–9 employees 16.8 3.0 6.0 13.7
Estabs. with 10–19 employees 5.9 0.7 1.9 4.5
Estabs. with 20–49 employees 2.3 0.0 0.6 1.8
Estabs. with 50–99 employees 0.6 0.0 0.0 0.4
Estabs. with over 100 employees 0.4 0.0 0.0 0.1
B. Bookstores
Mean 25%ile Median 75%ile
Total establishments 35.4 9.8 17.8 36.6
Estabs. with 1–4 employees 15.3 4.3 8.3 17.8
Estabs. with 5–9 employees 9.9 2.8 5.0 10.5
Estabs. with 10–19 employees 6.0 1.4 2.8 6.2
Estabs. with 20–49 employees 3.1 0.6 1.4 3.5
Estabs. with 50–99 employees 1.0 0.0 0.3 1.1
Estabs. with over 100 employees 0.2 0.0 0.0 0.1
C. Auto Dealers
Mean 25%ile Median 75%ile
Total establishments 73.5 24.8 41.7 82.0
Estabs. with 1–4 employees 11.3 3.2 5.8 12.1
Estabs. with 5–9 employees 5.2 1.7 3.3 6.7
Estabs. with 10–19 employees 10.5 3.6 7.2 13.5
Estabs. with 20–49 employees 24.8 9.0 15.5 28.3
Estabs. with 50–99 employees 15.4 3.9 7.4 16.3
Estabs. with 100-249 employees 6.0 0.4 1.9 5.6
Estabs. with over 250 employees 0.4 0.0 0.0 0.2
1
Table 2: Market Structure Patterns: Travel Agencies
A. Establishment Counts: US Aggregates
Employment Category
Year Total 1–4 5–9 10–19 20–49 50–99 100+
1994 28,118 18,186 6,774 2,121 759 169 109
1995 28,099 18,089 6,710 2,212 802 176 110
1996 28,735 18,654 6,724 2,181 859 200 117
1997 29,452 19,183 6,758 2,332 834 206 139
1998 28,776 18,460 6,755 2,325 861 212 163
1999 27,390 17,611 6,281 2,276 821 225 176
2000 25,975 16,783 5,836 2,091 845 234 186
2001 24,654 16,050 5,306 2,000 853 243 202
2002 21,079 14,281 4,151 1,581 681 201 184
2003 18,860 12,865 3,556 1,430 653 182 174
B. Local Market Structure and Fraction Purchasing Online
ln(total ln(total ln(establishments) by employment category
emp.) estabs.) 1–4 5–9 10–19 20–49 50–99 100+
N 3449 3449 3426 3306 2548 1740 783 538
R
2
0.96 0.98 0.97 0.94 0.91 0.89 0.83 0.84
Fraction -0.932* -1.117* -0.906* -1.538* -0.870* -0.357* 0.072 0.161
Online (0.047) (0.026) (0.036) (0.047) (0.065) (0.070) (0.106) (0.137)
C. Local Market Structure and Fraction Purchasing Online, with Year Fixed E?ects
ln(total ln(total ln(establishments) by employment category
emp.) estabs.) 1–4 5–9 10–19 20–49 50–99 100+
N 3449 3449 3426 3306 2548 1740 783 538
R
2
0.96 0.99 0.97 0.94 0.91 0.89 0.84 0.84
Fraction 0.278 0.033 0.029 -0.075 -0.178 0.180 -0.218 -0.195
Online (0.165) (0.084) (0.138) (0.161) (0.226) (0.251) (0.509) (0.592)
Notes: All regression speci?cations include CEA market ?xed e?ects and control for (logged) overall employment in
the market-year. Robust standard errors in parentheses. An asterisk denotes signi?cance at the ?ve percent level.
2
Table 3: Market Structure Patterns: Bookstores
A. Establishment Counts: US Aggregates
Employment Category
Year Total 1–4 5–9 10–19 20–49 50–99 100+
1994 13,520 6,625 3,840 2,198 708 102 47
1995 13,403 6,234 3,985 2,165 806 154 59
1996 13,134 5,916 4,039 1,940 966 211 62
1997 12,301 5,254 3,753 2,021 933 286 54
1998 12,151 5,031 3,588 2,025 1,088 357 62
1999 11,957 4,878 3,467 2,063 1,076 410 63
2000 11,662 4,641 2,953 2,349 1,163 485 71
2001 11,559 4,678 3,100 2,023 1,276 411 71
2002 12,178 5,494 2,777 2,089 1,275 475 68
2003 11,036 4,493 2,900 1,909 1,237 428 69
B. Local Market Structure and Fraction Purchasing Online, with Year Fixed E?ects
ln(total ln(total ln(establishments) by employment category
emp.) estabs.) 1–4 5–9 10–19 20–49 50–99 100+
N 3448 3448 3386 3338 3031 2400 1275 423
R
2
0.94 0.96 0.91 0.89 0.86 0.86 0.81 0.74
Fraction -0.307* -0.316* -0.161 -0.398* -0.817* 0.220 0.485 0.003
Online (0.148) (0.115) (0.172) (0.187) (0.210) (0.208) (0.357) (0.377)
Notes: All regression speci?cations include CEA market ?xed e?ects and control for (logged) overall employment in
the market-year. Robust standard errors in parentheses. An asterisk denotes signi?cance at the ?ve percent level.
3
Table 4: Market Structure Patterns: New Auto Dealers
A. Establishment Counts: US Aggregates
Employment Category
Year Total 1–4 5–9 10–19 20–49 50–99 100–249 250+
1994 24,130 2,715 1,724 4,142 9,017 4,853 1,601 78
1995 24,230 2,850 1,653 3,882 8,927 5,063 1,755 100
1996 24,639 3,320 1,691 3,735 8,757 5,155 1,866 115
1997 26,208 3,848 1,941 3,825 9,065 5,376 2,022 131
1998 26,216 4,117 1,971 3,777 8,873 5,421 1,931 126
1999 26,117 4,287 1,948 3,611 8,616 5,437 2,083 135
2000 26,225 4,440 1,841 3,505 8,380 5,592 2,303 164
2001 26,444 4,759 1,863 3,462 8,373 5,493 2,332 162
2002 25,625 4,176 1,723 3,282 8,202 5,600 2,451 191
2003 26,707 4,654 1,891 3,394 8,237 5,768 2,532 231
B. Local Market Structure and Fraction Purchasing Online, with Year Fixed E?ects
ln(total ln(total ln(establishments) by employment category
emp.) estabs.) 1–4 5–9 10–19 20–49 50–99 100–249 250+
N 3423 3425 3300 3021 3311 3423 3363 2455 643
R
2
0.99 0.99 0.88 0.82 0.91 0.95 0.95 0.94 0.80
Fraction 0.155* 0.130* 0.200 -0.081 -0.063 -0.231* 0.230* 0.595* 0.530
Online (0.054) (0.042) (0.187) (0.215) (0.146) (0.091) (0.117) (0.175) (0.602)
Notes: All regression speci?cations include CEA market ?xed e?ects and control for (logged) overall employment in
the market-year. Robust standard errors in parentheses. An asterisk denotes signi?cance at the ?ve percent level.
4

doc_873547838.pdf