Professional Documents
Culture Documents
Watson.2009 - 27-11 PDF
Watson.2009 - 27-11 PDF
Volume LVII
June 2009
0022-1821
No. 2
I. INTRODUCTION
in the variety
of products that they offer to consumers. Consider a retail market in which
each store sells many horizontally differentiated varieties of a single class of
good, for example, music CDs, books, clothes, or video rentals. Consumers
in such markets typically have idiosyncratic preferences over the different
available styles of the good. They may need to search across multiple
retailers to nd the outlet that sells their preferred combination of style and
price, in which case they are naturally drawn to sellers with a broader range
of available varieties. A stores choice of product variety is then a strategic
variable, depending endogenously on the variety choices of its competitors.
Thus a music store manager choosing whether to add CDs to his stock
weighs the costs of additional inventory and display space against the
increased probability that customers nd a good match for their musical
tastes there, rather than at a rival outlet elsewhere.
This paper is based on chapter 3 of my doctoral dissertation. For advice and guidance at
various stages of this project I am grateful to Michael Mazzeo, Robert Porter and Asher
Wolinsky. Thanks for comments are also due to David Barth, Avi Goldfarb, Shane Greenstein,
Ithai Lurie, Brian Viard and Robert Vigfusson, to the Editor and anonymous referees, and to
seminar participants at Northwestern, Kyoto IER, Tokyo, Texas, Melbourne, Analysis Group
and the 2004 IIOC and NASMES conferences. Financial support is gratefully acknowledged
from a Northwestern University Graduate School Graduate Research Grant and from the
Universitys Centre for the Study of Industrial Organization. All errors herein are my
responsibility.
w
Authors afliation: Department of Economics, University of Texas at Austin,
1 University Station C3100, Austin, Texas, 78712 U.S.A.
e-mail: watson@eco.utexas.edu
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics. Published by Blackwell Publishing, 9600 Garsington Road, Oxford OX4 2DQ, UK, and 350 Main Street, Malden, MA
02148, USA.
217
218
RANDAL WATSON
1
Redstone [2001]. Results from the experiment encouraged Blockbuster to implement a new
business model based on expanded retail inventories and revenue sharing with movie studios.
2
Note that terms like product range are not meant to connote distances in an explicit space of
product characteristics. Rather they refer to the number of different product lines carried by a
retailer.
3
Previous empirical studies which analyze other aspects of the eyecare industry include
Benham [1972], Kwoka [1984], Haas-Wilson [1989], and Parker [1995]. The Federal Trade
Commission has had a longstanding interest in the effects of state regulations on the practice of
optometry see, e.g., FTC [2003].
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
219
business stealing: all else equal, having more rivals in a market reduces the
marginal revenue (i.e., the probability of sale) associated with each item in a
sellers inventory, and equilibrium product ranges therefore fall.
A more complicated picture emerges when I move to rm level
regressions, taking into account the spatial differentiation between sellers
in a market. If a seller already has more than a few rivals nearby, then more
competition still has a baseline negative effect: per-rm variety falls when the
number of rivals increases, or when the distance to these rivals is reduced.
However this response is signicantly less negative when the incumbent
faces relatively little local competition, with few or no proximate
competitors. Suppose for example that a certain seller X initially operates
alone at a particular site, or with just one or two nearby rivals. Then the
empirical results indicate that shifting a rival to join X from elsewhere in the
market is more likely to lead to an expansion in Xs product range than to a
contraction.
Theories of agglomeration behaviour (or clustering) might explain this
nding. Models in this literature show how clusters of independent sellers
(e.g., shopping malls) attract more business by allowing consumers to
economize on search costs.4 In the present context a positive response of perrm variety to nearby competition is consistent with this behavior. Holding
xed a rms cost of stocking inventory, such a response suggests an increase
in the marginal revenue of the items initially held in stock. If price-cost
margins do not increase in response to more competition, then the increased
marginal revenue of an inventory item must reect an expectation of higher
quantities sold. With unit-demand consumers, this implies the attraction of
shoppers away from other locations in the market.
This evidence is circumstantial rather than direct; hence I consider an
additional test for clustering effects. If consumers plan shopping trips to
economize on search costs, then the different rm locations within a market
are effectively in competition with each other. All else equal, a cluster with
more sellers should attract more shoppers. When a new cluster forms
(because one rm moved to join another) it is then less likely to attract
business from other rms in the market if those other rms are themselves
concentrated at, e.g., a mall. To examine this effect in the empirical analysis I
construct a measure of the dispersion of rms elsewhere in the market, and
interact it with the response of a rms variety to more nearby competition. I
nd that the interaction is indeed signicantly positive. That is, a rm is more
likely to raise variety in response to more competition when other rms are
less concentrated, which is again suggestive of the existence of clustering
effects in these markets.
4
For theoretical analyses of clustering by single-variety rms see, e.g., Wolinsky [1983],
Dudey [1990], Konishi [2005], and Fujita and Thisse [2002, Ch. 7]. I am not aware of any
equivalent spatial analyses for multi-variety rms.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
220
RANDAL WATSON
Since these inferences are based on the estimated effects of the competition
measures, it is important to control in the empirical analysis for any
endogeneity in these measures. My initial approach to this problem is similar
to the selection-correction methodology employed in Mazzeo [2002]. I apply
a two-stage model of competition to the markets in my data. In the rst
stage, the entry behavior of sellers in each market is modelled using Seims
[2006] framework for endogenous location choice. That incompleteinformation entry framework is particularly suited to the analysis of spatial
differentiation, since it admits a relatively rich set of potential product
locations. From the entry model I generate selection corrections to be
incorporated into the second-stage regressions of product ranges on the
congurations of competitors. To check the sensitivity of this approach to
particular assumptions used in the rst-stage entry model, I also employ a
traditional instrumental-variables methodology. Neither approach produces results that contradict my earlier inferences.
The existing empirical literature on product-variety competition is fairly
sparse. Berry and Waldfogel [2001] study the relationship between per-rm
variety and market concentration in U.S. radio broadcasting, focusing on
the impacts of recent regulatory changes. They nd that greater concentration of ownership in a market raises variety per rm, where variety is
measured by the number of different programming formats on air.5
Alexander [1997] looks at the variety-concentration relationship in the U.S.
recorded-music market. Working with a measure of aggregate (i.e., marketwide) variety in popular songs, he nds a non-monotonic trend over several
decades, with aggregate variety reaching a maximum at medium levels of
concentration and declining thereafter. This non-monotonic pattern is
similar to the results presented below. However Alexanders focus is
somewhat different to that of the present analysis, since he uses market-level
measures of variety and concentration. I aim at deriving a richer picture of
the competitive effects by eliciting the relationship between rm-level
product ranges and differentiation from rivals. My results suggest that this
differentiation is indeed an important element of variety competition.
Somewhat closer to the present work is the study of Olivares and Cachon
[2007], who examine the relationship between sales, inventory and
competition at auto dealerships in a cross-section of small U.S. markets.
Among other results they nd a non-linear increasing relationship between
dealer inventories and competition the concave shape of this relationship is
again consistent with the patterns observed in my data.6
5
Related studies of product variety in media markets may be found in George and Waldfogel
[2003] and George [2002].
6
A marketing study looking at the retailing of yoghurt is Draganska and Jain [2005], who use
supermarket scanner data to incorporate consumers tastes for greater variety (number of
avors) into a differentiated-goods model of interrm competition. See also Bayus and Putsis
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
221
222
RANDAL WATSON
8
The industry wholesale catalogue Frames Data (Jobson Medical Information) reports
specications for eyeglass frames from more than 100 different manufacturers.
9
My survey data indicate that an eyeglasses seller implementing a one-standard-deviation
increase in its product range would need to stock around 300 extra styles. If the seller carries two
pairs in each style and the wholesale cost per pair is $60, then the additional inventory costs are
the interest costs on $36,000, a non-trivial burden for a small business.
10
Starting with Rosenthal [1980], various models have shown that price-cost margins in
oligopoly can increase in the number of competitors. Bagwell and Ramey [1994] and Watson
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
223
[2006] show how this effect can arise in models with multiproduct rms however the increase
in margins in those models only obtains if rms product ranges simultaneously fall.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
224
RANDAL WATSON
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
225
Table I
List of Markets in Sample
State
Illinois
Indiana
Iowa
Michigan
Minnesota
Ohio
Markets
Bloomington, Carbondale, Danville, Decatur, De Kalb, Freeport, Galesburg,
Kankakee, Quincy, Springeld, Urbana
Bloomington, Columbus, Kokomo, Lafayette, Marion, Richmond, Terre Haute
Ames, Burlington, Cedar Rapids, Clinton, Dubuque, Fort Dodge, Iowa City, Mason
City, Waterloo
Battle Creek, Benton Harbor, Holland, Jackson, Muskegon
Mankato, Rochester, St. Cloud
Ashtabula, Findlay, Lancaster, Lima, Manseld, Marion, Newark, Sandusky,
Zanesville.
Table II
Market Characteristics
Description
No. of Census tracts in market
No. of eyeglass sellers in market
Population of market (in 1,000s)
Proportion black (%)
Log (per capita income in $1,000)
Weighted average of median age (years)
Weighted average of median rent ($)
No. of malls in market with area 150,000 sq. ft.
No. of hospitals
No. of ophthalmology (MDs) practices
No. of interstates
Mean
StDev
Min
Max
22
13
86
6.5
3.0
35.6
500
1.3
1.8
2.6
0.9
9.5
4.3
36
5.5
0.1
2.9
58
0.6
0.7
1.5
0.7
8
4
31
0.4
2.8
30.0
401
0
1
0
0
48
23
170
27.6
3.2
39.9
610
2
3
6
2.5
Note: Counts of eyeglass sellers, MDs, and hospitals are from directories and a telephone survey. Demographic
data are from the 2000 Census: population is non-institutionalized population; black population is persons
reporting African-American race only; age and rent measures are market averages of tract-level medians,
weighted by tract population. Shopping centre data are from National Research Bureau [1998] areas for malls
are gross leasable areas. The interstate measure is the number of interstate highways passing through the market
an interstate that terminates in a market counts for an increment of one half.
in Seim [2006] are used to dene the set of markets. A town or group of towns
was initially included in the sample if it comprised a continuous built-up area
with total population in the range 25,000200,000, located entirely within (one
or more of) the six states under study. I dropped a market from this set if any
one of its principal business centers was less than 20 miles from a business
center in a separate built-up area of population 25,000 or greater. Thus the
sample excludes markets close to big metropolises in favour of regional centers
at least 20 miles from the next major town.13 Table I lists all the markets in
the study area and Table II summarizes some of their characteristics.
Column 1 of panel I in Table III classies the 571 sellers of eyeglasses in the
44 sample markets by type of outlet.14 Most (465) are specialist sellers,
13
The Rand McNally Marketing Atlas [2001] was used to dene built-up areas and locate
business centers.
14
Sellers were initially located from listings in telephone directories (American Business Disc,
www.yahoo.yp.com), and were then telephoned to conrm location and current operation.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
226
RANDAL WATSON
Table III
Sellers of Eyeglasses
Number of sellers
Category/sub-category
I. By type
(A) Eyecare specialists:
Comprising:
Optician
Optometrist
Ophthalmologist
(B) Dept.-store sellers
Total of (A) and (B)
II. By afliation (eyecare specialists only)
Lenscrafters, Pearle Vision
Other wide-area chain
Local chain or unafliated
Total
(1) Overall
465
277
19
358
88
106
571
10
212
55
64
341
40
55
370
465
22
25
230
277
Note: The frames subsample has 100% of sellers in IL, 50% elsewhere. All department store sellers provide
eyecare through an optometrist. Wide-area means operating in more than three markets in the sample.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
227
The second column of Table III categorizes this subsample into the different
classes of seller dened above. At each of these businesses the measure of
product variety is a simple count (usually done by me) of the number of
different frames on display for adult prescription lenses (excluding sunglasses
and safety glasses). About 5% of the observations in the frames subsample
are missing due to seller non-response. For any such outlet, the number of
frames is imputed to be the average for similar sellers in the same town.
Table IV and Figure 1 summarize the numbers of eyeglass frames stocked
by each kind of seller in the subsample. From the gure, the distribution of
Table IV
Number of Eyeglass Frames
No. of frames per seller
Category/sub-category
Mean
I. By type
(A) Eyecare specialists:
661
Comprising:
Optician
510
Optometrist
676
Ophthalmologist
631
(B) Dept.-store sellers
471
All in (A) and (B)
625
II. By afliation (eyecare specialists only)
Lenscrafters, Pearle Vision
1142
Other wide-area chain
718
Local chain or unafliated
604
Median
Std. dev.
1st quartile
3rd quartile
600
326
438
793
503
603
612
434
537
166
340
285
138
309
410
451
396
385
411
613
807
793
499
749
1055
670
551
423
206
270
825
587
410
1420
815
738
Note: Frames data are for the subsample with 100% of sellers in Illinois, 50% elsewhere.
See notes for Table III.
Figure 1
. Distribution of Sellers, by Type and Number of Frames
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
228
RANDAL WATSON
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
229
the set of locations (or cells) in each market.18 Within each market a census
tract is counted in the set of possible seller locations if its populationweighted centroid is within ten miles of the center of the markets main urban
area. For simplicity, all stores and other features of a tract are assumed to be
located at its population-weighted centroid, and the distance between tracts
is taken to be the distance between these centroids.
Let Vmki denote the equilibrium number of varieties sold by the i-th rm,
i 5 1, . . ., Nm, when it is located in cell k 5 1, . . ., Km of market m. Here Vmki
is measured as the number of different styles of spectacle frames on display at
an eyewear seller. It is assumed that the logarithm of Vmki is determined by
the following reduced-form relationship:
1
The vector Zmk contains exogenous demographic and location characteristics of cell mk, including both tract-specic characteristics (e.g., median age
of residents, population within one mile) and market-level variables (e.g.,
state dummies). Some specications could add to Zmk characteristics of rm
i such as chain afliation, but my baseline model avoids this because of the
potential for endogeneity between such rm-specic attributes and a rms
location.
Effects of competition on rm is product range are captured by the
function g(Amk, a). Here a is a vector of parameters and the vector Amk
classies rm is rivals according to their distance from cell mk. Section II
discussed the competing roles of agglomeration and business stealing in
determining the effects of competition on a rms equilibrium product
variety. To allow for these opposing effects, my specication of g(.) admits
non-monotonic effects of nearby rivals. For simplicity I use just two distance
classications (or bands), with A1mk representing the number of nearby
rival sellers within one mile (including rivals in the same tract), and A2mk
representing the number of distant rivals more than a mile away. I also
let Admk represent the presence of the rst three nearby rivals, i.e.,
Admk 1A1mk > 0 1A1mk > 1 1A1mk > 2, where 1(.) represents the
indicator function. Then Amk A1mk ; A2mk ; Admk , and g(.) is:
2
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
230
RANDAL WATSON
third-closest rivals. In other words, (2) is piecewise linear in A1mk with slope of
a1 ad for A1mk 3 and a1 for A1mk 4.19
With this specication of g(.), Conjecture 1 in section II corresponds to a
test of H0: a2 5 0. If a2 is signicantly negative, then the conjecture is
conrmed: competition from distant rivals reduces per-rm variety.
Conjecture 2 indicates that the agglomeration effects of nearby competition
should be more pronounced when distant rivals are more dispersed, and
when there are relatively few incumbent rms at a site, e.g., for A1mk 3. To
test this conjecture, I include in the regression an interaction between Admk
the sum of the indicators for the presence of the rst three nearby rivals and
a measure of the dispersion among distant competitors.
This measure is constructed by determining, for each cell mk in a market,
the set Dmk of distant rivals for that cell, i.e., the set of rms in tracts more
than a mile away from the given tract. Thus Dmk has A2mk elements. For each
rm j in Dmk, I count sjmk, the number of distant rivals in the same set, i.e., the
number of other rms in Dmk in tracts that are themselves more than a mile
away from js tract. I set smk to be the average across all j in Dmk of sjmk and
dene DISPERSmk smk =A2mk 1. (If A2mk 1 then I arbitrarily set
DISPERSmk to one. In the present application there are very few rms with
A2mk 1, and changing DISPERSmk to zero for these rms would not make
any great difference to the results.)
To interpret DISPERS, say, for example, that all A2mk distant rivals are
clustered at a single location. Then we have DISPERSmk 5 0, since sjmk 5 0
for all j in Dmk. On the other hand if the distant rivals all operate at separate
locations, each a mile away from the other, then we have DISPERSmk 5 1,
since sjmk A2mk 1 for all j in Dmk. Thus, DISPERSmk is a number between
zero and one, with higher values representing increases in the overall
dispersion of rivals. Conjecture 2 then corresponds to a test of the
signicance of a coefcient on DISPERSmk Admk .
Observation 1 may be applied to the data by testing H0 : a1 ad a2 5 0
and H0 : a1 a2 5 0. The former hypothesis represents the effect of
relocating a distant rival nearby, when a rm already has two or fewer
competitors in close competition. The latter represents the effect of such a
relocation when the rm has three or more rivals nearby. In the absence of
any theoretical structure, rejection of either hypothesis simply indicates that
spatial differentiation matters in variety competition. If the informal
structure of section II is taken on board then the rejection of either
hypothesis in favor of a positive response is interpreted as evidence for the
presence of agglomeration effects.
19
It can be seen that (2) may equivalently be written as: g 1A1mk 3A1mk aa
1A1mk 43aa A1mk 3ab A2mk a2 , where a1 ab and ad aa ab.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
231
far far
near near
gm on dmk
a
1nrm > 3dmk
a ;
where n nrm , m 5 1, . . ., M. The rst term on the RHS of (3) is a dummy for
the observed outcome of nrm in market m. Spatial differentiation is captured
far
near
by the second and third terms, in which dmk
and dmk
respectively represent
the average distance from the rm to its three closest rivals, and to all other
rivals. (Once again distances are measured in miles between tract centroids.)
Thus anear shows the notional effect on a rms variety, for a given total
number of rivals, of a marginal increase in the average distance to its three
closest rivals. A similar interpretation applies to afar
mk , for a marginal increase
in the average distance to all the rms more distant rivals.20
Interpretation of the parameters in (3) is not always straightforward. In
far
near
particular, the average distances dmk
and dmk
are each functions of the
20
The third term in (3) is preceded by the indicator 1nrm > 3 to allow for the possibility that a
market only has nrm 3 rivals in total.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
232
RANDAL WATSON
The function in (4) is again piecewise linear in the number of nearby rivals,
but because of the small numbers in group B, I assume that the kink occurs
between B1mk 1 and B2mk 2. Thus bd, the coefcient on Bdmk , is the extra
slope coefcient for the rst nearby group B rival.
V. ACCOUNTING FOR THE ENDOGENEITY OF LOCATIONS
233
i.e., for the expected value of the error given the observed location choices of
all rms in the market. These conditional expectations (which are linear in
parameters) can be added to the RHS of (1) in the same manner as ordinary
sample-selection corrections.
Further explanation of this entry model and the associated selection
corrections appears in the online appendix. For present purposes it should
be noted that this approach uses a particular assumption to deal with the
potential endogeneity in the competition measures in Amk. The Seim model
assumes for identication purposes that rms do not see any shared
location-specic unobservables when simultaneously choosing locations.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
234
RANDAL WATSON
That is, rms private information about their potential protability at each
tract is independent of their rivals information. Hence, conditional on the
observables, rms actual location choices are by assumption independent of
each other. In (1) this implies that, after conditioning on Zmk and any
additional observables from the rst stage, the indicators of rivals locations
in Amk are uncorrelated with umki. Thus, while the sample-selection
corrections employed in this approach allow for strategic interactions in
the rst-stage actions, they assume away some types of second-stage
endogeneity.
To deal with this issue, I consider also an instrumental-variables
approach, to control for any conditional correlation between the competition measures in Amk and the errors tm and umki. In this approach, the
sample-selection corrections used previously are retained, to control as
before for any endogeneity in the regressors in Zmk. As instruments for the
competition measures in Amk, I use four variables: a dummy indicator of a
tracts proximity to major arterial roads, an interaction of this dummy with
the tracts population density, a measure of the tracts distance from the
geographic centre of the market, and a count of the number of interstate
highways passing through the market.
Instruments are required to be conditionally correlated with rivals
location decisions but uncorrelated with the unobserved component in a
rms product range equation. Thus they should be factors affecting costs or
revenues that are unrelated to product range choices. Indicators of zoning
restrictions would be good candidates. Since direct observation of such
restrictions is difcult, I instead use a tracts proximity to major roads as a
proxy. It seems likely that tracts containing such roads would be zoned for
commercial use (at least in part). From the Census TIGER/Line les, I
constructed a dummy variable equal to one if a tract contains (or directly
abuts) a U.S. highway or interstate highway. About two thirds of all tracts
(whether or not they contain eyewear sellers) in all 44 markets contain such
thoroughfares. To control for rural tracts on the outskirts of a market
(through which an interstate might pass) I also include the interaction of this
major road dummy with the logarithm of a tracts population density.
Justication for the third instrument arises from the fact that rival rms in
the sample seem to be located closer to a markets centre than would be
predicted by its observed locational characteristics. That is, A1mk , the number
of nearby rivals, has a strong negative correlation with the distance from a
rms location to the geographic centre of its market, even after controlling
for demographics such as population density and proximity to malls and
large stores. This may reect persistence in rms location choices and the
migration over time of commercial activity out of a towns downtown area.
Casual observation suggests that eyewear sellers in these downtowns are
mainly older establishments. Proximity to the centre of a town may thus be a
proxy for a sellers date of entry into a market. If enough present-day
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
235
VI. RESULTS
VI(i).
Market-Level Regressions
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
236
RANDAL WATSON
TableV
OLS Regressions of AverageVariety at the Market Level, for Group-A
Sellers
Dependent variable: mean of log(number of frames) in market. N 5 44.
RHS variable
Constant
No. of malls
No. of hospitals
No. of ophthalmologists
Population (10,000s)
Propn. black (%)
Log(per capita income)
Weighted ave. of median age (years)
Weighted ave. of median rents ($100)
No. of interstates
No. of group-B rivals
No. of group-A rivals
(1)
(2)
(3)
(4)
7.1
(1.1)
0.056
(0.057)
0.063
(0.046)
0.0061
(0.019)
0.0011
(0.014)
0.0071
(0.0059)
0.047
(0.51)
0.018
(0.014)
0.0072
(0.11)
0.035
(0.048)
0.059z
(0.034)
0.027
(0.010)
7.4
(0.95)
0.066
(0.053)
0.070z
(0.041)
0.0057
(0.018)
0.00025
(0.012)
0.0056
(0.0055)
0.15
(0.46)
0.014
(0.013)
0.052
(0.098)
0.030
(0.044)
7.1
(1.2)
0.053
(0.063)
0.062
(0.050)
0.0072
(0.020)
0.0042
(0.015)
0.0086
(0.0068)
0.053
(0.55)
0.018
(0.015)
0.0052
(0.11)
0.045
(0.052)
0.063
(0.042)
0.027
(0.011)
7.6
(1.1)
0.079
(0.058)
0.060
(0.045)
0.0033
(0.019)
0.0044
(0.013)
0.0083
(0.0062)
0.25
(0.51)
0.012
(0.014)
0.061
(0.11)
0.040
(0.047)
0.19
(0.085)
0.34
(0.095)
0.38
0.11
0.47
0.02
0.087
(0.21)
0.030
(0.056)
0.0083
(0.053)
0.26
(0.12)
0.36
(0.10)
0.021
(0.20)
0.050
(0.051)
0.032
(0.049)
0.39
0.26
0.49
0.05
()
signicant at 1% level.
signicant at 5% level.
(z)
signicant at 10% level.
Note: Group B is a group of four department stores which have optical shops in almost all their outlets see
section IV. Average frames counts are taken over all group-A sellers with a non-missing observation in the
frames subsample. Average distances are calculated from all group-A sellers, regardless of missing frames
observations.
Standard errors (in brackets) are robust to heteroskedasticity.
()
237
22
I also estimated versions of models 3 and 4 allowing for non-monotonic distance effects.
However the non-monotonicities were not statistically signicant.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
238
RANDAL WATSON
TableVI
Summary Statistics for Firm-level Regressions
Description
Mean
Market-level variables
3.0
9.5
35.2
Tract-level variables
Median rent in tract, 2000 ($100)
5.1
No. of hospital beds in band 1
83
Black popn. in band 1 (10,000s)
0.08
Black popn. in band 2 (10,000s)
0.61
Dummy for mall in band 1
0.37
Area of malls in band 1 (100,000 sq. ft.)
6.1
No. of malls in band 2
0.97
1
No. of group-B sellers in band 1 Bmk
0.5
No. of group-B sellers in band 2 B2mk
1.4
d
Bmk
0.35
1
2.3
No. of group-A sellers in band 1 Amk
No. of group-A sellers in band 2 A2mk
8.9
d
1.8
Amk
Dist. to closest group-B rival (miles)
1.9
Ave. dist to other group-B rivals (miles)
3.0
Ave. dist. to 3 closest group-A rivals (miles)
0.9
Ave. dist. to other group-A rivals (miles)
3.1
Dispersion in distant rivals (DISPERS)
0.75
Log (per capita income in $1,000)
Population (10,000s)
Weighted ave. of median age (years)
StDev
Min
Max
0.1
3.7
3.0
2.8
3.1
30.0
3.2
17.0
39.9
1.1
206
0.12
0.57
0.48
2.1
0.66
0.7
1.1
0.48
2.0
3.9
1.2
1.9
1.8
1.1
1.5
0.19
2.2
0
0.00
0.02
0
1.6
0
0
0
0
0
1
0
0
0
0
0.7
0
8.3
1195
0.80
2.19
1
11
2
3
5
1
7
19
3
11.2
8.1
7.7
10.8
1
Note: See notes to Tables II and V. Variables are averaged over the 301 sellers in the variety regression
subsample. See text for denitions of Admk , Bdmk , DISPERS. Band 1 refers to the band 0-1 miles, band 2 to
anything further. Mall areas averaged only over those sellers with a mall in band 1.
A rivals in Table V are likely to be even more negative than the values
reported. The overall conclusion then is that at the market level, the
dominant effect of competition between non-big-store eyeglass sellers is one
of business stealing. From columns 3 and 4 of the table, one might be
tempted to venture the further conclusion that interrm distance is a
statistically irrelevant factor in competition in these markets. However the
following analysis of competition at the tract level shows this conclusion to
be misleading.
VI(ii).
Firm-Level Regressions
Table VI reports summary statistics for the explanatory variables in the rmlevel regressions. These statistics are calculated over the set of 301 group A
sellers included in these regressions. This set represents a random 50%
(100% in Illinois) of the group A sellers in each market. Equivalent statistics
for the set of all 571 sellers are quite similar to those in the table.23 In addition
23
There is a difference between this regression subsample and the set of sellers upon which
the summary variety statistics in Tables III and IV are based. The former comprises a random
50% of the group-A sellers in each market (100% in IL), rather than 50% of all sellers in each
market. However the differences in the summary statistics will be relatively small because
group B only accounts for around 13% of all sellers.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
239
240
RANDAL WATSON
TableVII
Firm-level OLS Variety Regressions, for Group-A Sellers
Dependent variable: log(number of frames). N 5 301.
RHS variable
Log (per capita income in market)
Market popn. (10,000s)
Weighted market ave. of median age (years)
Median rent in tract ($100)
No. of hospital beds in band 1 (100s)
(1)
(2)
0.17
(0.36)
0.012
(0.014)
0.021
(0.013)
0.041
(0.032)
0.024
(0.016)
0.34
(0.22)
0.056
(0.059)
0.15
(0.17)
0.055z
(0.028)
0.16
(0.065)
0.032
(0.088)
0.25
(0.12)
0.099
(0.047)
0.095
(0.032)
0.090z
(0.047)
0.043
(0.016)
0.12
(0.37)
0.024
(0.015)
0.026
(0.013)
0.044
(0.033)
0.024
(0.016)
0.21
(0.25)
0.095
(0.060)
0.19
(0.16)
0.060
(0.026)
0.16
(0.066)
0.043
(0.091)
0.26
(0.12)
0.11
(0.047)
0.095
(0.029)
0.040
(0.064)
0.17
(0.067)
0.042
(0.016)
YES
NO
0.17/0.00
YES
NO
0.18/0.00
(3)
0.016
(0.036)
0.032
(0.016)
0.48
(0.34)
0.14
(0.36)
0.056z
(0.030)
0.45
(0.31)
0.052
(0.10)
0.28
(0.14)
0.22
(0.61)
0.0094
(0.018)
0.0062
(0.021)
0.024
(0.034)
0.032
(0.015)
0.30
(0.27)
0.24
(0.15)
0.24
(0.18)
0.059
(0.025)
0.068
(0.12)
0.10
(0.05)
0.17
(0.076)
(4)
YES
0.26/0.00
0.054
(0.022)
0.0016
(0.030)
0.042
(0.037)
0.061
(0.030)
YES
NO
0.22/0.00
()
signicant at 1% level.
signicant at 5% level.
(z)
signicant at 10% level.
Note: See notes to Tables II, V, VI. In (2), coefcients on variables marked
(w)
show the effect of moving one unit from band 2 to band 1. Model 4 contains a full set of dummies for each
possible value of the numbers of group-A rivals, and of group-B rivals, in the whole market.
S.e.s (in brackets) robust to heteroskedasticity and within-tract correlation. Constant not shown.
()
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
241
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
242
RANDAL WATSON
26
Removing one rival from the distant set will affect the value of DISPERS. When there are
nine distant rivals this effect is small, and may be ignored for present purposes.
27
There are various ways to construct DISPERS, depending on how one treats the group-A
and group-B rivals. In Table VII DISPERS is constructed just w.r.t. group-A rms. An
alternative construction measures the dispersion in all distant rivals, regardless of type. Using
this measure gives similar results to those in Table VII.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
243
raise product ranges, indicating that their market expansion effects are not
conned to a single location, but are spread across the whole market.
Model 3 in table 9 adds market dummies to the model of column 1.
Coefcients on market-level variables will no longer be identied. Variables
which are measured as counts over nearby tracts (number of hospital beds,
black population, area of malls, numbers of group A and group B rivals)
may now be interpreted as the effect of moving a unit of that variable from
band 2 to band 1. Thus, as mentioned in section IV, the coefcient on A1mk ,
for example, is now a1 a2, the baseline effect of bringing a distant group A
rival closer. Note that the piecewise linear specication of the competition
effects is still retained, through the dummies Admk and Bdmk . The coefcients
on those dummies are still ad and bd, and the sum of the coefcients on Admk
and A1mk , for example, gives a1 ad a2, the effect of relocating one of the
rst three group A rivals nearby from far away.
Effects of exogenous location characteristics and of competition from
group B rivals show similar estimates both with and without market
dummies. Estimated effects of spatial differentiation between group A rivals
are of the same sign in both models, but are somewhat more pronounced in
model 3 than in model 1. Thus the baseline effect of moving a group A rival
closer is estimated to be 10% in model 3 (signicant at the 5% level),
compared with 5.2% previously. However if the relocation involves one
of the rst three nearby group A rivals then the variety effect is reversed, to
become 0.17 0.10 7.1%. Furthermore this effect now becomes signicant at the 5% level previously this effect was a marginally signicant
3.8%. The null hypothesis of linear competition effects (H0: ad 5 0) is
rejected with 95% condence. Overall the inferences drawn from model 1
still apply, but if anything with more force: the underlying effect of more
competition is one of business stealing, but this seems to be tempered by
agglomeration effects in small groups of rms.
Clearly the preceding inferences rely on a fairly simple specication for the
form of the competition effects g(.). I tested the robustness of this
specication in two ways. First, I ran two alternative versions of model 1,
redening the cutoff distance between bands 1 and 2 to be 0.5 miles, or 1.5
miles, instead of one mile.28 Using a Davidson-McKinnon J test of nonnested hypotheses (Greene [1997], p. 366) I tested the null that model 1 alone
explains the data against each of the alternatives. In each case H0 was
rejected, but only at the 10% signicance level. I also ran the tests in reverse,
with the 0.5 mile and 1.5 mile models as the nulls and model 1 as the
alternative in each case. Again H0 was rejected, but much more strongly, at
the 0.1% level in each case. On this basis it seems clear that the one-mile
cutoff is the preferred specication among the three models.
28
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
244
RANDAL WATSON
29
Furthermore, the distance measures are interacted with dummies so that, e.g., average
distance to 4th-closest-and-further group-A rivals only switches on if the rm actually has four
or more such rivals.
30
Model 4 predicts a rise of 0.16% for every mile when the average distance to other
^ b
^ 6:7% for the similar experiment
group-B rivals is reduced. Model 1 predicts a fall of b
1
2
of relocating a second-or-later distant rival near a rm. Since neither effect is statistically
signicant these results are not necessarily in contradiction.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
245
just the effects on the sets of closest and furthest rivals, but also the effects on
the average distances to these rivals.31
VI(iii).
The analysis so far assumes that the competition measures A1mk , Admk , and A2mk
may be taken as exogenous, i.e., that the spatial conguration of rms in each
market is well explained by the exogenous location characteristics included in
each regression. As discussed in section V, in practice rms location choices
may be affected by unobservable factors which cause the competition measures
to be endogenous. To account for this possibility, I model rms observed
locations as arising from a rst-stage entry game, deriving from this game a
selection correction to be included in the second-stage variety regression. This
correction is of the form in (5): a conditional expectation of the composite error
tm umki. It is itself the weighted sum of two subsidiary conditional
expectations: one for a tract and rm-specic unobserved prot effect, and
one for a prot effect shared by all rms in a market.
Model 2 in Table VIII includes these selection corrections among the
regressors from the baseline model 1 of Table VII. For ease of comparison,
the estimates for that earlier regression are repeated as model 1 in Table VIII.
The table shows that when the selection corrections are included, they are
not statistically signicant, and the estimated coefcients and standard
errors for the other explanatory variables are essentially unchanged.
(Standard errors are corrected for the sampling error in the two new
generated regressors an explanation of how to correct the standard errors
is in the online appendix.) On this basis there is no apparent problem of
selection-induced endogeneity in the RHS variables. However, as explained
in section V, the entry model on which the selection corrections are based
assumes that there are no tract-specic prot unobservables common to all
potential entrants. If such unobservables are in fact present, then the
competition measures in Amk might still be correlated with tm and umki, even
after the inclusion of the selection corrections.
To address this problem I turn to an instrumental variables technique,
using as instruments for the competition measures the four variables for
which summary statistics are shown in Table IX. That table also shows
estimated coefcients on these instruments from rst-stage regressions in a
2SLS IV procedure. (Coefcients on the other exogenous variables are
omitted for brevity.) Note in particular the strong correlation between
31
I also estimated a version of model 1 in which a rms closest rival can have an effect on
variety that differs from the effects of its second and third nearby rivals. That is, in place of (2)
I estimated a specication where g A1mk a1 A2mk a2 a0 d 1A1mk > 0 a00d 1A1mk > 1
a0d 0:110 (s.e. 5 0.085) and ^
a00d 0:076 (s.e. 5 0.075).
1A1mk > 2g. This regression yielded ^
Since H0:a 0 d 5 a00 d cannot be rejected (p-value 5 0.79), the use of this restriction in model 1
seems a reasonable abstraction.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
246
RANDAL WATSON
TableVIII
Variety Regressions Allowing for Endogenous Competition Measures
Dependent variable: log(number of frames). N 5 301.
RHS variable
(1) OLS
(2) OLS
(3) 2SLS
0.17
(0.36)
0.012
(0.014)
0.021
(0.013)
0.041
(0.032)
0.024
(0.016)
0.34
(0.22)
0.056
(0.059)
0.15
(0.17)
0.055z
(0.028)
0.16
(0.065)
0.032
(0.088)
0.25
(0.12)
0.099
(0.047)
0.095
(0.032)
0.090z
(0.047)
0.043
(0.016)
0.16
(0.37)
0.011
(0.015)
0.023z
(0.013)
0.041
(0.032)
0.025
(0.016)
0.36
(0.24)
0.058
(0.062)
0.15
(0.17)
0.058
(0.028)
0.17
(0.067)
0.038
(0.091)
0.25
(0.12)
0.10
(0.047)
0.099
(0.031)
0.092z
(0.049)
0.048
(0.017)
0.0097
(0.035)
0.032
(0.089)
YES
NO
0.20
(0.49)
0.038
(0.18)
0.064
(0.30)
0.046
(0.67)
0.024
(4.4)
0.63
(0.54)
0.033
(0.18)
0.20
(0.27)
0.076
(0.20)
0.28
(0.17)
0.14
(0.17)
0.30z
(0.17)
0.23
(0.20)
0.26
(2.2)
0.24
(0.65)
0.17
(0.21)
0.073
(0.11)
0.38
(0.55)
YES
NO
YES
NO
()
signicant at 1% level.
signicant at 5% level.
(z)
signicant at 10% level.
Note: See notes to Tables II, V, VI, VII. Model (1) repeats (1) from Table VII.
Instruments for (3) are in Table IX. See appendix for explanation of the selection corrections. S.e.s (in brackets)
robust to heteroskedasticity and within-tract correlation, and in (2) and (3) are corrected for sampling error in
generated regressors. Constant not shown.
()
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
247
Table IX
Instruments for 2SLS Regression
I. Summary Statistics
StDev
Min
Instrument
Mean
Max
2.1
1.3
0.056
8.0
0.72
0.45
1.3
1.2
1.9
1.1
0.72
0.19
(0.054)
0.073
(0.19)
0.17
(0.12)
0.51
(0.42)
4.3
0.18
(0.14)
0.12
(0.081)
0.088
(0.18)
2.5
0.024
(0.19)
0.12
(0.12)
0.031
(0.25)
Note: Estimated coefcients in panel II are from regressions of the endogenous variables on the instruments and
all other exogenous variables in Table VIII. (Other coefcients not shown.) S.e.s are robust to
heteroskedasticity and within-tract correlation, but are not corrected for the sampling error in the generated
regressors.
()
Mean and s.d. calculated over tracts where the dummy is not zero.
level. Note however that the pattern of signs on the estimated effects of
group A competition is the same as in model 1. That is, the baseline effect of
such competition on per-rm product variety is negative, and is more
negative for nearby competitors. (Compare ^a1 0:26 and ^
a2 0:17.) In
a small cluster of rms, however, the effect of new entry by such competitors
is less negative, and if the rival is relocating from elsewhere in the market
then the net effect on variety is positive, since ^
a1 ^
ad ^
a2 0:15 > 0.
A test of overidentifying restrictions does not reject the exogeneity of the
instruments in this procedure. Nor does a Hausman-type test reject the
exogeneity of the competition effects in Amk.33 Overall, the results of the IV
analysis thus do not contradict the previous inferences from model 1. Of
course, the validity of this conclusion rests on the suitability of the
instruments. In particular, the number of freeways is potentially problematic as an instrument. Since it is a market-level variable, it might be
correlated with unobserved market-level effects which are in turn correlated
with umki. To deal with this issue, I re-ran the 2SLS model with market
dummies, thus dispensing with the need for an instrument for the number of
33
To implement these tests, the model was estimated by efcient two-step GMM, using a
weighting matrix robust to both heteroskedasticity and clustering in the errors. (The Stata
procedure ivreg28 (Baum, Schaffer and Stillman [2007]) is very useful for this purpose.) The
efcient GMM estimates are very close to the 2SLS estimates. I did a limited investigation of the
effects on these tests of correcting for the sampling error in the generated regressors (i.e., in the
selection corrections). The p-values for the tests seemed if anything to increase, which is
consistent with the idea that (in cross-section analysis) standard errors on individual
coefcients are always underestimated if such sampling error is ignored.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
248
RANDAL WATSON
entrants Nm. The results are again noisy, but the pattern of signs on the
estimated competition effects supports the inferences from Table VII.
Finally, since I could not observe zoning regulations, I included in the entry
stage of my two-stage approach a measure of the density of local businesses as a
proxy.34 It was highly signicant, indicating that it is at least a reasonable proxy
for the unobserved regulations. However when this measure is included in the
second-stage variety regression, it is not signicant, and furthermore the
coefcients and standard errors on the other RHS variables (in particular, on
the competition effects) are more or less unchanged. I conclude that the
inuence of zoning on product variety seems to be captured by the other
locational characteristics already included in the model.
VI(iv).
This measure was a count of the number of chain fast-food outlets in band 1.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
249
Aguirregabiria, V., 2005, Retail Stockouts and Manufacturer Brand Competition, (Mimeo,
University of Toronto, Toronto, Canada).
Alexander, P. J., 1997, Product Variety and Market Structure: A New Measure and a
Simple Test, Journal of Economic Behavior and Organization, 32, pp. 207214.
Anderson, S. P. and De Palma, A., 1992, Multi-Product Firms: A Nested Logit
Approach, Journal of Industrial Economics, 40, pp. 261276.
Anderson, S. P. and De Palma, A., 2006, Market Performance with Multi-product
Firms, Journal of Industrial Economics, 54, pp. 95124.
Bagwell, K. and Ramey, G., 1994, Coordination Economies, Advertising, and Search
Behavior in Retail Markets, American Economic Review, 84, pp. 498517.
Baum, C. F.; Schaffer, M. E. and Stillman, S., 2007 ivreg28: Stata Module for Extended
Instrumental Variables/2SLS and GMM, http://ideas.repec.org/c/boc/bocode/
S4254011.html
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.
250
RANDAL WATSON
251
Lee, L.-F., 1982, Some Approaches to the Correction of Selectivity Bias, Review of
Economic Studies, 49, pp. 355372.
Lee, L.-F.; Maddala, G. S. and Trost, R. P., 1980, Asymptotic Covariance Matrices of
Two-Stage Probit and Two-Stage Tobit Methods for Simultaneous Equations Models
with Selectivity, Econometrica, 48, pp. 491503.
Mazzeo, M. J., 2002, Competitive Outcomes in Product-Differentiated Oligopoly,
Review of Economics and Statistics, 84, pp. 716728.
Morgan, E., 2007, Buying Strategy: Do You Have One?, Eyecare Business, 21, pp. 106112.
National Research Bureau, 1998, Shopping Center Directory: Midwest Volume (Chicago,
Illinois, U.S.A.).
Olivares, M. and Cachon, G. P., 2007, Competing Retailers and Inventory: An Empirical
Investigation of U.S. Automobile Dealerships, (Mimeo, Columbia University, New
York, New York, U.S.A.).
Orhun, A. Y., 2005, Spatial Differentiation in the Supermarket Industry, (Mimeo,
University of Chicago, Chicago, Illinois, U.S.A.).
Parker, P. M., 1995, Sweet Lemons: Illusory Quality, Self-Deceivers, Advertising, and
Price, Journal of Marketing Research, 32, pp. 291307.
Rand McNally Commercial Atlas & Marketing Guide, 2001 (Rand McNally, Chicago,
Illinois, U.S.A.).
Redstone, S., 2001, A Passion to Win, (Simon and Schuster, New York, New York, U.S.A.),
Rosenthal, R. W., 1980, A Model in which an Increase in the Number of Sellers Leads to
a Higher Price, Econometrica, 48, pp. 15751579.
Seim, K., 2006, An Empirical Model of Firm Entry with Endogenous Product-Type
Choices, RAND Journal of Economics, 37, pp. 619640.
Van Ryzin, G. J. and Mahajan, S., 1999, On the Relationship between Inventory Costs
and Variety Benets in Retail Assortments, Management Science, 45, pp. 14961509.
Watson, R., 2003, Competition in Product Variety: Theory and Evidence, unpublished
PhD dissertation, Northwestern University, Evanston, Illinois, U.S.A.
Watson, R., 2005, Entry and Location Choice in Eyewear Retailing, Mimeo, University
of Texas-Austin, http://www.eco.utexas.edu/
watson/.
Watson, R., 2006, Search, Availability and Competition in Product Ranges, Mimeo,
University of Texas-Austin. http://www.eco.utexas.edu/
watson/.
Wolinsky, A., 1983, Retail Trade Concentration Due to Consumers Imperfect
Information, Bell Journal of Economics, 14, pp. 275282.
Wooldridge, J. M., 2002, Econometric Analysis of Cross Section and Panel Data, (MIT
Press, Cambridge, Massachusetts, U.S.A.).
Supporting Information
Additional supporting information may be found in the online version of
this article:
APPENDIX
A(i). Generating selection corrections.
A(ii). Correcting the standard errors.
Please note: Wiley-Blackwell is not responsible for the content or
functionality of any supporting materials supplied by the authors. Any
queries (other than missing material) should be directed to the corresponding author for the article.
r 2009 The Authors. Journal compilation r 2009 Blackwell Publishing Ltd. and the Editorial Board of The Journal of Industrial
Economics.