Embed Size (px)
Citation preview
Ž .Journal of Urban Economics 50, 191�229 2001
doi:10.1006�juec.2001.2230, available online at http:��www.idealibrary.com on
The Determinants of Agglomeration1
Stuart S. Rosenthal
Department of Economics, Center for Policy Research, Syracuse University, Syracuse,New York 13244-1020
E-mail: [email protected]
and
William C. Strange
Faculty of Commerce and Business Administration, University of British Columbia,Vancouver, British Columbia V6T 1Z2, Canada
E-mail: [email protected]
Received May 23, 2000; revised April 23, 2001; published online July 20, 2001
This paper examines the microfoundations of agglomeration economies for U.S.manufacturing industries. Using industries as observations, we regress the Ellison�
Ž .Glaeser G. Ellison and E. Glaeser, 1997, J. Polit. Econ. 105, 889�927 measure ofspatial concentration on industry characteristics that proxy for the presence of knowl-edge spillovers, labor market pooling, input sharing, product shipping costs, and naturaladvantage. The analysis is conducted separately at the zipcode, county, and state levels.Results indicate that proxies for labor market pooling have the most robust effect,positively influencing agglomeration at all levels of geography. Proxies for knowledgespillovers, in contrast, positively affect agglomeration only at the zipcode level. Relianceon manufactured inputs or natural resources positively affects agglomeration at the statelevel but has little effect on agglomeration at lower levels of geography. The same istrue for the perishability of output, a proxy for product shipping costs. � 2001 Academic
Press
1. INTRODUCTION
A growing empirical literature has established that the spatial concentrationŽof manufacturing activity enhances productivity and growth e.g., Moomaw
Ž .An earlier version of this article was published on IDEAL First www.idealibrary.com but wasretracted because it contained inaccurate estimates caused by a programming error. The articlepublished here contains the corrected estimates.
1 We gratefully acknowledge the financial support of the Social Sciences and HumanitiesResearch Council of Canada, the UBC Centre for Real Estate and Urban Land Economics, the RealEstate Foundation of British Columbia, and the Center for Policy Research at Syracuse University.We thank David Audretsch for providing us with data on innovations and Jan Brueckner, VernonHenderson, and two anonymous referees for helpful comments. We also thank Peter Howe forvaluable research assistance and Esther Gray for help in preparing the manuscript.
191
0094-1190�01 $35.00Copyright � 2001 by Academic Press
All rights of reproduction in any form reserved.
ROSENTHAL AND STRANGE192
� � � � � � � �18 , Sveikauskas 23 , Nakamura 19 , Henderson 12 , and Ciccone and Hall� �.3 . These studies show that localization economies�economies of scalearising from spatial concentration of activity within industries�are of particu-lar importance. Urbanization economies�economies of scale arising from citysize itself�although important, have smaller effects on productivity. Glaeser,
� � � �Kallal, Scheinkman, and Shleifer 9 and Henderson, Kuncoro, and Turner 13demonstrate the importance of these sorts of increasing returns. Ellison and
� �Glaeser 5 establish that the level of agglomeration varies considerably acrossindustries, as does the tendency of an industry to coagglomerate with otherindustries.
This literature has had relatively little to say about the causes of agglomera-� �tion. Two notable exceptions are Audretsch and Feldman 1 and Dumais,
� �Ellison, and Glaeser 4 . Audretsch and Feldman use a spatial Gini coefficientto measure geographic concentration. They show that innovative activity issubstantially more concentrated than overall production and that industries thatemphasize research and development tend to be more spatially concentrated.2
� �Dumais et al. 4 look at the microfoundations of agglomeration economies byconsidering which industries coagglomerate. They find that industries withsimilar labor mixes enjoy the largest benefit from proximity, suggesting theimportance of labor market pooling.
In contrast, theoretical work in this area has had much more to say about thecauses of agglomeration. It has been demonstrated that agglomeration economies
Ž � �.can arise from labor market pooling Helsley and Strange 11 , input sharingŽ � �. Ž � �.Goldstein and Gronberg 10 , and knowledge spillovers Glaeser 7 . See
� �Quigley 21 for a survey of the extensive theoretical literature on the micro-foundations of agglomeration economies.
This paper bridges the empirical and theoretical literatures. Utilizing the� �Ellison and Glaeser 5 index, we measure the level of spatial concentration
among manufacturing industries at the zipcode, county, and state levels in thefourth quarter of 2000. The Ellison�Glaeser index depends on both thegeographic distribution of employment and the intraindustry allocation ofemployment to establishments. The expected value of the Ellison�Glaeserindex is zero when the spatial allocation of employment is random. Thus, theindex has the appealing feature of permitting comparisons between the actualpattern of spatial concentration and the concentration that would be expected toarise from a random allocation of employment.3
Matching geographic concentration measures with data on industry character-istics, we seek to explain differences in the spatial concentration of industries.
2 � �A related result is obtained by Jaffe et al 14 , who identify a ‘‘paper trail’’ of knowledgespillovers in the location of patent citations. They show that patent citations are highly spatiallyconcentrated, with citations 5 to 10 times as likely to come from the same SMSA as control patents.
3 As will become apparent, the Ellison�Glaeser index converges to the more widely knownspatial Gini measure of agglomeration as the industry structure approaches that of a perfectlycompetitive market.
AGGLOMERATION DETERMINANTS 193
We focus on the three microfoundations of agglomeration that have been mostprominent in the theoretical literature: knowledge spillovers, labor marketpooling, and input sharing. Our approach is to regress the Ellison�Glaeserlocalization index on industry characteristics that proxy for the three causes oflocalization and on controls for product shipping costs and natural advantage.The regressions are carried out using 4-digit manufacturing industries asobservations. All of the regressions are carried out separately for concentrationat the zipcode, county, and state levels, since the causes of agglomeration couldwell differ at different levels of geographic aggregation.
Results provide evidence of the importance of all three sources of localiza-tion economies. The evidence is strongest for labor market pooling, withproxies having a positive impact on agglomeration at all levels of geography.The proxies for knowledge spillovers impact agglomeration positively only atthe zipcode level. Reliance on manufactured inputs or natural resources posi-tively affects agglomeration at the state level but has little effect on agglomera-tion at lower levels of geography. The same is true for the perishability ofoutput, a proxy for product shipping costs. In contrast, reliance on serviceinputs reduces state-level agglomeration. Taking all of these results together, an
Žinteresting pattern emerges, with shipping-oriented attributes manufactured.inputs, resources, perishability influencing agglomeration at the state level,
knowledge spillovers impacting highly localized agglomeration, and laborimpacting agglomeration at all levels of geography. These findings are largely
Ž .robust, holding for both ordinary least squares OLS and 2-digit standardŽ .industry classification SIC fixed-effect specifications, as well as for alterna-
Ž .tive Metropolitan Statistical Area MSA based measures of geography, andwhen industries are aggregated from the 4-digit to the 3-digit level.
The remainder of the paper is organized as follows. Section 2 considers thedegree to which industries agglomerate. Section 3 looks at the determinants ofan industry’s agglomeration. Section 4 concludes the paper.
2. THE EXTENT OF AGGLOMERATION
2.1. An Index of Agglomeration
This section addresses the degree to which industries agglomerate. There area number of statistics that one might employ to characterize the degree ofagglomeration. A natural candidate is the spatial Gini coefficient, defined as
Ž .2G � Ý x � s , where x is location i’s share of total employment and s isi i i i i
the location’s share of employment in a particular industry. This statistic is� � � �employed by Krugman 16 and Audretsch and Feldman 1 , among others. It
takes on a value of zero when an industry is allocated across space in exactlythe same way as for total employment. It takes on a value close to oneŽ .depending on the size of the industry itself when the industry is completelyconcentrated in one location.
ROSENTHAL AND STRANGE194
� �As Ellison and Glaeser 5 note, however, G � 0 does not necessarily implythat the industry in question is overconcentrated. Suppose that an industry ismade up of a small number of large plants and that there is no agglomerativeforce�either an externality or a natural advantage�leading to concentration.In this case, G will take on a large value simply because of the industrialorganization of the industry. In Ellison and Glaeser’s metaphor, tossing threedarts will leave most of the dartboard without any darts. The spatial Ginicoefficient, therefore, does not distinguish random concentration arising fromindustrial structure from concentration arising from agglomerative externalitiesor natural advantage.
To address this problem, Ellison and Glaeser propose the following index ofconcentration:
G � 1 � Ý x 2 HŽ .i i� � . 2.1Ž .21 � Ý x 1 � HŽ .Ž .i i
H � Ý z 2 is a Herfindahl index of the J plants in the industry, with zj j j
representing the employment share of the jth plant. For a perfectly competitiveindustry with a large number of small plants, H approaches zero and �
Ž 2 . 4approaches G� 1 � Ý x . In this case, G measures spatial concentrationi i
without any contamination associated with industrial organization. More gener-ally, � takes on a value of zero when an industry is as concentrated as onewould expect from a random location process, while a positive value of �indicates excess concentration. As Ellison and Glaeser take pains to point out,however, a positive � does not necessarily indicate that agglomerative external-ities are present. Instead, agglomerative externalities and natural advantage arein a sense observationally equivalent. Observing that the industry is concen-trated does not identify the cause of the concentration.
2.2. Which Industries Agglomerate?
We compute the Ellison�Glaeser index using information from Dun andŽ .Bradstreet D&B , included in the IMarket Inc. MarketPlace database for the
fourth quarter of 2000.5 The complete version of the data set contains establish-
4 Ž 2 . Ž .The 1 � Ý x term is included in order that the index have the property that E � � 0 wheni iŽ � �neither agglomerative spillovers nor natural advantage are present see Ellison and Glaeser 5 for
. Ž 2 .details . For the state, county, and zipcode levels that we consider, 1 � Ý x is close to one,i i
taking on values of 0.9997, 0.9954, and 0.9578, respectively.5 IMarket Inc. is a commercial data vendor. IMarket obtains the core data in the MarketPlace file
from Dun and Bradstreet, another commercial data vendor, then matches the D&B data with a widevariety of other data from other data vendors, and packages all of these data together in theMarketPlace file. The analysis in this paper is based solely on the D&B portion of the MarketPlacedatabase. In addition, although earlier versions of this paper were based on data from 1996, wefocus here on data from the fourth quarter of 2000. This is because representatives at IMarket Inc.advised us that the more recent data is of higher quality and somewhat more complete.
AGGLOMERATION DETERMINANTS 195
ment-level information on over 12 million establishments in the United States.We utilized a more manageable and affordable version of the data set in whichthe data were aggregated up to the zipcode level.6 In phone conversations withanalysts at D&B, we were advised that firms requesting not to be in thedatabase are omitted from the data file. Partly for that reason, the D&Bdatabase, while extensive, does not contain the entire universe of establishmentsin the United States. Nevertheless, the D&B analysts felt that the omissionsfrom the data set are sufficiently random that the D&B database is representa-tive of the spatial distribution of establishments in the United States.7
We calculate the location statistic � at the state, county, and zipcode levelsseparately for manufacturing industries using three different definitions ofindustries based on 2-, 3-, and 4-digit SIC codes. As is apparent in Table 1a, foreach level of geography, the average level of agglomeration increases as onegoes from 2- to 3-digit industries and from 3- to 4-digit industries. This occursbecause as industries become aggregated into ever broader and fewer cate-gories, spatial patterns of establishment locations eventually approach that ofthe entire economy, causing G and � to shrink toward zero.8 For this reason,the remainder of our discussion is based primarily on 4-digit-level industries,though we will on occasion examine features of 2- and 3-digit-level industriesfor comparison.
Focusing on the 459 4-digit manufacturing industries, at the state level the9mean level of agglomeration, � , is 0.0485. At the county and zipcode levelss
the means are � � 0.0193 and � � 0.0101, respectively. Since � equals zeroc z
when an industry is as concentrated as a random allocation, whenever � � 0there is excess concentration while � � 0 implies an excess diffusion of
6 Ž .Additional details on the Dun and Bradstreet D&B MarketPlace file are provided at the Dunand Bradstreet web site, www.dnb.com. As described by Dun and Bradstreet, there are severalimportant benefits to firms from listing themselves in the D&B database and obtaining a D-U-N-Sidentification number. These benefits arise primarily because of the incredible size of the D&Bdata file. Because the D&B file is such an effective source of information on firms throughout theeconomy, businesses use the D&B file to do market analysis and search out potential tradingpartners. Individual firms therefore have an incentive to list themselves with D&B in much the wayfirms have an incentive to voluntarily list themselves in the yellow pages. In addition, DUNSidentification numbers are rapidly becoming a standard identification device in the economy, andmany companies including the Federal Government require that clients obtain a D-U-N-S number
� �as a precondition for engaging in trade. As noted in the D&B website, ‘‘It the D-U-N-S number isnow the standard for all United States Federal Government electronic commerce transactions tohelp streamline and reduce federal procurement costs.’’
7 Ž . Ž .In contrast, the Census of Manufacturing CM and County Business Patterns CBP , the dataŽ .sets used by Ellison and Glaeser 1997 , are designed as representative surveys. However, the CM
and CBP both suffer from restrictions on the type of firms and employment data reported, includingtop-coding problems. There is no top-coding in the IMarket database.
8 In the limit, with a single industry category, industry employment is coincident with the entiremanufacturing sector, G equals zero, H approaches zero, and � goes to zero.
9 See Table 1a for additional summary statistics.
ROSENTHAL AND STRANGE196
TABLE 1a
Summary Measures of Agglomeration among Manufacturing Industries at theSIC 2-Digit, 3-Digit, and 4-Digit Levels
Correlation with gamma at the
� Mean SD Min Max State level County level
2-DigitState 0.0241 0.0449 0.0017 0.1946 1.00County 0.0099 0.0284 0.0010 0.1300 0.91 1.00Zip code 0.0029 0.0088 0.0002 0.0399 0.91 1.00
3-DigitState 0.0433 0.0706 �0.0024 0.4078 1.00County 0.0163 0.0416 �0.0002 0.3718 0.78 1.00Zip code 0.0082 0.0342 �0.0009 0.3515 0.70 0.87
4-DigitState 0.0485 0.0717 �0.0709 0.4993 1.00County 0.0193 0.0379 �0.0131 0.3718 0.82 1.00Zip code 0.0101 0.0275 �0.0046 0.3515 0.58 0.73
� �employment. As noted by Ellison and Glaeser 5 , it is not obvious how todecide what levels of � constitute significant departures from a randomallocation. Looking at values of state-level � for notably concentrated indus-tries like computers and automobiles, they define � � 0.05 as highly concen-trated and � � 0.02 as not very concentrated. Applying these benchmarks tothe average values of � for our data, it is apparent that there is clear evidenceof excess concentration at the state level. Moreover, only 18 of the 459
Žstate-level � values are below zero implying excess dispersion in those.industries . At the county and zipcode levels, the average level of agglomera-
tion across the industries is much lower, but once again only a small number ofŽindustries have negative � values 9 industries at the county level and 9 at the
.zipcode level . Finally, observe that the correlation between the state- andcounty-level � among 4-digit industries is 82% while the correlation betweenstate- and zipcode-level � is only 58%. Together, these results and those abovesuggest that the process generating state-level agglomeration is different thanthe one generating agglomeration at the county and zipcode levels, a theme thatwill recur at various points in the discussion to follow.
� �Because Ellison and Glaeser 5 examined agglomeration only down to thecounty level, our measures of zipcode-level concentration are new to theliterature, and some discussion of the pattern of agglomeration at that level iswarranted, especially for those industries whose agglomeration has become partof the geographic folklore. One such industry is the carpet industry, SIC 2273,
� �the history of whose localization was discussed by Krugman 16 . This industry
AGGLOMERATION DETERMINANTS 197
shows a considerable degree of agglomeration at every level of geography, with� � 0.406, � � 0.089, and � � 0.048. The motor vehicle industry is denoteds c z
SIC 3711. It has � � 0.089, � � 0.020, and � � 0.0027. In the case of motors c z
vehicles, there is excess agglomeration at the state level, but much lessagglomeration at the county level and little excess agglomeration at the zipcodelevel. Based on these comparisons and the results in Table 1a, it is apparent thatthere is always less agglomeration at the zipcode level, and notoriouslyagglomerated industries may not even appear to be agglomerated at a microgeo-graphic level, at least relative to a random allocation of employment acrossspace.
Table 1b lists the level of agglomeration at the state, county, and zipcodeŽ .levels for all twenty 2-digit industries. Although Tobacco SIC 21 and Textiles
Ž .SIC 22 display considerable excess agglomeration, especially at the statelevel, most of the industries display relatively little excess agglomeration,although all of the � values are positive. Of course, as noted above, the highdegree of industry aggregation at the 2-digit level obscures much of thevariation in spatial concentration across industries.
TABLE 1b
Agglomeration of Manufacturing Industries at the SIC 2-Digit Level
SIC Definition State � County � Zip code �
20 Food and kindred products 0.00347 0.00119 0.0002921 Tobacco manufactures 0.19457 0.13002 0.0398922 Textile mill products 0.09410 0.00601 0.0017723 Apparel and related products 0.01159 0.00653 0.0018424 Lumber and wood products, except furniture 0.01168 0.00284 0.0003425 Furniture and fixtures 0.01212 0.00297 0.0007426 Paper and allied products 0.00844 0.00213 0.0003527 Printing, publishing, and allied products 0.00527 0.00264 0.0003928 Chemicals and allied products 0.01047 0.00369 0.0006229 Petroleum refining and related products 0.03605 0.01040 0.0042830 Rubber and miscellaneous plastics products 0.00385 0.00102 0.0002331 Leather and leather products 0.01513 0.00640 0.0029832 Stone, clay, glass, and concrete products 0.00357 0.00209 0.0005233 Primary metal products 0.01438 0.00202 0.0004134 Fabricated metal products, except machinery
and transportation equipment 0.00447 0.00095 0.00021
35 Machinery, except electrical 0.00170 0.00112 0.0002936 Electrical and electronic machinery,
equipment, and supplies 0.00869 0.00352 0.0005037 Transportation equipment 0.02203 0.00462 0.0008438 Scientific and professional instruments;
photographic and optical goods; watches 0.01453 0.00429 0.0001839 Miscellaneous manufactured commodities 0.00666 0.00306 0.00055
ROSENTHAL AND STRANGE198
Table 1c provides a sharper picture by listing the 10 most concentratedindustries at the state, county, and zipcode levels for the 4-digit industries. Thistable has a number of interesting implications. First, some of the most agglom-erated industries may well be agglomerated because of natural advantage rather
Ž . Žthan because of a spatial externality. Fur goods SIC 2371 and Cigarettes SIC.2111 are examples of this. Second, many of the agglomerated industries are
the kinds of manufacturing industries where one might expect agglomerationŽ .economies to be important. Guided missiles & space vehicles SIC 3761 and
Ž .Office machines SIC 3579 are examples of this. Third, although there aresome industries that are highly agglomerated at more than one level ofgeography, for the most part the lists are distinct. This finding provides furthersupport for the idea that different processes generate agglomeration at differentlevels of geography.
3. THE DETERMINANTS OF AGGLOMERATION
3.1. Overview
The central goal of this section is to evaluate the degree to which agglomera-tive externalities explain interindustry differences in spatial concentration.Accordingly, our strategy is to regress � on proxies for three key sources ofagglomerative spillovers: knowledge spillovers, labor market pooling, and inputsharing. We also provide controls for natural advantages and product shippingcosts. Summary statistics and data sources are provided in Table 2a at the4-digit level for the manufacturing sector.
3.2. Controls for Natural Advantage and Transportation Costs
It has long been recognized that natural advantages can affect the locationdecisions of firms because of both the cost of shipping inputs to the factory andthe cost of shipping output to the market. From that observation, it is a shortstep to recognizing that natural advantage can also influence an industry’s
� �spatial concentration. Kim 15 estimates a state-level Rybczynski equationrelating employment to factor endowments, assuming that all factors of produc-tion are immobile, including labor. He argues that the residuals in this estima-tion are upper bounds on the strength of agglomeration economies. In a similar
� �way, Ellison and Glaeser 6 employ predicted state-level employment variablesto account for the importance of natural advantage in agglomeration. Both Kim� � � �15 and Ellison and Glaeser 6 conclude that natural advantage is important.
Ž .We use several variables from the 1992 Bureau of Economic Analysis BEAinput�output tables to control for the importance of natural advantages associ-ated with proximity to inputs. The variables Energy per $ shipment, Naturalresources per $ shipment, and Water per $ shipment measure energy inputcost, the cost of natural resources, and water-related costs respectively as
AGGLOMERATION DETERMINANTS 199
TA
BL
E1c
The
10M
ostA
gglo
mer
ated
Man
ufac
turi
ngIn
dust
ries
atth
eSI
C4-
Dig
itL
evel
Zip
code
leve
lC
ount
yle
vel
Stat
ele
vel
SIC
SIC
desc
ript
ion
�SI
CSI
Cde
scri
ptio
n�
SIC
SIC
desc
ript
ion
�
2371
Fur
good
s0.
352
2371
Fur
good
s0.
372
2397
Schi
ffli
mac
hine
embr
oide
ries
0.49
937
61G
uide
dm
issi
les
&sp
ace
0.26
023
97Sc
hiff
lim
achi
neem
broi
deri
es0.
251
3761
Gui
ded
mis
sile
s&
spac
e0.
434
vehi
cles
,par
tsve
hicl
es,p
arts
3579
Off
ice
mac
hine
s&
part
s0.
145
3761
Gui
ded
mis
sile
s&
spac
e0.
239
2284
Thr
ead
and
hand
wor
kya
rns
0.41
3ve
hicl
es,p
arts
2087
Flav
orin
gex
trac
ts&
syru
ps0.
142
2874
Phos
phat
icfe
rtili
zers
0.21
523
71Fu
rgo
ods
0.40
831
49Fo
otw
ear,
exce
ptru
bber
,n.e
.c.
0.13
938
61Ph
otog
raph
iceq
uipm
ent
0.19
122
73C
arpe
tsan
dru
gs0.
406
and
supp
lies
2335
Wom
ens’
and
mis
ses’
dres
ses
0.11
821
11C
igar
ette
s0.
191
2084
Win
es,b
rand
y,an
dbr
andy
spir
its0.
372
2381
Fabr
icdr
ess
and
wor
kgl
oves
0.11
431
49Fo
otw
ear,
exce
ptru
bber
,n.e
.c.
0.18
522
51W
omen
s’ho
sier
y,ex
cept
sock
s0.
371
3764
Mis
sile
and
rock
eten
gine
s0.
111
2043
Cer
ealb
reak
fast
food
s0.
180
3533
Oil
&ga
sfi
eld
equi
pmen
t0.
339
&pa
rts
3676
Ele
ctro
nic
resi
stor
s0.
086
2335
Wom
en’s
,mis
ses’
,and
juni
ors’
0.17
824
36So
ftw
ood
vene
eran
dpl
ywoo
d0.
328
3844
X-r
ayap
para
tus,
tube
s,&
part
s0.
084
2841
Soap
&de
terg
ents
0.16
721
41M
anuf
actu
red
toba
cco
0.30
5
ROSENTHAL AND STRANGE200
TA
BL
E2a
4-D
igit
Man
ufac
turi
ngIn
dust
ries
:Def
initi
ons,
Dat
aSo
urce
s,an
dSe
lect
edSu
mm
ary
Stat
istic
sfo
rE
xpla
nato
ryV
aria
bles
No. of
Var
iabl
ena
me
Def
init
ion
Dat
aso
urce
and
tim
epe
riod
obs.
Mea
nS
DM
inM
ax
�4
�3
Inno
vati
ons
from
firm
sN
umbe
rof
new
prod
ucts
in19
82tr
ade
1982
U.S
.SB
AIn
nova
tion
Dat
a45
97.
22�
101.
72�
100.
0000
0.01
46w
ith
few
erth
anm
agaz
ines
for
firm
sw
ith
�50
0B
ase
and
1992
Ann
ual
Surv
ey50
0w
orke
rsem
ploy
ees
divi
ded
bydo
llar
valu
eof
Man
ufac
ture
rsfr
omN
BE
RŽ
.pe
r$
ship
men
tof
ship
men
tsin
$1,0
00,0
00�
4�
3In
nova
tion
sfr
omfi
rms
Num
ber
ofne
wpr
oduc
tsin
1982
trad
e19
82U
.S.
SBA
Inno
vati
onD
ata
459
7.55
�10
1.84
�10
0.00
000.
0188
wit
hm
ore
than
mag
azin
esfo
rfi
rms
wit
h�
500
Bas
ean
d19
92A
nnua
lSu
rvey
500
wor
kers
empl
oyee
sdi
vide
dby
doll
arva
lue
ofM
anuf
actu
rers
from
NB
ER
Ž.
per
$sh
ipm
ent
ofsh
ipm
ents
in$1
,000
,000
Shi
pmen
tsne
tof
inpu
tsS
hipm
ents
�M
ater
ial
cost
s19
92A
nnua
lSu
rvey
of45
986
.270
.824
.097
5.6
aŽ
.pe
rw
orke
rin
$1,0
00,0
00di
vide
dby
Man
ufac
ture
rsfr
omN
BE
RŽ
.to
tal
empl
oyee
sin
1,00
0Ž
Man
ager
ial
shar
eT
otal
empl
oyee
s�
Pro
duct
ion
1992
Ann
ual
Surv
eyof
459
0.28
690.
1219
0.07
810.
8270
a.
ofw
orke
rsem
ploy
ees
divi
ded
byto
tal
Man
ufac
ture
rsfr
omN
BE
Rem
ploy
ees
Sha
reof
wor
kers
wit
hS
hare
ofem
ploy
edin
divi
dual
sw
ith
1995
Con
sum
erP
opul
atio
n42
70.
0092
0.02
030.
0000
0.17
65P
h.D
.or
prof
essi
onal
Ph.
D.o
rpr
ofes
sion
alde
gree
Surv
eyM
arch
file
ade
gree
from
Uni
con
Inc.
Sha
reof
wor
kers
wit
hS
hare
ofem
ploy
edin
divi
dual
sw
ith
1995
Con
sum
erP
opul
atio
n42
70.
0332
0.03
420.
0000
0.16
37a
Mas
ter’
sde
gree
Mas
ter’
sde
gree
Surv
eyM
arch
file
from
Uni
con
Inc.
Sha
reof
wor
kers
wit
hS
hare
ofem
ploy
edin
divi
dual
sw
ith
1995
Con
sum
erP
opul
atio
n42
70.
1198
0.06
860.
0000
0.34
55a
Bac
helo
r’s
degr
eeB
ache
lor’
sde
gree
Surv
eyM
arch
file
from
Uni
con
Inc.
AGGLOMERATION DETERMINANTS 201
Man
ufac
ture
din
puts
Cos
tof
inpu
tsob
tain
edfr
om19
92B
EA
inpu
t�ou
tput
tabl
es45
90.
3502
0.11
660.
0063
0.65
72pe
r$
ship
men
tm
anuf
actu
ring
firm
sŽ
.S
IC20
thro
ugh
39pe
rdo
llar
ship
men
tsN
onm
anuf
actu
red
Cos
tof
mat
eria
lsot
her
than
man
ufac
ture
d19
92B
EA
inpu
t�ou
tput
tabl
es45
90.
0819
0.06
75�
0.11
190.
6557
inpu
tspe
rin
puts
,ene
rgy,
natu
ral
reso
urce
s,an
dan
d19
92A
nnua
lSu
rvey
ofb
$sh
ipm
ent
wat
erpe
rdo
llar
ofsh
ipm
ents
Man
ufac
ture
rsfr
omN
BE
RN
atur
alre
sour
ces
Cos
tof
natu
ral
reso
urce
inpu
tspe
rdo
llar
1992
BE
Ain
put�
outp
utta
bles
459
0.03
570.
1026
0.00
00.
7923
expe
nses
per
ofsh
ipm
ents
$sh
ipm
ent
Ene
rgy
expe
nses
Cos
tof
ener
gyin
puts
per
doll
ar19
92B
EA
inpu
t�ou
tput
tabl
es45
90.
0208
0.02
320.
0022
0.24
05pe
r$
ship
men
tof
ship
men
tsW
ater
expe
nses
per
Cos
tof
wat
erin
puts
per
doll
ar19
92B
EA
inpu
t�ou
tput
tabl
es45
90.
0018
0.00
290.
0000
0.02
38$
ship
men
tof
ship
men
tsIn
vent
orie
sD
olla
rva
lue
ofen
d-of
-yea
rin
vent
orie
s19
92A
nnua
lSu
rvey
of45
90.
1446
0.06
340.
0209
0.50
53pe
r$
ship
men
tpe
rdo
llar
ofsh
ipm
ents
Man
ufac
ture
rsfr
omN
BE
Rc
Ž.
nonp
eris
habi
lity
aS
eeT
able
2bfo
rad
diti
onal
deta
ils.
bS
eeA
ppen
dix
Tab
leA
-2fo
rad
diti
onal
deta
ils.
c See
Tab
le2c
for
addi
tion
alde
tail
s.
ROSENTHAL AND STRANGE202
fractions of the value of shipments. These variables were available at the 4-digitlevel.10 To the extent that industries concentrate because of a desire to locateclose to the sources of their energy, natural resource, and water related inputs,we expect the coefficients on these variables to be positive.11
It has also long been recognized that the cost of transporting output canaffect location decisions. A tempting approach to control for such effects wouldbe to use readily available BEA data on actual product shipping costs byindustry. This, however, would not be suitable because industries for which theper mile cost of shipping the product is high would locate so as to minimizedistances to their markets and the related shipping costs. Instead, we proxy forthe per mile cost of shipping the product using Inventories per $ of shipment,defined as the value of end-of-year inventories divided by the value ofshipments. Industries that produce highly perishable products face high productshipping costs per unit distance and, therefore, will seek to locate close to theirmarkets, ceteris paribus. With multiple markets, such industries will tend todisplay less agglomeration. Conversely, industries that produce nonperishableproducts face lower product shipping costs and should display more agglomera-tion.12
Table 2b provides compelling support for using Inventories to proxy forperishability. The table displays the ten 4-digit industries with the highestvalues of Inventories and the ten industries with the lowest values of Invento-ries. Industries with very low inventory�shipment ratios include meat packingplants, newspapers, milk and cream, and other clearly perishable products.Industries with the highest inventory�shipment ratios include aircraft, wine andother liquors, machinery, and other clearly nonperishable products. These dataon Inventories were obtained from the 1992 Annual Survey of Manufactures
Ž .which was obtained at the NBER website www.nber.org . To the extent thatindustries concentrate when per-mile costs of shipping the product are low, weexpect the coefficients on this variable to be positive.
10 The URL for the 1992 BEA Input�Output file is http:��www.bea.doc.gov�bea�dn2�i-o.htm.The file is zipped and downloadable. The file name is ‘‘1992 Benchmark I�O Table Six-DigitTransactions’’ and contains the make table, use table, direct requirements coefficients table, and
Žestimates by commodity of transportation costs and of wholesale and retail margins 498-industry.detail . Once unzipped there are a number of files, including instructions on how to make an extract
from the data sets. In addition, the input�output tables are organized by product type rather than bySIC category. We obtained a concordance from BEA to match the product types to 4-digit SICcategories.
11 A detailed description of the SIC categories used to construct the category Natural Resourcesis provided in the appendix. Note that coal, crude petroleum, and natural gas are included in theEnergy variable rather than in Natural resources.
12 Of course, other factors besides perishability of the product affect optimal inventory�shipmentratios. For example, internal economies of scale create incentives for firms to produce in bulk andstockpile output for later shipment. It is worth pointing out that internal economies of scale alsodirectly influence agglomeration through their impact on the size distribution of establishments.However, that is already dealt with through the inclusion of the Herfindahl index in � .
AGGLOMERATION DETERMINANTS 203
TABLE 2b
Industries with the Lowest and Highest Inventory-to-Shipment Ratios
Lowest inventory-to-shipment ratio Highest inventory-to-shipment ratio
SIC SIC description Inv�ship SIC SIC description Inv�ship
2011 Meat packing plants 0.021 3721 Aircraft 0.5052813 Industrial gases 0.021 2084 Wines, brandy, and brandy 0.470
spirits2711 Newspapers 0.022 2085 Distilled and blended liquors 0.3472026 Fluid milk and cream, and 0.023 3541 Machine tools, metal cutting 0.342
related products types and parts2051 Bread and other bakery 0.025 3533 Oil and gas field equipment 0.342
products and parts2096 Potato chips, corn chips, 0.027 3262 Vitreous china table and 0.323
and similar products kitchenware3711 Motor vehicles and 0.027 2063 Beet sugar 0.314
passenger car bodies2021 Creamery butter 0.032 3542 Machine tools, metal 0.293
forming types2015 Poultry slaughtering and 0.035 3511 Steam, gas, and hydraulic 0.291
processing turbines2082 Malt beverages 0.035 3356 Extruded nonferrour metal 0.291
mill products
3.3. Controls for Agglomerative Externalities
Two variables are used to proxy for input sharing. Manufactured inputs per$ of shipment is the ratio of the cost of inputs purchased from the manufactur-ing sector�SIC codes 20 to 39�to the value of shipments. This variable wasobtained from the 1992 BEA input�output tables and measures the relativeimportance of manufactured inputs for the industry. Among industries forwhich Manufactured inputs is large, the gains from sharing inputs are likely toalso be large, creating incentives to concentrate spatially. For that reason, weexpect Manufactured inputs to have a positive coefficient. Similarly, we alsoinclude a variable Nonmanufactured inputs per $ of shipment, where Nonman-ufactured inputs is the value of materials other than those already notedŽ . 13manufactured inputs, energy, natural resources, and water . This category ofinputs includes such things as legal services, accounting and financial services,insurance, communication, repair, and janitorial services. There are two impor-
13 Nonmanufactured inputs is measured as a residual and is calculated by subtracting our otherinput measures and value added per dollar of shipments from unity since shipments are approxi-mately equal to value added plus expenditures on materials. A detailed list of the SIC categoriesthat comprise the Nonmanufactured inputs is provided in the appendix. In addition, data on thevalue added and shipments used to construct Nonmanufactured inputs was obtained from the 1992Annual Survey of Manufactures while the other variables used to construct Nonmanufacturedinputs were obtained from the 1992 BEA input�output tables as noted above.
ROSENTHAL AND STRANGE204
tant differences between manufactured and nonmanufactured inputs. First, scaleeconomies are likely to be stronger for manufactured inputs. Second, manufac-tured inputs are likely to exhibit greater industry specificity. For both of thesereasons, there is less reason for industries that rely heavily on nonmanufacturedinputs to agglomerate. Accordingly, we expect Nonmanufactured inputs tohave less impact on agglomeration than does Manufactured inputs.
The variable used to proxy for the importance of knowledge spillovers isInnovations per $ of shipment. Innovations are defined as the number of newproducts advertised in trade magazines in 1982, the only year for which suchdata were readily available. An essential input for innovation is new knowl-edge. In that regard, innovative activity is related to the importance of knowl-edge spillovers. In addition, although our innovation variable predates ouragglomeration measures by 18 years, it seems likely that most industries forwhich innovation was important in 1982 would continue to place importance oninnovation in the 1990s. Accordingly, we anticipate that Innovations per $ ofshipment will have a positive effect on our industry concentration measures.The innovation data were collected by the U.S. Small Business Administrationas part of its Innovation Database and were available at the 4-digit level. See
� � 14Audretsch and Feldman 1 for additional details on these data.There is reason to believe that the operation of knowledge spillovers is linked
� �to the industrial organization of an industry. Saxenian 22 , for instance, arguesthat the open managerial structure of the high-technology firms in SiliconValley gave it an advantage over the relatively closed structure typical of thelarge high-tech firms populating Boston’s Route 128. Consistent with that
� �argument, Rosenthal and Strange 20 find that smaller establishments have alarger effect on the attractiveness of a location than do larger establishments,
� �ceteris paribus. In addition, Audretsch, van Leeuwen, Menkveld, and Thurik 2find that small establishments are more productive than large establishments,ceteris paribus. To allow for the possibility that innovativeness has differenteffects on agglomeration depending on the size of the firms that innovate, wepartition the Innovations variable into innovations at firms with fewer than500 employees and innovations at firms with more than 500 employees.15
14 We are grateful to David Duretsch for providing these data.15 Two other variables were considered but rejected as proxies for the importance of knowledge
spillovers. The first is the number of patents. However, patents are not really the same asinnovations. In some industries, a single innovation can be associated with hundreds of patents. Inaddition, the U.S. Patent Office codes patents based on the product type, not the industry to whichthe innovating firm belongs. Thus, it is difficult to accurately match patent data to the SICdefinitions of industries. Another candidate variable as a proxy for the importance of informationspillovers would be industry expenditures on research and development. However, because manyinnovations are associated with business practice rather than the deliberate search for new productsor processes, this variable does not provide as precise a measure of the importance of informationspillovers as do the innovations. In addition, expenditures on research and development areindirectly related to the role of information spillovers in that they are an input rather than an output.
AGGLOMERATION DETERMINANTS 205
The most difficult of the Marshallian microfoundations to proxy is labormarket pooling. If pooling is possible, an industry benefits by agglomeratingbecause it is better able to hire workers with industry-specific skills. Theproblem in proxying for the importance of pooling in an industry is that it isdifficult to identify industry characteristics that are related to the specializationof the industry’s labor force. We therefore separately employ three differentproxies. The first is Net productivity, equal to the value of shipments less thevalue of purchased inputs, all divided by the number of workers in the industry.This measure of the productivity of labor is obtained by using the ASM data for1992 taken from the NBER website as described above. The second is the ratio
Ž .Management workers� Management � Production workers . This ‘‘brains tobrawn’’ variable measures the share of supervisory and support labor inproduction. If little of such labor is needed, then production is more likely to bea matter of routine, and specialized labor is likely to be less important. Thisvariable is also constructed using data from the 1992 ASM. The final approachto proxying for labor market pooling is to employ variables on workereducation, specifically the percentage of workers with Doctorates, Master’sDegrees, and Bachelor’s Degrees.16 These data are obtained from Consumer
Ž . 17Population Survey CPS data from 1995.It is worth noting that all of these proxies for the importance of labor market
pooling are positively correlated, as shown in Table 2c. For example, correla-tion between shipments net of inputs per worker and the other proxies for labormarket pooling range between 21 and 31%. Correlation between Managerialshare of workers and Share of workers with Master’s degrees is 53%. Giventhe strong positive correlation between these variables, the models to follow areall estimated separately for each of the three sets of labor market poolingproxies. In all cases these variables are expected to have positive coefficients.
16 It is important to note that while educated workers may indeed be specialized, these variablesdo not capture the degree to which less-educated workers may also have specialized industry-specific
Ž � � .skills i.e., Marshall’s 17 cutlery manufacturers .17 The CPS reports the industry of occupation for individual workers. We computed the
distribution of employed workers across such industry categories, and then matched industry codesŽto SIC categories using a correspondence table provided at the census website www.bls.census.
.gov� cps� bindcd.htm . It is worth noting that CPS industry codes correspond to 3-digit SIC codeswith the exception of two industry codes that match to 2-digit SIC codes, and one industry codethat matches directly to a 4-digit SIC code. In order to use these data for 4-digit-level analysis,therefore, we assigned the 3-digit SIC education values to 4-digit member subgroups in the SICclassification scheme. Unfortunately, this precludes using the education variables when 3-digit SICfixed effects are included in some of the models since the education variables do not vary within3-digit SIC classifications.
ROSENTHAL AND STRANGE206
TABLE 2caCorrelation Between Proxies for the Importance of Labor-Market Pooling
Share of Share of Share ofworkers with workers workers
Shipments net Managerial Ph.D. or with withof inputs per share of professional Master’s Bachelor’s
worker workers degree degree degree
Shipments net of inputs 1.00per worker
Managerial share 0.27 1.00of workers
Share of workers with 0.24 0.28 1.00Ph.D. or professionaldegree
Share of workers with 0.31 0.53 0.50 1.00Master’s degree
Share of workers with 0.21 0.57 0.45 0.56 1.00Bachelor’s degree
aSample size equals 427 4-digit industries.
3.4. Estimates of the Determinants of Agglomeration
The effect of agglomerative spillovers on the spatial concentration indexes ismeasured by estimating
� � � X � � , 3.1Ž .j , m m j , m
where � is the localization statistic for the mth industry at level ofj, m
geography j, X is the vector of industry characteristics with associatedm
coefficient vector � , and � is assumed to be an independent and identicallyj, mŽ .distributed error term. We estimate Eq. 3.1 separately for the three geographic
specifications, with � measured at the state, county, and zipcode levels.j
Before proceeding further, it is important to discuss identification. Becausethe role of natural advantages and product shipping costs in an industry is likelyto be exogenous to the level of agglomeration, coefficient estimates on thesevariables provide direct measures of their impact on concentration. For theremaining variables, the coefficients describe the equilibrium relationship be-tween industry characteristics and agglomeration: industry characteristics affectthe propensity to agglomerate, but agglomeration can influence industry charac-teristics. In both directions, however, these relationships are governed by thedegree to which agglomeration reduces costs. Specifically, agglomeration re-duces the cost of innovation by enhancing knowledge spillovers while also
AGGLOMERATION DETERMINANTS 207
reducing the cost of labor and intermediate inputs through labor market poolingand input sharing. Precisely for these reasons, industries sensitive to innovation,labor, and intermediate input costs are more likely to agglomerate. Thus,evidence of a positive relationship between agglomeration and these otherfactors confirms that tendencies to innovate, pool labor, and share inputs alllead to an increase in agglomeration.
Table 3a presents ordinary least-squares estimates of our model. As dis-cussed above, we estimate separate models for each level of geography�zipcode, county, and state�and for each set of labor-market poolingproxies�net shipments per worker, managerial share of workers, and educa-tion. In total, therefore, the table presents nine regressions, three for each levelof geography.
A set of results in Table 3a that warrants immediate discussion are theadjusted R2-values for each of the models. These range from near zero at thezipcode level of roughly 7% at the state level. On the surface, this suggests thatstate-level agglomeration is more closely related to agglomerative spilloversand natural advantages than are county- and zipcode-level agglomeration. Thisfinding will prove robust in the analyses to follow. At the same time, the verylow values for the adjusted R-squares suggest that our proxies for agglomera-tive spillovers and natural advantages explain only a fraction of the variation inagglomeration across industries. This raises the possibility that omitted industryattributes could bias our estimates.
To address that concern, Tables 3b and 3c provide a stringent set ofrobustness checks. Table 3b repeats the analyses in Table 3a but includes
Ž .2-digit SIC level fixed effects 20 in all , while Table 3c includes 3-digit SICŽ . 18level fixed effects 140 in all . With these fixed effects added to the models,
2 Ž .adjusted R -values range from 4 to 21.5% with 2-digit fixed effects Table 3bŽ .and from 28 to 40% with 3-digit fixed effects Tables 3c . Inclusion of these
fixed effects, therefore, controls for a host of potentially important omitteddeterminants of agglomeration. But, at the same time, it is important torecognize that the fixed effects potentially soak up much of the meaningfulvariation in the data, making identification difficult, especially when 140 fixedeffects are included in the model as in Table 3c. Bearing that tradeoff in mind,our discussion below emphasizes the OLS results in Table 3a but frequentreferences will also be made to the fixed-effects models as well.
An important result in Tables 3a, 3b, and 3c is the consistent evidence of apositive and significant influence of labor market pooling at all levels of
18 The latter model cannot be estimated when education is used to proxy labor-market poolingbecause the education variables are available up to the 3-digit level and, therefore, do not varywithin the 4-digit subclassifications.
ROSENTHAL AND STRANGE208
TA
BL
E3a
aO
LS
Gam
ma
Reg
ress
ions
for
Em
ploy
men
tatA
llE
stab
lishm
ents
atth
e4-
Dig
itSI
CL
evel
Zip
code
leve
lC
ount
yle
vel
Stat
ele
vel
Man
ager
ial
Man
ager
ial
Man
ager
ial
Net
labo
rsh
are
ofE
duca
tion
Net
labo
rsh
are
ofE
duca
tion
Net
labo
rsh
are
ofE
duca
tion
prod
uctiv
ityw
orke
rsof
wor
kers
prod
uctiv
ityw
orke
rsof
wor
kers
prod
uctiv
ityw
orke
rsof
wor
kers
Inno
vatio
nsfr
omfi
rms
�0.
8664
�1.
1384
�1.
2112
�1.
8782
�2.
3535
�2.
4088
�4.
1396
�4.
3866
�5.
0579
with
few
erth
an�
1.06
�1.
36�
1.34
�1.
72�
2.06
�1.
96�
2.01
�2.
08�
2.24
500
wor
kers
Inno
vatio
nsfr
omfi
rms
0.88
740.
9299
1.05
870.
7349
0.92
311.
0206
�0.
7731
�0.
3629
�0.
3011
with
mor
eth
an1.
171.
211.
320.
730.
880.
94�
0.41
�0.
19�
0.15
500
wor
kers
�5
�4
�4
Ship
men
tsne
tof
inpu
ts6.
300
�10
1.55
3�
102.
046
�10
per
wor
ker
3.27
6.05
4.24
Man
ager
ials
hare
0.01
690.
0210
�0.
0123
ofw
orke
rs1.
461.
33�
0.42
Shar
eof
wor
kers
with
�0.
0096
0.11
490.
0781
Ph.D
.or
prof
essi
onal
�0.
121.
040.
39de
gree
Shar
eof
wor
kers
with
0.12
850.
1662
0.37
76M
aste
r’s
degr
ee2.
472.
362.
92
AGGLOMERATION DETERMINANTS 209
Shar
eof
wor
kers
with
�0.
0308
�0.
0285
�0.
1236
Bac
helo
r’s
degr
ee�
1.23
�0.
84�
1.98
Man
ufac
ture
din
puts
0.01
640.
0101
0.00
690.
0229
0.00
130.
0011
0.11
990.
0794
0.09
53pe
r$
ship
men
t1.
220.
750.
491.
280.
070.
063.
562.
312.
69N
onm
anuf
actu
red
inpu
ts�
0.01
26�
0.01
95�
0.03
19�
0.03
33�
0.04
86�
0.07
40�
0.10
04�
0.11
71�
0.16
52pe
r$
ship
men
t�
0.65
�1.
00�
1.39
�1.
28�
1.81
�2.
38�
2.06
�2.
36�
2.89
Nat
ural
reso
urce
s0.
0046
0.01
030.
0050
�0.
0036
0.00
700.
0043
0.11
460.
1218
0.11
74ex
pens
espe
r0.
320.
710.
33�
0.19
0.35
0.21
3.21
3.32
3.10
$sh
ipm
ent
Ene
rgy
expe
nses
�0.
0387
�0.
0336
�0.
0449
�0.
0747
�0.
0838
�0.
0816
0.00
56�
0.04
860.
0531
per
$sh
ipm
ent
�0.
63�
0.53
�0.
62�
0.91
�0.
97�
0.83
0.04
�0.
310.
30W
ater
expe
nses
0.16
670.
2614
0.33
080.
3196
0.55
180.
6491
0.77
401.
0779
1.19
39pe
r$
ship
men
t0.
350.
550.
660.
500.
840.
950.
650.
890.
95In
vent
orie
s0.
0266
0.01
530.
0171
0.04
660.
0230
0.02
250.
1696
0.14
680.
1397
per
$sh
ipm
ent
1.27
0.73
0.77
1.67
0.79
0.75
3.23
2.74
2.53
Ž.
nonp
eris
habi
lity
2R
0.03
40.
016
0.03
00.
094
0.02
40.
049
0.10
40.
068
0.09
62
Adj
.R
0.01
50.
000
0.00
50.
076
0.00
40.
024
0.08
60.
050
0.07
2Sa
mpl
esi
ze45
945
942
745
945
942
745
945
942
7
at-
Rat
ios
are
belo
wth
eco
effi
cien
ts.
Con
stan
tsar
eno
tre
port
edin
orde
rto
cons
erve
spac
e.
ROSENTHAL AND STRANGE210
TA
BL
E3b
aSI
C2-
Dig
itFi
xed
Eff
ectG
amm
aR
egre
ssio
nsfo
rE
mpl
oym
enta
tAll
Est
ablis
hmen
tsat
the
4-D
igit
SIC
Lev
el
Zip
code
leve
lC
ount
yle
vel
Stat
ele
vel
Man
ager
ial
Man
ager
ial
Man
ager
ial
Net
labo
rsh
are
ofE
duca
tion
Net
labo
rsh
are
ofE
duca
tion
Net
labo
rsh
are
ofE
duca
tion
prod
uctiv
ityw
orke
rsof
wor
kers
prod
uctiv
ityw
orke
rsof
wor
kers
prod
uctiv
ityw
orke
rsof
wor
kers
Inno
vatio
nsfr
omfi
rms
�0.
4694
�0.
7951
�0.
5989
�1.
6353
�2.
0981
�1.
7462
�2.
8179
�3.
2661
�3.
0565
with
few
erth
an�
0.56
�0.
94�
0.65
�1.
49�
1.82
�1.
40�
1.42
�1.
61�
1.42
500
wor
kers
Inno
vatio
nsfr
omfi
rms
1.13
661.
3517
1.36
800.
9726
1.33
001.
3867
0.16
840.
5195
0.58
05w
ithm
ore
than
1.49
1.75
1.70
0.97
1.26
1.27
0.09
0.28
0.31
500
wor
kers
�4
�4
�4
Ship
men
tsne
tof
inpu
ts1.
049
�10
2.19
7�
102.
196
�10
per
wor
ker
4.34
6.89
3.80
Man
ager
ials
hare
0.04
210.
0389
0.03
57of
wor
kers
2.96
2.01
1.05
Shar
eof
wor
kers
with
0.04
770.
1695
0.16
47Ph
.D.o
rpr
ofes
sion
al0.
501.
310.
73de
gree
Shar
eof
wor
kers
with
0.20
430.
2003
0.50
22M
aste
r’s
degr
ee2.
942.
143.
09
AGGLOMERATION DETERMINANTS 211
Shar
eof
wor
kers
with
�0.
0013
0.02
94�
0.57
5B
ache
lor’
sde
gree
�0.
040.
74�
0.83
Man
ufac
ture
din
puts
0.00
46�
0.00
37�
0.00
550.
0131
�0.
0104
�0.
0102
0.05
910.
0351
0.04
32pe
r$
ship
men
t0.
29�
0.24
�0.
330.
64�
0.49
�0.
461.
580.
941.
12N
onm
anuf
actu
red
inpu
ts0.
0087
�0.
0035
�0.
0092
0.00
57�
0.01
76�
0.03
50�
0.02
19�
0.04
50�
0.06
12pe
r$
ship
men
t0.
40�
0.16
�0.
360.
20�
0.60
�1.
02�
0.43
�0.
87�
1.03
Nat
ural
reso
urce
sex
pens
es0.
0149
0.00
78�
0.00
570.
0257
0.00
49�
0.00
980.
1100
0.08
880.
0597
per
$sh
ipm
ent
0.85
0.44
�0.
311.
110.
20�
0.39
2.62
2.10
1.37
Ene
rgy
expe
nses
0.00
370.
0258
0.02
20�
0.00
200.
0161
0.02
140.
0186
0.03
480.
1125
per
$sh
ipm
ent
0.05
0.35
0.25
�0.
020.
160.
180.
110.
200.
55W
ater
expe
nses
0.44
710.
4918
0.69
670.
6364
0.63
961.
1237
2.17
952.
1767
2.88
53pe
r$
ship
men
t0.
840.
911.
220.
910.
871.
461.
711.
682.
16In
vent
orie
s�
0.00
01�
0.01
82�
0.01
560.
0031
�0.
0251
�0.
0283
0.11
690.
0893
0.09
06pe
r$
ship
men
t0.
00�
0.80
�0.
650.
10�
0.80
�0.
872.
171.
631.
60Ž
.no
nper
isha
bilit
y
2R
0.11
80.
098
0.11
30.
192
0.11
20.
138
0.25
70.
234
0.27
02
Rad
j.0.
060
0.03
90.
046
0.14
00.
053
0.07
30.
208
0.18
40.
215
Sam
ple
size
459
459
427
459
459
427
459
459
427
Num
ber
offi
xed
effe
cts
2020
2020
2020
2020
20
at-
Rat
ios
are
belo
wth
eco
effi
cien
ts.
Con
stan
tsar
eno
tre
port
edin
orde
rto
cons
erve
spac
e.
ROSENTHAL AND STRANGE212T
AB
LE
3ca
SIC
3-D
igit
Fixe
d-E
ffec
tGam
ma
Reg
ress
ions
for
Em
ploy
men
tatA
llE
stab
lishm
ents
atth
e4-
Dig
itSI
CL
evel
Zip
code
leve
lC
ount
yle
vel
Stat
ele
vel
Man
ager
ial
Man
ager
ial
Man
ager
ial
Net
labo
rsh
are
ofN
etla
bor
shar
eof
Net
labo
rsh
are
ofpr
oduc
tivity
wor
kers
prod
uctiv
ityw
orke
rspr
oduc
tivity
wor
kers
Inno
vatio
nsfr
omfi
rms
with
�0.
1646
�0.
2501
�1.
3302
�1.
4147
�2.
5586
�2.
5291
few
erth
an50
0w
orke
rs�
0.21
�0.
32�
1.17
�1.
22�
1.17
�1.
14In
nova
tions
from
firm
sw
ith1.
4545
1.55
491.
0547
1.17
84�
0.19
80�
0.12
37m
ore
than
500
wor
kers
2.16
2.26
1.06
1.16
�0.
10�
0.06
�4
�4
�4
Ship
men
tsne
tof
inpu
ts1.
211�
101.
784�
102.
195�
10pe
rw
orke
r3.
673.
662.
34M
anag
eria
lsha
reof
wor
kers
�0.
0116
�0.
0271
�0.
0661
�0.
65�
1.04
�1.
33M
anuf
actu
red
inpu
tspe
r0.
0062
�0.
0125
�0.
0049
�0.
0324
0.00
92�
0.02
45$
ship
men
t0.
32�
0.67
�0.
17�
1.18
0.17
�0.
47N
onm
anuf
actu
red
inpu
tspe
r�
0.01
80�
0.04
05�
0.05
84�
0.09
10�
0.10
94�
0.14
78$
ship
men
t�
0.72
�1.
64�
1.58
�2.
50�
1.54
�2.
14N
atur
alre
sour
ces
expe
nses
0.01
740.
0011
0.01
89�
0.00
620.
0736
0.03
94pe
r$
ship
men
t0.
840.
060.
62�
0.20
1.25
0.68
Ene
rgy
expe
nses
per
0.02
68�
0.02
060.
0402
�0.
0357
�0.
0064
�0.
1192
$sh
ipm
ent
0.35
�0.
270.
36�
0.31
�0.
03�
0.55
Wat
erex
pens
espe
r$
ship
men
t�
0.00
76�
0.03
050.
5661
0.53
283.
1269
3.08
74�
0.01
�0.
050.
680.
631.
961.
92In
vent
orie
spe
r$
ship
men
t0.
0005
�0.
0155
0.02
500.
0034
0.15
850.
1387
Ž.
nonp
eris
habi
lity
0.02
�0.
650.
710.
102.
352.
07
2R
0.59
40.
577
0.53
40.
516
0.51
70.
511
2R
adj.
0.39
90.
374
0.31
20.
285
0.28
60.
277
Sam
ple
size
459
459
459
459
459
459
Num
ber
offi
xed
140
140
140
140
140
140
effe
cts
at-
Rat
ios
are
belo
wth
eco
effi
cien
ts.
Con
stan
tsar
eno
tre
port
edin
orde
rto
cons
erve
spac
e.
AGGLOMERATION DETERMINANTS 213
geography, for all three proxies, and in both the OLS and fixed-effectsspecifications. The variable Shipments net of inputs per worker is alwayspositive and significant in all of the models; Managerial share of workers ispositive at the zipcode and county levels in the OLS and 2-digit fixed-effectmodels, though it is significant only for the 2-digit specification. Among theeducation variables, there is also a consistent pattern, with the Master’s degreevariable positive and at least marginally significant in all of the specifications.The consistency of these results provides strong evidence that labor marketpooling is associated with industrial agglomeration. That finding is consistent
Ž .with results from Dumais et al. 1997 who also report strong evidence of labormarket pooling.
The coefficients on Manufactured inputs are positive but insignificant in thezipcode and county models, providing at most weak evidence that industrieswith a propensity toward input sharing concentrate at these levels of geography.The state-level coefficients, on the other hand, are all positive and significant in
Ž .the OLS model Table 3a , though significance is reduced continuously as oneadds 2- and then 3-digit fixed effects to the model. Nevertheless, on balance,there is support for the idea that input sharing contributes to spatial agglomera-tion at the state level.
In contrast to the role of Manufactured inputs, the variable Nonmanufac-tured inputs has a negative coefficient in nearly all of the models and is
Ž .significant at the state level for the OLS Table 3a and 3-digit fixed-effectŽ .specifications Table 3c . Consistent with our priors, this suggests that the type
of inputs upon which an industry depends influences the propensity to agglom-erate. A reliance on manufactured inputs contributes to agglomeration. But,a reliance on service inputs�an important component of nonmanufacturedinputs�does not, perhaps because these inputs are produced under constantreturns or are not industry-specific and hence are available everywhere. Overall,
� �our results on input sharing are in the spirit of Marshall 17 .There is also suggestive evidence for the importance of knowledge spillovers,
but the evidence here is both mixed and weaker than for the other Marshallianmicrofoundations. At the county and state levels, Innovations from firms withmore than 500 workers is nearly always insignificant and in some instances hasa negative coefficient. However, at the zipcode level, large-firm innovation hasa positive coefficient in all of the different models, with the coefficient not
Ž .significant in the OLS specifications Table 3a , marginally significant in theŽ .2-digit fixed-effect specification Table 3b , and significant in the 3-digit
Ž .fixed-effect specification Table 3c . On the other hand, small-firm innovationhas consistently negative coefficients across the models, with the coefficientsignificant at higher levels of geography in the OLS specification. The resultthat large-firm innovation has a positive and significant effect only at thezipcode level is appealing given priors that knowledge spillovers attenuate
ROSENTHAL AND STRANGE214
rapidly. But the negative coefficients on small firm innovations are difficult toexplain, although these effects disappear with the inclusion of high-level fixedeffects.19 On balance, therefore, we characterize our results here as suggestingthat knowledge spillovers contribute to agglomeration at the local level, espe-cially when innovative activity is based in large, well-established firms. But thisconclusion should be viewed with caution, and further study is certainlywarranted.
The remaining variables in Tables 3a, 3b, and 3c proxy for the importance ofnatural advantages as discussed earlier. On the input side, it is notable thatindustries that rely heavily on natural resources exhibit greater agglomerationonly at the state level, with little effect at the zipcode and county levels.Specifically, the coefficients on the Natural resources variable are positive andsignificant at the state level but are insignificant at the other levels ofgeography. This result is quite apparent in the OLS and 2-digit fixed-effectmodels, but much less so in the 3-digit fixed-effect model. A similar resultholds for reliance on Water related resources, which is also positive andsignificant in the 2- and 3-digit fixed-effect models, but not significant in theOLS model. In contrast, Energy is not significant in any of the models. Overall,these findings are consistent with priors, and they suggest that industriesdependent on natural resources, such as timber and mining, are more likely toagglomerate because of a common need to locate close to the source of naturalresource inputs. Moreover, as with reliance on manufactured inputs, reliance onnatural resources contributes to agglomeration at the state level but is notevident at the zipcode and county levels.
The character of these findings is echoed in our estimates of the influence ofproduct shipping costs on agglomeration. The variable Inventories per $shipment always has a positive and significant impact on state-level agglomera-tion, regardless of the choice of labor pooling proxy and regardless of theinclusion of industry fixed effects. This variable is always insignificant, how-ever, at lower levels of geography. Given that Inventories is an inverse proxyfor product shipping costs, these results support the idea that industries withoutput that is costly to transport are more likely to locate close to their marketsand, as a result, exhibit less agglomeration.
19 � �Arguments from Saxenian 22 , for example, suggest that knowledge generated at a given firmis more likely to spill over to the local economy if that knowledge is generated at small as opposedto large firms. In addition, our state-level results are somewhat at variance with Audretsch and
� �Feldman 1 , who found that industries with large expenditures on research and development weremore likely to be concentrated at the state level. Of course, both the dependent and independentvariables are different in our specification.
AGGLOMERATION DETERMINANTS 215
Taking all of these results together, an interesting pattern emerges. Relianceon manufactured and naturally occurring inputs and the production of perish-able products serve to increase the importance of shipping costs in firm locationdecisions. That, in turn, positively affects state-level agglomeration but has littleeffect on agglomeration at lower levels of geography. In contrast, knowledgespillovers positively affect agglomeration at highly localized levels, while areliance on skilled labor affects agglomeration at all levels of geography.20
3.5. The Geographic Nature of Agglomeration
This section looks systematically at geographic differences in the determi-nants of agglomeration. We will focus on the degree to which the differences inthe geography of agglomeration discussed above are statistically significant. InTables 4a and 4b, we present OLS and 2-digit fixed-effect estimates of thedifference in agglomeration at the county�zipcode level, � �� , and at thec z
state�county level, � � � .21 Beginning once more with the adjusted R2-s c
values, a different pattern from Table 3 emerges. First, the adjusted R2-valuesare very small in both tables for the county�zipcode regressions, ranging from2 to roughly 9%. In addition, nearly all of the coefficients are individuallyinsignificant in the county�zipcode regressions. This suggests that there is littlesystematic difference in the determinants of agglomeration at the county levelrelative to the zipcode level. In contrast, the adjusted R2-values are compara-tively large for the state�county regressions, ranging from 8 to 9% for the OLSspecification and from 27 to 29% in the 2-digit fixed-effects specification.These findings suggest that there is considerable systematic variation in the
20 Two additional sets of robustness checks were carried out to evaluate the sensitivity of ourfindings to alternative specifications of the model. First, we experimented with using MSAs as thegeographic unit of analysis. This was done in two ways: by estimating over MSAs only, discardingdata from non-MSA locations, and treating each MSA as a separate geographic unit and byaugmenting this sample with the non-MSA counties. Interpreted broadly, results from the MSA-onlymodel are approximately a blend of those reported previously for the county- and state-levelmodels. This is as anticipated since MSAs are larger than counties but smaller than states.Similarly, results from the MSA plus non-MSA county model are very similar to the county model.Again, this is as anticipated since the geographic scopes of the two models in this instance aresimilar. Details of these regressions are presented in Tables A-3a and A-3b in the Appendix.
A second set of robustness checks reestimated Tables 3a and 3b, measuring � and theright-hand-side variables at the 3-digit SIC level. In general, results from those regressions supportthe principal findings presented above, with some variation. However, because the 4-digit modelsprovide 459 industries while the 3-digit models aggregate to just 140 industries, the 4-digit modelswere favored. Results from the 3-digit-level analyses are not provided in order to conserve space.
21 Estimates from the 3-digit fixed-effect model are generally weaker but do not change the basicconclusions below and are not reported in order to conserve space.
ROSENTHAL AND STRANGE216
TABLE 4aaOLS Gamma Difference Regressions�All Establishments, 4-Digit SIC Level
County�zipcode � State�county �
Managerial ManagerialNet labor share of Education Net labor share of Education
productivity workers of workers productivity workers of workers
Innovations from firms �1.0119 �1.2151 �1.1977 �2.2613 �2.0331 �2.6491with fewer than �1.59 �1.82 �1.69 �1.55 �1.38 �1.66500 workers
Innovations from firms �0.1525 �0.0067 �0.0381 �1.5080 �1.2860 �1.3217with more than �0.26 �0.01 �0.06 �1.11 �0.95 �0.94500 workers
�5 �5Shipments net of 9.230�10 4.930�10inputs per worker 6.19 1.44
Managerial share 0.0041 �0.0333of workers 0.45 �1.64
Share of workers with 0.1245 �0.0368Ph.D. or professional 1.95 �0.26degree
Share of workers with 0.0377 0.2113Master’s degree 0.93 2.31
Share of workers with 0.0023 �0.0951Bachelor’s degree 0.12 �2.15
Manufactured inputs 0.0066 �0.0088 �0.0059 0.0969 0.0780 0.0942per $ shipment 0.63 �0.81 �0.53 4.05 3.26 3.76
Nonmanufactured �0.0207 �0.0291 �0.0421 �0.0671 �0.0685 �0.0912inputs per �1.37 �1.85 �2.35 �1.93 �1.98 �2.26$ shipment
Natural resources �0.0082 �0.0033 �0.0007 0.1182 0.1147 0.1131expenses per �0.74 �0.29 �0.06 4.66 4.49 4.22$ shipment
Energy expenses �0.0360 �0.0503 �0.0367 0.0803 0.0352 0.1347per $ shipment �0.76 �1.00 �0.65 0.73 0.32 1.06
Water expenses 0.1528 0.2904 0.3184 0.4545 0.5260 0.5448per $ shipment 0.41 0.76 0.81 0.54 0.62 0.61
Inventories 0.0200 0.0077 0.0054 0.1230 0.1238 0.1172per $ shipment 1.23 0.46 0.31 3.29 3.32 3.00Ž .nonperishability
2R 0.096 0.019 0.045 0.099 0.100 0.1132R adj. 0.077 0.000 0.019 0.081 0.082 0.090
Sample size 459 459 427 459 459 427
at-Ratios are below the coefficients. Constants are not reported in order to conserve space.
AGGLOMERATION DETERMINANTS 217
TABLE 4baSIC 2-Digit Fixed Effect Difference Regressions�All Establishments, 4-Digit SIC Level
County�Zipcode Gamma State�County Gamma
Managerial ManagerialNet labor share of Education Net labor share of Education
productivity workers of workers productivity workers of workers
Innovations from �1.1659 �1.3030 �1.1473 �1.1826 �1.1680 �1.3103firms with fewer �1.77 �1.90 �1.56 �0.87 �0.86 �0.90than 500 workers
Innovations from �0.1639 �0.0217 0.0186 �0.8042 �0.8105 �0.8061firms with more �0.27 �0.04 0.03 �0.65 �0.65 �0.63than 500 workers
�4 �7Shipments net of 1.148�10 �1.530�10inputs per worker 6.01 0.00
Managerial share �0.0031 �0.0032of workers �0.27 �0.14
Share of workers 0.1218 �0.0048with Ph.D. or 1.59 �0.03professionaldegree
Share of workers �0.0040 0.3019with Master’s �0.07 2.74degree
Share of workers 0.0307 �0.0869with Bachelor’s 1.30 �1.85degree
Manufactured inputs 0.0085 �0.0067 �0.0048 0.0459 0.0456 0.0534per $ shipment 0.69 �0.53 �0.36 1.81 1.82 2.05
Nonmanufactured �0.0029 �0.0141 �0.0258 �0.0276 �0.0274 �0.0263inputs per �0.17 �0.80 �1.27 �0.79 �0.79 �0.65$ shipment
Natural resources 0.0107 �0.0029 �0.0041 0.0843 0.0839 0.0694expenses per 0.77 �0.20 �0.28 2.95 2.96 2.36$ shipment
Energy expenses �0.0057 �0.0098 �0.0006 0.0206 0.0187 0.0911per $ shipment �0.10 �0.17 �0.01 0.18 0.16 0.66
Water expenses 0.1893 0.1478 0.4270 1.5431 1.5372 1.7616per $ shipment 0.45 0.34 0.94 1.78 1.77 1.95
Inventories 0.0032 �0.0069 �0.00127 0.1138 0.1145 0.1189per $ shipment 0.18 �0.37 �0.66 3.10 3.11 3.11Ž .nonperishability
2R 0.143 0.071 0.090 0.315 0.315 0.3442R adjusted 0.087 0.011 0.021 0.271 0.271 0.294
Sample size 459 459 427 459 459 427Number of fixed 20 20 20 20 20 20
effects
at-Ratios are below the coefficients. Constants are not reported in order to conserve space.
ROSENTHAL AND STRANGE218
determinants of state-level agglomeration relative to the determinants of ag-glomeration at lower levels of geography.
Focusing on the state�county regressions, results support the most clear-cutfindings from the previous section. Manufactured inputs, Natural resources,and Inventories all have positive and highly significant effects in the OLS
Ž .model Table 4a and at least marginally significant effects following theŽ .inclusion of 2-digit SIC fixed effects Table 4b . In addition, the Water
expenses variable also has a positive and marginally significant effect once thefixed effects are added to the model. As noted above, these variables proxy forthe importance of locating close to output markets and to factor inputs that tendto be concentrated in a relatively small number of states. In contrast, Nonmanu-factured inputs has a negative and marginally significant effect in the OLSmodels and negative but not significant effects in the 2-digit fixed-effect model.Observe also that the various proxies for labor-market pooling are insignificantin all of the models with the exception of Masters and Bachelors degrees,which have opposite signs. Masters has a positive effect and Bachelors has anegative effect. As discussed above, labor pooling was found to positivelyinfluence agglomeration at all levels of geography. It is not surprising, there-fore, that reliance on skilled labor does not help to systematically explaindifferences in agglomeration at the different levels of geography.22
3.6. The Agglomeration of New Establishments
The patterns of agglomeration that we have studied thus far reflect decadesof economic decisions. It is interesting to compare those patterns to agglomera-tion arising from more recent decisions. Accordingly, in this section wemeasure agglomeration at the 4-digit level using employment at just thoseestablishments that were 5 years old or younger.23 An important initial findingis that for every level of geography the average � for employment at newestablishments is very similar to the average � for all employment. At the state,county, and zipcode levels, the averages for all employment are � � 0.0485,s
� c � 0.0193, and � � 0.0101. For new-establishment employment, the aver-r z
22 Note also that Innovations from both small and large firms is insignificant in all of themodels.
23 To our knowledge, this is the first time that anyone has measured the agglomeration ofemployment at such newly established enterprises.
AGGLOMERATION DETERMINANTS 219
24ages are � � 0.00384, � � 0.0177, and � � 0.0104. Thus far, it appearss c z
that new-establishment agglomeration is similar to all-establishments agglomer-ation.
Tables 5a and 5b present OLS and 2-digit SIC fixed-effect estimates of thedeterminants of new-establishment agglomeration using the same specificationas in Tables 3a and 3b.25 As before, the adjusted R2-values are very low for the
Ž .OLS specification Table 5a . In contrast to previous findings, however, theadjusted R2 remains low even after inclusion of 2-digit SIC fixed effects, withvalues ranging from 0 to 3%. The immediate conclusion, therefore, is that,compared to the agglomeration of all establishments, agglomeration of employ-ment at newly created establishments is not as strongly related to the Marshal-lian microfoundations of agglomerative spillovers and to natural advantages.This conclusion is further supported by examination of the individual coeffi-cients in Tables 5a and 5b. While the qualitative patterns are often similar toresults from Tables 3a and 3b, the level of significance for new-establishmentagglomeration is substantially reduced, especially for state-level agglomer-ation.26
There are two ways in which one might account for these results. First, newestablishments could differ systematically from older establishments. Thiswould be the case in a dynamic setting in which new establishments that choosesuboptimal locations are more likely to fail. In that case, surviving establish-ments would be more likely to be clustered in patterns that reflect the forcesand benefits of agglomeration economies and proximity to natural advantages.
A second interpretation is that the more random pattern of locations amongnewly established enterprises reflects a fundamental change in the tendency toagglomerate. Today’s business environment is in some ways quite differentfrom that of 50 years ago. This has led some to question whether cities willplay the same crucial role in the next millennium that they have in the one just
24 In addition, the median difference between � based on new versus all employment is veryclose to zero for each level of geography.
25 Results from 3-digit fixed-effect specifications do not change the general conclusions dis-cussed below and are not presented in order to conserve space.
26 The principal exception to this generalization is the Inventories variable, which is positive andsignificant for all levels of geography and for all specifications of the model. This may indicate thatnewly established enterprises are especially sensitive to the cost of shipping their product to marketwhen choosing their locations.
ROSENTHAL AND STRANGE220
TA
BL
E5a
aO
LS
Gam
ma
Reg
ress
ions
for
Em
ploy
men
tatU
nder
Age
5E
stab
lishm
ents
atth
e4-
Dig
itSI
CL
evel
Zip
code
leve
lC
ount
yle
vel
Stat
ele
vel
Man
ager
ial
Man
ager
ial
Man
ager
ial
Net
labo
rsh
are
ofE
duca
tion
Net
labo
rsh
are
ofE
duca
tion
Net
labo
rsh
are
ofE
duca
tion
prod
uctiv
ityw
orke
rsof
wor
kers
prod
uctiv
ityw
orke
rsof
wor
kers
prod
uctiv
ityw
orke
rsof
wor
kers
Inno
vatio
nsfr
omfi
rms
0.63
450.
5769
0.72
26�
0.13
16�
0.29
31�
0.07
57�
3.38
07�
3.64
22�
2.96
24w
ithfe
wer
than
0.66
0.59
0.69
�0.
11�
0.24
�0.
06�
1.26
�1.
35�
1.05
500
wor
kers
Inno
vatio
nsfr
omfi
rms
�0.
8512
�0.
8242
�0.
9549
�0.
5000
�0.
5276
�0.
6298
1.31
781.
3595
�0.
4635
with
mor
eth
an�
0.95
�0.
92�
1.03
�0.
44�
0.46
�0.
530.
530.
55�
0.19
500
wor
kers
�5
�5
�5
Ship
men
tsne
tof
2.04
0�
101.
650
�10
6.08
0�
10in
puts
per
wor
ker
0.90
0.57
0.97
�3
�2
�2
Man
ager
ials
hare
2.25
4�
101.
395
�10
1.62
1�
10of
wor
kers
0.17
0.81
0.43
Shar
eof
wor
kers
with
0.07
670.
1199
0.00
24Ph
.D.o
rpr
ofes
sion
al0.
810.
980.
01de
gree
AGGLOMERATION DETERMINANTS 221
Shar
eof
wor
kers
with
0.02
260.
0363
0.13
65M
aste
r’s
degr
ee0.
370.
460.
83Sh
are
ofw
orke
rsw
ith0.
0222
0.00
830.
0465
Bac
helo
r’s
degr
ee0.
770.
220.
59M
anuf
actu
red
inpu
ts0.
0124
0.00
940.
0170
0.04
950.
0508
0.05
570.
1290
0.12
290.
1451
per
$sh
ipm
ent
0.79
0.60
1.04
2.44
2.51
2.62
2.94
2.80
3.27
Non
man
ufac
ture
d�
0.00
25�
0.00
45�
0.01
10�
0.03
90�
0.04
16�
0.05
61�
0.04
28�
0.04
95�
0.05
73in
puts
per
�0.
11�
0.20
�0.
42�
1.32
�1.
42�
1.64
�0.
67�
0.78
�0.
80$
ship
men
tN
atur
alre
sour
ces
0.00
290.
0042
0.01
060.
0046
0.00
780.
0104
0.08
050.
0859
0.10
50ex
pens
espe
r0.
170.
240.
590.
210.
350.
451.
701.
802.
18$
ship
men
tE
nerg
yex
pens
es�
0.01
13�
0.01
30�
0.01
44�
0.04
90�
0.03
76�
0.04
68�
0.34
28�
0.33
79�
0.30
18pe
r$
ship
men
t�
0.16
�0.
18�
0.17
�0.
53�
0.40
�0.
43�
1.71
�1.
66�
1.34
Wat
erex
pens
es�
0.31
47�
0.28
41�
0.43
360.
0855
0.11
07�
0.03
331.
0124
1.10
391.
4125
per
$sh
ipm
ent
�0.
56�
0.51
�0.
750.
120.
15�
0.04
0.65
0.71
0.90
Inve
ntor
ies
0.05
350.
0504
0.05
150.
0957
0.09
070.
0942
0.15
230.
1413
0.13
85pe
r$
ship
men
t2.
172.
042.
013.
022.
862.
842.
222.
062.
00Ž
.no
nper
isha
bilit
y
2R
0.01
60.
015
0.02
70.
045
0.04
60.
056
0.04
60.
044
0.05
42
Adj
.R
0.00
00.
000
0.00
00.
026
0.02
70.
031
0.02
70.
025
0.02
9Sa
mpl
esi
ze45
945
942
745
945
942
745
945
942
7
at-
Rat
ios
are
belo
wth
eco
effi
cien
ts.
Con
stan
tsar
eno
tre
port
edin
orde
rto
cons
erve
spac
e.
ROSENTHAL AND STRANGE222
TA
BL
E5b
aSI
C2-
Dig
itFi
xed
Eff
ectG
amm
aR
egre
ssio
nsfo
rE
mpl
oym
enta
tUnd
erA
ge5
Est
ablis
hmen
tsat
the
4-D
igit
SIC
Lev
el
Zip
code
leve
lC
ount
yle
vel
Stat
ele
vel
Man
ager
ial
Man
ager
ial
Man
ager
ial
Net
labo
rsh
are
ofE
duca
tion
Net
labo
rsh
are
ofE
duca
tion
Net
labo
rsh
are
ofE
duca
tion
prod
uctiv
ityw
orke
rsof
wor
kers
prod
uctiv
ityw
orke
rsof
wor
kers
prod
uctiv
ityw
orke
rsof
wor
kers
Inno
vatio
nsfr
omfi
rms
0.49
220.
5394
0.79
10�
0.05
650.
0310
0.27
09�
3.59
46�
3.61
33�
2.81
57w
ithfe
wer
than
0.49
0.53
0.72
�0.
040.
020.
19�
1.29
�1.
30�
0.96
500
wor
kers
Inno
vatio
nsfr
omfi
rms
�1.
0628
�1.
0576
�1.
0518
�0.
7160
�0.
7660
�0.
8056
1.47
671.
4893
�0.
4592
with
mor
eth
an�
1.15
�1.
15�
1.10
�0.
61�
0.65
�0.
650.
580.
58�
0.18
500
wor
kers
�5
�5
�6
Ship
men
tsne
tof
inpu
ts3.
360
�10
�1.
690
�10
6.23
0�
10pe
rw
orke
r1.
12�
0.44
0.08
�2
�2
�3
Man
ager
ials
hare
�2.
011
�10
�1.
491
�10
2.43
5�
10of
wor
kers
�1.
18�
0.69
0.05
Shar
eof
wor
kers
with
0.14
230.
1496
0.00
55Ph
.D.o
rpr
ofes
sion
al1.
241.
020.
02de
gree
Shar
eof
wor
kers
with
�0.
0236
�0.
0491
0.21
22M
aste
r’s
degr
ee�
0.29
�0.
460.
96
AGGLOMERATION DETERMINANTS 223
Shar
eof
wor
kers
with
0.00
00�
0.02
810.
0059
Bac
helo
r’s
degr
ee0.
00�
0.62
0.06
Man
ufac
ture
din
puts
0.01
840.
0120
0.01
840.
0611
0.06
110.
0680
0.12
840.
1280
0.14
47pe
r$
ship
men
t0.
970.
640.
932.
532.
552.
682.
452.
472.
74N
onm
anuf
actu
red
inpu
ts0.
0154
0.01
310.
0056
�0.
0115
�0.
0093
�0.
0227
0.01
040.
0097
0.02
02pe
r$
ship
men
t0.
590.
510.
18�
0.35
�0.
28�
0.58
0.15
0.14
0.25
Nat
ural
reso
urce
sex
pens
es0.
0118
0.00
670.
0088
0.02
110.
0205
0.02
090.
1088
0.10
860.
1154
per
$sh
ipm
ent
0.54
0.31
0.39
0.76
0.74
0.71
1.81
1.81
1.89
Ene
rgy
expe
nses
per
$0.
0444
0.03
220.
0185
�0.
0181
�0.
0263
�0.
0486
�0.
1915
�0.
1903
�0.
1878
ship
men
t0.
510.
370.
18�
0.16
�0.
24�
0.37
�0.
80�
0.79
�0.
68W
ater
expe
nses
per
$0.
0581
0.01
200.
0396
0.73
200.
7088
0.72
621.
0201
1.02
301.
7423
ship
men
t0.
090.
020.
060.
890.
860.
830.
570.
580.
96In
vent
orie
spe
r$
ship
men
t0.
0499
0.05
100.
0501
0.09
260.
0970
0.09
990.
1172
0.11
620.
1136
Ž.
nonp
eris
habi
lity
1.83
1.86
1.74
2.66
2.78
2.71
1.55
1.54
1.48
2R
0.04
60.
046
0.05
00.
090
0.09
10.
099
0.09
00.
090
0.10
02
Rad
j.0.
000
0.00
00.
000
0.03
10.
031
0.03
10.
031
0.03
10.
031
Sam
ple
size
459
459
427
459
459
427
459
459
427
Num
ber
offi
xed
effe
cts
2020
2020
2020
2020
20
at-
Rat
ios
are
belo
wth
eco
effi
cien
ts.
Con
stan
tsar
eno
tre
port
edin
orde
rto
cons
erve
spac
e.
ROSENTHAL AND STRANGE224
ended.27 Additional research is needed to discriminate between these twocompeting explanations for our result.
4. CONCLUSION
This paper has considered an important but understudied question in theempirical literature on agglomeration: What are the microfoundations of ag-glomeration economies? Using zipcode-, county-, and state-level employmentdata for the fourth quarter of 2000, we compute the measure of agglomeration
� �developed by Ellison and Glaeser 5 . The agglomeration measure is thenmatched with various industry characteristics that proxy for the importance ofknowledge spillovers, labor-market pooling, input sharing, natural advantagesthat affect input shipping costs, and product shipping costs. We find evidenceof the importance of all of these determinants of agglomeration.
We also uncover an interesting geographic pattern that may well reflect theidiosyncratic characteristics of each of the determinants. Variables that proxyphysical input and product shipping costs�including reliance on naturalresources, manufactured inputs, and production of nonperishable output�allpositively affect state-level agglomeration but have little effect on agglomera-tion at lower levels of geography. The geographic scope of these effects
Žsuggests that state-level transportation modes i.e., train, truck, and barge.transport may play an important role in the location patterns of industries
sensitive to shipping costs. At the other extreme, knowledge spillovers posi-tively affect agglomeration only at the zipcode level, possibly because suchspillovers attenuate rapidly across space. Finally, reliance on skilled laborpositively affects agglomeration at all geographic levels. This latter result isparticularly robust and may reflect spillover benefits that arise when skilledworkers can seek out new job opportunities without having to move out ofcounty or out of state. Together, these patterns explain an important share of the
Žvariation in state- versus county-level agglomeration across industries up to.30% . Nevertheless, considerable unexplained variation in agglomeration re-
mains, suggesting a role for continued research in this area.We also find that employment at newly formed establishments is much less
systematically related to the microfoundations of agglomeration than is employ-ment at existing establishments. This could reflect a dynamic selection mecha-
27 � �As Glaeser 8 notes, there are many factors that will come together to determine the futurerole of cities. One of these is the importance of agglomeration economies. If our findings can beinterpreted to indicate that new firms agglomerate less and are less sensitive to Marshallian factors,then this would suggest a decline in the importance of cities. It is important to recognize, however,that there is a body of other evidence suggesting that agglomeration economies continue to exert
Ž � �.powerful attractions, even to new establishments see Rosenthal and Strange 20 .
AGGLOMERATION DETERMINANTS 225
nism, where only establishments that choose locations conducive to agglomera-tive spillovers and benefits from natural advantages survive. But our resultscould also reflect a fundamental change in the nature of establishment locationdecisions. Once again, further research is warranted.
APPENDIX
TABLE A-1
SIC Codes Used to Create Natural resources per $ Shipmentafrom the 1992 BEA Input�Output Tables
Industrycode Description of industry category
10200 Poultry and eggs10301 Meat animals10302 Miscellaneous livestock20100 Cotton20201 Food grains20202 Feed grains20203 Grass seeds20300 Tobacco20401 Fruits20402 Tree nuts20501 Vegetables20502 Sugar crops20503 Miscellaneous crops20600 Oil bearing crops20701 Forest products20702 Greenhouse and nursery products30001 Forestry products30002 Commercial fishing40001 Agricultural, forestry, and fishery services40002 Landscape and horticultural services50001 Iron and ferroalloy ores, and miscellaneous metal ores, n.e.c.60100 Copper ore60200 Nonferrous metal ores, except copper90001 Dimension, crushed and broken stone90002 Sand and gravel90003 Clay, ceramic, and refractory minerals90004 Nonmetallic mineral services and miscellaneous
100000 Chemical and fertilizer minerals
a Ž . ŽCoal Industry Code 07 and crude petroleum and natural gas Industry.Code 08 were included as part of the Energy variable rather than Natural
resources. The latter, in contrast, is comprised of output from mining, agricul-ture, etc., as indicated by the list above.
ROSENTHAL AND STRANGE226
TABLE A-2
Industry Codes Implicitly Used to Create Nonmanufactured input per $ shipmentafrom the 1992 BEA Input�Output Tables
Ž .32 Additional Categories Related to Government Services Are Omitted to Conserve Space .
Industry Industrycode Description of industry category code Description of industry category
650100 Railroads and related services 750003 Automobile parking and car washes650200 Local and suburban transit and 760101 Motion picture services and theaters
interurban highway passenger650301 Trucking and courier services, 760102 Video tape rental
except airŽ650302 Warehousing and storage 760201 Theatrical producers except motion
.picture , bands, orchestras650400 Water transportation 760202 Bowling centers650500 Air transportation 760203 Professional sports clubs and
promoters650600 Pipelines, except natural gas 760204 Racing, including track operation650701 Freight forwarders and other 760205 Physical fitness facilities and
transportation services membership sports and650702 Arrangement of passenger 760206 Other amusement and recreation
transportation services660100 Telephone, telegraph 770100 Doctors and dentists
communications, andcommunication services
660200 Cable and other pay television 770200 Hospitalsservices
670000 Radio and TV broadcasting 770301 Nursing and personal care facilities690100 Wholesale trade 770303 Other medical and health services690200 Retail trade, except eating 770304 Veterinary services
and drinking700100 Banking 770305 Other medical and health services700200 Credit agencies other than banks 770401 Elementary and secondary schools700300 Security and commodity brokers 770402 Colleges, universities, and
professional schools700400 Insurance carriers 770403 Private libraries, vocational schools,
and educational services700500 Insurance agents, brokers, 770501 Business associations and
and services professional membership710100 Owner-occupied dwellings 770502 Labor organizations, civic, social,
and fraternal associations710201 Real estate agents, managers, 770503 Religious organizations
operators, and lessors710202 Royalties 770504 Other membership organizations720101 Hotels 770600 Job training and related services720102 Other lodging places 770700 Child day care services720201 Laundry, cleaning, garment 770800 Residential care
services, and shoe repair720202 Funeral service and crematories 770900 Social services, n.e.c.720203 Portrait photographic studios, and 780100 U.S. Postal Service
other miscellaneous personal
AGGLOMERATION DETERMINANTS 227
TABLE A-2�Continued
Industry Industrycode Description of industry category code Description of industry category
720204 Electrical repair shops 780200 Federal electric utilities720205 Watch, clock, jewelry, and 780500 Other federal government
furniture repair enterprises720300 Beauty and barber shops 790100 State and local government
passenger transit730101 Miscellaneous repair shops 790200 State and local government
electric utilities730102 Services to dwellings and 790300 Other State and local government
other buildings enterprises730103 Personnel supply services 800000 Noncomparable imports730104 Computer and data processing 810001 Scrap
services730106 Detective and protective services 810002 Used and secondhand goods730107 Miscellaneous equipment rental 820000 General government industry
and leasing730108 Photofinishing labs and commercial 830001 Rest of the world adjustment to
photography final uses730109 Other business services 840000 Household industry730111 Management and public relations 850000 Inventory valuation adjustment
services730112 Research, development, and testing 880000 Compensation of employees
services, except noncommercial730200 Advertising 890000 Indirect business tax and
nontax liability730301 Legal services 900000 Other value added730302 Engineering, architectural, and 910000 Personal consumption
surveying services expenditures730303 Accounting, auditing and 920000 Gross private fixed investment
bookkeeping, and miscellaneousservices,
740000 Eating and drinking places 930000 Change in business inventories750001 Automotive rental and leasing, 940000 Exports of goods and services
without drivers750002 Automotive repair shops 950000 Imports of goods and services
and services
a Nonmanufactured inputs was calculated as a residual by forming 1 � Value added �Manufactured inputs � Natural resources � Energy � Water, where all of these variables are
Ž .measured per dollar of shipment. Value added deflated by shipments was obtained from theŽ .Annual Survey of Manufactures ASM , while the other variables were obtained directly from the
Ž .BEA input�output tables, the root source of which is the Census of Manufactures CM . Becausethe ASM is based on a subset of the universe of the manufacturing establishments, while the CMcovers the entire universe, some discrepancies occur. This and an imperfect matching of some ofthe BEA input�output industry codes to the 4-digit SIC code classifications account for a smallnumber of industries for which our calculated non-manufactured inputs variable is negative.
ROSENTHAL AND STRANGE228
TABLE A-3
SIC 2-Digit Fixed Effect Gamma Regressions for MSA Measures of Geography�aEmployment at All Establishments at the 4-Digit SIC Level
MSAs only MSAs plus non-MSA counties
Managerial ManagerialNet labor share of Education Net labor share of Education
productivity workers of workers productivity workers of workers
Innovations from firms �2.5334 �2.7657 �2.4749 �1.9882 �2.4475 �2.1143with fewer than �1.43 �1.52 �1.25 �1.75 �2.05 �1.65500 workers
Innovations from firms 0.8681 1.1438 1.3789 1.0178 1.3751 1.4812with more than 0.54 0.69 0.80 0.98 1.26 1.32500 workers
�4 �4Shipments net of 2.421� 10 2.217� 10inputs per worker 4.71 6.71
Managerial share �0.0194 0.0376of workers �0.63 1.87
Share of workers with 0.4812 0.1854Ph.D. or professional 2.34 1.39degree
Share of workers with 0.1291 0.2176Master’s degree 0.87 2.25
Share of workers with �0.1137 0.0158Bachelor’s degree �1.80 0.38
Manufactured inputs 0.0442 0.0105 0.0058 0.0126 �0.0114 �0.0110per $ shipment 1.33 0.31 0.17 0.59 �0.52 �0.48
Nonmanufactured �0.0203 �0.0433 �0.0918 �0.0026 �0.0261 �0.0503inputs per �0.45 �0.93 �1.68 �0.09 �0.85 �1.42$ shipment
Natural resources 0.0885 0.0582 0.0338 0.0186 �0.0026 �0.0157expenses per 2.37 1.54 0.85 0.77 �0.11 �0.61$ shipment
Energy expenses 0.0235 0.0076 0.0364 �0.0079 0.0093 0.0015per $ shipment 0.15 0.05 0.20 �0.08 0.09 0.01
Water expenses 1.5195 1.4083 1.8511 1.0142 1.0142 1.5559per $ shipment 1.35 1.22 1.52 1.40 1.33 1.96
Inventories 0.0158 �0.0029 �0.0163 0.0054 �0.0227 �0.0257per $ shipment 0.33 �0.06 �0.32 0.18 �0.70 �0.77Ž .nonperishability
2R 0.170 0.128 0.152 0.204 0.127 0.1572R adj. 0.116 0.071 0.088 0.152 0.070 0.093
Sample size 459 459 427 459 459 427Number of fixed 20 20 20 20 20 20
effects
at-Ratios are below the coefficients. Constants are not reported in order to conserve space.
AGGLOMERATION DETERMINANTS 229
REFERENCES
1. D. B. Audretsch and M. Feldman, R&D spillovers and the geography of innovation andŽ .production, American Economic Review, 86, 630�640 1996 .
2. D. B. Audretsch, G. van Leeuwen, B. Menkveld, and R. Thurik, Market dynamics in theNetherlands: Competition policy and the response of small firms, International Journal ofIndustrial Organization, forthcoming.
3. A. Ciccone and R. E. Hall, Productivity and the density of economic activity, AmericanŽ .Economic Review, 86, 54�70 1996 .
4. G. Dumais, G. Ellison, and E. Glaeser, Geographic concentration as a dynamic process, NBERŽ .Working Paper 6270 1997 .
5. G. Ellison and E. Glaeser, Geographic concentration in U.S. manufacturing industries: AŽ .dartboard approach, Journal of Political Economy, 105, 889�927 1997 .
6. G. Ellison and E. Glaeser, The geographic concentration of an industry. Does natural advantageexplain agglomeration, American Economic Association Papers and Proceedings 89,
Ž .311�316 1999 .Ž .7. E. Glaeser, Learning in cities, Journal of Urban Economics, 46, 254�277 1999 .
Ž .8. E. Glaeser, Are cities dying?, Journal of Economic Perspectives, 12, 139�160 1998 .9. E. Glaeser, H. Kallal, J. Scheinkman, and A. Shleifer, Growth in cities, Journal of Political
Ž .Economy, 100, 1126�1152 1992 .10. G. Goldstein and T. Gronberg, Economies of scope and economies of agglomeration, Journal
Ž .of Urban Economics, 16, 91�104 1984 .11. R. W. Helsley and W. C. Strange, Agglomeration economies and matching in a system of
Ž .cities, Regional Science and Urban Economics, 20, 189�212 1990 .12. J. V. Henderson, Efficiency of resource usage and city size, Journal of Urban Economics, 19,
Ž .47�70 1986 .13. J. V. Henderson, A. Kuncoro, and M. Turner, Industrial development in cities, Journal of
Ž .Political Economy, 103, 1067�1085 1995 .14. A. Jaffe, M. Trajtenberg, and R. Henderson, Geographic localization of knowledge spillovers as
Ž .evidenced by patent citations, Quarterly Journal of Economics, 108, 577�598 1993 .15. S. Kim, Regions, resources and economics geography: Sources of U.S. regional comparative
Ž .advantage, 1880�1987, Regional Science and Urban Economics, 29, 1�32 1999 .Ž .16. P. Krugman, ‘‘Geography and Trade,’’ MIT Press, Cambridge, MA 1991 .
Ž .17. A. Marshall, ‘‘Principles of Economics,’’ Macmillan, London 1920 .18. R. L. Moomaw, Productivity and city size: A critique of the evidence, Quarterly Journal of
Ž .Economics, 96, 675�688 1981 .19. R. Nakamura, Agglomeration economies in urban manufacturing industries: A case of Japanese
Ž .cities, Journal of Urban Economics, 17, 108�124 1985 .20. S. Rosenthal and W. C. Strange, Geography, industrial organization, and agglomeration, Center
Ž .for Policy Research Working Paper 14, Syracuse University 1999 .21. J. Quigley, Urban diversity and economic growth, Journal of Economic Perspectives, 12,
Ž .127�138 1998 .22. A. Saxenian, ‘‘Regional Advantage: Culture and Competition in Silicon Valley and Route
Ž .128,’’ Harvard University, Cambridge, MA 1994 .23. L. Sveikauskas, The productivity of cities, Quarterly Journal of Economics, 89, 393�413
Ž .1975 .
