Journal of Economic Geography 9 (2009) pp. 853–868 doi:10.

1093/jeg/lbp007
Advance Access Published on 9 March 2009

Localization economies and establishment size:

was Marshall right after all?y
Octávio Figueiredo*, Paulo Guimarães** and Douglas Woodward***

Abstract
This article re-examines the relationship between industry localization and the employ-
ment size of establishments. To measure localization, we use an approach that builds
on Ellison and Glaeser’s (1997, Journal of Political Economy, 105: 889–927) dartboard
location model. Our localization index is based on plant (establishment) counts and
more closely reflects the firm externalities of Marshallian industrial districts, in contrast
with broader measures of localization that include an establishment’s internal
economies, such as Holmes and Steven’s (2002, The Review of Economics and
Statistics, 84: 682–690) analysis based on employment location quotients. In line
with Alfred Marshall’s theory of external economies in industrial districts, we find evidence
that plants located in areas where an industry exhibits localization, or excess concen-
tration, are smaller than plants in the same industry outside such areas.

Keywords: localization, external economies, plant size, industrial districts

JEL classifications: R12, R39, L11
Date submitted: 22 February 2008 Date accepted: 22 January 2009

1. Introduction
Economic geographers have long known that some industries are highly localized; that
is, concentrated rather than dispersed across space. Classic studies include Hoover
(1937), which examined the US shoe and leather industries, and Lichtenberg (1960),
which documents industry concentration in New York City. Ellison and Glaeser (1997)
provide evidence that this phenomenon is prevalent in many US industries.
The widely accepted explanation for localization (or industry-specific agglomeration)
is Alfred Marshall’s notion of spatial externalities—economic benefits external to the
firm, but internal to an industry in a particular region. Marshall’s view of these positive
interfirm externalities stands in contrast to internal firm economies of scale. In the
Principles of Economics (Book IV, Chapter XII) he wrote:

Looking more closely at the economies arising from an increase in the scale of production of
any kind of goods, we found that they fell into two classes–those dependent on the general
y
An earlier version of this article was presented at the 2007 North American Meetings of the Regional Science
Association International, Urban Economics sessions.
*Faculdade de Economia, Universidade do Porto, 4200 Porto, Portugal and CEMPRE, Universidade do
Porto, 4200 Porto, Portugal email 5octavio@fep.up.pt4.
**Division of Research and Department of Economics, Moore School of Business, University of South
Carolina, Columbia, SC 29208, USA, and CEMPRE, Universidade do Porto, 4200 Porto, Portugal and IZA
Bonn, Germany. email 5guimaraes@moore.sc.edu4.
***Division of Research and Department of Economics, Moore School of Business, University of South
Carolina, Columbia, SC 29208, USA.
email 5woodward@moore.sc.edu4

ß The Author (2009). Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oxfordjournals.org
854 . Figueiredo et al.

development of the industry, and those dependent on the resources of the individual houses of
business engaged in it and the efficiency of their management; that is, into external and internal
economies.

Marshall proposed three sources of external economies to explain why firms cluster
within the same industry. The first is the presence of an array of input providers,
allowing for productivity gains resulting from vertical disintegration and specialization.
The second is related to labor market pooling, where agglomeration improves each
firm’s productivity because it facilitates the firm-worker matching process. The third
is a firm’s ability to capture industry-specific knowledge spillovers that take place in
an industrial district. In Marshall’s view, these positive externalities lead to increased
specialization and higher productivity allowing firms to operate at a smaller scale.
In some illuminating passages from Industry and Trade, Marshall stated:

For long ages industrial leadership depended mainly on the number and extent of centres
of specialized skill in which [. . .] external economies abounded [. . .] Each single business was
on a small scale; and though it had access to many of the economies of production on a large
scale, these were external to it, and common to the whole district.[. . .] Thus each firm, though
of moderate size, might reasonably hope to obtain most of the advantages in production, which
would be accessible only to vast businesses, if each had been mainly dependent on its own
resources. Under these conditions, a very large capital in the aggregate was distributed over
many firms of moderate size, each with its own individual life, its own power of initiative, and its
own personal relations with its employees [...] Its own (Internal) economies were not great: but
it took its part in affording a large market for firms in branches of manufacture, which supplied
it with made or half-made materials: and in developing (External) economies of general
organization, which gradually became common property. (Marshall, 1919, 114–115, 206)

This idea is mirrored in the industrial district literature, where Marshallian clusters
are typically associated with small firms. In the late 20th-century writing on economic
geography, ‘new industrial districts’ connote regional clusters of small-scale firms,
exemplified by the ‘flexible manufacturing’ areas of northern Italy and other European
regions (Piore and Sabel, 1984; Goodman and Bamford, 1989). Notably, Markusen’s
(1996) often cited survey on the typology of districts stressed that in the industrial
districts literature many case studies indicate that the Marshallian districts are areas
composed of a vast array of firms with low internal scale economies. Essentially, these
districts, as delineated by the Markusen’s survey, are ‘sticky places’ of small firms
held together by positive spatial externalities. She clearly stated that ‘. . .Marshall
envisioned a region where the business structure is comprised of small, locally owned
firms’ (p. 297).
The empirical relationship between industry agglomeration and size was addressed
systematically for the first time in a paper by Holmes and Stevens (2002). Using
different levels of geographic aggregation and US data for 1992, the study looked at
the correlation between (industry-standardized) plant employment size and a broad
measure of agglomeration across regions. Surprisingly, and seemingly contrary to
Marshall’s theory of external economies in industrial districts, they found a positive
association. Additional evidence for this result—that firms are larger in areas where
industries cluster–was provided by Barrios et al. (2006), which makes use of panel data
for Irish counties spanning from 1973 to 2000. Using a different approach (point-
pattern methodology), Duranton and Overman (2008) found that larger establishments
 Localization economies and establishment size . 855

in the UK exhibit higher localization than smaller establishments, confirming Holmes

and Stevens (2002). Yet they also uncover localization behavior for numerous industries
comprised of small establishments. Lafourcade and Mion (2007), using Italian data
and two different methodologies, also found mixed evidence. In a first step, the
authors divide their sample into ‘small’ and ‘large’ plants and calculate Ellison and
Glaeser’s (1997) employment concentration measure for the two subsamples. They then
complement this analysis with an indicator of spatial autocorrelation (the Moran
index). According to the results based on the Ellison and Glaeser’s (1997) employment
concentration measure, large plants are more concentrated than small plants. However,
considering distance-based patterns via spatial autocorrelation, the authors found that
small establishments in Italy exhibit a greater tendency to co-locate in adjacent areas.
Holmes and Stevens (2004) and Manning (2007) advanced new explanations and
models to bring together theory and evidence about firm size in areas of industry
agglomeration. Holmes and Stevens (2004) draw on the new economic geography to
develop a theoretical model that can account for the regularity found in Holmes and
Stevens (2002) and Barrios et al. (2006). From a different research perspective (labor
economics), Manning (2007) attempts to explain Holmes and Stevens (2002) result
(which he calls the plant size-place effect), building a theoretical model that crucially
relies on the assumption of monopsonistic labor markets.
In this article, we propose to look at firm employment size and localization through
a different empirical lens. Common to Holmes and Stevens (2002) and Barrios et al.
(2006) is the use of Florence’s (1939) employment location quotient as the measure
of agglomeration of an industry at a given location. This statistic has an intuitive
appeal, but lacks a theoretical underpinning. Moreover, as we argue in this article,
the employment location quotient broadly captures agglomeration, but it does not
specifically measure Marshallian industry localization. The reason is that it
encompasses both internal scale economies and Marshallian external economies.
As an alternative, we develop a localization measure grounded on the dartboard
location model of Ellison and Glaeser (1997). We then use this statistic to re-examine
the link between localization and firm size. Applied to Portuguese data, the analysis
reveals a consistent, statistically significant negative relation between scale and
localization. This result upholds the predictions of Marshall and the industrial district
literature regarding firm size inside areas of industry agglomeration.
The plan for the rest of the article is as follows. In the next section, we review measures
of localization, highlighting both conceptual and measurement problems with the
employment location quotient. In Section 3 we present our theoretical framework,
derive the measure of localization and discuss related econometric issues. Section 4
presents our data, while Section 5 discusses our empirical findings. In Section 6 we offer
a conclusion.

2. Measures of localization
Seventy years after it was first introduced, Florence’s (1939) employment location
quotient is the most widely used statistic to evaluate industrial agglomeration at a given
location [among many examples, see Kim (1995)]. Once the spatial scale of analysis
is identified, location quotients may be calculated for each industry and region by
computing the ratio between the regional employment share for the industry and the
856 . Figueiredo et al.

industry’s national share of total employment.1 The main advantages of the measure
are computational simplicity and the availability of regional data by industry, which
means it can be applied in many contexts. Despite its popularity, however, the location
quotient lacks a theoretical foundation. Moreover, it is not a precise measure of
localization in the Marshallian sense of an industrial district. One fundamental prob-
lem is that the employment location quotient is unable to differentiate between external
and internal scale economies. The location quotient will be the same whether the
regional employment share for an industry results from a single large establishment or
from a cluster of smaller establishments. Clearly, a large employment location quotient
that results from one large plant does not reflect external agglomeration economies
of any type. In that case, we do not have an industrial district. Thus, geographic
concentration (as measured by the employment location quotient) is entirely explained
by internal returns to scale.
Another problem stems from the potential inability of the location quotient to
capture the randomness of the underlying plant location decisions, which alone can
explain some degree of spatial concentration. Because of the discrete nature of the
phenomenon being measured it is possible to observe spurious concentration, that is,
concentration that occurs by chance alone.
Ellison and Glaeser (1997) developed a measure of industry localization that
overcomes these pitfalls. Importantly, they provide a theoretical foundation for their
measure, proposing an index that is based on the firm location model of Carlton (1983).
This location model builds on McFadden’s (1974) random utility (profit) maximization
framework—the workhorse for empirical research on industrial location decisions.
Second, because the index is based on a probabilistic model, it naturally accounts for
the inherent randomness (lumpiness) that will be observed if location decisions are
determined by chance alone. Ellison and Glaeser (1997) also claimed that their method
expurgates the effect of internal scale economies from the industry localization measure.
However, more recently, Guimarães et al. (2007) showed by example that applications
of the employment-based Ellison and Glaeser (1997) index could lead to counter-
intuitive results because it does not completely purge the effect of industrial
concentration resulting from internal scale economies. That paper demonstrates that
applications of the Ellison and Glaeser statistic relying on employment data instead of
plant count data (as proposed by Ellison and Glaeser) offered no statistical advantage
and would lead to increased imprecision as a measure of Marshallian localization based
on external economies.2 Guimarães et al. (2007) proposed an alternative statistic (based
on plant counts) that is consistent with the theoretical framework of Ellison and Glaeser
(1997) but removes the effect of internal scale economies. This alternative index has the
added advantage of offering a statistical test of significance for the existence of
localization economies. Recent developments on the measurement of industrial
localization include the D-index of localization of Mori et al. (2005) and the approaches

1 Formally, the location quotient is Qjk ¼ ðxjk =xj Þ=ðxk =xÞ, where xjk is employment in region j and industry
k, xj and xk are employment in region j and in industry k, respectively, and x is total employment in the
economy.
2 The use of plant counts can also be justified from a theoretical point of view. In the location decision
model of Carlton (1983), the starting point for Ellison and Glaeser’s (1997) dartboard model, individual
workers do not scatter according to a pattern. It is the location of plants or establishments that is chosen
by decision makers.
 Localization economies and establishment size . 857

of Marcon and Puech (2003) and Duranton and Overman (2005) that deal with the
modifiable areal unit problem.
It should be recognized that the papers reviewed here have dealt with the question of
measuring ‘an industry’ level of localization but not with the problem of quantifying (as
is the case for the location quotient) the level of localization of ‘an industry at a given
location’.3 Thus, despite these recent developments, we still do not have an adequate
method to evaluate localization in a specific region. In the next section we develop such
measure, building on the dartboard location model of Ellison and Glaeser (1997).

3. Methodology
3.1. Random utility framework
Consider an economy with J distinct regions. In that economy firm location decisions
are based on profit maximization behavior. As in Ellison and Glaeser (1997) dartboard
model, we assume that a firm in a particular industry, say firm i in industry k, evaluates
potential profits at every location and selects the location with the highest profit. Profits
(in log form) are given by,
log ijk ¼ log j þ jk þ "ijk ; ð3:1Þ
where j reflects the expected profitability of locating in region j that is common to all
industries. The term jk is a variable that captures external economies (along with
natural advantages) of region j that also affects the firm’s profit function for firms in
industry k. Finally, the "ijk is a random effect that picks all other nonsystematic factors
affecting firm i’s profits. Assuming that "ijk is an identically and independently
distributed random term with an Extreme Value Type I distribution,4 then we can apply
McFadden’s (1974) result to obtain an expression that gives the probability that a firm
from that particular industry will locate in region j:
expðlog j þ jk Þ
pjk ¼ PJ : ð3:2Þ
j¼1 expðlog j þ jk Þ

In this framework, conditional on their location choices, firms in industry k will on

average have the same level of profits at each location (for a demonstration of this result
see Appendix A). That is, despite the unevenness in the distribution of firms across
space it must be true that, on average, profits are identical in every region. If we are
willing to assume that regional labor markets are perfectly competitive and that in areas
of agglomeration firms experience a productivity bonus, then it must be the case that
firms are smaller in those areas. This argument is well demonstrated in Manning (2007),
where he shows that the most common forms of popular theories of agglomeration—
those based on knowledge spillovers, gains on some inputs resulting from firm
specialization and the division of labor within agglomerations, or improved match

3 This is the same distinction as between the coefficient of location [proposed for the first time by Hoover
(1936)] and Florence’s (1939) location quotient. Ellison and Glaeser (1997), Mori et al. (2005)
and Guimarães et al. (2007) are developments of the former which gives (for each industry) a
P x
single statistic measuring the
x
 respective degree of localization. Formally, the coefficient of
location is Ck ¼ 12 Jj¼1  xjkk  xj .
4 In the discrete-choice literature, this distribution has been referred to by other names such as Gumbel
and (mistakenly) Weibull.
858 . Figueiredo et al.

quality in dense labor markets—all predict that establishment size should, on average,
be smaller in areas of concentration of an industry (Appendix B).5

3.2. An empirical measure of localization

In their work, Ellison and Glaeser (1997) assumed that the variables that captured
external economies and natural advantages of region j for industry k, jk, were random
variables with a distribution that satisfied the following assumption:
A1 : Eðpjk Þ ¼ wj ; ð3:3Þ
and where the wj are the elements of a reference distribution of overall economic
activity. Following the suggestion of Ellison and Glaeser (1997), we assume that wj
equals the share of total manufacturing employment; that is,
xj xj
w j ¼ PJ  ;
j¼1 xj x

where xj denotes total manufacturing employment in location j.6 Assumption A1

conveys the idea that on average and across industries, the distribution of firms
replicates the distribution of overall manufacturing activity. We retain this assumption
and add a new one, assumption A2, requiring that the distribution of the industry-
specific effects be such that,
j
A2 : Eðpjk Þ ¼ PJ ; ð3:4Þ
j¼1 j

meaning that on average the spatial distribution of firms also reflects the spatial
distribution of expected profits that are common to all industries. From a practical
standpoint this assumption is equivalent to assuming a relation of proportionality
between j and xj. Note that assumption A2 is compatible with the framework of
Ellison and Glaeser (1997).7 This means that we can now rewrite (3.2) as,
expðlog xj þ jk Þ xj expðjk Þ
pjk ¼ PJ ¼ PJ : ð3:5Þ
j¼1 expðlog xj þ jk Þ j¼1 xj expðjk Þ

Looking back at (3.1), we see that estimates of jk would provide an ideal measure of
the degree of localization of an industry at a given location. Based on expression (3.5)
we can derive such estimator for jk . The idea is to treat the jk as constants that need
to be estimated. To implement this approach, consider a given industry with nk firms

5 Under different assumptions, it is possible to build theoretical models that predict the opposite. As
indicated in the introduction, Manning (2007; emphasizing monopsony in labor markets) and (Holmes
and Stevens (2004; drawing on the ideas from the new economic geography) developed models that
predict firms inside agglomerations to be larger than outside such areas. However, these are not models
in the Marshallian tradition, which stresses external economies based on knowledge spillovers, firm
specialization, and better matching in the labor market as the sources of agglomeration.
6 We can think of regions with more overall manufacturing employment as having the potential to
generate higher profits for any industry. Ellison and Glaeser (1997) suggest that other variables, such as
population, could also be used to ‘proxy’ profit levels.
7 For an example see Guimarães et al. (2007). In that paper, the authors proposed a specific distribution
for jk, the gamma distribution. That distribution satisfies both assumptions A1 and A2 as well as the
other assumption in Ellison and Glaeser (1997) regarding the variance of pjk.
 Localization economies and establishment size . 859

spatially distributed as ðn1k ; n2k ; . . . ; nJk Þ. Now, the likelihood of observing that
particular distribution of plants is given by:
Y
J
n
Lk ¼ pjkjk :
j

Taking logs, maximizing with respect to exp(jk) and solving the first-order conditions,
we obtain the following set of equations:
 
njk =nk
expðjk Þ ¼ A ¼ Qzjk A; ð3:6Þ
xj =x
P
where A ¼ x1 Jj¼1 xj expðjk Þ is a constant specific to each industry. Interestingly, the
likelihood estimates of the expðjk Þ are equal to the product of a location quotient
calculated in terms of plant counts, Qzjk , and an unknown constant. This means that
we can use Qzjk as a measure of the intensity of agglomeration economies of an industry
at given locations but those numbers are not comparable across industries. Note
that in expression (3.6), we could have obtained the traditional Florence (1939)
x =x
employment location (Qxjk ¼ xjkj =xk ) had we assumed that the location probabilities of
firms in an industry were weighted by industry employment instead of number of
plants. However, we argue here, as in Guimarães et al. (2007), that Marshallian
localization economies should be measured using establishments (or plants), for the use
of employment leads to a statistic that picks both internal and external economies of
scale.8 Rooted in the dartboard location model, the above expression provides a
theoretically sound alternative to the traditional Florence’s (1939) employment location
quotient, the measure used in Holmes and Stevens (2002) to test the relationship
between establishment scale and localization. To our knowledge, this is the first time
that this interpretation is given to the location quotient.

3.3. Econometric issues

To explore the relation between localization and establishment size, Holmes and
Stevens (2002) regressed an industry-standardized measure of establishment size on the
employment location quotient. The study considered two levels of analysis: a ‘plant
level’, with individual data for each establishment, and a ‘location level’, with data
grouped by industry and region. The variables for the two levels of analysis were
computed differently because in the ‘plant level’ regression the inclusion of the own
plant employment on the variables on both sides of the regressions could bias the
results. For the ‘location level’ regressions the measure of plant size and the location
quotient were computed as Qsjk ¼ ðxjk =njk Þ=ðxk =nk Þ and Qxjk ¼ ðxjk =xj Þ=ðxk =xÞ, respec-
tively. In the ‘plant level’ regressions the own plant employment was excluded from
the calculations and the corresponding measures were Qesijk ¼ ðxi Þ=ððxk  xi Þ=ðnk  1ÞÞ

8 Note that, when measuring geographic concentration with employment measures, the level we obtain for
a given industry in a region results from two components: ‘agglomeration of firms’ (the number of firms
in that industry and region) and ‘industrial concentration’ (the size of plants in the industry and region).
To measure agglomeration, we need to look at geographic concentration conditional on industrial
concentration. Industrial concentration is a matter of internal economies of scale, while agglomeration
concerns external economies. Since we are using plant counts instead of employment in the numerator of
our localization measure, industrial concentration, and thus internal scale economies, are by definition
excluded from the localization measure.
860 . Figueiredo et al.

and Qexijk ¼ ððxjk  xi Þ=ðxj  xi ÞÞ=ððxk  xi Þ=ðx  xi ÞÞ. After computing these variables,
the authors implemented simple log–log linear regressions of Qsjk on Qxjk and Qesijk
on Qexijk .
Like Holmes and Stevens (2002), we regress industry-standardized establishment
size (Qs) on the measure of agglomeration derived in (3.6). Thus, our regression for
the ‘location level’ analysis is:
log Qsjk ¼  þ  logðQzjk Ak Þ þ ij ;

which rearranged yields,

log Qsjk ¼ k þ  log Qzjk þ ij : ð3:7Þ

The specification above requires the introduction of industry-fixed effects. To obviate

the small sample bias that comes from having the own plant contributing to both left
and right hand side variables and to maintain comparability with Holmes and Stevens
(2002), in our plant-level analysis we use the regression in (3.7), replacing Qs by Qes and
Qz by a location quotient in terms of plant counts in which the current plant is excluded
from the calculation ðQezijk ¼ ½ðnjk  1Þ=ðnk  1Þ=½ðxj  xi Þ=ðx  xi ÞÞ.9

4. Data
As found in Guimarães et al. (2000, 2007) and Cabral and Mata (2003), among other
scholarly studies, we were fortunate to have access to a detailed annual survey provided
by the Portuguese Ministry of Employment—the Quadros do Pessoal database. This
survey collects information for all the establishments operating in Portugal, except
family businesses without wage-earning employees. The survey includes information
on establishment location, sector of activity, actual employment level and since 1995
firm start-up date. We restricted our analysis to manufacturing plants and used data
spanning from 1995 to 2005 (the most recent available year). Relying on the 5-digit (325
industries) classification of the Portuguese Standard Industrial Classification system
(CAE rev.2), we make use of the 275 Portuguese concelhos as the spatial units of
analysis.10 In Table 1 we provide summary statistics of our variables of interest.
Throughout this 11-year period, we have in our final dataset a total of 106,810 plants
with an average size of 17.7 employees (standard deviation of 60.2). These correspond
to 439,092 observations at the plant level.11 Table 2 reports correlations between the
main variables used in the regressions. As expected, because both statistics are picking
geographic concentration of economic activity, our localization measure is positively
correlated with Florence’s employment location quotient. However, the correlation
between establishment size and the two alternative measures of localization show
different signs. This contrasting result serves as a prelude for the regression analysis
that follows.

9 This small sample problem is not as relevant in our paper as it is in Holmes and Stevens (2002). Without
the correction the results remain practically unchanged.
10 The concelho is a Portuguese administrative region roughly equivalent to a US county, but with a
smaller average area. Over time, there were some minor changes in the number of concelhos. To
maintain compatibility, we used the spatial breakdown of 275 concelhos that was valid (for the Portugal
mainland) from 1995 through 1998. These have an average area of 322.5 square kilometers.
11 Cabral (2007) noted that the size distribution of Portuguese manufacturing plants is similar to what is
observed for other industrialized countries, even though plants are on average smaller.
 Localization economies and establishment size . 861

Table 1. Summary statistics

Plant level Location level

log Qesijkt log Qexijkt log Qezijkt log Qsjkt log Qxjkt log Qzjkt

Mean 0.72 0.64 0.85 0.68 0.41 1.09

Standard deviation 1.18 1.45 1.28 1.16 1.81 1.50
Minimum 6.32 7.42 4.22 6.33 6.80 4.29
Maximum 5.17 6.77 7.00 4.89 9.44 9.44
N 439,092 109,788

Note: The t subscript is for year.

Table 2. Spearman rank correlation between variables of interest

log Qexijkt log Qezijkt

Plant level
log Qesijkt 0.023* 0.061*
log Qezijkt 0.872* –
N 439,092

log Qxjkt log Qzjkt

Location level
log Qsjkt 0.554* 0.074*
log Qzjkt 0.750* –
N 109,788

*Statistical significance at the 1% level; the t subscript is for year.

5. Results
Tables 3–6 present our results. All regressions were estimated with a clustered-robust
correction for the standard errors. The correction accounts for possible unobservable
correlation between repeated observations throughout the 11-year period of our sample
and tends to produce rather conservative t-statistics.12
With regressions at the plant and location levels, Table 3 uses Florence employment
location quotients to measure localization. At the plant level, the regression in column 1
implements the specification in Holmes and Stevens (2002). As in Holmes and Stevens
(2002) and Barrios et al. (2006), we find unequivocal evidence of a positive and
statistically significant relationship between firm size and agglomeration.13 Then, in
the second regression (column 2), we introduce industry-fixed effects accounting for

12 For a discussion of cluster-robust standard errors, see Cameron and Trivedi (2005). In our application,
this always generates more conservative estimates than conventional heteroskedasticity-robust (White)
standard errors. For the ‘plant level’ regressions, the clusters are defined at the establishment level. In
the ‘location level’ analysis, we set the clusters at the region/industry level.
13 For comparability purposes, we estimated regressions without industry-fixed effects. However, as
shown earlier, it is more appropriated to include these effects in the regressions because location
quotients should not be compared across sectors.
862 . Figueiredo et al.

Table 3. Regression estimates with employment location quotients

Plant level Location level

(1) (2) (3) (4) (5) (6)

log Qex 0.029 (9.9) 0.045 (13.8) 0.048 (15.2) 0.020 (6.9) – –
log Qx – – – – 0.361 (84.8) 0.438 (100.9)
Fixed effects
Year Yes Yes Yes Yes Yes Yes
Industry No Yes Yes – No Yes
Age No No Yes – – –
Plant No No No Yes – –
N 439,092 439,092 422,719 439,092 109,788 109,788

Note: t-Statistics associated with cluster-robust standard errors are given in parenthesis.

industry-specific characteristics that may affect establishment size. When we do this,

the evidence in favor of a positive relationship becomes more compelling. Both the
estimated coefficient and the t-statistic increase. Next, we add an additional control to
account for the well-established positive relation between the firm size and age (Mata
et al., 1995; Cabral and Mata, 2003). Age was categorized in five classes and introduced
as a fixed effect.14 Again, we observe (column 3) increases in the size of the coefficient
and its t-statistic. Finally, in the fourth regression (column 4) we introduce the more
stringent control, an establishment-specific fixed effect that accounts for all establish-
ment characteristics that may influence an establishment’s size.15 Even in this case
where we are relying exclusively in temporal variation to estimate the model, we still
observe the same qualitative result between size and agglomeration. For comparison
purposes, Table 3 also displays location-level regressions (columns 5 and 6). Here,
as in Holmes and Stevens (2002), both the estimated coefficients and the t-statistics
increase substantially possibly indicating the existence of an aggregation bias. In short,
the Portuguese data show the same robust positive association between localization
(measured with employment quotients) and scale as found in Holmes and Stevens
(2002) for the USA and Barrios et al. (2006) for Ireland.
Table 4 implements the same regressions as in Table 3 but using as independent
variable the location quotient based on plant counts derived in (3.6). As previously
argued, this measure is more closely related to Marshallian external economies than
Florence’s employment location quotient. The results contradict the previously found
positive association between localization and establishment size. In this case (column 1),
the estimates for the -coefficient turned out to be negative and statistically significant.
To be sure, this effect is relatively small. Keeping everything else constant, if a plant
moved from a region with a (count-based) location quotient of 1 to another region with
a location quotient of 2 that plant would see a decrease of 3.5% in its employment. For
the average size plant in our dataset (17.7 employees) this would represent a decrease of

14 These five classes of age are: [0,1], [2,5], [6,10], [11,25] and more than 26 years.
15 Besides age, there are other firm-specific characteristics that have been linked to firm size. For example,
Bernard and Jensen (1999) have shown that exporting firms tend on average to be larger.
 Localization economies and establishment size . 863

Table 4. Regression estimates with plant count location quotients

Plant level
Location level
(1) (2) (3) (4)

log Qez 0.035 (9.3) 0.027 (7.4) 0.017 (4.5) –

log Qz – – – 0.051 (8.0)
Fixed Effects
Year Yes Yes Yes Yes
Industry Yes Yes – Yes
Age No Yes – –
Plant No No Yes –
N 439,092 422,719 439,092 109,788

Note: t-Statistics associated with cluster-robust standard errors are given in parenthesis.

0.6 workers, while for the largest plant (4229 employees) that decrease would amount to
148 workers.
As before, we tried several different specifications controlling for industry, age and
plant-specific effects. Table 4 shows that the negative relation between scale and
localization is robust across specifications and that the estimated coefficients are quite
stable even for the location-level regression. These results uphold the predictions of
Marshall and the industrial district literature regarding firm size inside areas of industry
agglomeration.
As additional tests of robustness we also ran regressions using different levels of
sectorial and regional disaggregation.16 The regressions shown in Tables 5 and 6 include
both the specification of Holmes and Stevens (2002) and ours. In Table 5, we use
the 3-digit (103 industries) classification of the Portuguese Standard Industrial
Classification system (CAE rev.2). As before, we make use of the 275 Portuguese
concelhos as the spatial units of analysis. In turn, for the regressions in Table 6 we
maintain the 5-digit sectorial breakdown of the CAE but rely on a spatial breakdown
in distritos, a higher Portuguese administrative region level composed of several
adjacent concelhos.17 As shown in these tables, the essential results still hold. The only
exception is the regression for logQez shown in column 3 of Table 6. In that case the
coefficient is not statistically significant. Note, however, that this is a tricky regression,
since we are relying solely in temporal variation to estimate the -coefficient. Also,
the distrito is a large geographic unit that encompasses substantial heterogeneity
within itself. For that reason, it is not the ideal spatial unit of analysis to account
for agglomerations in Portugal.

16 We also tried another robustness test not reported here. Namely, we re-ran all regressions dropping five
sectors that had a very small number of plants. In some cases the location quotients for these sectors
assumed large values. However, regression results were not affected by these few extreme values.
17 The Portuguese mainland is divided in 18 distritos and the average area of each distrito in our dataset is
4926 square kilometers. For comparison purposes note that average county size in the USA is
approximately 2650 square kilometers.
864 . Figueiredo et al.

Table 5. Regression estimates for concelhos and 3-digit SIC industries

Plant level estimates with Qex

(1) (2) (3) (4)

log Qex 0.039 (12.4) 0.047 (14.2) 0.052 (16.0) 0.016 (5.4)
Fixed effects
Year Yes Yes Yes Yes
Industry No Yes Yes –
Age No No Yes –
Plant No No No Yes
N 470,920 470,920 453,469 470,920

Plant level estimates with Qez

(1) (2) (3)

log Qez 0.037 (9.6) 0.026 (7.0) 0.007 (1.7)

Fixed effects
Year Yes Yes Yes
Industry Yes Yes –
Age No Yes –
Plant No No Yes
N 470,920 453,469 470,920

Note: t-Statistics associated with cluster-robust standard errors are given in parenthesis.

Table 6. Regression estimates for distritos and 5-digit SIC industries

Plant level estimates with Qex

(1) (2) (3) (4)

log Qex 0.053 (11.8) 0.055 (11.8) 0.057 (12.5) 0.048 (9.6)
Fixed effects
Year Yes Yes Yes Yes
Industry No Yes Yes –
Age No No Yes –
Plant No No No Yes
N 486,743 486,743 468,837 486,743

Plant level estimates with Qez

(1) (2) (3)

log Qez 0.053 (10.1) 0.046 (8.8) 0.005 (0.7)

Fixed effects
Year Yes Yes Yes
Industry Yes Yes –
Age No Yes –
Plant No No Yes
N 486,743 468,837 486,743

Note: t-Statistics associated with cluster-robust standard errors are given in parenthesis.
 Localization economies and establishment size . 865

6. Conclusion
This article re-examines the relationship between the localization of an industry and
establishment employment size. To assess localization, we propose an approach
following Ellison and Glaeser’s (1997) dartboard location model. Given several
advantages emphasized in the article, our localization measure provides an alternative
to Florence’s (1939) still commonly used employment location quotient. Like the
index in Ellison and Glaeser (1997), our statistic has a theoretical foundation. Because
the measure is based on a probabilistic model, it can also account for spurious
geographic concentration (i.e. concentration that occurs by chance alone). Importantly,
our method purges the effect of internal economies of scale from the localization index.
In turn, it more closely reflects the Marshallian notion of external economies.
Using this new measure, we tested the relationship between localization and
establishment employment scale with Portuguese data. It appears that Marshall was
right, at least in the context of establishment (rather than employment) concentration.
We find evidence that plants located in areas where an industry exhibits localization
tend to be smaller than plants in the same industry outside such areas. The results,
which fall in line with Marshall’s theory of external economies in industrial districts,
are robust across several tests. Previous studies on establishment scale have not actually
tested the Marshallian externalities hypothesis because they use localization measures
that combine internal scale economies and local industry externalities. While our
analysis supports Marshall and the industrial districts literature in the case of Portugal,
further study of this relationship is needed in other contexts.
These findings should also be situated within the large body of literature on
agglomeration that compares the productivity of plants in concentrated areas with
the productivity of plants outside these areas (Henderson, 1986; Ciccone and Hall,
1996; Ciccone, 2002). Upholding Marshall, these studies found a positive effect on
productivity across regions from the density of economic activity. Our results also imply
that plants inside industrial districts benefit from an external productivity bonus over
more geographically dispersed plants.

Acknowledgements
The authors acknowledge the support of FCT, the Portuguese Foundation for Science and
Technology. We would also like to thank the Ministry of Employment, Statistics Department, for
the permission to use the Quadros do Pessoal database and Miguel Portela and João Cerejeira for
their help in accessing the data.

References
Barrios, S., Bertinelli, L., Strobl, E. (2006) Geographic concentration and establishment scale:
an extension using panel data. Journal of Regional Science, 46: 733–746.
Bernard, A. B. and Jensen, J. B. (1999) Exceptional export performance: cause, effect, both?
Journal of International Economics, 47: 1–25.
Cabral, L. M. B. (2007) Small firms in Portugal: A selective survey of stylized facts, economic
analysis, and policy implications. Portuguese Economic Journal, 6: 65–88.
Cabral, L. and Mata, J. (2003) On the evolution of the firm size distribution: facts and theory.
American Economic Review, 93: 1075–1090.
Cameron, A. C. and Trivedi, P. K. (2005) Microeconometrics: Methods and Applications.
Cambridge, NY: Cambridge University Press.
866 . Figueiredo et al.

Carlton, D. (1983) The location and employment choices of new firms: an econometric model
with discrete and continuous endogenous variables. The Review of Economic and Statistics, 65:
440–449.
Ciccone, A. (2002) Agglomeration effects in Europe. European Economic Review, 46: 213–227.
Ciccone, A. and Hall, R. E. (1996) Productivity and the density of economic activity. American
Economic Review, 86: 54–70.
Duranton, G. and Overman, H. (2005) Testing for localisation using microgeographic data.
Review of Economic Studies, 72: 1077–1106.
Duranton, G. and Overman, H. (2008) Exploring the detailed location patterns of UK
manufacturing industries using microgeographic data. Journal of Regional Science, 48:
213–243.
Ellison, G. and Glaeser, E. (1997) Geographic concentration in U.S. manufacturing industries:
a dartboard approach. Journal of Political Economy, 105: 889–927.
Florence, P. S. (1939) Report on the Location of Industry. London, UK: Political and Economic
Planning.
Goodman, E. and Bamford, J. (1989) Small Firms and Industrial Districts in Italy. New York:
Routlege.
Guimarães, P., Figueiredo, O., Woodward, D. (2000) Agglomeration and the location of foreign
direct investment in Portugal. Journal of Urban Economics, 47: 115–135.
Guimarães, P., Figueiredo, O., Woodward, D. (2007) Measuring the localization of economic
activity: a parametric approach. Journal of Regional Science, 47: 753–774.
Henderson, J. V. (1986) Efficiency of resource usage and city size. Journal of Urban Economics,
19: 47–70.
Holmes, T. and Stevens, J. (2002) Geographic concentration and establishment scale. The Review
of Economics and Statistics, 84: 682–690.
Holmes, T. and Stevens, J. (2004) Geographic concentration and establishment size: analysis
in an alternative economic geography model. Journal of Economic Geography, 4: 227–250.
Hoover, E. M. (1936) The measurement of industrial localization. The Review of Economic and
Statistics, 18: 162–171.
Hoover, E. M. (1937) Location Theory and the Shoe and Leather Industries. Cambridge, MA:
Harvard University Press.
Kim, S. (1995) Expansion of markets and the geographic distribution of economic activities: the
trends in U.S. regional manufacturing structure, 1860–1987. Quarterly Journal of Economics,
110: 881–908.
Lafourcade, M. and Mion, G. (2007) Concentration, agglomeration and the size of plants.
Regional Science and Urban Economics, 37: 46–68.
Lichtenberg, R. (1960) One-Tenth of a Nation: National Forces in the Economic Growth of the
New York Region. Cambrige, MA: Harvard University Press.
Manning, A. (2007) The Plant Size–place Effect: Agglomeration and Monopsony in Labour
Markets. London: CEP Discussion Paper No. 773.
Marcon, E. and Puech, F. (2003) Evaluating the geographic concentration of industries using
distance-based methods. Journal of Economic Geography, 3: 409–428.
Markusen, A. (1996) Sticky places in slippery space: a typology of industrial districts. Economic
Geography, 72: 293–313.
Marshall, A. (1919) Industry and Trade, 3rd edn. 1920, London: Macmillan.
Mata, J., Portugal, P., Guimarães, P. (1995) The survival of new-plants: Start-up conditions
and post-entry evolution. International Journal of Industrial Organization, 13: 459–481.
McFadden, D. (1974) Conditional logit analysis of qualitative choice behavior (chapter 4).
In P. Zarembka (ed.) Frontiers in Econometrics, pp. 105–142. NY: Academic Press.
Mori, T., Nishikimi, K., Smith, T. (2005) A divergence statistic for industrial localization.
The Review of Economic and Statistics, 87: 635–651.
Piore, M. J. and Sabel, C. F. (1984) The Second Industrial Divide. New York: Basic Books.
 Localization economies and establishment size . 867

Appendix A
The expected profit of a firm conditional on its decision to locate in a particular
location, say e
j, is:

Eðlog ijk j e
j Þ ¼ log e þ e þ Eð"ijk j e
jÞ ðA1Þ
j jk

To calculate Eð"ijk j e
j Þ; we first need to derive the distribution of "ijk j e
j. By assumption "ijk
is i.i.d. with Extreme Type Value I distribution. We can use Bayes theorem to derive:
" #
1
Prð ej \ "ijk Þ ¼ expð"ijk Þ exp  expð"ijk Þ ;
Prð e
jÞ
and from there obtain:
" #
1 1
Prð"ijk j e
jÞ ¼ expð"ijk Þ exp  expð"ijk Þ :
Prð e
jÞ Prð e
jÞ

It then follows that:

Z " #
þ1
1 1
Eð"ijk j e
jÞ ¼ x expðxÞ exp  expðxÞ dx
Prð e
jÞ 1 Prð e
jÞ
¼  lnðPrð e
j ÞÞ þ ; ðA2Þ

where  is the Euler–Mascheroni constant ( ¼ 0:577215 . . .). Replacing (A2) into (A1)
and simplifying we obtain:
X
J
Eðlog ijk j e
jÞ ¼ expðlog j þ jk Þ þ ;
j¼1

which shows that regardless of location, expected profits are constant and identical
for all firms in the same industry.

Appendix B
Following Manning (2007), suppose that the revenue function for the typical firm
located in region j may be expressed as AjR(N) where Aj includes all ways in which
agglomeration affects profits and R(N) is a function of employment. It is assumed that
other factors affecting profits have been concentrated out. This means that we can write
the ‘reduced-form’ profit function as:
j ¼ Aj RðNÞ  Wj N; ðB1Þ
where Wj is the wage in region j. The first-order condition for profit maximization
with respect to employment yields:
Aj R0 ðNÞ  Wj ¼ 0;
allowing us to derive a labor demand curve at the employer level,
N ¼ Nd ðWj =Aj Þ; ðB2Þ
868 . Figueiredo et al.

which must be decreasing in its argument. Thus we can replace (B2) into (B1) to obtain:
j ¼ Aj d ðWj =Aj Þ;
where d (.) is a function that must also be decreasing in its argument. If profits are
identical across regions as suggested by the Random Utility Maximization framework,
then in regions with higher agglomeration (higher Aj) it must also be true that Wj/Aj is
higher. This in turn implies that firms have smaller N in regions of agglomeration.
Through a similar argument, Manning (2007) shows that the same result is obtained if
we assume that the productivity effect of agglomeration is labor-augmenting, that is,
that the revenue function is expressed as R(AjN). For a more detailed discussion of
these (and other) models see Manning (2007).