Accepted Manuscript

Why are some regions more innovative than others? The role of small firms in
the presence of large labs

Ajay Agrawal, Iain Cockburn, Alberto Galasso, Alexander Oettl

PII: S0094-1190(14)00026-6
DOI: [Link]
Reference: YJUEC 2930

To appear in: Journal of Urban Economics

Received Date: 9 April 2013

Revised Date: 8 March 2014

Please cite this article as: Agrawal, A., Cockburn, I., Galasso, A., Oettl, A., Why are some regions more innovative
than others? The role of small firms in the presence of large labs, Journal of Urban Economics (2014), doi: http://
[Link]/10.1016/[Link].2014.03.003

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
 Why are some regions more innovative than others?
The role of small firms in the presence of large labs∗
Ajay Agrawal Iain Cockburn
University of Toronto and NBER Boston University and NBER
[Link]@[Link] cockburn@[Link]
Alberto Galasso Alexander Oettl
University of Toronto and CEPR Georgia Institute of Technology
[Link]@[Link] [Link]@[Link]

March 26, 2014

Abstract
We study the impact of small firms on innovation in regions where large labs are present.
Small firms generate demand for specialized services that lower entry costs for others. This
effect is particularly relevant in the presence of large firms that spawn spin-outs caused by
innovations deemed unrelated to the firm’s overall business. We examine MSA-level patent
data during the period 1975-2000 and find that innovation output is higher in regions
where both a sizable population of small firms and large labs are present. The finding is
robust to across-region as well as within-region analysis and the effect is stronger in certain
subsamples in a manner that is consistent with our explanation.

JEL Classifications: 047, 018, L16.

Keywords: innovation, spin-outs, cities, patents, firm size diversity, externalities

∗
We thank the Editor, Ed Glaeser, the anonymous referees, as well as Ashish Arora, Lee Branstetter, Bruno
Cassiman, Joshua Gans, Avi Goldfarb, Brent Goldfarb, Steve Klepper, Nico Lacetera, Saul Lach, Matt Marx,
Carlo Menon, Matt Mitchell, Mark Schankerman, Will Strange, Heidi Williams, Michael Wyrwich, and semi-
nar participants at the University of Toronto, the London School of Economics, the London Business School,
Carnegie Mellon University, the NBER Productivity meeting, the Wharton Technology Miniconference, the
Queen’s Conference on Innovation and Entrepreneurship, the REER conference at Georgia Tech, the CEPR
conference at the University of Cyprus, the HEC workshop on Entrepreneurship, the ZEW, the Conference on
Entrepreneurship and Innovation at Northwestern University, the HBS Strategy Conference, and the EARIE
conference in Stockholm. We gratefully acknowledge generous financial support from the Martin Prosperity
Institute, the Centre for Innovation and Entrepreneurship at the Rotman School of Management, and the Social
Sciences and Humanities Research Council of Canada for funding this research. All errors and omissions are our
own.
1 Introduction
A striking feature of economic geography is the large variation in innovation productivity across
regions. Silicon Valley and Boston are popular examples of regions that are significantly more
productive than others in terms of innovation. In Figure 1, we illustrate a broader cross section
of such variation using patent data on US computers and communications. Even Metropolitan
Statistical Areas (MSAs) of a similar size in terms of the number of local inventors often differ
substantially in terms of their innovation productivity (number of citation-weighted patented
inventions per inventor). For example, Rochester, NY and Portland, OR had a similar number
of innovators working in the computer and communications industry in 1995, but Portland
inventors generated almost double the number of citation-weighted patents.
Regional productivity disparities have led to a variety of policies focused on enhancing
local innovation. Such initiatives often focus either on encouraging entrepreneurship (e.g., San
Diego, CA, New York, NY, and St. Louis, MO) or on attracting large corporate labs (e.g., Flint,
MI, Greenville, SC, and Shelby, AL).1 We argue that effective regional innovation policymaking
requires an understanding of how the structure of local R&D manpower is related to innovation
productivity.
In this paper, we study how local innovation is affected by the organization of R&D man-
power in that region. For over six decades, since Schumpeter (1942), innovation scholars have
tried to understand the relationship between product market industry structure and innova-
tion (Geroski (1990), Grossman and Helpman (1991), Cohen and Klepper (1996), and Aghion
et al. (2005)). Other research has focused on the relationship between innovation and regional
industrial diversity, for example, comparing innovative output from cities focused primarily on
one industry (e.g., automobiles) with industrially diverse cities (e.g., electronics, chemicals, and
textiles) (see Jacobs (1969), Glaeser et al. (1992), Feldman and Audretsch (1999), and Delgado
et al. (2010)). Despite this extensive literature, the effect of R&D labor organization on local
innovation has so far attracted little empirical and theoretical attention. We fill this gap by
combining insights from urban economics and entrepreneurship.2

1
The 2008 Brookings Institute’s Blueprint for American Prosperity offers a comprehensive overview of such
regional initiatives. For example, in San Diego, the CONNECT program has helped the development of more
than 2,000 small firms in the hi-tech and bio-tech sectors since 1985. New York recently launched the NYC High-
Tech CONNECT program modeled on San Diego’s CONNECT. Similarly, St. Louis implemented a number of
policies to promote regional entrepreneurship. Flint, Greenville, and Shelby focused instead on attracting large
firms - GM, BMW, and Mercedes, respectively.
2
See Cohen (2010) for an excellent survey of the ‘neo-Schumpeterian’ empirical literature. Our paper also
contributes to the literature on spin-out formation. While this literature has explored the impact of parent firm
characteristics on spin-out performance (Franco and Filson (2006)) and contrasted spin-outs to other entrants
(Chatterji (2009)), our paper is to our knowledge the first to examine the impact of regional R&D manpower
organization on local spin-out formation.

1
 A number of previous studies provide guidance to our analysis of the impact of R&D
labor organization on regional innovation. First, Vernon (1960) and Chinitz (1961) argue that
an increasing number of small firms “thicken” local markets for ancillary services and thus re-
duce the cost of spin-out formation. Second, Schumpeter (1942) and Galbraith (1952) suggest
that large firms may have an advantage in the production of ideas.3 Third, Cassiman and
Ueda (2006) argue that large firms only commercialize innovations that “fit” with their estab-
lished research activities. However, if potentially profitable, then spin-outs may commercialize
“misfit” inventions that do not fit with the assets, mandate, or strategy of the parent firm.4
These forces indicate that the manner in which regional R&D manpower is organized may
have an impact on local innovation. In particular, they suggest that innovation productivity is
greater in MSAs where a sizeable population of small labs is present together with at least one
large lab. This is because spin-out formation requires the presence of large labs and small firm
market thickness lowers the cost of entry, rendering a spin-out more profitable. This suggests
that spin-out formation is enhanced when numerous small labs and at least one large lab are
present. Because spin-outs allow innovators to commercialize inventions that would otherwise
be abandoned since they are not a good fit with their employer’s research activities, the number
of commercialized inventions also increases when both types of labs are present.
We test these empirical implications using a 26-year panel dataset at the MSA-technology-
year level.5 The data show a substantial regional innovation premium in MSAs “diverse” in
firm-sizes, which we define as MSAs where numerous small labs coexist with at least one large
lab compared to MSAs of a similar size without many small labs or a large lab. For example,

3
This typically arises when the lab can spread R&D fixed costs over a larger number of innovations (see Cohen
and Klepper (1996) for a micro-foundation). Empirical evidence of such an advantage is provided in Klette
(1996), Henderson and Cockburn (1996), and Cockburn and Henderson (2001). Alternatively, scale advantages
may arise from division of labor efficiencies (Arora and Gambardella (1994)) or human capital complementarities
(Jones (2008)).
4
Prominent examples of such spin-outs include: Intel, founded by Andy Grove, Bob Noyce and others to
make a product that Fairchild was unwilling to make; Lotus Development, founded by Mitch Kapor, that left
Digital Equipment Corporation; and FreeMarket, founded by a General Electric (GE) engineer after GE rejected
his initial proposal. In 2002, the Wall Street Journal reported that in 2001 GE’s researchers suggested more
than 2,000 new products but only five proposals were accepted for product development (see Cassiman and Ueda
(2006) and Klepper and Sleeper (2005) for additional examples).
5
A number of case studies also provide support for our theory. For example, consider Portland, OR versus
Rochester, NY (lack of small firms) and Atlanta, GA versus Seattle, WA (lack of large firms) in 1995. In terms of
Portland and Rochester, the number of inventors patenting in the “computers and communications” technology
class is very similar in the two cities (roughly 1,000 inventors). Nonetheless, Portland outperforms Rochester,
obtaining almost 50% more patents and about twice the number of citation-weighted patents than Rochester.
While both cities register a similar presence of large labs, the number of small labs is substantially different:
Portland has more than five times the number of small labs as Rochester. On the other hand, in the “chemicals”
technology class, Seattle and Atlanta have a similar number of small labs (38 and 36, respectively) and also
a similar number of overall inventors (457 and 484, respectively), but only Atlanta has a large lab (Kimberly
Clark). The difference in innovation output: Atlanta has 37% more citation-weighted patents.

2
focusing on between-region variation, we find that in 1995 “diverse” regions have an average
47% innovation productivity premium five years later.
The empirical variation we exploit in these regressions is mostly driven by changes in
the population of small labs in regions where at least one large lab is present. This is because
regions with a sizeable population of small labs typically have large labs as well. Thus, we
interpret a switch to one in the “diverse” indicator as an increase in the number of small labs
where a large lab is present.
We approach the cross-sectional correlations with caution because firm size composition
and regional innovation are surely endogenous. In other words, although our focus is on whether
and how firm size composition influences region-level innovation, regional innovation likely
influences local firm size composition. For example, regions that are more innovative, perhaps
due to large companies and/or universities that spend heavily on R&D, likely generate more
new small firms that increase the likelihood that those regions will be “diverse.” In addition,
small firms that are especially innovative are more likely to either grow or attract large firms
into their region, increasing the likelihood that those regions become diverse. Furthermore,
unobserved characteristics of a region may affect both the local allocation of R&D labor as
well as innovation. For example, a positive shock in the value of technologies produced in the
MSA-class (e.g., regional variation in expertise in software development for mobile devices at
the time of the arrival of the first iPhone) may lead to an increase both in the entry of small
firms and in the likelihood of innovation and lead to an upward bias in the OLS estimates.
Downward bias is possible too. For example, successful innovation may induce incumbents to
deter entry of new firms, making diversity less likely.
Thus, we take a series of steps to reject the null hypothesis that small firms in the
presence of a large lab do not play a role in influencing regional innovation. First, we employ
an estimation approach that controls for MSA-class specific attributes (with MSA-class fixed
effects) and general technology trends (with class-year effects). When we focus on within-
region-technology class variation over time (1975-2000) and use MSA-technology fixed effects
(our baseline specification), we find that in periods where at least one large lab and numerous
small labs co-exist, MSAs experience a 17% increase in citation-weighted patent counts per
inventor relative to periods when those MSAs have below this threshold level of diversity.
In addition, to address the concern that our diversity measure is simply capturing vari-
ation in regional product market competition that is correlated with innovation, we show that
our results are robust to introducing additional measures of industry rivalry. Similarly, our find-
ings are not affected when we control for more detailed measures of agglomeration, suggesting
that diversity is not simply a proxy for regions with a large number of inventors.
Next, we turn to exploring the mechanism that links firm size diversity to regional inno-
vation. First, we show that diversity is associated with a 28% increase in spin-out formation.

3
Second, we expect that any barrier to spin-out formation will reduce the beneficial effect of firm
size diversity. We show that the effect of firm size diversity on innovation is indeed reduced
by the presence of strong non-compete laws. Third, since spin-out formation is predicated
on ideas produced by large labs that are subsequently deemed unrelated, we expect that re-
gions with large labs that maintain a narrower focus and thus produce more “misfit ideas”
will benefit more from firm size diversity. We show that the effect of firm size diversity on
innovation is indeed higher in regions with more narrowly focused large labs. We exploit these
mechanism-related findings to assess the validity of our theory against competing explanations
for the positive association between regional firm size diversity and innovation. Several alter-
native theories are consistent with a subset of the correlations we report but none with the
whole set. Our approach is not intended to rule out these alternate explanations, but rather to
provide evidence suggesting these alternatives do not alone drive the estimated correlation and
at the same time reject the null hypothesis that firm size diversity does not influence regional
diversity.
To further support our theory, we show that the increase in innovation generated by
diversity has long-lasting effects and is predominantly driven by the patenting activity of small
firms. We also show that the positive correlation between diversity and innovation is robust to
the inclusion of MSA-year effects that capture unobserved time-varying heterogeneity at the
regional level.
To further address endogeneity, we show that this result is robust to using lagged income
tax rates, which vary at the state level, as an instrument for regional firm size diversity. High
income tax rates induce entry since they increase the effective income of an owner relative to an
employee (Cullen and Gordon (2007); Poterba (1989)) and thus may affect subsequent regional
firm size diversity. We use an eight year lag between the income tax rate, which we exploit
as an instrument, and innovation outcomes. Because in most technology fields the speed of
technology advances renders patents obsolete during such a long time window (Caballero and
Jaffe (2002)), we expect such lagged taxes to be correlated with current diversity levels but
exogenous to current innovation decisions. Finally, and in addition, we show that the main
result is robust to: 1) disaggregating the firm size diversity measure into separate measures
for large and small labs, 2) focusing on a smaller sample of just large MSAs, 3) focusing on
a smaller sample that drops California MSAs, 4) applying different measures of diversity with
different cutoffs for large and small firms, and 5) an alternate measure of diversity that uses
County Business Patterns Census data rather than patent data.

4
2 Theoretical Framework
The idea that large labs may have an advantage in the production of ideas goes back to Schum-
peter (1942) and Galbraith (1952). Several justifications for a positive effect of firm size on in-
novative activity have been offered in the literature (see Cohen (2010) for a survey). First, large
labs can spread the R&D fixed costs over a larger number of innovations. Second, large firms
may be able to better exploit complementarities between R&D and other non-manufacturing
activities (e.g., finance or marketing). Third, in the presence of capital market imperfections,
large firms may have an advantage in financing risky R&D projects using internally-generated
funds. Finally, large firms may reduce the risk associated with the prospective returns to
innovation through diversification.
Nevertheless, large labs may not exploit all the ideas that are produced inside their
boundaries, and employees may consider leaving their employer and commercializing the inno-
vation through a spin-out. This may occur because the limited commercialization capacity of
large firms may induce them to focus only on innovations that narrowly fit their established
commercial activities (Christensen (1997); Cassiman and Ueda (2006)). Imperfect intellectual
property protection may also induce scientists to leave large labs without revealing the ideas
generated in previous employment (Anton and Yao (1995); Gans and Stern (2002)). Finally,
disagreement among members of management teams may be more likely in large labs, which
may prompt dissidents to leave (Klepper and Thompson (2010)).
The tension between advantages in idea production and disadvantages in idea exploita-
tion is consistent with the mixed empirical evidence on the relationship between firm size and
R&D productivity. The results in Klette (1996), Henderson and Cockburn (1996), and Cock-
burn and Henderson (2001) indicate an advantage in innovation activities for large firms. At the
same time, a variety of other studies show that innovative output may increase proportionally
or less than proportionately than firm size (Cohen (2010)).
This trade-off also suggests that large labs have an ambiguous impact on regional innova-
tion. If large labs develop only few ideas that “fit” with their established research trajectories
and if it is difficult for employees to leave their employer to commercialize “misfit” inventions,
then the presence of large labs may not generate any premium in regional innovation. However,
when spin-out formation is not too costly, the commercialization of “misfit” ideas may generate
a positive association between the presence of large labs and regional innovation.
Since Vernon (1960) and Chinitz (1961), the urban economics literature has argued that
spin-outs are facilitated in geographic areas characterized by the presence of a large number
of small firms. The idea is that a sizable mass of small firms will “thicken” local markets for
ancillary services and thus reduce entry costs. Helsley and Strange (2002) provide a micro-
foundation for this idea, developing a model where a large number of small labs generate a dense

5
network of input suppliers that facilitates spin-out formation. Natural candidates for industry
specific ancillary services are early-stage capital (angel investors and venture capitalist tend to
focus on specific industries) as well as specialized real estate, legal, and financial services. The
effect also may be driven by industry-specific intermediate products that large firms can produce
in-house but small firms cannot (Helsley and Strange (2002)). An alternative mechanism is
suggested by Glaeser and Kerr (2009), who argue that, in the presence of small firms, an area
develops a culture of entrepreneurship that induces employees of large firms to leave their
employer and start their own firms.
The above discussion suggests that large and small firms are complementary and that
regional innovation may be enhanced when they coexist. Intuitively, the presence of large firms
increases the number of ideas generated in the region and the expected number of spin-outs.
At the same time, a large pool of small firms lowers the cost of entry for spin-outs and thus
allows ideas that are not a good fit for a large firm to be commercialized.
In the appendix, we present a mathematical formalization of this argument. We develop
a theoretical framework where an MSA is endowed with scientists who are divided into large
and small labs. Large labs have the capability to commercialize more than one innovation,
but small labs do not (only one). We assume that the probability of a large lab discovery
increases with the number of scientists employed and that ideas differ in their “fit” with the
large lab. Large lab scientists negotiate with their labs about the destiny of their discoveries
and may commercialize the innovation outside the lab by opening a new small lab (spin-out) at
a cost that is decreasing in the presence of other small labs in the MSA. We show that regional
innovation is maximized in “diverse” MSAs, which we define as MSAs having at least one large
lab and a number of small labs above a threshold.

3 Data
We focus on two units of analysis. First, we study cross-region variation and use MSA-
technology class as our unit of analysis (e.g., Rochester, NY - chemicals). Then, we turn
our attention to within-region variation and use MSA-class-year as our unit of analysis.
In constructing our sample, we begin with the set of 268 MSAs defined in 1999 by the
US Office of Management and Budget6 and the set of six one-digit technology classes described
in Hall et al. (2001).7 This generates 1608 MSA-class observations. We then construct our
panel dataset, which includes 26 years (1975-2000) and thus contains 41,808 MSA-class-year

6
[Link]
7
This classification scheme includes: chemicals, computers and communications, drugs and medical, electrical
and electronic, mechanical, and other.

6
observations.
Patent data from the United States Patent and Trademark Office (USPTO) offer the only
practical innovation measures available for large-scale studies by providing detailed information
both on regional patenting activity and on the affiliation and location of patenting inventors
in a region. Nonetheless, certain qualification should be kept in mind. First, not all inventions
are patented, but the innovation literature has shown that technologies with greater impact on
social welfare and economic growth are more likely to be patented (Pakes and Griliches (1980);
Griliches (1990)). Second, the coding of location, affiliation, and identity of the inventors is
likely to generate random measurement error in our constructs. The associated attenuation
bias will cause us to underestimate the impact of diversity on innovation, so our estimates are
conservative in this sense. We measure innovative activity, our main dependent variable, using
a citation-weighted count of US patents:

Weighted Patentsjkt+5 : the forward citation weighted sum of distinct patents with
primary technology classification k and application year t + 5 where at least one inventor is
located in MSA j.

We use inventor address information to assign a patent to an MSA, exploiting the US

National Geospatial-Intelligence Agency dataset to match cities and townships to counties
and ultimately MSAs. If a patent has at least one inventor from a particular MSA, then we
increment the counter for that MSA by one. Thus, a patent with three inventors located in
three different MSAs increments the patent counter for each of those MSAs by one. However,
if all three inventors are located in the same MSA, then the counter for that MSA is only
incremented by one.8
We construct this measure using all patents applied for (and subsequently granted) be-
tween 1975 and 2000, with at least one inventor with a US address. We exclude patents that
cannot be attributed to an MSA (due to incomplete address information or a location out-
side an MSA) and patents assigned to universities and hospitals. The USPTO is the original
source of our patent data. We complement these data with classification data from the NBER
(technology classification, assignee name).
Patent citations identify prior knowledge upon which a patent builds, and prior literature
(starting with Pakes and Griliches (1980)) has often employed the number of forward-citations
received by a patent as an indirect measure of patent value.
We also consider an unweighted patent count as an additional innovation metric:

Patentsjkt+5 : the number of distinct patents with primary technology classification k

and application year t + 5 where at least one inventor is located in MSA j.

8
Results are robust to constructing this variable using only data from the first inventor.

7
 Our second dependent variable is a measure of spin-out formation, which we use in the
latter part of our analysis when we turn our attention to the mechanism through which firm
size diversity influences innovation output. We define a spin-out as a particular type of lab.
We define a lab in MSA j, technology class k, in year t as follows. First, we take all assignees
who, within a five-year window (year t and the four preceding years), apply for at least one
(eventually granted) patent in technology class k. Second, using this list of assignees, we
identify labs in MSA-class jk if there are at least three different inventors located in MSA j
who are named in class k patents of that assignee during the five-year window. Thus, if a firm
has operations (i.e., R&D labs or facilities) in n different MSA-classes, our procedure will treat
it as n distinct entities. We thus compute the number of patents attributed to such “labs” in
any given year by aggregating by (standardized) assignee name, location, and application date.
We define a spin-out as a new lab in MSA-class jk if among the patents applied for
during the first year of lab activity we identify at least one inventor who previously patented
in one of the large labs in MSA jk.9 We construct the following variable to measure spin-out
formation:

log(Spin-Outsjkt ): logarithm of one plus the count of spin-outs in year t MSA-class jk.

Finally, our main explanatory variable is a measure of diversity. We construct this

variable by first identifying the distribution of lab sizes in each technology class k and year t.
By construction, lab size has a lower bound of three (each lab has at least three inventors).
Across the various class-years, the median size is about five inventors, the 75th percentile is
about nine inventors, and the 97th percentile is roughly 54 inventors. We use this distribution
to define large and small labs. A large lab is a lab where the number of inventors is above the
97th percentile in the technology class-year distribution. We define a lab as small if the number
of inventors is below the 75th percentile. We arbitrarily choose the size thresholds for large
and small labs (97th and 75th percentile), but we perform robustness checks in the Appendix.

Diversejkt : dummy variable equal to one if in year t MSA-class jk has at least one
active large lab and at least 139 active small labs.

The number of small labs (139) corresponds to the 99th percentile of the distribution of
the number of small labs across the entire sample. Also, for this cut-off, we perform a number
of robustness checks in the Appendix. This measure intends to capture MSAs where both large
labs and many small labs coexist. Notice that the diverse dummy is distinct from a count of the
number of labs because it considers lab sizes. It is also distinct from traditional concentration
measures (Herfindahl or share of top four labs) because intermediate values of these measures

9
Results are robust to the exclusion of spin-outs that share the same name of large labs in other MSA-classes
jk.

8
may emerge both with and without diversity.10
Table 1 presents descriptive statistics for the main variables. The average MSA-class has
approximately 91 inventors and eight labs. On average, the inventors of an MSA-class apply
for about 28 patents per year, and these patents receive roughly 466 forward citations. The
distributions of these variables are highly skewed. The median MSA-class in our sample has
only one lab and five inventors who apply for three patents per year and receive 33 cites.
The diverse dummy equals one for about 1% of the sample. The small fraction of diverse
observations is mostly due to the fact that a large number of MSA-classes do not have enough
inventors to display a positive diversity measure. In Table 2 we focus on MSA-class-years
that are “at risk” of becoming diverse.11 The diverse dummy equals one for roughly 16%
of this smaller sample. On average, these MSA-classes have approximately 105 labs and 1348
inventors, who apply for about 381 patents per year that receive roughly 6681 forward citations.
Table 3 illustrates the variation in diversity status that will be exploited in our empirical
analysis. We observe a switch in diversity status for 36 of the 149 MSA-classes that are at risk
of becoming diverse during the sample period. The first MSA-class to become diverse is New
York (in the chemical technology area) in 1976, although switches occur throughout the entire
sample period. Eleven unique MSAs experience a switch in at least one technology class, and
all technology classes experience at least four switches.

4 Methodology
Our main econometric model focuses on the relationship between count-based measures of
innovative activity Yjkt+5 in MSA-class jk in period t + 5 and the indicator for firm size
diversity diversejkt in MSA-class jk in period t. We typically model the conditional expectation
of innovative activity as:

E[Yjkt+5 ] = exp(αdiversejkt + xjkt β + γjk + λtk ) (1)

where xjkt is a vector of control variables, γjk is an MSA-class specific idiosyncratic effect, and
λtk is a vector of technology class time-period effects. Notice that all the dependent variables
are lagged five periods to account for simultaneity concerns.

10
To see this, consider the following example. MSA A has four identical labs, each employing a quarter of
the local inventors. MSA B has one lab employing half of the local inventors as well as a very large number
of small labs. In MSA A, the Herfindahl index is equal to 1/4, and in MSA B the Herfindahl index is also
(approximately) 1/4. However, MSA B has large firm–small firm coexistence, while MSA A does not.
11
The lab size distribution is bounded below by three (by definition each lab has at least three inventors), and
across the various classes-years the 97th percentile is roughly 54 inventors. Therefore, our constructed cut-off is
equal to 3x139+54=471 inventors.

9
 Equation (1) uses the log-link formulation due to the non-negative and highly skewed
nature of our count-based dependent variables. Following Wooldridge (1999), we adopt the
Poisson quasi maximum-likelihood estimation that yields consistent estimates as long as the
conditional mean is correctly specified. We cluster the standard errors to allow for arbitrary
heteroskedasticity and autocorrelation.
When x in Equation (1) includes measures of the number of inventors working in the
MSA-class, α indicates whether MSA-classes with a diverse configuration receive more citation-
weighted patents per inventor; therefore, it is a measure of MSA-class productivity.

5 Results
5.1 Firm Size Diversity and Innovation
5.1.1 Across–Region Variation

Table 4 contains our first set of results, which show a robust positive association between firm
size diversity and innovation in cross-section regressions. We estimate all models in Table 4
using Poisson, with robust standard errors clustered at the MSA-class level to account for over-
dispersion. In Columns (1) to (3), the dependent variable is the citation-weighted patent count
or, equivalently, the total forward citation count for issued patents applied for by all inventors
in the MSA-class in year t + 5.
Column (1), focusing on year 1995, shows a large positive correlation between size di-
versity and innovation. In Column (2), we include a control for the number of inventors in the
MSA-class. The positive coefficient on diverse now indicates that diverse MSAs obtain more
innovation per inventor. In Column (3), we show that the correlation is similar when we con-
trol for technology effects. Exponentiation of the coefficient in Column (3) implies that diverse
MSA-classes have a 47% innovation premium over non-diverse MSA-classes.12 In Column (4),
we show a similar positive correlation measuring innovation with un-weighted patent counts.

5.1.2 Within–Region Variation

Focusing on year 1995, the cross-section regressions reported in Table 4 indicate a positive
association between diversity and innovation. Column (1) of Table 5 confirms this result in a
pooled cross section that exploits the entire sample period. While this correlation is consistent
with our theory, the result may be due to unobserved MSA-class heterogeneity that is correlated
with diversity and innovation. Moreover, the previous regressions include MSA-classes that do
not have enough inventors to display a positive diversity measure. To address these concerns,

12
This follows because e0.382 − 1 = 0.47.

10
in Column (2) we move to a fixed-effects Poisson estimator (Hausman et al. (1984)) with MSA-
class fixed effects and technology class-year effects and drop all the MSA-classes that are too
small to be diverse. This specification isolates the within MSA-class co-variation of diversity
and innovation. The estimated coefficient implies an increase in the citation-weighted patent
count of about 17% in periods where MSAs are diverse.
Column (3) shows that the qualitative and quantitative results are robust to introducing
additional controls, including the Herfindahl concentration index for the labs in the MSA-class
and the number of active labs. This confirms that the diversity measure is not simply capturing
lab concentration or fragmentation. In Columns (4) and (5), we show that results are similar in
the larger dataset that includes small MSA-classes. The magnitude and statistical significance
of the coefficients are similar with longer or shorter lag structures.13 Finally, in Column (6), we
look at unweighted patent counts. We still find a positive and significant correlation between
diversity and innovation, but the magnitude of the effect drops to less than 9%, suggesting that
diversity has a greater impact on the quality-adjusted measure than on the number of patents.14
Overall, the results in Table 5 document a strong positive association between diversity and
innovation. Results are also robust to introducing MSA-year effects (coefficient = 0.089 and
p-value<0.05).15
The diversity dummy is a natural empirical construct to assess the correlation between
innovation and diversity predicted by our theoretical model. In Appendix 2, we show this
correlation is present in regressions with both direct effects and interaction (Table A.2), con-
firming that an innovation premium is associated with the co-existence of large and small labs
and that the presence of only large or small labs does not drive these results.

5.2 Mechanism
5.2.1 Firm Size Diversity and Spin-Out Formation

The previous analysis illustrates a positive and robust effect of firm size diversity on innovation.
Our theoretical framework indicates that the main channel through which diversity affects

13
We re-estimate the effect of diversity on regional innovation exploring different time lags for the cite-weighted
patent counts. The increase in innovation becomes statistically significant in the year that follows the switch
and its effect appears stable over time. The effect persists during the ten years that follow the switch.
14
In our model, for simplicity, we assume that all innovations have the same quality and differ only in their
fit with the existing research activity. Introducing quality heterogeneity would still generate the prediction that
diversity is associated with an increase both in quality (total number of cites) and in the number of patents.
Whether the effect is stronger on quality-adjusted patents than on the number of patents will depend on the
correlation between the quality and the fit parameters.
15
We obtain similar positive correlations in: (1) panel regressions in which we use the 36 two-digit categories
defined by Hall, Jaffe, and Trajtenberg (2001) as the level of technology disaggregation and (2) regressions in
which the dependent variable does not include cites obtained by large labs.

11
innovation is spin-out formation. This is consistent with previous literature indicating that
spin-out innovation is superior to those of other entrants (Agarwal et al. (2004), Franco and
Filson (2006), Chatterji (2009)).16 To assess the importance of this mechanism, in Table 6 we
investigate the relationship between diversity and spin-out formation.
Because spin-outs form in only 4% of the MSA-class observations, our preferred specifi-
cation is an OLS regression with MSA-class fixed effects, technology class-year effects, and a
dummy variable for instances with no spin-outs. The dependent variable is the logarithm of
one plus the number of spin-outs.17
The coefficient in Column (1) implies a 39% increase in the number of spin-outs when an
MSA-class becomes diverse, controling for the number of inventors in large labs and a dummy
variable in instances of no spin-outs. In Column (2), we introduce additional controls for
the total number of inventors in the MSA-class. In this specification, the estimated diversity
coefficient implies a 30% increase in spin-outs. Columns (3) and (4) confirm the result in
the full sample that includes small MSA-classes. The estimates from these models show that
diversity is associated with a 105% increase in spin-out formation. The difference in elasticities
is due to a substantial difference in the number of spin-outs between the two samples.18
Overall, the results in Table 6 provide direct evidence of a positive correlation between
diversity and spin-out formation that is consistent with our theoretical framework. It is impor-
tant to notice that these regressions exploit a restrictive measure of spin-outs that relies only
on patent data and requires inventors to patent both in large and new small labs. To provide
additional support to the idea that spin-outs are the main mechanism through which diversity
affects innovation, in the next sub-sections we present further indirect evidence consistent with
our theory.

5.2.2 Firm Size Diversity and Non-Compete Agreements

Because in our theoretical framework the main channel through which diversity increases in-
novation is spin-out formation, an additional implication of our theory is that when spin-outs
cannot be formed, the beneficial effect of diversity disappears and a single large lab maximizes

16
Our data confirms this. For example, in 1995, spin-out patents receive 30% more citations than patents by
other new entrants (average across all technology classes and MSAs)
17
The conditional maximum likelihood estimation of the Poisson model drops MSA-classes in which the de-
pendent variable is zero for the entire time period. This is equivalent to dropping about 84% of our data. With
Poisson estimations in this smaller sample, we obtain a positive but statistically insignificant correlation between
diversity and number of spin-outs.
18
Nonetheless, the absolute magnitude of the diversity effect is similar in the two samples. OLS estimates
show that diversity is associated with an increase in the number of spin-outs of 2.02 in the small sample and
with an increase of 6.31 spin-outs in the full sample with overlapping confidence intervals

12
innovation. This result suggests the correlation between innovation and size diversity is likely
to be smaller in settings with substantial impediments to spin-out formation. To explore this
idea empirically, we interact diverse with the extent to which MSAs are located in states where
non-compete agreements are strongly enforced. If high enforcement of non-compete laws pre-
vents spin-out formation, then we expect the organization of the local R&D labor market to
have no impact on innovation.
To construct a non-compete enforcement index, we follow Garmaise (2011) and construct
a measure based on the twelve enforcement dimensions studied by Malsberger (2004). The index
assigns one point to each dimension in which the jurisdiction law exceeds a given threshold; its
value varies from zero to 12.
The original index constructed by Garmaise (2011) covers the period 1992-2005. We
extend the time period from 1975 to 1992 by collecting information on changes in non-compete
laws. The only change we identify for the period 1975-1992 occurs in Michigan in 1985.19 Using
this extended dataset, we construct a dummy variable, high enforce, that equals one when the
index value is greater than or equal to six.20
We present regressions in which the diversity dummy is interacted with the high en-
forcement dummy in Table 7. As expected, we find negative and significant coefficients on the
interaction terms, indicating that the effect of diversity is reduced by the presence of strong
non-compete laws. The results are robust to introducing additional controls and to using the
full sample that includes small MSAs.21
In interpreting these results, it is important to notice that the effect of non-compete
clauses will depend on the relationship between fit and competition. Low fit between an
innovation and the existing lab activity does not imply an absence of competition between the
spin-out and the large lab. Christensen (1997) provides a series of examples of companies not
adopting new technologies because of opportunity costs (e.g., low initial profit margins or low
attractiveness to the company’s best customers) and eventually suffering from competing with
these new technologies. If ideas that do not fit also have a competitive effect on large labs,
then the innovation premium generated by diversity is lower when non-compete agreements
are strongly enforced.

19
Prior to 1985, Michigan outlawed non-compete agreements, but in 1985 it passed legislation that enforced
them. In the Garmaise (2011) data, the score for Michigan is five. We assign an index equal to zero before 1985
and an index of five after 1985.
20
Following Stuart and Sorenson (2003) and Marx et al. (2010), we also generate a state-level indicator variable
with a value of one if the state generally precludes, through statute or precedents, the enforcement of non-compete
covenants. Also, with this measure, the results strongly support the predictions of the model.
21
We cannot estimate the direct effect of highenf orce on innovation because the dummy is constant for
all states over the entire sample period and thus collinear with the MSA-class fixed effects. In non-reported
regressions, we also find that the mediating effect of non-compete enforcement holds in spin-out regressions.

13
5.2.3 Firm Size Diversity and Large Lab Focus

We expect spin-outs to be formed to commercialize ideas that are not a good fit with the main
research activity of a large lab. This implies that diversity has a larger impact on innovation
the greater the number of misfit ideas. Empirically, this suggests an additional test to highlight
the role of spin-outs as an important mechanism linking R&D market structure and innovation:
diversity should have a differential impact on innovation depending on the likelihood that ideas
fit in large lab research trajectories. We expect MSA diversity to have a greater impact on
innovation when large labs follow narrow research trajectories and thus their ideas are more
likely to be misfits. Similarly, when large labs follow broad research strategies, few spin-outs
will be formed and the presence of small labs will be less beneficial to MSA innovation.
To construct a measure of narrowness for large lab research trajectories, for each large
lab we compute the share of patents accounted for by the top four classes (C4 index) based on
its patenting activity across three-digit technology classes (the 426 different USPTO n-classes)
in a four-year window. For MSA-class-year cells with multiple large labs, we construct the
mean value of the index for the labs in the cell. We generate an indicator variable, large lab
focus, that equals one when the MSA-class-year C4 index is above the median relative to other
MSAs within that class-year.
As an example of the variation in the data, consider chemical patenting activity in St.
Louis, MO and San Diego, CA in 1994. Our data indicate the presence of only one large lab
in both MSAs. For the St. Louis lab, the top four three-digit technology classes account for
only 37% of its patenting, and the corresponding MSA-class large lab focus indicator is equal
to zero. For the San Diego lab, the C4 index is approximately 85% and large lab focus is equal
to one for this MSA-class.
In Table 8, we interact this variable with the diversity measure. Column (1) shows
that the effect of diversity is stronger in MSA-classes where large labs follow narrow research
trajectories. Column (2) show that the correlation is robust to the inclusion of additional
controls. Columns (3) and (4) show that results are similar in the larger sample that includes
small MSAs.22

6 Alternative Explanations and Robustness Checks

There are many reasons why co-existence of large and small firms may be associated with an
increase in innovation. We first assess several such reasons in light of their consistency with the
empirical correlations and provide additional results that help us rule out alternative theories

22
In non-reported regressions, we find that results are both quantitatively and qualitatively similar when we
exclude large lab patents from our dependent variable.

14
that might be consistent with a subset, but not the whole set, of the correlations we report. We
then describe a variety of additional extensions and robustness checks for interested readers in
Appendix 2.

6.1 Product Market Competition

Aghion et al. (2005) develop a theoretical model in which an increase in the intensity of compe-
tition may strengthen or reduce the incentives to innovate depending on whether firms compete
neck-and-neck or whether there is a leader and a laggard. In their model, the steady state dis-
tribution of technology gaps generates an inverted-U shape relationship between innovation
and competition. Empirical evidence of this non-linear inverted-U relationship was originally
documented by Scherer (1967) and replicated by Levin et al. (1985) and Aghion et al. (2005).
In light of these studies, concern remains that firm size diversity may simply capture
intermediate levels of product market rivalry. This is because the co-existence of large and small
firms generates a market structure that is likely to be associated with lower competitive rivalry
than environments with only small firms. At the same time, diversity may generate greater
competition than market structures with only a few large firms. Notice that an important
difference between the studies documenting an inverted-U relationship and our setting is that
we conduct our analysis at the MSA-technology class level and not at the industry level. We
expect most of the labs in our sample, especially the largest labs in diverse MSAs, to compete
industry-wide at the national and international levels. Therefore, we do not expect industry-
wide product prices to be substantially affected by the local lab structure. Moreover, the
theoretical framework of Aghion et al. (2005) focuses on a duopoly and has no predictions on
spin-out formation. The inverted-U shape theory is also difficult to reconcile with our findings
that diversity has a greater impact when non-compete agreements are not strongly enforced and
when large labs operate in few technology areas. As a further robustness check, in Column (1) of
Appendix Table A.3, we introduce the square term for the Herfindahl concentration index. We
control for polynomial terms of the Herfindahl index to capture any non-linear effect between
regional market concentration and innovation. If the diversity dummy is simply capturing
intermediate values of concentration, introducing such polynomial terms should substantially
reduce the correlation between diversity and innovation. Instead, we find that results are very
similar when we introduce the square of the concentration index. In non-reported regressions
we obtain similar results with higher-order polynomials.

6.2 Agglomeration Economies

Firm size diversity may be correlated with agglomeration economies that increase innovation
productivity of labs in the MSA-class (Ellison et al. (2010)). In all our regressions, we control

15
for the number of inventors working in the MSA-class, but this linear control can be inadequate
if agglomeration economies arise only for very large MSAs.
Notice that if only agglomeration generates the diversity premium, then it is not clear why
its impact is greater when non-compete agreements are not strongly enforced and when large
labs operate in few technology areas. Nonetheless, to further address this concern, in Column
(2) of Table A.3, we introduce polynomial controls for the number of inventors in the MSA-
class and the total number of inventors in the MSA. If the diversity dummy is only capturing
agglomeration economies at the MSA or MSA-class level, introducing these polynomial terms
should substantially reduce the correlation between diversity and innovation. Instead, we find
that results are very similar when we introduce the square terms.
In unreported regressions, we obtain similar results using higher-degree polynomials.
Results are also robust to controlling more flexibly for the number of MSA inventors by intro-
ducing a dummy for each decile of the distribution, and there is essentially no change in the
diversity coefficient in regressions that exploit the sample of large MSAs.23

6.3 Large Lab Demand for Innovation and Licensees

Another competing explanation is that the role of large labs as consumers of small lab in-
novations drives the correlation between diversity and innovation. In other words, large labs
generate a demand for technologies that induces entry by small firms and increases innovative
output. Alternatively, the correlation may be driven by the role of small labs as licensees of
large lab innovations that do not fit with large lab commercialization strategies.
While we expect these forces to have an impact on MSA innovation (Agrawal and Cock-
burn (2003)), alone these alternative theories cannot explain the entire set of correlations that
we observe in the data. In particular, they are difficult to reconcile with the strong association
between diversity and spin-out formation and with the finding that the effect of diversity is
weaker when non-compete agreements are strongly enforced. Finally, we should expect large
labs to boost innovation demand more strongly when they operate in a variety of technology
areas, which is the opposite of what we find in our data.

6.4 Strategic Patenting

Ziedonis (2004), Lanjouw and Schankerman (2004), and Noel and Schankerman (2006) show
that firms tend to expand their patent portfolios in response to potential hold-up problems

23
We also collect information on population across the different MSAs in 1995 from the US Census. Adding
this control to the cross-section regressions does not affect our estimates (coefficient= 0.396 with p-value<0.05).
Results also show that our baseline panel regression is robust to the inclusion of MSA-year effects that flexibly
capture agglomeration forces varying at the MSA level.

16
generated by “thickets” in the market for technologies. If regional firm size diversity is associ-
ated with patent thickets, then the increase in innovation that we register may be due to the
presence of overlapping and fragmented patent rights and not by the interplay of large and
small firm externalities.
Previous studies, however, have documented the strategic patenting effect at the tech-
nology level and not at the regional level. Because our regressions include technology class time
trends, they control for variation in “thickets” over time. Moreover, the defensive patenting
explanation would suggest an increase in the number of patents but a decline in the quality of
the patents in the presence of regional diversity, which is the opposite of what we find in our
data. Finally, strategic patenting is hard to reconcile with greater spin-out formation (in the
same technology) and with the stronger impact of diversity of innovation when non-compete
agreements are not enforced.

6.5 Innovation Affects Lab Structure

There is the concern that the correlation between diversity and innovation is due to changes in
innovation outcomes rather than driven by changes in lab structure. In other words, “reverse
causality” may take place if potential future innovation generates expansion of large labs and
entry of small labs.
To address this concern, throughout our empirical analysis we use a five-year lag for the
control variables. The IV regressions presented in Table 9 also address this concern by exploiting
exogenous variation in MSA-class diversity. Finally, note that if an innovation shapes the lab
configuration, then we must explain why its impact is greater when non-compete agreements
are not strongly enforced and when large labs operate in few technology areas, neither of which
are obvious.

6.6 Instrumental Variable: Lagged Income Tax Rates

In Table 5, we introduce MSA-class fixed effects to control for time-invariant heterogeneity
affecting both local R&D market structure and innovation. To identify the causal effect of
diversity on innovation, we also need to address the potential bias arising from the correlation
between time variant unobservable heterogeneity and firm size diversity. This correlation can
arise in a variety of ways. A positive shock in the value of the technologies produced in the
MSA-class may lead to an increase both in the entry of small firms and in the likelihood of
innovation. Similarly, capital market shocks (e.g., the dotcom boom in the late 1990s) may
facilitate expansion of large firms and the entry of new firms as well as increase the overall
investment in R&D. Human capital shocks (e.g., an increase in the H1B visa cap) may also
lead to an inflow of scientists in the MSA that may cause a change in both lab structures

17
and innovation. Similarly, local productivity shocks due to availability of information and
communication technologies (Agrawal and Goldfarb (2008); Bloom et al. (2012)) may impact
both the organization of R&D labor and innovation. Finally, successful innovation may lead
empire-building CEOs to engage in takeovers and eliminate small firms from the MSA-class.
To address these endogeneity concerns, we need an instrumental variable that affects
firm size diversity but does not directly affect innovation. We exploit variation in income tax
rates over the sample period as an instrument for diversity. Many papers have documented a
positive relationship between income tax rates and self-employment (among others, see Long
(1982), Poterba (1989), Bruce (2000), Gentry and Hubbard (2004), and Cullen and Gordon
(2007)). A common explanation for this finding is that self-employment offers tax-sheltering
opportunities through the deduction of certain types of expenses (Cullen and Gordon (2007)).
This idea is also discussed in Lerner (2012), who argues that economists have widely accepted
the idea that high income tax rates spur corporate employees to found companies.
If high income taxes induce some scientists to establish their own small labs, then we
should expect income taxes to be positively correlated with diversity. Additionally, as long as
states do not change income taxes in response to innovation shocks, we should expect variation
in income tax rates to be uncorrelated with unobservable heterogeneity affecting innovation.
Nonetheless, a possible problem with this instrument is that high income tax rates may also
affect spin-out formation, thus complicating the interpretation of the estimates. To address
this concern, we control for both contemporaneous personal income and corporate tax rates
exploiting the lagged (three years) personal income tax rates as an instrument. Intuitively, we
expect income tax rates in year t − 3 to be correlated with market structure (diversity) in year t
because of their effect on past entry decisions. At the same time, because we control for current
personal income and corporate taxes, we expect lagged income taxes to be uncorrelated with
the current decision to form a spin-out and with innovation in year t + 5.24
Following Galasso et al. (2013), we obtain information on state and federal income and
capital gain taxes from the NBER Taxsim database described in Feenberg and Coutts (1993).25
Our main tax variable is the combined (state plus federal) tax rate that a representative
household faces in a specific MSA in a given year. Appendix Table A.1 illustrates the variation
of income tax rates averaged across states for five-year time periods. There is a substantial
decline in tax rates in the late 1980s and an increase in the early 1990s. The differences between

24
The IV generates exogenous variation in the presence of small firms that is only one side of regional diversity.
In our data, most of the variation in diversity is driven by variation in the presence of small labs in regions where
large labs are present.
25
This data set contains income tax rates by year and state for an additional $1,000 of income for a represen-
tative household (with $500,000 of wage income split evenly between husband and wife).

18
the lowest and highest also suggest variation across US states.26
We present estimates of the instrumental variable regressions in Table 9. Following
Bertrand et al. (2004), we block bootstrap the standard errors at the state level to account
for serially correlated unobserved state shocks. Column (1) reports the baseline OLS esti-
mates with diversity non-instrumented. The coefficient on the diversity dummy implies a 20%
increase in innovation that is very similar to the effect estimated with the baseline Poisson
model. Column (2) introduces controls for personal income and corporate taxes. The mag-
nitude of the diversity coefficient diminishes slightly to 0.161 but remains highly significant.
Column (3) presents coefficients of the first-stage regression indicating that lagged income taxes
are positively associated with diversity. The estimated impact of taxes on diversity is large: a
10 percentage point increase (e.g., moving from a 40% rate to a 50% rate) in lagged income tax
rates increases the probability of being diverse by approximately 50%. This is not surprising
given that previous literature has documented a substantial impact of taxes on entrepreneur-
ship (e.g., Cullen and Gordon (2007) find that a five percentage point income tax cut leads
to a 40% fall in entrepreneurial risk taking). The regression shows that current income and
corporate taxes are not correlated with diversity. In unreported regressions, we also find no sta-
tistically significant correlation between spin-out formation and lagged income taxes (p-value
=0.34). Together, these results suggest that lagged income taxes may serve as a valid instru-
ment for diversity. Column (4) reports the second-stage estimates with the diversity dummy
instrumented by lagged income taxes. Qualitatively, we find that diversity is still positively
associated with innovation. The coefficient of diversity is more than three times greater than
the one in the OLS estimate, but the standard error is also larger and the two confidence inter-
vals overlap.27 In unreported regressions, we also find that instrumenting diversity with lagged
income taxes has a quantitatively and qualitatively similar impact on spin-out formation.
While most of the previous literature studying the impact of taxes on entrepreneurship
focuses only on income tax rates, it is plausible that the effect of income taxes on entry depends
on non-income taxes that new firms are required to pay. To this end, in unreported regressions
we control for capital gains tax rates. Because reductions in capital gain taxes may be associated
with other pro-business policies that affect innovation, we include this variable on both stages
of the IV regressions. We find that results are robust to the inclusion of capital gains taxes.
The coefficient on diversity is still about three times larger than the one in the OLS estimates,

26
For the period 1978-1998, our data indicate 253 changes in state-level income tax. About half of the changes
(128 events) are tax hikes and with an average increase of 0.82 percentage points. For the tax cuts (125 events),
the average reduction is 1.01 percentage points.
27
We obtain very similar estimates with a five-year lag for income taxes. Similarly, including taxes in both
year t-3 and year t-5 as instruments leads to similar results. Results are qualitatively similar with the logarithm
of the tax variables.

19
suggesting that endogeneity generates a downward bias.28 Results are similar if we introduce
additional controls, such as the number of labs in the MSA-class, lagged corporate tax rates,
gross state product per capita, or the Herfindhal index.

6.6.1 Additional Robustness Checks

We conduct a number of additional empirical exercises to document the robustness of our main
empirical results to alternative specifications and measurement strategies.
First, Appendix Table A.2 provides evidence supporting the correlation between diversity
and innovation in a series of regressions that distinguish between the direct effect of the presence
of a large lab, the presence of numerous small labs, and their interaction. Columns (1) and
(2) show that the presence of a large lab has no significant impact on innovation, whereas
the presence of numerous small labs is associated with a greater number of cites per inventor.
Columns (3) and (4) show that the two direct effects are not significant once we introduce
the interaction between the two variables (the diversity dummy). These results confirm that
the co-existence of at least one large lab and numerous small labs boosts innovation, not the
presence of only one of these two factors.
Concern remains that California MSAs, accounting for a large fraction of our switches
in diversity, drive the results. In Column (3) of Appendix Table A.3, we show that we obtain
similar results using the smaller sample of non-California MSAs.29
We also examine the technology specificity of the small firm externality. Specifically, we
look at whether the diversity effect is driven by the presence of numerous small firms in the
same MSA-class of a large lab compared to the presence of many small labs in the same MSA
of a large lab (irrespective of class). Column (4) of Appendix Table A.3 exploits a diversity
measure that uses a small firm number cutoff at the MSA level rather than the MSA-class
level. The correlation between this alternative diversity measure and innovation is small and
not statistically significant, suggesting that externalities tend to be technology-specific.30
We use patent data to construct the diversity measure, which may be associated with
measurement error. To address this concern, we use an alternative diversity measure based on

28
In the first stage, we notice that capital gains taxes have a negative impact on diversity. This is intuitive
because high capital gain taxes reduce the profits that can be made by selling firm assets and shares and therefore
reduce self-employment incentives. Moreover, high capital gain taxes may reduce the supply of venture capital,
which in turn will reduce entry of new firms.
29
To conduct this exercise, we rescale the diversity measure using the distributions of large lab size and of
small firms in this new sample.
30
We also examine whether MSA-classes with numerous small firms benefit from the existence of large labs
operating in the same technology area but located in different MSAs in the same state. We find that this is not
the case: co-existence of large and small labs is associated with an innovation premium only if large and small
labs are located in the same MSA-class.

20
County Business Patterns Census data that reports detailed information on firm size across US
MSAs. Because these data are available only for a sub-period of our panel, we focus on 1995 and
construct a dummy (diversecbp) that equals one if the MSA-class has at least one establishment
with more than 1,000 employees and a number of establishments with 5-49 employees above the
99th percentile. Column (5) of Table A.3 shows that results are robust to using this alternative
diversity measure.31
In Appendix Table A.4, we show that results are similar if we employ different measures
of diversity. We start by changing the definition of a large lab, moving its cutoff to the 98th
and 99th percentile of the size distribution. We also alter the definition of a small lab, from
exploiting all the labs below the 75th percentile to using labs in the 50th-75th range and the
50th-97th range. Results are similar if we reduce the threshold level of small labs in the diversity
measure by 20% and 30% (from 139 to 111 and 97, respectively).
In unreported regressions, we also alter the definition of the large MSA sample. In the
previous tables, we use the entire sample and MSA-class years where the number of inventors
is above 471 (the minimum required to have diversity). Results are robust to considering
MSA-classes that have more than 471 inventors for the entire sample period or to having an
MSA-class entering the sample when it passes the 471 threshold. Finally, results are robust to
dropping technology class 6 (Miscellaneous).

7 Conclusion
Our results suggest that the way in which R&D labor is organized in a region is associated
with its dynamism and growth, at least insofar as this is captured by rates of patenting. Our
findings point to a potentially important role for market structure in driving the performance
of local innovation economies. Like Aghion et al. (2005) but on a different dimension (R&D
labor market rather than product market), we find that extreme structures are not optimal for
innovation.
The implications for policy makers and managers are subtle. The returns to particular
regional innovation policies (e.g., attracting “anchor tenants” or cultivating entrepreneurial
ventures) are state dependent. In other words, our results suggest that there is no universal
“best” policy, but rather the optimal policy depends on the organizational structure of R&D
labor in a region at a given point in time. In simplistic terms, a region with large firms
but few young entrepreneurial firms may benefit more from policies designed to cultivate new
ventures rather than to attract more large firms, whereas regions without local large firms may

31
We obtain similar results using spin-outs as the dependent variable. There is no change in the diversity
coefficient if we control for the Census measure of the number of employees in the MSA-class.

21
benefit most from attracting those.32 Our results also suggest that managers at large firms
may benefit from taking into account the structure of local R&D manpower because it has
important effects both on the mobility of R&D employees and on the innovative performance
of firms. For example, in regions with many young entrepreneurial ventures, employee retention
may require more attention. Also, firms that are located in regions that are more diverse have
a greater potential for inventions to be commercialized outside the firm. That has both positive
and negative implications for managers. On the positive side, managers have more options for
commercialization but on the negative side it may be more costly to keep inventions inside the
firm.
Following an approach that is common in the empirical literature on the determination
of industries’ innovative activity (Aghion et al. (2005), Bloom et al. (2010), Cohen (2010)), we
examine the interplay between firm size diversity and innovation by performing comparative
statics in the regional organization of R&D labor. Because innovation and regional lab struc-
ture are mutually endogenous, our empirical analysis exploits fixed effects, lagged dependent
variables, and instrumental variables to identify the impact of diversity on innovation. We leave
for future research an analysis of the impact of innovation on R&D labor structure, wages, and
their evolution over time.
Our results also point to additional opportunities and challenges for further research.
While our characterization of market structure is richer than one-dimensional measures such
as the Herfindahl index or concentration ratios, it still represents only a first step. We note
that even on the somewhat easier turf of the product market, decades of empirical research
have yet to establish robust relationships between traditional measures of market structure and
outcomes such as price-cost margins. Furthermore, while we distinguish between incumbents’
and entrants’ contributions to overall innovation activity, this represents only initial progress
towards understanding the impact of local innovation market structure on the intensive versus
extensive margin of innovation by different types of firms.
Finally, our empirical analysis is sparing in its use of explanatory variables and also
relies heavily on only one noisy measure of innovative output, patents. We would prefer to
capture other factors that potentially directly affect innovation performance, such as Glaeser-
type “amenities,” rather than through fixed effects. Linking the types of patent statistics used
here to other sources, such as demographics or data on the production economy, may help with
external validation as well as provide the basis for a richer investigation of local innovation
markets.

32
An additional implication of our results is that policies focused on facilitating spin-out formation (e.g.,
amending non-compete enforcement laws) also may have a beneficial effect on regional innovation. It remains
for future research to study how different policies may interact with the organization of R&D labor organization
and affect regional innovation.

22
References
Agarwal, R., R. Echambadi, A. M. Franco, and M. Sarkar (2004). Knowledge transfer through inheri-
tance: Spin-out generation, development, and survival. The Academy of Management Journal 47 (4),
pp. 501–522.

Aghion, P., N. Bloom, R. Blundell, R. Griffith, and P. Howitt (2005, May). Competition and innovation:
An inverted-u relationship. The Quarterly Journal of Economics 120 (2), 701–728.

Agrawal, A. and I. Cockburn (2003, November). The anchor tenant hypothesis: exploring the role of
large, local, r&d-intensive firms in regional innovation systems. International Journal of Industrial
Organization 21 (9), 1227–1253.

Agrawal, A. and A. Goldfarb (2008). Restructuring research: Communication costs and the democra-
tization of university innovation. American Economic Review 98 (4), 1578–90.

Anton, J. and D. Yao (1995). Start-ups, spin-offs, and internal projects. Journal of Law, Economics
and Organization 11, 362–78.

Arora, A. and A. Gambardella (1994). The changing technology of technological change: General and
abstract knowledge and the division of innovative labour. Research Policy 23 (5), 523–532.

Bertrand, M., E. Duflo, and S. Mullainathan (2004). How much should we trust differences-in-differences
estimates? The Quarterly Journal of Economics 119 (1), 249–275.

Bloom, N., C. Propper, S. Sailer, and J. V. Reenen (2010). The impact of competition on management
quality: Evidence from public hospitals. NBER Working Paper # 16032.

Bloom, N., R. Sadun, and J. V. Reenen (2012). Americans do i.t. better: Us multinationals and the
productivity miracle. American Economic Review, Forthcoming 102 (1), 167–201.

Bruce, D. (2000). Effects of the united states tax system on transitions into self-employment. Labour
Economics 7, 545–574.

Caballero, R. J. and A. B. Jaffe (2002). How high are the giants’ shoulders: An empirical assessment
of knowldge spillovers and creative destruction in a model of economic growth. In A. B. Jaffe and
M. Trajtenberg (Eds.), Patents, Citations, and Innovations: A Window into the Knowledge Economy.
Cambridge: MIT Press.

Cassiman, B. and M. Ueda (2006). Optimal project rejection and new firm start-ups. Management
Science 52, 262–75.

Chatterji, A. K. (2009). Spawned with a silver spoon? entrepreneurial performance and innovation in
the medical device industry. Strategic Management Journal 30 (2), 185–206.

Chinitz, B. (1961). Contrasts in agglomeration: New york and pittsburgh. The American Economic
Review 51 (2), pp. 279–289.

Christensen, C. (1997). Innovator’s Dilemma: When New Technologies Cause Great Firms to Fail.
Harvard Business Press Books.

Cockburn, I. M. and R. M. Henderson (2001, November). Scale and scope in drug development: Unpack-
ing the advantages of size in pharmaceutical research. Journal of Health Economics 20 (6), 1033–1057.

23
Cohen, W. M. (2010). Fifty years of empirical studies of innovative activity and performance. In B. H.
Hall and N. Rosenberg (Eds.), Handbook of The Economics of Innovation, Vol. 1, pp. 129 – 213.
North-Holland.

Cohen, W. M. and S. Klepper (1996, July). A reprise of size and r&d. Economic Journal 106 (437),
925–51.

Cullen, J. B. and R. H. Gordon (2007). Taxes and entrepreneurial risk-taking: Theory and evidence for
the u.s. Journal of Public Economics 91, 1479–1505.

Delgado, M., M. E. Porter, and S. Stern (2010). Clusters and entrepreneurship. Journal of Economic
Geography 10 (4), 495–518.

Ellison, G., E. Glaeser, and W. Kerr (2010). What causes industry agglomeration? evidence from
coagglomeration patterns. American Economic Review 100 (3), 1195–1213.

Feenberg, D. and E. Coutts (1993). An introduction to the taxsim model. Journal of Policy Analysis
and Management 12 (1), 189–194.

Feldman, M. P. and D. B. Audretsch (1999). Innovation in cities: Science-based diversity, specialization

and localized competition. European Economic Review 43 (2), 409 – 429.

Franco, A. and D. Filson (2006). Spin-outs: knowledge diffusion through employee mobility. RAND
Journal of Economics 37 (4), 841–860.

Galasso, A., M. Schankerman, and C. Serrano (2013). Trading and enforcing patent rights. RAND
Journal of Economics 44, 275–312.

Galbraith, J. K. (1952). American capitalism: The concept of countervailing power. New York: Houghton
Mifflin.

Gans, Joshua, D. H. and S. Stern (2002). When does start-up innovation spur the gale of gale of creative
destruction? RAND Journal of Economics 33, 571–586.

Gans, J. and S. Stern (2000). Incumbency and r&d incentives: Licensing the gale of creative destruction.
Journal of Economics and Management Strategy 9, 485–511.

Garmaise, M. J. (2011). Ties that truly bind: Noncompetition agreements, executive compensation,
and firm investment. Journal of Law, Economics, and Organization 27 (2), 376–425.

Gentry, W. and G. Hubbard (2004). Success taxes, entrepreneurial entry and innovation. NBER working
paper #10551.

Geroski, P. A. (1990, July). Innovation, technological opportunity, and market structure. Oxford
Economic Papers 42 (3), 586–602.

Glaeser, E. L., H. D. Kallal, J. A. Scheinkman, and A. Shleifer (1992, December). Growth in cities.
Journal of Political Economy 100 (6), 1126–52.

Glaeser, E. L. and W. R. Kerr (2009, 09). Local industrial conditions and entrepreneurship: How much
of the spatial distribution can we explain? Journal of Economics & Management Strategy 18 (3),
623–663.

Griliches, Z. (1990). Patent statistics as economic indicators: A survey. Journal of Economic Litera-
ture 28 (4), 1661–1707.

24
Grossman, G. M. and E. Helpman (1991, May). Quality ladders and product cycles. The Quarterly
Journal of Economics 106 (2), 557–86.

Hall, B. H., A. B. Jaffe, and M. Trajtenberg (2001, October). The NBER Patent Citation Data File:
Lessons, Insights and Methodological Tools. Working Paper 8498, National Bureau of Economic
Research.

Hausman, J., B. Hall, and Z. Griliches (1984). Econometric Models for Count Data with an Application
to the Patents-R&D Relationship. Econometrica 52 (4), 909–938.

Helsley, R. W. and W. C. Strange (2002). Innovation and input sharing. Journal of Urban Economics 51,
25–45.

Henderson, R. and I. Cockburn (1996, Spring). Scale, scope, and spillovers: The determinants of research
productivity in drug discovery. RAND Journal of Economics 27 (1), 32–59.

Jacobs, J. (1969). The Economy of Cities. Random House: New York.

Jones, B. (2008). The burden of knowledge and the “death of the renaissance man”: Is innovation
getting harder? The Review of Economic Studies 76 (1), 283–317.

Klepper, S. and S. Sleeper (2005). Entry by spinoffs. Management Science 51, 1291–1306.

Klepper, S. and P. Thompson (2010). Disagreement and Intra-industry spinoffs. International Journal
of Industrial Organization 28, 526–538.

Klette, T. J. (1996, Autumn). R&d, scope economies, and plant performance. RAND Journal of
Economics 27 (3), 502–522.

Lanjouw, J. O. and M. Schankerman (2004, April). Protecting intellectual property rights: Are small
firms handicapped? Journal of Law and Economics 47 (1), 45–74.

Lerner, J. (2012). The Architecture of Innovation: The Economics of Creative Organizations. Harvard
Business School Press.

Levin, R., W. Cohen, and D. Mowery (1985). R&d appropriability, opportunity,and market structure:
New evidence on some schumpeterian hypotheses. American Economic Review 75, pp. 20–25.

Long, J. (1982). The income tax and self-employment. National Tax Journal 35, 31–42.

Malsberger, B. M. (2004). Covenants Not to Compete : A State-by-State Survey. Washington, DC:

BNA Books.

Marx, M., J. Singh, and L. Fleming (2010). Does non-compete enforcement create a brain drain? HBS
working paper.

Noel, M. and M. Schankerman (2006). Strategic patenting and software innovation. CEPR Discussion
Paper 5701.

Pakes, A. and Z. Griliches (1980). Patents and R&D at the firm level: A first report. Economics
Letters 5 (4), 377–381.

Poterba, J. M. (1989). Capital gains tax policy toward entrepreneurship. National Tax Journal 42 (3),
375–389.

25
Scherer, F. (1967). Market structure and the employment of scientists and engineers. American Economic
Review 57, pp. 524 –531.

Schumpeter, J. A. (1942). Capitalism, Socialism, and Democracy. New York, Harper.

Stuart, T. E. and O. Sorenson (2003). Liquidity events and the geographic distribution of entrepreneurial
activity. Administrative Science Quarterly 48 (2), pp. 175–201.

Vernon, R. (1960). Metropolis 1985. Harvard University Press Cambridge, MA.

Wooldridge, J. M. (1999). Distribution-free estimation of some nonlinear panel data models. Journal of
Econometrics 90 (1), 77–97.

Ziedonis, R. H. (2004). Don’t fence me in: Fragmented markets for technology and the patent acquisition
strategies of firms. Management Science 50 (6), pp. 804–820.

26
 Figure 1. Variation in Regional Innovation

20  

18   SEATTLE,  WA    

WASHINGTON    DC    
16  
BOISE  CITY,  ID  
PORTLAND,  OR    
14  
SAN  DIEGO,  CA  
Cita%on  Wighted  Patents  per  Inventor  

12   ATLANTA,  GA   DALLAS,  TX      

SACRAMENTO  ,  CA   BOSTON,  MA  
RALEIGH,  NC  
CHICAGO,  IL    
10   DENVER,  CO    
BINGHAMTON,  NY  

8   DETROIT,  MI    

TUCSON,  AZ  
6  
ROCHESTER,  NY  
 

4  

2  

0  
0   500   1000   1500   2000   2500   3000  
Number  of  Inventors    

Note: We do not include the “Silicon Valley” MSA on this scatterplot since doing so would compress
the other observations (7869 inventors, 13.95 cites per inventor).

27
 Table 1. Summary Statistics; N = 41,808
Variables mean median Std. Dev. min max
Weighted Patentsjkt+5 465.72 33 2556.94 0 161861
Patentsjkt+5 27.73 3 118.42 0 6436
Diverse 0.01 0 0.08 0 1
LargeLab 0.11 0 0.32 0 1
# Small Labs 6.43 1 23.37 0 995
# Inventors 91.16 5 411.66 0 23689
# LargeLab Inventors 33.62 0 199.81 0 11725
# Labs 8.08 1 29.76 0 1318
Herf 0.27 0 0.38 0 1
Spin-out 0.08 0 0.26 0 1
# Spin-outs 0.18 0 1.80 0 158
High Enforce 0.27 0 0.44 0 1
No Enforce 0.21 0 0.41 0 1
Large Lab Focus 0.06 0 0.23 0 1

28
 Table 2. Summary Statistics; N = 1,870
Variables mean median Std. Dev. min max
Weighted Patentst+5 6680.68 4166.0 10093.74 418.00 161861
Patentst+5 381.71 268.0 413.65 72.00 6436
Diverse 0.16 0.0 0.36 0.00 1
LargeLab 0.96 1.0 0.20 0.00 1
# Small Labs 83.03 63.0 73.10 2.00 995
# Inventors 1347.22 905.0 1425.11 472.00 23689
# LargeLab Inventors 574.62 366.5 747.89 0.00 11725
# Labs 105.08 80.0 94.01 2.00 1318
Herf 0.15 0.1 0.16 0.01 1
Spin-out 0.57 1.0 0.50 0.00 1
# Spin-outs 2.35 1.0 5.62 0.00 111
High Enforce 0.24 0.0 0.43 0.00 1
No Enforce 0.30 0.0 0.46 0.00 1
Large Lab Focus 0.45 0.0 0.50 0.00 1

Table 3. Variation in Diversity (Year of the Switch)

MSA Chemical Computers Drugs Electronics Mechanical Other
Boston 1993 1995 1994 1992 1996 1996
Chicago 1996 2000 1979 1978
Dallas 1999
Detroit 1994
LA 1998 1995 1995 1993 1978 1980
NYC 1976 1995 1989 1987 1978 1977
Philadelphia 1996 1998
San Diego 2000 1998
San Francisco 1990 1989 1992 1987 1995 1994
Seattle 2000
DC 1998

29
 Table 4. Diversity and Innovation (Cross-Section)
(1) (2) (3) (4)
Sample: Full, 1995 Full, 1995 Full, 1995 Full, 1995
Dependent Variable: Weighted Weighted Weighted
Patentsjkt=2000 Patentsjkt=2000 Patentsjkt=2000 Patentsjkt=2000
∗∗ ∗ ∗
diversejkt=1995 3.762 0.406 0.382 0.296∗∗
(0.333) (0.185) (0.165) (0.106)
logInventorsjkt=1995 0.937∗∗ 0.891∗∗ 0.861∗∗
(0.050) (0.027) (0.017)
Constant 6.005∗∗ 1.841∗∗ 1.348∗∗ −0.328∗∗
(0.140) (0.284) (0.140) (0.109)
Class FE X X
Observations 1608 1608 1608 1608
log likelihood −1260012.43 −303371.15 −168774.27 −12157.30
Notes: Observations are at the MSAj -classk level. All specifications are Poisson regressions estimated by maximum
likelihood. Weighted patents is the forward citation weighted sum of distinct patents with primary technology
classification k and application year t + 5 where at least one inventor is located in MSA j. The main independent
variable, diverse, equals 1 if MSA j in class k in 1995 has at least one large lab and more than 139 active small labs, and
0 otherwise. Loginventors is the log of the number of distinct active inventors in MSA j, class k in 1995.
Robust standard errors clustered at the MSA-class level in parentheses.
+ p < 0.10, ∗ p < 0.05, ∗∗ p < 0.01

30
 Table 5. Diversity and Innovation
(1) (2) (3) (4) (5) (6)
Sample: Full At-Risk At-Risk Full Full Full
Dependent Variable: Weighted Weighted Weighted Weighted Weighted
Patentsjkt+5 Patentsjkt+5 Patentsjkt+5 Patentsjkt+5 Patentsjkt+5 Patentsjkt+5
∗∗ ∗∗ ∗∗ ∗ ∗
diversejkt 0.309 0.161 0.137 0.146 0.127 0.085∗
(0.116) (0.053) (0.050) (0.063) (0.058) (0.043)
logInventorsjkt 0.911∗∗ 0.644∗∗ 0.518∗∗ 0.293∗∗ 0.200∗∗ 0.334∗∗
(0.014) (0.077) (0.097) (0.034) (0.054) (0.040)
logNumLabsjkt 0.187 0.205∗∗ 0.061
(0.138) (0.071) (0.052)
herfjkt −0.763 −0.074 −0.108∗∗
(0.563) (0.052) (0.035)
Year X Class FE X X X X X X
MSA X Class FE X X X X X
Observations 41808 1864 1864 41262 41262 41444

31
Num. Groups 140.00 140.00 1587.00 1587.00 1594.00
log likelihood −3781751.18 −143844.93 −140280.30 −1345383.49 −1332410.79 −120051.25
Notes: Observations are at the MSAj -classk -yeart level. All specifications are estimated with Poisson with Fixed Effects Quasi maximum likelihood (QML).
Weighted patents is the forward citation weighted sum of distinct patents with primary technology classification k and application year t + 5 where at least one
inventor is located in MSA j. The main independent variable, diverse, equals 1 if MSA j in class k in year t has at least one large lab and more than 139 active small
labs, and 0 otherwise. Loginventors is the log of the number of distinct active inventors in MSA j, class k in year t, logNumLabs is the log of the number of distinct
active labs in MSA j, class k in year t, and herf is the Herfindahl concentration index for labs in MSA j, class k in year t. Column 3 and 4 only includes
MSA-class-years that have enough inventors to be at-risk of displaying a positive diversity measure.
Robust standard errors clustered at the MSA-class level in parentheses.
+ p < 0.10, ∗ p < 0.05, ∗∗ p < 0.01
 Table 6. Diversity and Spin-Out Formation
(1) (2) (3) (4)
Sample: At-Risk At-Risk Full Full
Estimation: OLS OLS OLS OLS
Dependent Variable: ln Spin-outsjkt ln Spin-outsjkt ln Spin-outsjkt ln Spin-outsjkt
∗∗ ∗∗ ∗∗
diversejkt 0.329 0.266 0.718 0.717∗∗
(0.071) (0.069) (0.104) (0.104)
logInventorsLargejkt 0.017 −0.034∗ 0.010∗∗ 0.009∗∗
(0.023) (0.016) (0.002) (0.002)
No Spin-Out Dummy −0.963∗∗ −0.952∗∗ −1.026∗∗ −1.026∗∗
(0.028) (0.028) (0.015) (0.015)
logInventorsjkt 0.507∗∗ 0.004∗∗
(0.141) (0.001)
Year X Class FE X X X X
MSA X Class FE X X X X
Observations 1873 1873 41808 41808
Num. Groups 149.00 149.00 1608.00 1608.00
R2 0.74 0.75 0.74 0.74
Notes: The dependent variable, ln Spin-outsjkt is the logarithm of one plus the count of spin-outs in MSA j, class k in
year t.
Robust standard errors clustered at the MSA-class level in parentheses.
+ p < 0.10, ∗ p < 0.05, ∗∗ p < 0.01

32
 Table 7. Innovation and Non-compete Clauses
(1) (2) (3) (4)
Sample: At-Risk At-Risk Full Full
Dependent Variable: Weighted Weighted Weighted Weighted
Patentsjkt+5 Patentsjkt+5 Patentsjkt+5 Patentsjkt+5
diversejkt 0.189∗∗ 0.168∗∗ 0.175∗ 0.157∗
(0.064) (0.061) (0.072) (0.066)
diversejkt X high enforcej −0.152+ −0.174∗ −0.194∗ −0.206∗∗
(0.091) (0.082) (0.084) (0.079)
logInventorsjkt 0.630∗∗ 0.475∗∗ 0.291∗∗ 0.197∗∗
(0.078) (0.101) (0.034) (0.054)
logNumLabsjkt 0.227 0.209∗∗
(0.139) (0.070)
herfjkt −0.730 −0.073
(0.546) (0.052)
Year X Class FE X X X X
MSA X Class FE X X X X
Observations 1864 1864 41262 41262
Num. Groups 140.00 140.00 1587.00 1587.00
log likelihood −142174.66 −138097.89 −1342338.58 −1328971.60
Notes: All specifications are estimated by Poisson with Fixed Effects QML. The variable high enforce is equal to 1 if
MSA j is in a high-enforcement state in year t. A state is high-enforcement if it has a value of 5 or greater on the
13-point scale (0-12) based on Garmaise (2011). This variable is time-invariant during our sample period, and so its
effect is identified through the interaction with the time-varying variable diverse.
Robust standard errors clustered at the MSA-class level in parentheses.
+ p < 0.10, ∗ p < 0.05, ∗∗ p < 0.01

33
 Table 8. Innovation, Large Lab Focus and Diversity
(1) (2) (3) (4)
Sample: At-Risk At-Risk Full Full
Dependent Variable: Weighted Weighted Weighted Weighted
Patentsjkt+5 Patentsjkt+5 Patentsjkt+5 Patentsjkt+5
diversejkt 0.118∗ 0.105+ 0.077 0.071
(0.060) (0.055) (0.062) (0.059)
large lab focusjkt −0.028 −0.020 −0.007 0.005
(0.030) (0.029) (0.031) (0.029)
diversejkt X large lab focusjkt 0.096∗ 0.077∗ 0.146∗∗ 0.121∗
(0.041) (0.039) (0.055) (0.050)
logInventorsjkt 0.613∗∗ 0.507∗∗ 0.284∗∗ 0.190∗∗
(0.070) (0.095) (0.032) (0.051)
largelabjkt 0.052 0.047 0.019 0.029
(0.069) (0.066) (0.037) (0.035)
logNumLabsjkt 0.164 0.203∗∗
(0.133) (0.066)
herfjkt −0.733 −0.070
(0.579) (0.051)
Year X Class FE X X X X
MSA X Class FE X X X X
Observations 1864 1864 41262 41262
Num. Groups 140.00 140.00 1587.00 1587.00
log likelihood −142142.56 −139157.24 −1339608.32 −1327459.97
Notes: All specifications are estimated by Poisson with Fixed Effects QML. The variable large lab focus is equal to 1 if
the C4 concentration ratio of the patenting activity across three-digit technology classes of the large labs in MSA j, class
k in year t is above the median value. This measure captures MSAs whose large labs patent in more focused (narrow)
technology areas.
Robust standard errors clustered at the MSA-class in parentheses.
+ p < 0.10, ∗ p < 0.05, ∗∗ p < 0.01

34
 Table 9. Innovation and Lagged Income Tax IV
(1) (2) (3) (4)
Sample: At-Risk At-Risk At-Risk At-Risk
Estimation: OLS OLS First Stage 2SLS
Dependent Variable: ln Weighted ln Weighted ln Weighted
Patentsjkt+5 Patentsjkt+5 diversejkt Patentsjkt+5
diversejkt 0.182∗∗ 0.161∗∗ 0.700∗
(0.049) (0.042) (0.346)
logInventorsjkt 0.505∗∗ 0.353∗∗ 0.216+ 0.230∗
(0.109) (0.124) (0.111) (0.116)
IncomeTaxjt 0.056∗∗ 0.014 0.034∗∗
(0.016) (0.021) (0.012)
CorpTaxjt −0.024+ −0.000 −0.023
(0.013) (0.024) (0.017)
IncomeTaxjt−3 0.057∗
(0.024)
Year X Class FE X X X X
MSA X Class FE X X X X
Observations 1835 1606 1606 1597
Num. Groups 149.00 149.00 149.00 140.00
First Stage F-statistic 17.52
R2 0.94 0.94 0.41 0.93
Notes: Column 1 replicates the specification estimated in Table 3, Column 4 with OLS. This specification serves as a
baseline for results when estimated with linear least squares. Column 2 introduces two controls. IncomeTaxjt is the
personal income tax in year t in MSA j’s state and CorpTaxjt is the corporate income tax in year t in MSA j’s state.
Column 3 presents the first stage regression with diverse as the dependent variable. IncomeTaxjt−3 is the personal
income tax in year t − 3 in MSA j’s state. Column 4 presents 2SLS results with diverse instrumented by IncomeTaxt−3 .
Block (state-level) bootstrapped standard errors in parentheses.
+ p < 0.10, ∗ p < 0.05, ∗∗ p < 0.01

35
Appendix 1: Theoretical Model
Consider an MSA with T scientists divided into J large labs and N small labs. A small lab
employs only one scientist and cannot commercialize multiple innovations. Large labs employ
at least two scientists and have the capability to commercialize more than one innovation.
Moreover, each large lab j has an existing research activity that generates profits λj . We indicate
as Sj > 1 the number of scientists working in large lab j and denote with SL = Ji=1 Sj the
P

total number of scientists working in large labs. Notice that T = SL + N .

We assume there are economies of scale in innovation: each small lab discovers an inno-
vation with probability 1/∆, and each scientist in a large lab of size Sj discovers an innovation
with probability Sj /∆ with ∆ ≥ T . 33

If commercialized in a small lab, an innovation generates profits equal to π with 1 ≥

π > 0. For a large lab, the profits from the commercialization of an innovation discovered
internally are equal to π + a. The parameter a ∈ R captures the “fit” of the innovation with
the existing research activity of the lab (Cassiman and Ueda (2006)). A positive value for a
may arise because of scale or scope economies. Conversely, a negative value for a indicates
the existence of coordination costs between the innovation and the existing research activity.
Alternatively, a < 0 may indicate the opportunity cost of deploying resources from the existing
research activity. Notice that an innovation is profitable for a large lab j if the profits from
commercialization (π+a+λj ) exceed those without commercialization (λj ), i.e., when π+a > 0.
A scientist working in a large lab can implement the discovered innovation outside the
lab by opening a new small lab (spin-out). The cost of opening a new lab is equal to k(N )
with k(0) > 1, k 0 (N ) < 0, and limN →∞ k(N ) = 0. The entry cost is decreasing in N because
the presence of other small labs generates a positive externality.34
The timing of the game is as follows. In the first period, innovations are discovered.
In the second period, large lab scientists negotiate with their labs about the destiny of their
innovations. The possible strategies are: (i ) internal commercialization (with profits π+a+λj ),
(ii ) spin-out formation (with profits π − k(N )), and (iii ) non-commercialization (with profits
λj ). In the third period, spin-outs are formed and ideas are commercialized.

33
This is equivalent to assuming an innovation production function that is quadratic in lab size, i.e. the
expected number of innovations in a lab of size Sj is Sj2 /∆. We impose this functional form for tractability but
results hold under milder assumptions on the magnitude of the scale economies. In particular, diversity strictly
maximizes innovation if large lab scientists discover an innovation with probability xSj /∆ + (1 − x)/∆ for any
x > 0. Moreover, diversity weakly maximizes innovation even if there are no scale economies (i.e. when x = 0).
34
A possible micro-foundation of this cost can be provided building on the model of Helsley and Strange
(2002) where a large number of small labs generates a dense network of input suppliers that facilitates spin-out
formation. An alternative interpretation for k(N ) is that in the presence of small firms, the MSA develops a
culture of entrepreneurship that induces employees of large labs to leave their employer and start their own firms
(Glaeser and Kerr (2009)).

36
 Commercializing an innovation (internally or through a spin-out) may require the de-
velopment of subsequent technologies necessary to extract profits from the idea. If this oc-
curs, commercialized innovations generate new research trajectories. Conversely, the non-
commercialization strategy produces orphan ideas that do not generate follow-on research.35
To solve the bargaining problem at period 2, we assume that the lab makes a take-it-or-
leave-it offer to the scientist.

Proposition 1 Established small labs commercialize internal innovations with π ≥ 0.

Large labs commercialize internal innovations with π + a ≥ 0 and a ≥ −k(N ). A spin-out is
formed if π ≥ k(N ) and a ≤ −k(N ).
Proof In a small lab, it is efficient to commercialize an innovation if π ≥ 0. To avoid
spin-out formation, large labs need to offer the scientist π − k(N ). This implies that spin-outs
are formed if π + a − (π − k(N )) ≤ 0 or a ≤ −k(N ). 

In equilibrium, when the cost of establishing a spin-out is large, scientists do not find
it profitable to form a spin-out and thus the large lab commercializes the innovations if the
fit with the established business is high enough. When the cost of forming a spin-out is low
(k ≤ π), then inventors form spin-outs to commercialize innovations that are not a good fit
with the existing research activity.
We assume that a is distributed with a continuous and differentiable cumulative distri-
bution F (a). Under this assumption, a spin-out is formed with a positive probability as long
as π ≥ k(N ) that is satisfied if N ≥ N where N is defined as π − k(N ) = 0. Therefore, an
innovation of a large lab scientist will generate a spin-out with probability:

 F (−k(N ))
 if N ≥ N
Pr(spin-out) = .

0 if N <N


Similarly, an innovation discovered in a large lab is commercialized (either internally or

through spin-outs) with probability:


 1 if N ≥ N
Pr(Comm. from Large)=

1 − F (−π) if N <N


whereas each scientist in a small lab discovers an innovation with probability 1/∆. Because in

35
Empirically, this implies that greater commercialization may not be associated with greater patenting (be-
cause orphan ideas may also be patented) but is likely to be associated with greater citation-weighted patent
counts (because orphan ideas do not generate new research trajectories).

37
large lab j each scientist discovers an innovation with probability Sj /∆, the expected number
of innovations in lab j is Sj2 /∆. Exploiting these results, we obtain the following proposition.

Proposition 2 The expected number of spin-outs, N S, is maximized in the presence of

one large lab and N ≥ N . If T is large enough, then the expected number of commercialized
innovations, NI, is maximized in the presence of one large lab and N small labs.
Proof We denote with N S the expected number of spin-outs in the MSA.

J
X Sj2
NS = Pr(spin-out)
∆
j=1
 2
SL S
 ∆ H F (−k(N )) if N ≥ N


= . (2)



0 if N < N

J
where H S = (Sj /SL )2 is the Herfindahl concentration index of large labs. Differentiation of
P
j=1
(2) shows that N S is maximized when H S = 1, which implies that at most one large lab will
be present. Because there are no spin-outs if N < N , the number of small labs that maximizes
spin-out formation is the largest between N and N ∗ where we define N ∗ as

N ∗ = arg max ϕ(N )

N

(T − N )2
where ϕ(N ) ≡ F (−k(N )). We now consider the total innovation activity in the
∆
MSA. The total number of commercialized innovations is:
J
X Sj2 N
NI = Pr(Comm. from Large)+
∆ ∆
j=1

SL2 S N
if N ≥ N


 H +

 ∆ ∆
= . (3)
SL2

 N
H S (1 − F (−π)) +


 if N <N
∆ ∆

N I increases in H S so it is maximized when H S = 1. Consider now the case in which

N < N . Evaluated at H S = 1, NI has the following functional form:

(T − N )2 N
g(N ) = (1 − F (−π)) +
∆ ∆

38
that is decreasing in N for T large enough. When N ≥ N and H S = 1, NI can be written as:

(T − N )2 N
z(N ) = +
∆ ∆

that is also decreasing in N . It is easy to see that z(N ) > g (N ), so innovation is maximized
with diversity if z(N ) > g(0), which we can rewrite as:

(T − N )2 + N > T 2 (1 − F (−π))

or
T [T F (−π) − 2N ] > −N 2 − N

that is satisfied for T large enough. 

Proposition 2 indicates that the presence of large and small labs are complementary
forces affecting both innovation and spin-out formation. Intuitively, a greater concentration
among large labs increases the number of ideas generated and the expected number of spin-outs.
A larger pool of existing small labs facilitates the formation of spin-outs and thus allow ideas
that are not a “good fit” for a large lab to be commercialized. There is a positive interaction
between the two effects. A large number of small labs has no effect on spin-out formation in
the absence of large labs. At the same time, an increase in large lab concentration generates a
greater number of spin-outs in the presence of small lab externalities.36
Let us label an MSA as diverse if there are N ≥ N small labs and only one large lab. By
showing that spin-out formation and the number of commercialized innovations are maximized
with diverse MSA configurations, Proposition 2 implies that any non-diverse MSA configuration
is dominated by at least one diverse configuration in terms of spin-out formation or innovation.

Comparison of Diverse and Non-Diverse MSAs

Take an arbitrary “diverse” MSA with N ≥ N small labs and one large lab of size (T − N ). We
first show that this configuration generates more commercialized ideas than any non-diverse
MSA with N < N . Notice that when N < N , we maximize the amount of commercialized
ideas by allocating all scientists to a single large lab: T 2 (1 − F (−π))/∆. This amount of

36
For simplicity, we do not consider: (i) cannibalization and complementarity effects between the innovations
developed by the spin-outs and those commercialized by the large lab and (ii) imperfect IP protection. In a
previous draft of the paper, we showed it possible to obtain similar results in a model with cannibalization,
complementarity, and imperfect IP protection as in Gans and Stern (2000) and in a model where spin-outs are
formed because of disagreements between large lab scientists as in Klepper and Thompson (2010). Intuitively,
as long as the IP regime allows spin-outs to be profitable, the presence of local externalities encourages spin-out
formation (formal proofs available upon request).

39
commercialized ideas is lower than the one in the diverse MSA if:

(T − N )2 + N > T 2 (1 − F (−π))

that is satisfied for T large enough. Consider now a non-diverse MSA with N ≥ N and
multiple large labs, each of a size not exceeding T − N . Among all these MSAs, we maximize
the amount of commercialized innovation by allocating (T − N − 1) to one large lab, (N − N )
scientists to a second large lab, and N scientists to small labs. The amount of commercialized
ideas of this MSA is lower than the one in the diverse MSA if:

(T − N )2 + N > (T − N − 1)2 + (N − N )2 + N
2(T − N ) > (N − N )2 − (N − N ) + 1

that is satisfied when T is large enough. This shows that the diverse MSA dominates all the
non-diverse MSAs with N ≥ N and multiple large labs, each of a size not exceeding T − N .
We now consider spin-out formation. Because no spin-out takes place if N < N , the
diverse MSA generates more spin-outs than all the non-diverse MSA with N < N . Consider
now a non-diverse MSA with Ne ∈ [N, N ] small labs and multiple large labs, each of a size not
exceeding T − N . The MSA generates the maximum
  amount of spin-outs when we allocate
(T − N − 1) scientists to one large lab and N − N
e scientists to a second large lab. This
configuration will generate less spin-outs than a diverse MSA if
h i
(T − N )2 F (−k(N )) ≥ (T − N − 1)2 + (N − N
e )2 F (−k(N
e)

that is satisfied for T large enough because F (−k(N )) ≥ F (−k(N

e ). This shows that the diverse
MSA dominates all the non-diverse MSAs with N e ∈ [N, N ] small labs.

Relaxing the Definition of Diversity

Take an arbitrary “diverse” MSA with N ≥ N small labs and multiple large labs of size Sj with
j = 1, ..., J. We first show that this configuration generates more commercialized ideas than
any non-diverse MSA with N < N . Notice that when N < N , we obtain the maximum amount
of commercialized ideas by allocating all scientists to a single large lab: T 2 (1 − F (−π))/∆. The
amount of commercialized ideas is lower than the one in the diverse MSA if:
X
Sj2 + N > T 2 (1 − F (−π)). (4)
j

40
Call k the largest lab and notice that Sk 2 + N < Sj2 + N. Then a sufficient condition to
P
j
have (4) satisfied is:
Sk 2 + N > T 2 (1 − F (−π))
p
that is satisfied if Sk > S ≡ T 2 (1 − F (−π)) − N .

41
Appendix 2: Additional Tables

Table A.1. Income Tax Rates

Period Mean Std. Dev Min. Max.
1977-1979 52.80 2.26 50 59.9
1980-1984 56.68 1.95 50 56.9
1985-1989 42.02 9.67 28 56.25
1990-1994 38.27 5.46 28 48.15
1995-1999 44.03 1.84 40.79 46.89
2000-2005 40.99 2.63 36.05 46.28

42
 Table A.2. Innovation and Diversity - Robustness with Levels
(1) (2) (3) (4)
Sample: Full Full Full Full
Estimation: QML QML QML QML
Dependent Variable: Weighted Weighted Weighted Weighted
Patentsjkt+5 Patentsjkt+5 Patentsjkt+5 Patentsjkt+5
largelabjkt −0.002 0.012 0.031
(0.039) (0.035) (0.032)
Smallthreshjkt 0.146∗ −0.080 −0.064
(0.064) (0.068) (0.067)
largelabjkt × Smallthreshjkt † 0.227∗∗ 0.193∗
(0.083) (0.077)
logInventorsjkt 0.288∗∗ 0.292∗∗ 0.290∗∗ 0.191∗∗
(0.034) (0.034) (0.033) (0.051)
logNumLabsjkt 0.211∗∗
(0.070)
herfjkt −0.063
(0.051)
Year X Class FE X X X X
MSA X Class FE X X X X
Observations 41262 41262 41262 41262
Num. Groups 1587 1587 1587 1587
log likelihood −1356561.51 −1346174.30 −1346108.58 −1333246.93
Notes: † This interaction is the same as our diversity measures, diversejkt .
Robust Standard errors clustered at MSA-class in parentheses.
+ p < 0.10, ∗ p < 0.05, ∗∗ p < 0.01

43
 Table A.3. Additional Robustness Checks I
(1) (2) (3) (4) (5)
Sample: Full Full No California‡ Full Full, 1995
Dependent Variable: Weighted Weighted Weighted Weighted Weighted
Patentsjkt+5 Patentsjkt+5 Patentsjkt+5 Patentsjkt+5 Patentsjkt=2000
diversejkt 0.123∗ 0.119∗ 0.095∗∗
(0.057) (0.052) (0.035)
diversemsajt 0.031
(0.042)
diversecbpjkt=1995 0.279∗∗
(0.099)
logInventorsjkt 0.225∗∗ 0.192∗∗ 0.250∗∗ 0.286∗∗ 0.916∗∗
(0.054) (0.058) (0.039) (0.034) (0.043)
logNumLabsjkt 0.177∗
(0.071)
herfjkt −0.496∗∗

44
(0.18)
herf2jkt 0.407∗
(0.162)
logInventors2jkt −0.010+
(0.006)
logMSAInventorsjt 0.182∗∗
(0.052)
logMSAInventors2jt 0.018∗∗
(0.006)
Year X Class FE X X X X
MSA X Class FE X X X X
Class FE X
Observations 41262 41262 38792 41262 1608
Num. Groups 1587.00 1587.00 1492.00 1587.00
log likelihood −1331011.70 −1290636.07 −1193672.34 −1356126.58 −172699.85
Notes: diversecbpt corresponds to a diversity measure constructed with data from the Census County Business Patterns dataset (1995). logMSAInventorst
corresponds to the log of inventors in MSA j across all classes k.
‡ We construct diverse in this specification using the rescaled distribution of firms outside California.

Robust standard errors clustered at the MSA-class in parentheses.

+ p < 0.10, ∗ p < 0.05, ∗∗ p < 0.01
 Table A.4. Additional Robustness Checks II
(1) (2) (3) (4) (5) (6)
Sample: Full Full Full Full Full Full
Diverse Definition: LargeLabs LargeLabs Small Labs Small Labs # Small Labs # Small Labs
> 98th p > 99th p 50th -75th p 50th -97th p > 111 > 97
Dependent Variable: Weighted Weighted Weighted Weighted Weighted Weighted
Patentsjkt+5 Patentsjkt+5 Patentsjkt+5 Patentsjkt+5 Patentsjkt+5 Patentsjkt+5
diversejkt 0.147∗ 0.134∗ 0.140∗ 0.135∗ 0.116+ 0.110+
(0.064) (0.068) (0.058) (0.061) (0.062) (0.058)
logInventorsjkt 0.292∗∗ 0.292∗∗ 0.291∗∗ 0.291∗∗ 0.291∗∗ 0.290∗∗
(0.034) (0.034) (0.034) (0.034) (0.034) (0.034)
Year X Class FE X X X X X X
MSA X Class FE X X X X X X
Observations 41262 41262 41262 41262 41262 41262
Num. Groups 1587.00 1587.00 1587.00 1587.00 1587.00 1587.00
log likelihood −1346082.01 −1348483.79 −1346467.99 −1347062.73 −1348758.11 −1349537.11
Notes: Robust Standard errors clustered at MSA-class in parentheses.
+ p < 0.10, ∗ p < 0.05, ∗∗ p < 0.01

45