Content uploaded by Justin Tumlinson
Author content
All content in this area was uploaded by Justin Tumlinson
Content may be subject to copyright.
1
Does Social Proximity Enhance Business Partnerships?
Theory and Evidence from Ethnicity’s Role in US Venture Capital*,†
Deepak Hegde
New York University – Stern School of Business
Justin Tumlinson
Ifo Institute at the University of Munich
Abstract
We develop a formal model to understand the selection and influence effects of social proximity between
business partners. Consistent with the model’s predictions, we find that US venture capitalists (VCs) are
more likely to select startups with coethnic executives for investment, particularly when the probability of
the startups’ success appears low. Ethnic proximity between VCs and the startups they invest in is
positively related to performance, measured by the probability of the companies’ successful exit through
acquisitions and initial public offerings (IPOs), and net income after IPO. Twostage regression estimates
suggest that the positive effect of coethnicity is largely due to influence, that is, superior communication
and coordination between coethnic VCs and startup executives after the investment. To the extent that
VCs expect to work better with coethnic startups, they invest in coethnic ventures that are of lower
observable quality than noncoethnic ventures.
JEL classification: G24, Z1
Keywords: Venture Capital; Entrepreneurship; Social Capital; Social networks; Homophily
* We thank seminar participants at New York University, University of California at Berkeley, Duke
University, Rutgers University, University of Munich, National Bureau of Economics Research 2011 Winter
Meetings, Indian Institute of Management (Ahmedabad and Bangalore), and SRI/ASQ’s 2011 Professional
Development Workshop for helpful suggestions. We thank Jörg Claussen for his assembly of the LinkedIn data
set. This research was supported by grants from the Kauffman Foundation (#20110112) and the Institute of
Business and Economic Research at the University of California, Berkeley.
† Address correspondence to dhegde@stern.nyu.edu and tumlinson@ifo.de
2
1. Introduction
In 2004, Vinod Khosla, Indian billionaire and cofounder of Sun Microsystems, started Khosla Ventures.
By 2011, the Silicon Valleybased venture capital firm’s portfolio included US companies founded or
cofounded by: Ramesh Chandra (MokaFive), Srini Devadas (Verayo), Yogi Goswami (Sunborne),
Sandeep Gulati (Zyomed), Siraj Khaliq (WeatherBill), Ramu Krishnan (Ramu Inc.), Ashok
Krishnamurthi (Xsigo), Hosain Rahman (Aliph), Anil Rao (Sea Micro), Mulpuri Rao (Soladigm), Bindu
Reddy (MyLikes), Mohit Singh (Seeo), and Adya Tripathi (Tula). If we added CEOs’ and Directors’
names, the list of executives of Indian origin in Khosla’s portfolio of companies would grow longer still.
Khosla Ventures does not advertise a preference for investing in companies started by ethnic Indians, but
casual observation suggests that it has one. How does social proximity affect the choice of business
partners and their performance?
This study models the interaction of two conceptually distinct mechanisms that shape the performance of
socially proximate business partnerships: selection and influence. Individuals may have better access to,
and superior information about, opportunities within their social networks—social proximity may thus
facilitate business partner selection. After forming a partnership, shared norms and discourse may
improve coordination and monitoring among socially close individuals—hence, proximity may positively
influence the partnership after formation. We formalize these mechanisms and generate testable
propositions about the circumstances under which socially proximate agents are likely to partner and
succeed. That social proximity should produce superior economic outcomes is neither certain nor
obvious. Preferences for interactions with socially close individuals may cause agents to discriminate
against superior opportunities outside their social networks and laxly monitor partners within their social
networks; thus, tastebased selection and influence could undermine the economic success of socially
proximate relationships.
We test our model’s predictions over the social proximity induced by shared ethnicity in the context of
the business partnerships formed between VC partners (VC “partners” are principals who make, and
monitor, investments) and startup executives using a sample of almost all US venturebacked deals
between 1991 and 2010. We assemble the names of 22,000 USbased VC partners and 85,000 USbased
startup executives from the rosters of 2,687 VCs and 11,235 startups they funded and classify each
partner and executive, based on their family name (surname) and given name, as belonging to one of ten
distinct ethnic groups. Then, for each investment, we compute a binary measure of coethnicity between
the investing VC and funded startup indicating whether the VC and the company have toplevel personnel
of the same ethnicity. We also construct control variables for investment, VC and company
3
characteristics, including investment amount, geographic clustering, as well as VC and company industry
specialization that may correlate with ethnic proximity and investment payoffs.
We find that VCs are more likely to invest in a startup when the VC and company have toplevel
personnel of the same ethnicity. Coethnicity’s predictive power is highest for earlystage investments,
which have lower ex ante success rates than more mature companies. Ordinary Least Squares (OLS)
estimates suggest that coethnicity both increases the probability of successful exit through IPO or
acquisition and augments startups’ postIPO performance. We purge unobserved heterogeneity among
VCs in VCfixed effects regressions and find that even within a given VC’s portfolio, ethnically closer
startups perform better. The estimated positive effects of coethnicity are driven by executives belonging
to the less common but more distinct ethnic communities in the US (i.e., individuals not of AngloCeltic
or West European origin).
These findings are based on correlations obtained after controlling for the observable characteristics of
VCs and companies but do not distinguish between the effects of ethnicitybased selection of highquality
investments and coethnicity’s influence on performance through enhanced coordination between investors
and entrepreneurs. We try to isolate the influence effects (“treatment effect” in econometric parlance) of
coethnicity by employing three separate strategies: (a) an instrumental variables (IV) approach which
accounts for omitted variables, such as unobserved VC and company quality, that affect performance
through selection; (b) a method developed by Ackerberg and Botticini (2002), also based on IVs, that
isolates the effect of exogenous market characteristics unrelated to the influence effects of coethnicity on
performance; and (c) a twostage Heckman model that corrects for a broader set of factors that affect
selection (including unobserved quality) while predicting performance. These three approaches all yield
estimates of coethnic influence substantially larger than OLS estimates and suggest that coethnicity
improves performance through strong postinvestment influence.
Our finding that ethnic proximity facilitates VCcompany matching particularly during early funding
rounds, when the probability of the startups’ success is low, taken together with our twostage estimates,
imply that VCs select coethnic companies (over noncoethnic ones) even when they appear to be of lower
observable quality. This seemingly counterintuitive implication is consistent with our theoretical
prediction that VCs require lower quality signals from ethnically proximate companies to trigger
investment, both because VCs read the signals from coethnic companies more precisely, and because VCs
anticipate coethnicity’s postinvestment positive influence.
We subject our findings to a battery of robustness tests. First, we check whether reverse causality could
be driving our estimated correlations—that is, by VCs appointing coethnic executives to their portfolio
4
companies after they perform well. We obtain the tenure of company executives for a subsample of our
data from LinkedIn, the world’s largest professional networking site, and find that the performance results
strengthen when we limit company executives to those present at the time of investment. Second, the
estimated positive effect of coethnicity on performance is stronger when the startup founder and VC
partner who sits on the startup’s executive board (and thus monitors the investments) are of the same
ethnicity. Third, the performance benefits of coethnicity are particularly strong for firsttime
entrepreneurs. Fourth, for a subsample of our data, we show that controlling for previous school ties
between VC partners and their portfolio executives, a potential correlate of ethnic closeness, does not
qualitatively alter our findings. Fifth, we find that coethnic VCs neither invest more money nor take more
time, than noncoethnic VCs to achieve successful exits. Hence, coethnicity’s positive influence appears
to stem from coordination efficiencies, not from VCs expending additional resources to ensure the
success of coethnic investments.
Our study contributes to the research on social associations in at least three ways. First, we formally
derive the implications of selection and influence on business partnerships that vary in the strength of
their social associations. Our propositions apply to associations based on attributes other than ethnicity
(such as geographic proximity or industry specialization), and to a variety of partnerships including those
between employer and employee, mentor and apprentice, and, even husband and wife.1 Second, we build
on an approach for identifying ethnic information based on individuals’ publicly available names
pioneered by Kerr (2008) and Agrawal et al. (2008) to extract a finegrained classification of ethnic
groups in a representative sample of US executives and confirm the viability of this approach for large
sample studies of ethnic origins. Third, a growing body of research explores the influence of social
networks on economic transactions conducted either across national boundaries (e.g., Gould 1994,
Bottazzi 2012) or within individual ethnic conclaves (e.g., Chung and Kalnis 2006). This research leaves
open the possibility that the benefits ascribed to social proximity are not due to proximity per se but
rather the parties’ specialized knowledge, such as a superior understanding of foreign institutions. We
uniquely demonstrate that proximity improves performance, very likely by reducing postselection
coordination costs, even within a country.
2. Theory
2.1 Model
1 Incidentally, Bratter and King (2008) provide evidence that interracial marriages are more likely to end in
divorce compared to marriages within an ethnic community.
5
This section presents a formal model of the core partnership tasks, selection and influence, as a function
of social proximity. We build on Morgan and Várdy’s (2009) model of hiring under statistical
discrimination to include influence effects and analyze proximity’s effect on partnership performance. To
maintain consistency with our empirical context, we label the parties of the model as VC and company
engaged in an investment partnership. But since the model abstracts away from any particular activity
specific to the VC industry, it is general enough to illuminate the consequences of proximity based on any
number of individuals’ social attributes such as culture, gender, race, ethnicity, alumni networks, and so
on, and for different types of partnerships. The key elements of the model are that a party desires a
partner with whom to develop a successful relationship by (i) searching for a suitable partner, (ii)
screening potential candidates based on observable signals of quality, (iii) selecting a partner, and (iv)
influencing the relationship after commitment. Social proximity facilitates screening and influence, and
thus has strategic consequences for search, selection, and ultimately the relationship’s performance.
In our context, VCs invest in companies they expect to be successfully sold, either to the public or to
another firm. Success is a function of two attributes: (i) the company’s unobservable quality ∈0,1,
where 1 indicates a high quality company and occurs with prior probability , independent of all
else, and (ii) the quality of the postinvestment relationship between the VC and the company.
VCs and companies reside on an dimensional metric space, where location represents relative
composition in the space of social associations—the th coordinate is strength of social affiliation . In the
context of venture capital, this may be the proportion of the VC or company personnel with social
affiliation . VCs evaluate potential deals one at a time, by targeting search at a specific location on the
metric space—the next potential company is more likely to reside close to the targeted location than
further away. Formally, the density , of discovered companies at location decreases about .
A VC then observes a signal of the discovered company’s quality , where is the observable
social distance between company and VC, and noise is distributed 0,, such that its
variance increases in distance (i.e.d
⁄0). In other words, VCs get more precise signals of
proximate companies’ unobservable quality, allowing them to screen proximate companies better.
Social distance also negatively affects a company’s success probability, conditional on quality—it
succeeds with probability , such that 0.2 Thus, proximity positively influences success. A
VC decides to accept the company or reject it, incurring a fixed search cost and targeting a new search.
2 Many activities occur in an investment partnership, which we do not explicitly model, but decreasing
sufficiently captures. For example, in a model where VC partners and company executives exert effort with
6
We explicitly model the strategic behavior of just one party, in our case, the VC, in relationships. In many
partnerships both parties are engaged in symmetric searching, screening, selecting and influencing
activities. Since social distance in bilateral partnerships is reduced for one party if and only if it is reduced
for the other, and qualitatively the effects of distance are the same for both parties in all respects, our
simplification to model only one side of bilateral partnerships is without loss of generality.3
2.2 Analysis
Now, suppose a VC has finished researching a company (i.e. and are observed) and must decide to
either invest in the company or reject it in favor of searching for and evaluating another. Let and Φ
respectively be the pdf and cdf of the standard normal distribution. From Bayes’ rule,
1
,
1
1
1 (1)
is the posterior probability that a company is high quality. Thus, the probability that an investment in a
socially distant company with signal will be successful is . Inverting this, the signal
1
2ln1
(2)
indicates that a ethnically distant company has a probability of success. Note that for all
there is no signal high enough: if the desired success probability exceeds the probability that the post
investment influence alone is successful , then even a guaranteed high quality company will have a
lower than probability of success. Since the value of searching again is always identical ex ante, the
cutoff signal above which the VC accepts the company always signifies the same optimal posterior
cost increasing in social distance (say because distant parties are more difficult to communicate with) and both
benefit from investment success, it is straightforward to show that in equilibrium, mutual effort and probability
of success increases in social proximity, which we do in the Theoretical Appendix. Since these background
activities are unobservable in our empirical setting, we abstract from them for parsimony in the main text.
3 Our simplification does not cover partnerships with more than two decision makers, because being close to
one (potential) member of the partnership does not necessarily imply being close to all other members of the
partnership. Search strategies could be more complex in such situations, because the closeness to other
(potential) partners would need to be considered by all parties.
7
success probability ∗. Formally, ∀,∗, or equivalently, ∀,
∗.4 Thus, without loss
of generality we can assume ∗, because any company not meeting this requirement will be
automatically rejected by the VC.
Taking the derivative of (2) with respect to and evaluating at ∗ yields
∗2
ln1
∗
∗′
∗ (3)
The first term of (3) is positive if and only if the target posterior probability exceeds the prior probability
of success (∗), and the second is always positive.5 Equation (3) positive implies that the
threshold signal to trigger investment decreases in social proximity. Intuitively, the first term denotes the
screening advantage VCs have in evaluating proximate firms—the closer the company is, the more
reliably a favorable signal indicates high quality (assuming ∗). The second term is the
influence effect, the advantage that a close company has in exiting successfully, independent of quality. If
influence is large, it is optimal for the VC to accept proximate companies of probabilistically lower
quality than it would accept among those further away. Thus,
Lemma 1: VCs set lower acceptance criteria for socially proximate companies if the prior probability
that a distant company will succeed () is less than ∗, the threshold success probability,
regardless of distance.
That is, if VCs reject most companies regardless of distance, which we assume and is generally accepted,
then VCs accept lower valued quality signals from close companies.6,7 A casual observer might perceive
this as tastebased discrimination, but it is not—VCs set the same minimum success probability for
companies at all locations. The quality signal denoting this minimum probability is lower for close
companies, both because when a close company sends a “high” quality signal it indicates a high quality
4 Explicit calculation of ∗ would require additional assumptions on the search costs of VCs, which our data
does not inform. Nevertheless we can compute comparative statics on the quality and performance of
investments with respect to social proximity without introducing additional assumptions.
5 The argument of the logarithm exceeds one if and only if ∗.
6 “The typical venture organization receives many dozens of business plans for each one it funds.” (Gompers
and Lerner 2004, p 7).
7 The threshold falls as social proximity’s postinvestment influence strengthens (i.e. ′ becomes more
negative). In the limit, as the treatment effect diminishes to 0 (i.e. approaches a constant), the statement
of Lemma 1 becomes “if and only if.” To see this, observe that the second term of equation (3) goes to 0.
8
company with greater certainty than when a distant company does so, and because the VC knows it can
compensate for low quality, to an extent, with positive influence after investment. This means that a VC
has generally observed quality signals from its socially close investments that are lower. So, it is
reasonable to ask, “How does social distance affect the performance of actual investments that the VC
makes?” Define the probability that an accepted company of distance is high quality
1
∗,1Φ∗1
1Φ∗1
1Φ∗
1
Thus, the success rate of accepted companies is . In the Theory Appendix we show
Proposition 1: Socially proximate investments are more likely to succeed (i.e.
⁄0 ).
VCs search their own social circles, because close investments are more likely to succeed. Hence, VCs
also evaluate disproportionately more companies with whom they have close social associations. Since
quality is location independent, Lemma 1 implies that, given any candidate stream to evaluate, those near
the VC will be overrepresented in its portfolio. Thus,
Proposition 2: VCs are disproportionately more likely to invest in socially proximate companies.
2.3 Discussion
Since signals of quality are unbiased, an independent auditor’s expected signal of a portfolio company
equals the conditional probability that the investment is high quality. Thus, the auditor’s expected signal
varies with ethnic distance exactly as the probability of high quality, conditional on investment. Formally,

1,1,01
which Lemma 4 (in the Theory Appendix) calculates.8 Lemma 4 shows that the sign of
⁄ turns on
the sum of two terms: the first corresponds to screening and is negative (i.e. proximity raises quality),
while the second (i.e.
2′
∗⁄ ) corresponds to influence and is positive (i.e. proximity
permits lower quality).10 Thus, if influence were absent (i.e. ′0), the VC would screenin higher
quality, socially close companies—empirically manifesting as a positive selection effect of proximity.
However, as the influence effect increases (i.e. ′ becomes more negative), proximity increases the
8 The errors of the auditor’s signal have zero mean, because they are independent of the VC investment decision.
10 In particular, see equation (10).
9
probability of selecting high quality companies less. Intuitively, the VC tolerates probabilistically lower
quality, but socially close firms, to capitalize on anticipated postinvestment influence. If influence is
strong enough, its positive effect overwhelms screening’s negative effect, and the VC’s close investments
are actually lower quality (though still have overall higher probability of success)—empirically
manifesting as a negative selection effect of proximity. This does not imply that proximitybased
selection benefits are absent. On the contrary, the VC searches to recruit more ethnically proximate
companies and screens them more precisely but, anticipating positive postinvestment influence, tolerates
lower expected quality to maximize final probability of success. Both screening and influence effects may
be arbitrarily strong, but the sign of the empirical selection effect depends on their relative strength.
Understanding these subtle links between selection and influence helps interpret our empirical findings.
Figure 1 graphically illustrates the model’s mechanism. Suppose that the VC accepts only companies with
probability of success greater than 13
. Consider first the case where social distance’s role is limited to
screening: high quality investments succeed with probability 23
⁄, regardless of social distance.11
Thus, the VC invests if and only if the posterior probability that the investee is high quality is at least 12
,
or alternatively, weakly greater than the posterior probability that the investee is low quality. From
Bayes’ Rule we can write the required condition as 1,0,1. 12 The
LHS is the density of signals the VC sees from distant companies that are high quality, scaled by the
prior probability that the company is high quality. The RHS is the density of signals the VCs sees from
distant companies that are low quality, scaled by the prior probability 1 that the company is low
quality. The LHS and RHS appear as the dark gray and light gray bellcurves respectively in Figure 1,
where the upper pair represent the scaled signal densities of near companies and the lower pair the scaled
signal densities of far companies. The scaled densities intersect where a company is equally likely to be
high or low quality.13 Investments in companies with this signal succeed with probability
1
,13
. Companies with lower signals (which the VC refuses to invest in) succeed less
often, while companies with higher signals succeed more. Notice, though, that this threshold signal is
11 Formally, for all ,23
.
12 Following equation (1) this is more completely derived
1,1
1
0
,
13 Formally, Prθ1θ
,y12
Prθ0θ
,y.
10
greater for socially distant companies (0.99 when near versus 1.6 when close), because noisier
signals cause the VC to weight its priors stronger, which are that companies are typically low quality.
Now, let postinvestment influence also depend on social proximity. Suppose that high quality
companies now succeed with 100% probability if they are socially close to their VC but with only 50%
probability if far away (i.e. 1 and 12
). To achieve a 13
probability of success the
VC can either invest in a close company that is high quality with a probability 13
or a distant company
that is high quality with probability 23
. Under our example parameterization, this is equivalent to
observing a close company with a signal 0.68 or a distant one with a signal 2.29. Dashed
vertical lines denote these new threshold values in Figure 1. Although marginal close and marginal far
companies succeed with equal probability, the former now is more likely to be low quality, because the
VC anticipates superior postinvestment influence and tolerates probabilistically lower quality.
Figure 1 here
Our model treats selection and influence effects abstractly enough to analyze investment performance in
the presence of varied social associations, but how does coethnicity practically convey these advantages
in venture capital investing? These two broad effect classes could drive performance in at least four
distinct ways: (i) VCs and entrepreneurs may search each other and meet at lower cost owing to being
part of the same ethnic network—a preinvestment selection effect; (ii) Once in contact, communication
advantages, such as shared language or mutual understanding of the significance of qualifications,
markets and opportunities, may reduce asymmetric information and facilitate mutual screening—a pre
investment selection effect; (iii) The communication advantages may continue postinvestment, making
monitoring and coordination less costly. Expectations regarding punctuality, worklifebalance, employer
employee loyalty, hierarchy, collective versus individual responsibility and so on vary with culture. Thus,
when unforeseen circumstances arise, coethnic parties may act more compatibly—a postinvestment
influence effect; and (iv) Misbehavior by either party is more likely to be observed by shared social
networks, communicated and punished within the networks. Thus, the shared ethnic community may
curtail opportunistic behavior and reduce monitoring costs—a postinvestment influence effect. Although
our empirical tests cannot distinguish among the above channels of influence, we attempt to test the
model’s propositions and isolate the postinvestment influence effects of coethnicity.
3. Empirical specification and sample
3.1 The empirical specification
11
Here, we empirically assess the three main predictions of the model stated in Lemma 1, Proposition 1, and
Proposition 2. Rather than following the sequence of derivation, we test Proposition 2 and Lemma 1 first,
since they pertain to investment selection, and then Proposition 1, which pertains to postinvestment
influence. We test Proposition 2 (and Lemma 1) by estimating the probability that a given VC invests in a
company (Pr,1) as a function of company c’s characteristics (denoted by the row vector C), VC
v’s characteristics (denoted by the row vector V) and VCcompany pair characteristics (denoted by the
row vector CV). An ideal test of Proposition 2 should estimate the probability of investment as a function
of coethnicity for all companies that a VC evaluated and check whether the VCs disproportionately
invested in coethnic companies. But we do not have data on the set of companies that VCs evaluated, and
so test whether VCs are more likely to invest in coethnic companies, relative to counterfactual
opportunities based on the observable characteristics of VCs and companies. For these estimations,
implemented both through a conventional multivariate regressions and Propensity Score Matching (PSM)
methods, we construct a sample of all actual VCcompany pairs (for which ,1) and counterfactual
VCcompany pairs (for which ,0). That is, we estimate,
Pr,1,, (4)
We describe the construction of the counterfactuals in Section 4.
We then test Proposition 1 after specifying binary measures of performance (,∈0,1) of VC
company pairs as a function of variables in C, V, and CV. Each observation is a unique VCcompany pair
such that the VC invested at least once in the company (i.e., ,1). Hence, we estimate,
Pr,1,1,, (5)
The independent variable of interest in equations (4) and (5) measures the ethnic proximity of the VC
company pair (an element of CV). We chose the VCcompany pair as our unit of analysis, rather than the
VC partnercompany executive pair, because investment decisions and contracts are made at the firm
(VC/company) level, and our data naturally incorporates information about firmlevel attributes that
influence the firms’ investment decisions (such as industry preference, location, size, and roundlevel
investments). We calculate ethnic proximity between VCcompany pairs using information on VC
partners’ and company executives’ ethnic origins. We describe this variable, as well as other elements of
C, V and CV, after describing our estimation sample in detail below.
3.2 The sample and variables
We collect data on VCs and their investments from VentureXpert, a proprietary database of Venture
Economics owned by Thomson Reuters. Venture Economics assembles data on deals between VCs and
12
their portfolio companies from the quarterly reports of VCs and other institutional investors and
supplements this data with information collected from trade publications, company web pages, mailedout
surveys and telephone contacts with VCs and companies. The coverage of deals in VentureXpert is more
comprehensive than in other databases: Gompers and Lerner (1999) conclude that it contains over 90% of
all venture investments, especially for the later years of their study, and Kaplan et al. (2008) report that it
covers 85% of all deals.
VentureXpert’s information on VCs and companies includes their founding dates, geographic location,
industry category and the names of VC partners and companies’ toplevel executives. Although
VentureXpert covers 290,000 unique deals between 1969 and the present from across the globe, the data
on our variables of interest is more complete for the investments of USbased VCs and companies started
after 1990. Hence, we restricted our sample to deals covering companies started between 1991 and June
2, 2010 and funded by USbased VCs. This restriction and cleaning the raw data left 2,687 unique US
based VCs and 11,235 unique USbased companies involved in 73,916 (roundlevel) deals. The average
company in our sample received funding from 2.8 VCs (Figure 2 presents a histogram of the number of
VCs funding each company in our sample), and the deals covered 32,017 unique actual VCcompany
pairs (pairs for which ,1). The following paragraphs describe the construction and sample
characteristics of our explanatory and control variables.
Figure 2 here
3.2.1 Companyspecific variables
(a) Ethnic origins of top executives: VentureXpert lists VC firms’ partners and their portfolio companies’
toplevel executives by given and family names.14 We assign each executive a most likely ethnicity based
on the executive’s family name and given name. Origins Info Ltd., a commercial vendor of namebased
classification services for ethnically targeted marketing campaigns provided the assignment. It uses a
proprietary database constructed from a variety of sources, such as the American Dictionary of Family
Names and international telephone directories, to identify the most likely ethnic origin for over 1,800,000
family names and 700,000 given names. Origin Info’s classification assigns an ethnicity to each name
based on the family name first, and when family names are inadequate for accurate identification (e.g. for
family names like Lee), uses a combination of family name and given name to identify ethnicity (e.g.
14 VentureXpert does not record the entry and exit of company executives and VC partners, and the list of
names we obtained reflects the ethnic composition of companies and VCs when VentureXpert last updated this
data. To overcome this limitation, we incorporate information on executives’ entry and exit dates from
LinkedIn, the professional networking site, for a subset of our data in a robustness check.
13
Seungjun Lee is classified as Korean and Keith Lee as AngloCeltic).
Although several studies have validated the accuracy of inferring ethnic origins from names in large
samples (see Webber 2007), the approach suffers from several limitations, including that it undercounts
the size of ethnic groups whose individuals assume names common among other ethnic groups (e.g.
personnel of Jewish origin frequently assume AngloSaxon, German and East European names and are
undercounted by our study) and overcounts the size of ethnic groups that provide such assumed names.
To the extent that classification errors cause us to miss actual coethnic matches, we expect the errors to
make it less likely for us to observe a positive relationship between ethnic proximity and VCcompany
matching/performance if the relationship actually exists.15
The 11,235 US companies in our sample employed a total of 85,168 toplevel executives, 13,598 of
which were also VC partners, typically listed as nonmanaging board members in the companies. We
dropped these executives from the sample of company executives and retained them in the sample of VC
partners. Our ethnic classification scheme then assigned each company executive to one of the following
ten most common ethnic groups in the US: AngloCeltic, West European, East European, North
European, South European, Chinese, Indian, Japanese, Jewish and Korean. Executives not belonging to
one of the ten ethnic groups were assigned to a miscellaneous “Others” category.16 Table 1 notes the sub
ethnic groups and nationalities (e.g. English, Irish, and Welsh) that comprise the ethnic groups. Kerr
(2008) classifies US inventors using similar ethnic groupings.
Given the completeness of VentureXpert’s coverage, our sample distribution of ethnic origins should
represent the actual ethnic distribution in US venturebacked companies, subject to the caveats noted
above. Table 1 compares the fraction of each ethnicity in the overall US population to the fractions for the
executives of USbased startups. Top executives of USbased companies are primarily of AngloCeltic
and WestEuropean origin, more or less comparable to their proportions in the overall US population.
15 To see this, assume that coethnicity is, in fact, associated with positive performance. This means, ceteris
paribus, a company executive belonging to ethnicity X is more likely to perform better when she receives
investment from a VC partner also belonging to ethnicity X. Classification errors could be of three different
types: (i) both the VC and the executive are incorrectly classified as belonging to ethnicity Y, (ii) the executive
belonging to X is incorrectly classified as belonging to ethnicity Y, and (iii) the VC belonging to X is
incorrectly classified as belonging to ethnicity Y. (i) will not affect the estimated average effect of coethnicity,
but both (ii) and (iii) will result in matches that are noncoethnic and a higher probability of superior
performance, thus biasing our estimates of the positive effect of coethnicity downwards.
16 Table A1 of the Data Appendix lists the ten most common surnames for each ethnic group in our sample of
USbased company executives.
14
Jewish, Chinese, and Indian individuals are overrepresented as executives relative to their overall
populations in the US.
Table 1 here
For each portfolio company c in our sample, a unit vector ,,,,…, indicates its position in
11dimensional ethnic space (one element each for the ten ethnic groups plus one for Others). Each
coordinate indicates the proportion of the company’s top executives belonging to the corresponding
ethnicity. We calculate the ethnic proximity for each VCcompany pair using and include the vector in
our estimations to control for the proportion of different ethnicities within firms and VCs.
(b) Number of top executives: Portfolio companies in our sample list 8.55 toplevel executives, on
average (SD = 5.26; Range = 156). The executives are most commonly designated Chief Executive
Officer, Chief Financial Officer, Founder, President, Director, Board Member, and VicePresident. We
use the number of executives belonging to each firm to control for its size and capital requirements, both
of which may influence the firms’ ethnic preferences and performance.
(c) Founding year: An “average year” in our sample from 1991 to 2010 produces 690 startups. The surge
of startup companies in 1999 and 2000 (1,218 and 971 startups, respectively) reflects the “dotcom
boom,” and the steep drop in foundings during 20082009 (348 and 132 startups, respectively) reflects the
economic downturn and perhaps truncated coverage in recent years (VentureXpert collects data about a
company when a VC reports funding it, typically two to three years after its start date).
(d) Industry: VCs invest primarily in the Internet (21.7% of sample companies), Computer Software
(20.7%), Medical/Health (12.7%), Communications (8.1%) and Biotechnology (7%) industries.17
Dummy variables for 18 industries control for unobserved industryspecific features.
3.2.2 VCspecific variables
(a) Ethnic origins of VC partners: We classified the 22,110 partners of the 2,687 USbased VCs in our
sample by ethnic origin as described in 3.2.1. Column 3 of Table 1 reports the fraction of VC partners in
our sample by ethnic origin. Most partners are of European heritage (AngloCeltic and West European
ethnicities together account for nearly 70% of the sample’s VC partners). Jewish, Indian, Chinese,
Korean, and Japanese individuals are overrepresented as partners relative to their overall US populations.
For each VC v in our sample, we generate an ethnic position vector ,,,,…, to calculate
the ethnic proximity of each VCcompany pair and control for VCs’ ethnic composition.
17 Table A2 of the Data Appendix reports the industrydistribution of the sample companies.
15
(b) Number of partners: VCs in our sample have 8.2 partners on average (S.D. = 12.23, Range = 1246).
The number of VC partners imperfectly proxies for VC size and the depth of its pockets, which may
influence both the ethnic composition of the VCs and the probability of investing in any given company.
(c) Founding year: The VCs in our sample are, on average, older than the companies; 51% were founded
before 1991. An “average year” between 1991 and 2010 produces 2.5 new VCs; however, increased VC
foundings accompany the surge of startups in 1999 and 2000 (nine and six new VCs, respectively).
Founding year dummies control for yearspecific economic activity that may influence both investments
and the ethnic composition of VCs (such as the boom of software startup companies during the late 1990s
that may have increased both VC investments and the fraction of relevant Chinese and Indian personnel).
3.2.3 CompanyVC pair specific variables
(a) Geographic distance: VCs tend to invest in geographically close companies, because collocation, like
coethnicity, arguably facilitates superior monitoring and management of investments (Lerner 1995;
Sorenson and Stuart 2005). To the extent that ethnic communities tend to cluster in space, geographic
proximity may correlate to both ethnic proximity and investment performance (see Agrawal 2008 and
Kerr 2008). To control for the geographic clustering of ethnic communities, we measure geographic
distance between each VCcompany pair by converting the headquarter addresses reported by
VentureXpert to longitude and latitude via the Google Geocoding API and compute greatcircle (“asthe
crowflies”) distances between VCs and companies using the Haversine formula (first published by
Sinnott 1984, though long known by navigators).
(b) Industry distance: VCs find it easier to make and monitor investments in industries in which they
have prior experience (Hellmann 2000). VCs and entrepreneurs belonging to certain ethnicities may also
share industry aptitude and experience. If so, shared industry expertise may correlate with ethnic
proximity and matching/performance. To control for this, we construct a variable of industry distance as
the percentage of investments that the VC has made in industries other than the one in which the paired
company operates (we use VentureXpert’s assignment of each company to one of 18 industries). This
measure of industry distance ranges from 0, when all of a VC’s prior investments were in the matched
company’s industry, to 1, when the VC has no other investments in the company’s industry.
(c) Ethnic distance: Depending on our empirical context and objective, we use three different measures
of ethnic proximity. The first measure allows us to compute and compare coethnicity’s effect for each of
the ten major ethnic groups. The measure is a vector of ten binary variables, and, for each ethnic group,
indicates whether the VC and the company each have an individual of the ethnicity. For example, a VC
company pair comprised of a VC with two partners of Indian origin and one partner of Chinese origin,
16
and a company with three executives of Indian origin and one partner of Jewish origin, will have
COETHNIC INDIAN turned on to one, while the other nine elements of the pair’s coethnicity vector will
be set to zero. 91.2% of the 32,017 unique VCcompany pairs in the sample had at least one AngloCeltic
employee each, and 56.6% of the dyads shared West European heritage (Column 1 of Table 2 reports the
relative frequency of coethnic VCcompany pairs for all ethnic groups in the sample).
Second, some of our estimations require a more parsimonious measure of coethnicity (e.g. for use in
2SLS regressions, which instrument for coethnicity). Hence, we create a binary measure of coethnicity
indicating whether or not the paired VC and company both had personnel belonging to any one of the
eight ethnic groups with distinct identities in the US (i.e. the variable is set to one if any VC partner and
any company executive of the VCcompany pair share the same ethnicity other than AngloCeltic, West
European, or Others). According to this measure, 46.6% of the VCcompany relationships were based on
at least one coethnic partnership. In comparison, a more inclusive measure that indicted any shared
ethnicity between VCs and companies (i.e. including VCcompany pairs consisting of coethnic
individuals of AngloCeltic, West European, or Other origins) would result in 97% of all sample VC
Company pairs being marked as “coethnic” and eliminate the variation (in ethnic proximity) required to
identify the effects of coethnicity.
Table 2 here
Third, we calculate a continuous measure of ethnic distance between each VCcompany pair as the
Mahalanobis distance between their ethnicity position vectors, and described under Sections 3.2.1.a
and 3.2.2.a above. Formally, Mahalanobis distance, ,, where
vectors and represent the ethnic positions of VCs and companies, respectively, S is the covariance
matrix, and T the matrix transpose operator. The advantage of the Mahalanobis measure is that, unlike
our binary measure of coethnicity, it accounts for the statistical prevalence of the different ethnicities in
the sample as well as cooccurrence of the different ethnicities in the sample. In all our regressions, we
specify the Mahalanobis ethnic distance, geographic distance, the number of company executives, and the
number of VC partners in logs to soften the effect of outliers.
4. Does ethnic proximity affect VCcompany matching?
4.1 Proximity and matching
Our matching analysis requires constructing a sample of VCcompany pairs, both actual, for which the
investment happened, and counterfactual, for which investment could have happened but did not. Since
our sample has 2,687 VCs and 11,235 companies, there are over 30 million theoretically possible pairs, of
17
which 32,017 are actual and the rest counterfactual. To distill this to a computationally manageable
number, we eliminate pairs for which the VC never invests in the company’s industry—such matches are
unlikely by revealed preference. We also eliminate pairs for which the VC was not active in a oneyear
window on either side of the (first and last) date on which the company received funding. This retains all
actual matches but eliminates nearly 50% of the counterfactual ones, leaving about 15 Million
counterfactual pairs. We work with random samples drawn from this set of actual and counterfactual pairs
due to computational constraints. We draw a 10% random sample (1,300,761 pairs, of which 3,520 were
actual matches) and, for each pair, calculate ethnic distance, geographic distance, industry distance as
well as other company and VC characteristics. We do not observe whether the startups in our data
approached certain VCs for investment but were turned down, and hence cannot predict the probability of
receiving funding as a function of proximity. Instead, we aim to test whether matched VCs and company
executives are more likely to be coethnic than unmatched ones, and our set of counterfactuals,
representing a sample of random (but feasible) unrealized matches serves this objective.
Table 2 reveals that coethnic personnel are, on average, more likely for actual VCcompany pairs than
counterfactual pairs: the difference in matching likelihood is statistically significant (at p<0.05) for all ten
ethnic groups. Next, we formally investigate the relationship between ethnic proximity and the probability
of VCcompany match (i.e. Pr,1 in equation 4) by estimating multivariate Maximum Likelihood
Probit regressions. Table 3 reports the results—Probit estimates and corresponding marginal effects of the
influence of the explanatory variables on the probability of a VCcompany match appear under Panel A.
Table 3 here
Columns 1 and 2 of Panel A confirm that after controlling for geographic distance, industry distance,
foundingyear effects of VCs and companies, the proportion of different ethnic individuals in VCs and
companies and industryspecific effects, coethnicity is positively related to the probability of a VC
company match for all ethnic groups (except for individuals of AngloCeltic origin). The positive effect
of coethnicity is statistically significant (at p<0.05) for Chinese, Indian, Jewish, and South European
ethnicities (the South European group is more homogenous than other European groups and is composed
primarily of individuals with origins in Italy and Spain). This finding confirms and extends the result first
reported in Bengtsson and Hsu (2010) that Chinese and Indian VCs in the US disproportionately invest in
companies started by members of their own community.
Column 4 shows that the average marginal effect of a single coethnic pair on matching for members of
distinct ethnic groups (Chinese, Indian, Japanese, Jewish, Korean, East European, North European and
South European) is nearly four times coethnicity’s effect for the “indistinct” groups (AngloCeltic, West
European and Others); in fact, coethnicity’s estimated effect for the latter does not statistically differ from
18
zero. The magnitude of the marginal effects may appear small (a single coethnic pair increases the
probability that a VC invests in the given company by 0.04 percentage points), but the unconditional
probability of a VCcompany pair match in our sample is 0.25%, implying that an additional coethnic pair
is associated with a 16% higher probability of a match—an economically substantial effect. Columns 5
and 6 confirm the positive effect of ethnic proximity using our Mahalanobis measure of ethnic distance.
One drawback of the above method is that it estimates coethnicity’s effect by comparing the
characteristics of actual VCcompany matches to that of many counterfactual matches. A number of these
counterfactual matches may not be comparable to the actual matches along characteristics that affect the
probability of matching. Hence, we further refine the set of counterfactual VCcompany matches by
calculating assignment probabilities, or propensity scores, for VCcompany matches. The propensity
scores are obtained from a Logit regression that predicts the probability of a match based on the set of
observable VC characteristics, company characteristics, and VCcompany pair characteristics (except
ethnic proximity) described in Section 3.2. We then compare the average ethnic proximity of the actual
VCcompany pairs to the average ethnic proximity of the counterfactual sample that retains only those
pairs with matching probabilities comparable to the actual matches.18
Panel B of Table 3 reports results from this PSM exercise. The estimates suggest that the “average
treatment effect on the treated group” (ATT), which is conceptually equivalent to the marginal effect of
coethnicity estimated by the Probit regressions above, is 0.027 percentage points. This effect is
statistically significant, but lower in magnitude than the marginal effect of coethnicity estimated by the
Probit regression (0.04 percentage points), perhaps because PSM compares against a more plausible set of
counterfactuals.19 The estimated ATT translates to a 11% higher probability of matching for VC
companies with coethnic individuals. These results empirically support Proposition 2: VCs match more
with (i.e. are more likely to invest in) ethnically close companies.
4.2 Proximity, matching and quality signals
18 We experimented with various PSM techniques including nearest neighbor matching, kernel matching, and
caliper matching to construct the appropriate control group and confirmed that the results are not sensitive to
the technique used. The results reported here are the most conservative ones (i.e., yield the lowest estimates of
the effect of coethnicity) and are based on caliper matching, which uses a prespecified tolerance level on the
maximum propensity score distance (“caliper”) to minimize the risk of bad matches.
19 We check and confirm that the covariates are identical and balanced across our control and treatment groups.
Any differences between the groups are within the acceptable bounds prescribed in Rosenbaum and Rubin
(1985). Table A4 of the Appendix reports the results of our balancing tests.
19
According to Lemma 1, VCs screen coethnic investments less stringently, both because VCs are surer that
the coethnic company they are evaluating is of the indicated quality and because they know that the
positive influence effects of coethnicity will compensate for lower quality at the time of investment.
Although the econometrician cannot measure the quality signals observed by the VCs when it invested,
one can check whether VCs are more likely to invest in coethnic startups associated with lower quality
signals by using information ex ante generally correlated with startup success.
Rather than providing all the capital required by startups upfront, VCs inject capital into their portfolio
companies in successive stages or “rounds.” This staged infusion allows VCs to learn about the quality
and prospects of startups, while preserving their option to discontinue funding if the venture appears
unlikely to succeed (e.g. Bergemann and Hege 1998, Wang and Zhou 2004). Hence, the average success
probability of startups at first round funding (R1) is lower than the success probability of startups that
receive second round funding (R2), which is lower than the success probability of startups that survive to
the third round (R3), and so on.20 If VCs are more likely to select coethnic ventures in earlier rounds, then,
this will provide evidence that VCs tolerate lower quality signals from coethnic startups.
We construct the actual and counterfactual matches anew to incorporate roundlevel information. Since,
in our sample, all VCs that fund a given company in round also fund it in round 1 if the round
occurs, we restrict actual pairs to newly formed matches and remove counterfactual matches which were
actual matches in previous rounds. As before, we only retain plausible matches, based on the companies’
industry and the VCs’ revealed industrypreference, and the relevant window of investment opportunity
during which VCs were active. We estimate a separate model for each of the three subsamples
representing the first three rounds. Clearly, since survival to round 1 requires survival to round ,
and we only consider the new matches in each round, the sample size decreases from round to round. Our
computing resources constrained us to work with 30% random samples of the actual and plausible
counterfactual matches for each round. The dependent variable for this analysis equals one if the given
VC financed the company in the corresponding round, and zero if it did not.
Figure 3 here
Companies in our sample receive 3.3 rounds of funding, on average, and 75% of firms that experience
exit events (IPOs, acquisitions, mergers, LBOs and bankruptcies) do so with five or fewer funding rounds
(see Figure 3). So, we focus on the relationship between VCcompany matching and ethnic proximity for
the first four funding rounds. Table 4 presents the corresponding Probit estimates. Because some
20 In our sample, firms that received funding in R1, R2, R3 and R4 had IPO probabilities of 7.7%, 9.4%,
11.1%, and 12.2% respectively.
20
coethnic groups (e.g. Japanese and Korean) lack sufficient numbers of actual coethnic pairings in each
round to precisely estimate their effects, we estimate and report results obtained by using the binary
variable that indicates the presence coethnic personnel belonging to any of the eight distinct ethnic groups
(i.e. coethnic pairs classified as AngloCeltic, West European or Others do not set the variable to one).
Table 4 here
The estimates in Table 4 suggest that ethnic proximity plays a more significant role in matching VCs to
companies during earlier rounds, when VCs face the highest search and screening costs. Panel A shows
that an additional coethnic pair is associated with an increase in the probability of matching by 0.03% in
the first round (both at p<0.01); for second and third rounds, the effect drops to 0.01% (p<0.05) and does
not statistically differ from zero for the fourth round. Although we do not report the estimates for later
rounds, we find that the estimated effect of coethnicity for rounds R5 and higher were not statistically
different from zero. Roundlevel PSM results also confirm this decay in the estimated average treatment
effect on the treated (ATT) with the progression of rounds (see Panel B of Table 4). The estimated effects
of ethnic proximity follow a similar pattern when measured by the continuous Mahalanobis distance
metric. Interestingly, the estimated effects of geographic and industry proximity also follow a similar
pattern, consistent with the explanation that search and selection advantages conferred by collocation and
cospecialization become less salient as noise about companies’ quality decreases.
The probability of startups’ success also depends on their lifestage. As a startup matures, ideas become
tangible products, business plans translate to verifiable costs and revenues, expansion plans can be better
evaluated, and the probability of subsequent failure diminishes. Thus, an alternative test for Lemma 1
suggests that coethnic VCs should be more likely to invest in less mature (i.e. lower ex ante quality)
companies. Since the progress of startups along their lifecycle correlates highly with the number of
investment rounds received, we limit attention to the first time the startups receive venture funding—do
coethnic VCs invest in less mature companies in R1? Of the 10,134 startups in our R1 sample, 21% were
denoted as “Seed Stage,” 41.7% as “Early Stage,” 16.4% as “Expansion Stage,” 3.7% as “Late Stage,”
and 17.3% as “Buyout and Acquisition Stage.” The estimates in Table 5 confirm that ethnic proximity
most significantly predicts VCstartup matching during the first round of investment for Seed Stage, Early
Stage, and Expansion Stage companies (estimated effect of 0.03% at p<0.01 in each case), and has no
statistically significant effect for either Late Stage or Buyout and Acquisition Stage, when the probability
of company failure is relatively low.21
21 In our sample of firms that received R1 funding, those in the Buyout and Acquisition phase had an IPO
probability of 13%, while firms in the earlier stages had IPO probabilities in the 5.7%8.3% range.
21
Table 5 here
Finally, the distribution of company age at the time of initial venture investment also indicates that VCs
accept lower quality signals from ethnically closer companies. The average startup company that closes
its first funding round with a noncoethnic VC (as before, “coethnic” denotes shared ethnicity among
individuals belonging to one of the eight distinct groups) does so 985 days after incorporation compared
to 901 days (nearly a full quarterofayear later) for one funded by a coethnic VC. Hence, coethnic
investments appear to be associated with lower quality signals as suggested by Lemma 1.
5. Is proximity related to superior performance?
5.1 Successful exits through IPOs and acquisitions
Much of the mentoring provided by VC partners to startups aims to maximize the likelihood of IPO,
because VCs earn the highest average returns through this exit channel (Cochrane 2006; Hochberg et al.
2007). Unlike startup survival, an alternative measure of success which could be driven by VCs’ tastes
for keeping ethnically close companies afloat, IPOs require public markets to evaluate company
prospects. Therefore, in the absence of investmentlevel rateofreturn data, IPOs are the clearest available
signal of investment success. Although VCs tend to approach the acquisitions market either as a second
best option to going public or when they want to exit a business through “fire sales,” previous work
suggests acquisitions also generate positive returns for VCs and startups (see Gompers and Lerner 2000).
Hence, our primary measure of investment success indicates companies’ successful exits via IPOs and
acquisitions (the Data and Methods Appendix shows that our key results reported below hold, and in
some cases, strengthen when we restrict our performance measure to indicate exits via IPOs alone). We
also use information on companies’ financial performance after IPOs to measure performance for a subset
of the companies for which such data are available.
VentureXpert identifies companies that have exited through IPOs, mergers, LBOs, acquisitions, and
bankruptcies, but the rest are classified as “Private.” Among the companies denoted “Private” are two
types: (i) companies that failed to either go public or be acquired and were eventually written off by the
VCs22, and (ii) companies that were started during the later years of our sample and have not yet had the
time to exit or be abandoned. Since many companies designated “Private” may be defunct and written off
by the VCs, we eliminate all “Private” companies from our analysis that, as of December 31, 2010, had
received no funding in more than four years. We chose this threshold because 95% of the companies in
22 Unlike exits through IPOs, mergers or acquisitions, company exits via writeoffs or abandonments by VCs
are not recorded by VentureXpert. For less than 2% of the companies in the sample, VentureXpert recorded
Chapter 7 and Chapter 11 filings, but these numbers do not capture firms that are less officially defunct.
22
our sample that went public or were acquired did so within four years from their last date of financing.
This left 5,950 unique companies funded by 2,121 VCs and 17,418 observations (unique VCcompany
pairs) in the estimation sample. We then verified and considered only those events as successful exits for
which VentureXpert’s indication of IPO and acquisition events were also present in Securities Data
Company’s (SDC) Global New Issues database and SDC Platinum’s M&A database.
Table 6 here
Table 6 shows that 22.2% of the 5,950 companies in our sample exited through IPOs and about the same
percentage exited through acquisitions.23 Thus, overall, 44.5% of the companies are considered successful
exits. These companies appear to share a higher proportion of coethnic personnel with their VCs than
those that exited through other means or stayed private. Figure 4 also reveals that the distribution of
ethnic distances for successful exits, IPOs in particular, is concentrated at lower values.
Figure 4 here
Table 7 presents Probit MLE estimates of the relationship between proximity and successful exits
measured by a binary dependent variable equal to one if the company went public or was acquired, and
zero for all other outcomes. The estimations control for the proportion of VC and company personnel that
belong to each of the ethnic groups, industryspecific effects, company and VC founding year effects, size
of the business partners, total investment by the VCs in the companies, and geographic and industry
distance between VCs and their portfolio companies. As with the matching regressions, we first estimate
the effect of coethnicity separately for each of our eleven different ethnic categories. Column 1 shows that
shared ethnicity is positively associated with the probability of successful exit for each of our distinct
ethnic groups, except for Korean, although coethnicity’s positive effect is only statistically significant for
the Jewish and South European groups (the lack of statistical significance for the other distinct groups is
because of the relatively small number of observations associated with these groups; pooling the
information in these coethnic indicator variables into one variable, as our COETHNIC DISTINCT
GROUPS variable does, increases the number of observations and yields more precise estimates).
Interestingly, coethnicity appears to negatively relate to performance for the indistinct groups (Anglo
Celtic and West European). The coefficient on the binary variable which indicates the average effect of
shared coethnicity across the eight distinct groups suggests that switching the ethnicity of one VC partner
to that of a company executive increases the probability of successful exit by 3.1% (Column 4).
23 These proportions are similar to the ones presented by Cochrane (2006). In our “raw data,” which retains
active firms too young to experience exit events, the proportion of firms with IPOs and acquisition events is
much lower—12% and 13% respectively.
23
Table 7 here
Next, we control for the unobserved quality of VC partners by incorporating VCfixed effects (which
control not only for VCquality but also other unobserved VC characteristics, which may influence their
investment performance, such as access to syndicates of coinvestors, managerial talent pools, reputation,
stage preferences and access to capital). Rather than Probit MLE, we estimate VCfixed effects
regressions as Linear Probability Models (LPM) for two reasons. First, previous research has shown that
slope estimates of nonlinear models with fixedeffects, such as Probit, can be biased (Heckman 1981);
and, second, our maximum likelihood algorithms for Probit fail to converge with the additional 2,121
dummy variables in the VCfixed effects model (this methodological issue, called the “incidental
parameters problem,” is well documented—see, e.g. Greene 2001). Although the LPM has its limitations,
it produces point estimates of the effect of explanatory variables very close to the estimates produced by
Probit MLE regressions. Column 5 of Table 7 shows that in the model with VCfixed effects, the
estimated average effect of a coethnic pair (for distinct ethnic groups) on the probability of successful exit
(2.5%) is comparable to the estimated marginal effect of coethnicity without (3.1%).24 Thus, even within
a given VC’s portfolio, startup companies that are ethnically closest to the VCs perform best. To
graphically illustrate the effect of coethnicity on performance we predict the performance of all
companies using the full set of controls (i.e. all explanatory variables except VCcompany coethnicity,
and then plot the difference between actual and predicted performances separately for coethnic and non
coethnic VCcompany pairs in Figure 5. The density of these residuals for coethnic VCcompany pairs is
clearly “rightshifted” suggesting they perform much better than expected.
Figure 5 here
Although any of a startup’s top executives seeking VC funding and several VC partners may be involved
in evaluating and selecting investments, interactions between VCs and startups after the investment
mostly occur between a couple of VC partners and company executives that sit on the company’s board.
Hence, we recalculate our measure of coethnicity by considering only founders and CEOs of startup
companies and VC partners who sit on their portfolio companies’ board of directors and estimate the
performance regressions with this refined measure. The companies in our sample list 1.4 top executives
on average as either founders or CEOs, and VCs, on average, placed an average of 1.15 partners on
companies’ boards, conditional on placement. Since only 28% of the VCcompany pairs in our estimation
sample placed VC partner(s) on the company’s board, we calculate the ethnic proximity measure between
24 In models with VCfixed effects, VCspecific observables such as number of VC partners and percentage of
ethnic personnel in VC are subsumed by the VCspecific dummies and drop out of the estimations.
24
founders and CEOs of startup companies and all VC partners for the cases with no VC partners on the
company’s board and control for the presence of VC partners on the board. Incorporating these changes
yield estimates of the effect of coethnicity on the probability of successful exit of 2.8% (p< 0.01) for an
additional coethnic pairing (see Column 6), slightly higher than the estimate (2.5%) obtained using the
unrestricted set of company executives and VC partners. If we recalculate the coethnicity measure based
on the company personnel that are listed as founders alone (only 22% of our sample pairs were associated
with companies that listed at least one founder on their rosters; hence, we also control for the
idiosyncratic effect of companies that list the founder), the estimated effect of coethnicity jumps to 3.7%
(p<0.01). Since founders could not have been hired at later stages, our estimates of coethnicity’s effect
on performance are unlikely to be driven VCs adding coethnic personnel after investment.
We estimate each specification above by restricting successful exits to IPOs alone, as well as IPOs and
“good” acquisitions.25 We find that the corresponding effects of coethnicity are comparable to the ones
obtained by including all acquisitions (the corresponding VCfixed effects estimates are reported in Table
A5 of the Data and Methods Appendix). The effects of ethnic proximity on performance obtained by
using the Mahalanobis measure of proximity are qualitatively similar to the ones obtained by using the
binary variable (Table A6 of the Data and Methods Appendix tabulates the corresponding results). As one
might expect, the amount of funding received by the companies, VC and company size, geographic
proximity and industry proximity are all positively related to the probability of successful exit.
5.2 Isolating ethnic proximity’s influence effects
The above estimations show that ethnic proximity positively relates to performance, even within a given
VC’s portfolio. Yet, this effect, identified through conditional correlations, could reflect either selection
of higher quality companies by coethnic VCs, or positive influence due to coethnic partners’ lower
coordination costs after investment. Although our matching results suggest that VCs invest in coethnic
ventures generally associated with lower quality signals, we do not observe the true quality of the startups
and cannot immediately conclude that coethnicity has a positive influence on performance.
5.2.1. Instrumental variables: The main challenge to isolating the influence effects of coethnicity stems
from omitted variables, such as the unobserved quality of startups, which affect ethnicitybased selection
of startups and performance. The ideal experiment to identify coethnicity’s postinvestment influence
would randomly assign startup companies to different VCs, and then measure differences in the
25 We defined “good acquisitions” as those acquisitions for which the transaction value of the acquisition
reported by SDC Platinum’s Mergers and Acquisition’s database exceeded total VC investments in the startup.
69% of the acquisitions in our data were considered “good” by this measure.
25
performance of coethnic and noncoethnic relationships. However, both our model and first set of
empirical findings suggest that VCs do not select startups randomly with respect to ethnicity—in fact, we
have seen that VCs systematically favor investments in earlystage coethnic startups. Alternatively, we
could utilize quasinatural experiments, such as natural disasters or wars, which lead to the migration of
ethnic communities into the US and generate exogenous variation in the availability of coethnic
investment opportunities, but no such largescale “experiments” are available during the period of our
study. Thus, we employ instrumental variables to isolate exogenous variation in the probability of
coethnic investment; that is, variables which affect the propensity of VCs to invest in coethnic startups
but do not directly bear upon the postinvestment performance of coethnic relationships.
We propose and implement two IVs: (i) the probability of coethnic investments in a company’s market,
and (ii) state, industry and yearfixed effects at the time of the startup’s founding. Marketlevel
characteristics are natural candidates for IVs in our case because, to an extent, they exogenously
determine the availability of coethnic partners and thus the likelihood of coethnic matches. Intuitively,
after controlling for geographic proximity and industry specialization of the potential partners, a VC of
Indian origin is more likely to encounter, and invest in, a coethnic startup in California (where VCs and
entrepreneurs of Indian origin abound) than in New York (where VCs and entrepreneurs are drawn from a
broader pool of ethnic backgrounds). However, conditioned on encountering a coethnic entrepreneur, an
Indian VC is no more likely to enjoy screening advantages in New York than in California. Further,
while the average (preinvestment) quality of Indian entrepreneurs may be higher in California than in
New York, there is no reason to believe that the average quality of coethnic Indian investments will be
higher in California than in New York (unless VCs preferences to form coethnic matches and the quality
thresholds they set to initiate investments differs across states). We can thus examine how, after
partialling out the effect of geographic and industryproximity and other observable VCcompany
characteristics, the variation in the propensity to form coethnic VCcompany matches predicted by market
elements shapes the performance of coethnic matches. Researchers have previously used similar
instruments based on marketlevel aggregates and fixedeffects to identify the treatment effects of
variables such as investor experience and geographic proximity on investor performance (e.g. Sørensen
2007, Botazzi et al. 2008, and Tian 2011).
Our first instrument captures variation across markets in VCs’ propensity to invest in coethnic companies.
We define each “market” as the given company’s state, industry and funding year triplet, and calculate the
mean ethnic proximity between VCs and companies in each of the 2,875 unique markets in our sample
(we exclude the focal VC and company from the calculation of the corresponding market’s mean). To
facilitate ease of interpretation and to avoid tabulating multiple sets of coefficients, we report OLS rather
26
than Probit estimates for the secondstage equations. We start with the baseline OLS estimates with VC
fixed effects (reported in Column 5 of Table 7 but repeated again in Column 1 of Table 8 for easy
comparison with the IV estimates). We find that our proposed instrument (average coethnicity of VC
company pairs in the focal firm’s market) is strongly related to the probability of the pair sharing common
ethnicity (= 0.522 at p<0.01; tstat = 49.89), after controlling for other factors. Column 2 of Table 8
shows that the effect of coethnicity obtained by using this instrument through 2SLS estimation (0.121) is
substantially larger than the OLS estimate (0.025) and is statistically significant.
Table 8 here
Botazzi et al. (2008) adapt a general IVbased approach proposed by Ackerberg and Botticini (2002) to
explain the matching of companies to experienced investors, based on the assumption that the matching of
VCs to companies depends on the exogenous marketspecific availability of VCs and startups. Similarly,
the propensity for coethnic matching should differ across markets based on factors unrelated to
coethnicity’s influence on a given VCcompany relationship. Thus, market fixed effects, together with
interaction effects among market factors, serve as appropriate instruments to isolate the effects of
coethnicity on performance. Our data includes companies located in 50 states, in 18 different industries
and funded in the 20 years between 1991 and 2010. We include fixed effects for the states, industries, and
years, as well as fixed effects for the interactions of stateindustry and industryfunding years. This
results in 1,275 binary variables (49171950172017 that subsume the effect of
quasinatural experiments such as changes in visa policy, waves of immigration of certain ethnic
communities into the US, or industryspecific macro trends that arguably influence the probability of
coethnic investments, but not their quality, conditional on investment. Column 3 of Table 8 shows that the
effect of coethnicity obtained through this approach (0.169) is also substantially larger than the OLS
estimate and is statistically significant. Both IV estimations employ standard errors clustered at the state
industryyear levels for the statistical tests. The confidence intervals associated with IV estimates of
coethnicity’s influence, although estimated with larger standard errors, do not overlap with the confidence
intervals around the OLS estimate and are statistically different from the latter (at p<0.05).
We conduct three wellknown tests to check the validity of our instruments. First, weak instruments, or
instruments that are not sufficiently correlated with the endogenous regressor (the existence of a coethnic
bond for a given VCcompany pair in our case), will not only fail to correct the biases of OLS estimates,
but result in incorrect tests of significance (Bound, Jaeger, and Baker 1995). The “strength” or relevance
of instruments can be checked by testing for the joint significance of the excluded instruments in the first
stage; in particular, Stock and Yogo (2005) recommend firststage Fstatistics in the range of 1025 for
instrument relevance. We find that the firststage Fstatistics for the excluded instruments in our two IV
27
estimations are 641.8 (p<0.01) and 202.4 (p<0.01) respectively. These values are well above the critical
values and pass Stock and Yogo’s test for instrument relevance. Second, the Anderson canonical
correlation statistics of 1540.4 (p<0.01) and 755.05 (p<0.01) associated with our two IV regressions also
lead us to reject the hypothesis of underidentification, confirming the strong correlation between our
excluded instruments and the endogenous regressor. Third, the DurbinWuHausman test, which involves
comparing the coefficients obtained by OLS and IV rejects the null hypothesis (15.52 at p<0.01 and 11.51
at p<0.01) that the effect of the endogenous regressor is orthogonal to the error term in OLS regressions,
thus validating the superiority of estimates obtained through our IV approaches. Still, these IV approaches
have their shortcomings: one could argue that coethnic investments in some markets are of systematically
higher (preinvestment) quality for reasons that are not adequately captured by our control variables.
Subject to this caveat, our IV tests suggest that the influence of unobserved variables, such as quality, on
the performance of coethnic partnerships is negative, further supporting the possibility that VCs select
coethnic firms of lower (preinvestment) quality than noncoethnic firms.
5.2.2. Heckman selection correction: We next follow the approach proposed in Heckman (1979), which
explicitly estimates a firststage selection equation that predicts the sorting and matching of VCs and
companies and then incorporates this information in an outcome equation that estimates the treatment
effect of coethnicity. To implement Heckman’s twostage model, we return to the sample of possible VC
company matches used to estimate the effect of coethnicity on the probability of matching. As before,
after eliminating nonplausible, counterfactual matches as described before, we estimate a firststage
equation with the full set of company, VC and pair characteristics (explained in Section 4.1). We also add
an additional variable to the matching equation that will not be part of the secondstage outcome equation:
the percentage of coethnic matches in each company’s founding yearstateindustry. This variable relates
closely to the matching of VCcompany pairs for the reasons explained above, does not directly affect the
influence of coethnicity on performance for a given VCcompany pair, and thus imposes the exclusion
restriction. We then use the parameter estimates obtained by this matching equation to compute the
Inverse Mills Ratio (IMR) for each observation. Finally, using the “selected” sample—observations for
which VCs invested in the company—we estimate the outcome equation with our usual set of control
variables and the IMR as an additional explanatory variable to correct for selection bias.26 Column 4 of
Table 8 reports the corresponding estimates of the performance equation. The estimated effect of
26 Our Heckman correction model does not include VCfixed effects both because they do not meaningfully
belong in the firststage which predicts matching of VCs to startups and because the small number actual
matches available to us in the secondstage (because they are obtained from the 10% random sample of
possible VCcompany pairs used in the firststage) does not permit the estimation of VCspecific intercepts.
28
coethnicity (0.13), which we interpret as its influence effect (or “treatment effect”) is again significantly
larger than the OLS estimates.
Table 9 here
Table A7 of the Data and Methods Appendix confirms that the strong positive influence effect of
coethnicity holds when we restrict the definition of success to exits through IPOs alone. Moreover, the
strong positive postinvestment effects of coethnicity persist even after successful exit. Table 9 shows
that companies that are ethnically closer to their VCs continue to flourish even after IPO: in the model
with VCfixed effects, an additional coethnic pair is associated with, on average, a $0.1 million higher
market capitalization and $0.009 million higher net income one year after IPO for the startups. Thus, we
find no evidence that ethnically close VCs and companies “hoodwink” public markets in their IPOs.
These tests collectively confirm coethnicity’s strong positive influence on performance. Also, while we
have not directly established the magnitude (or direction) of coethnicity’s selection effect, the fact that
OLS estimates, which include the effects of selection, are substantially lower than the IV estimates that
control for unobserved quality, suggests that VCs tend to invest in relatively lower quality coethnic
companies. This selection of startups associated with lower quality signals is consistent with our finding
that coethnic partnerships are particularly likely during early funding rounds and when companies are
young. These results align with the intuition exposed by our model: VCs search to recruit more
ethnically proximate companies, and screen them more precisely, but anticipating positive post
investment influence, tolerate lower expected quality to maximize final probability of success.
5.3 Alternative explanations and robustness checks
(i) We find a strong positive relationship between VCs’ investments in coethnic ventures and investment
success and interpret this as evidence for the positive influence of coethnic partners. But one might ask
whether reverse causality drives our estimates: when a company performs well, VC partners replace its
top executives with their ethnic brethren. Interviews with VCs and entrepreneurs, however, suggest just
the opposite—neither party wishes to alter a successful partnership. So, such replacements in thriving
firms are rare. Our estimates obtained after limiting company personnel to founders, who are unlikely to
be VC chosen replacements, should also mitigate this concern. Still, we examined this issue directly by
assembling data on the entry and exit dates of the company executives in our sample from LinkedIn, the
world’s largest professional networking database. We were able to identify and gather data for 5,272
(6.1%) company executives in our sample that had employment records on LinkedIn. We then
recalculated our ethnic distance measure for each VCcompany pair using only those executives that were
actively employed by the startup company when the company was first funded by the corresponding VC.
29
This reduces the number of VCcompany pairs for which we can compute ethnic distance to 1,306 or
7.4% of the full list of actual VCcompany pairs. The estimated effects of ethnic proximity obtained by
fitting the successful exitprobability regressions to this restricted sample are 22.5 times larger than those
obtained from the corresponding fullsample OLS/Probit estimates (Columns 1 and 2 of Table A8 in the
Data and Methods Appendix display the corresponding estimates).27
(ii) Our analysis estimates the effect of ethnic proximity after controlling for the most salient VC and
company characteristics known to affect performance. But our coethnicity measure may pick up the effect
of other ethnicityrelated social associations, including school ties between VC partners and company
executives (e.g. Bengtsson and Hsu 2010 show that VC partners from elite US universities tend to fund
ventures founded by executives with degrees from elite US universities; Rider 2012 finds that social
associations, including school ties, affect VCs’ partnership decisions). We investigated the effect of
common school ties by assembling data on the educational institutions attended by VC partners and
company executives from LinkedIn. We coded a binary “school ties” variable indicating whether one of
the VC partners attended the same institution as any of its portfolio company’s executives (we could
construct this variable for 31% of the VCcompany pairs in our sample after identifying educational
institution affiliations for about 6% of company executives and VC partners). We then reestimated our
performance regressions including the school ties variable. Columns 3 and 4 of Table A8 of the Data and
Methods Appendix show that although school ties have a strong positive relationship on the probability of
successful exit, they do not qualitatively alter the estimated effects of coethnicity.28
(iii) We have argued that the observed positive relationship between proximity and investment
performance stems from coethnicity’s strong influence effects; that is, coethnic VC partners and company
executives work together better due to reduced coordination and monitoring costs. One might argue still
that the positive relationship of coethnicity could be driven by VCs allocating more time and resources to
coethnic companies. Table A9 of the Data and Methods Appendix shows that companies which
successfully exit when paired with coethnic VCs (a) do not require more time to exit (measured from the
27 We cannot estimate VCfixed effects regressions for this subsample due to the small number of observations
relative to the number of VCs.
28 Column 3 of Table A8 suggests that the effect of coethnicity on successful exits (IPOs and acquisitions) is
not statistically significant on the inclusion of the school ties variable, but this is an artifact of the subsample
for which LinkedIn data on school ties is available, rather than due to correlation between school ties and
coethnicity. We estimated the identical regression by omitting the ties variable using the LinkedIn subsample
and found that the estimated coefficient on coethnicity (0.018) was also not statistically significant in this
subsample. However, the effect of coethnicity on IPOs remains robust and significant (Column 4 of Table A8).
30
first funding round); (b) do not go through more funding rounds; and (c) do not receive more funding.
These findings are inconsistent with the argument that VCs inefficiently subsidize their coethnic
investments to inflate their probability of success.
(iv) Serial entrepreneurs have prior histories of company founding or success, give off more precise
signals of quality, and may need less communication and monitoring from coethnic VCs after investment.
Thus, startups with serial entrepreneurs are comparable to more mature companies, and we expect
coethnicity to play less of a role in the presence of such experienced entrepreneurs. We identify all
company executives that appeared on the rosters of two or more companies in our dataset, as serial
entrepreneurs. 13% of our sample companies listed such individuals most likely to be serial entrepreneurs.
As expected, excluding the companies associated with these individuals yielded estimates of coethnicity
higher than the ones obtained from the full sample (Table A10 of the Data and Methods Appendix
presents the corresponding estimates. We also find that companies associated with the serial entrepreneurs
in our sample are nearly twice as likely to successfully exit, ceteris paribus).
5.4 Effect of ethnic proximity on VCs’ payoffs
We find that ethnic proximity of VCs and entrepreneurs is associated with a higher probability of the
portfolio investment going public or being acquired. How much is this increased likelihood of IPO or
acquisition worth to VCs? We compute the positive impact of an increase in IPO or acquisition
probability on the ex ante expected rate of return for an investment as the derivative of expected rate of
return with respect to IPO or acquisition probability. First, condition the expected rate of return (r) on the
IPO or acquisition event (abbreviated below):
1 (6)
where p is the probability of an IPO or acquisition. Then, the derivative with respect to p is just the
difference between the expected rates of return when an IPO or acquisition occurs and when it does not:
 (7)
Since data on rates of return for individual investments, which do not exit via IPO or acquisition are not
generally available, we isolate from (6) and substitute it into (7):

Simplifying further,

(8)
31
Estimates for the three parameters on the RHS of the above can be found in Cochrane (2005). Cochrane
estimates mean returns of 698% on VCs investments that exit in IPOs or acquisitions. Accounting for the
selection that occurs prior to a successful exit, Cochrane estimates overall ex ante expected returns to VC
investments () of 59%. In Cochrane’s sample, 41.9% (p) of firms IPO or are acquired (not including
an additional 3.7% registered for IPO), comparable to the 44.46% of firms that exit through IPOs or
acquisitions in our sample. Substituting these values into (8) yields
⁄11.00. This implies that
our conservatively observed increase in the probability of successful exit of 2.5% (Column 5 of Table 7)
associated with an additional executive who shares ethnicity with a VC partner increases the expected rate
of return around 27.5% at the time of investment. These IRR estimates show that the economic returns of
coethnic partnerships are substantial, but should be interpreted cautiously—they rely on Cochrane’s
finding that VCs, on average, enjoy 698% returns from successful exit events.
6. Conclusion
Our formal model highlights the subtle interaction between the selection and influence effects of social
associations in business partnerships. It can be applied to many settings where the association between
potential partners can be described with a distance metric. The model proposes that if proximity improves
(selection relevant) information and most potential candidates are unsuitable, then increased confidence in
their evaluation will cause evaluators to set lower acceptance thresholds over observable quality signals
for nearby candidates. If proximity also improves performance after the partnership’s formation, then
anticipating this, evaluators will drop thresholds for close opportunities further, even to the point that
close candidates of lower quality will be accepted. But this is not taste based discrimination—for these
close relationships will perform better on average than distant ones. Thus, agents will target their searches
for potential partners nearby and partner disproportionately with social neighbors.
Our empirical analysis confirms the model’s predictions. We show that conditional on investment, ethnic
proximity between VCs and company executives is positively related to the probability that the venture
exits in an IPO or acquisition, and to postIPO market capitalization and net income. We also show that
VCs are more likely to select ventures led by coethnic executives for investment, and the effect of
proximity on investment selection is particularly salient for earlystage startups. Thus, our findings
suggest that in the VC industry, favoritism toward one’s ethnic brethren brings superior economic
payoffs. According to the National Venture Capital Association, “In 2008, [US] venture capitalbacked
companies employed more than 12 million people and generated nearly $3 trillion in revenue (NVCA
2009, p 2).” If the ethnicity of a single executive can substantially affect the probability of investment
from a particular VC, of growing to sale on public markets, and postIPO income, as we have found, we
can conclude that individuals’ social associations have profound economic consequences.
32
In our study, ethnic proximity proxies for a complex web of social ties that include linguistic, religious,
and many other associations that bind together members of the same ethnic group. Individuals may
choose to tap into certain associations borne out of a common ethnicity and not others. In teasing apart the
effects of shared location, industry preferences, and educational background from lessdistinct aspects of
ethnic proximity that plausibly affect investments, we have only taken a first step in identifying the true
effects of ethnic proximity and the channels through which they operate.
References
Ackerberg, D., Botticini, M., 2002. Endogenous matching and the empirical determinants of contract
form. Journal of Political Economy 110, 564–591.
Agrawal, A., Kapur, D., and McHale, J. 2008. How Do Spatial and Social Proximity Influence
Knowledge Flows? Evidence from Patent Data, Journal of Urban Economics, 64: 258269.
Bengtsson, O. and Hsu, D. H. 2010. How Do Venture Capital Partners Match with Startup Founders,
Wharton Working Paper Series dated August 2010
Bergemann, D., Hege U., 1998. Venture capital financing, moral hazard, and learning. Journal of
Banking and Finance 22, 703–735.
Bottazzi, L., Da Rin, M., Hellmann, T., 2008. Who are the active investors? Evidence from venture
capital. Journal of Financial Economics 89, 488–512.
Bottazzi, L., Da Rin, M. and Hellmann, T. 2012. The Importance of Trust for Investment: Evidence from
venture capital, University of British Columbia, Working paper
Bratter, J. L. and King, R. B. 2008. But Will It Last? Marital Instability Among Interracial and Same
Race Couples. Family Relations, 57: 160–171.
Bound, John, David A. Jaeger, and Regina Baker. 1995. Problems with Instrumental Variables Estimation
When the Correlation Between the Instruments and the Endogenous Explanatory Variables is Weak.
Journal of the American Statistical Association, 90(430): 443–50.
Cochrane, J. 2006. The Risk and Return of Venture Capital, Journal of Financial Economics, 75: 352.
Gompers, P. A., and Lerner, J. 1999. The Venture Capital Cycle. MIT Press.
Gordon, R.D. 1941. “Values of Mill's ratio of area to bounding ordinate of the normal probability integral
for large values of the argument,” The Annals of Mathematical Statistics, 12, 364366.
Gould, D. M. 1994. Immigrant Links to the Home Country: Empirical Implications for U.S. Bilateral
Trade Flows, Review of Economics and Statistics 76:2, May, p.302316.
33
Gompers, P. A. and J. Lerner. The venture capital cycle. MIT press, 2004.
Greene, W. 2001. Estimating Econometric Models with Fixed Effects. Unpublished paper available at
http://www.stern.nyu.edu/~wgreene
Heckman, J., 1979. Sample selection bias as a specification error. Econometrica 47, 591–530.
Heckman, J. 1981. The Incidental Parameters Problem and the Problem of Initial Conditions in
Estimating a Discrete TimeDiscrete Data Stochastic Process, in C. Manski and D. McFadden, eds.,
Structural analysis of Discrete Data with Econometric Applications, MIT Press: Cambridge, p. 179196.
Hellmann, T. 2000. Venture Capitalists: The Coaches of Silicon Valley” in The Silicon Valley Edge: A
Habitat for Innovation and Entrepreneurship, eds. W. Miller, C.M. Lee, M.Gong Hanock & H. Rowen,
Stanford University Press
Hochberg, Y., A. Ljungqvist and Y. Lu. 2007. Whom You Know Matters: Venture Capital Networks and
Investment Performance, Journal of Finance, 1: 266301.
Kalnis, A. and Chung. W. 2006. Social Capital, Geography, and the Survival: Gujarati Immigrant
Entrepreneurs in the U.S. Lodging Industry. Management Science 52(2):233247.
Kaplan, S., B. Sensoy and P. Stromberg. 2008. How Well Do Venture Capital Databases Reflect Actual
Investments? University of Chicago, Ohio State University, Swedish Institute for Financial Research,
Working Paper.
Kerr, W. 2008. Ethnic Scientific Communities and International Technology Diffusion. Review of
Economics and Statistics, 90: 518537.
Lerner, J. 1995. Venture Capitalists and the Oversight of Private Firms, Journal of Finance, 50: 301–318.
Morgan, J., Várdy, F. 2009. “Diversity in the Workplace,” American Economic Review, 99(1), 472485.
NVCA. 2009. Venture Impact. The Economic Importance of Venture CapitalBacked Companies to the
US Economy. National Venture Capital Association. Fifth Edition.
Puri, M. and Zarutskie, R. 2012. On Life Cycle Dynamics of VentureCapital and NonVentureCapital
Backed Firms. Journal of Finance, forthcoming.
Rider, C. I., 2012. Employees’ prior affiliations constrain organizational network change: Evidence from
U.S. venture capital and private equity. Forthcoming, Administrative Science Quarterly.
Rosenbaum, Paul R. and Donald B. Rubin. 1985. Constructing a Control Group Using Multivariate
Matched Sampling Methods That Incorporate the Propensity Score. The American Statistician 39:1 33
38.
Samford, M.R. 1953. “Some inequalities on Mill's ratio and related functions,” The Annals of
Mathematical Statistics, 24, 132134.
34
Sinnott, R. W. 1984. The virtues of the Haversine. Sky and Telescope 68 (2): 159.
Sorenson, O., and Stuart, T. 2001. Syndication Networks and the Spatial Distribution of Venture Capital
Investments, American Journal of Sociology, 106: 154688.
Stock, J., Yogo, M., 2005. Testing for weak instruments in IV regression. In Andrews D.W.K. and
Stock, J.H. (Eds.), Identification and Inference for Econometric Models: Essays in Honor of Thomas
Rothenberg, Cambridge University Press, Cambridge, p. 80–108.
Tian, X. 2011. The causes and consequences of venture capital stage financing. Journal of Financial
Economics 101, 132159.
Wang, S., Zhou, H., 2004. Staged financing in venture capital: moral hazard and risks. Journal of
Corporate Finance 10, 131–155.
Webber, R. 2007, Using names to segment customers by cultural, ethnic or religious origin, Journal of
Direct, Data and Digital Marketing Practice, Vol 7, No 3.
35
Tables and Figures
Table 1: Ethnic origins of USbased VC partners and executives of startup companies
The three columns display the percentage of individuals belonging to each of the ten different ethnic origins
(and a miscellaneous “others” category) in the following three samples: (a) the US population, (b) top
executives of USbased startup companies funded by USbased VCs, and (c) USbased VCs. The numbers in
(a) were provided by OriginsInfo Ltd. based on their records of individuals in the US population; (b) is based
on the names of 85,168 toplevel executives at 11,235 USbased companies started between 1991 and 2010
and funded by USbased VCs; and (c) is based on the names of 22,110 partners working at 2,687 USbased
VCs.
Column # 1 2 3
Ethnic group US Overall USbased
executives
USbased VC
partners
ANGLOCELTIC(a) 58.45 52.68 50.47
WEST EUROPEAN(b) 14.3 18.33 18.44
SOUTH EUROPEAN(c) 9.76 7.07 6.95
EAST EUROPEAN(d) 3.32 4.2 4.32
NORTH EUROPEAN(e) 3.2 3.64 3.34
JEWISH 0.99 3.62 3.53
CHINESE 0.74 1.82 2.96
INDIAN 0.66 3.53 3.74
KOREAN 0.37 0.47 1.09
JAPANESE 0.24 0.5 0.95
OTHERS(f) 7.97 4.14 4.22
(a) ANGLOCELTIC includes individuals with origins in England, Australia, Ireland, Scotland and Wales.
(b) WEST EUROPEAN includes individuals with origins in Belgium, Germany, France, the Netherlands and
Switzerland.
(c) SOUTH EUROPEAN includes individuals with origins in Greece, Italy, Portugal and Spain.
(d) EAST EUROPEAN includes individuals with origins in Albania, the Balkans, Bosnia and Herzegovina,
Bulgaria, Croatia, the Czech Republic, Estonia, Hungary, Georgia, Latvia, Poland, Romania, Russia, Serbia,
and Ukraine.
(e) NORTH EUROPEAN includes individuals with origins in Denmark, Finland, Iceland, Norway and
Sweden.
(f) OTHERS is a miscellaneous category and includes individuals with origins in MiddleEastern, South
American, and South Asian countries not captured by the remaining groups.
36
Table 2: Ethnic proximity and probability of VCcompany match
The table compares sample means for the different measures of coethnicity for actual VCcompany pairs
(Column 1), counterfactual VCcompany pairs (Column 2), and the difference between the two (Column 3).
All differences are statistically significant at 95% confidence levels.
Column # 1 2 3
Ethnic group Actual VC
Company pairs
Counterfactual
pairs Difference
COETHNIC ANGLOCELTIC 0.912 0.857 0.055
COETHNIC WEST EUROPEAN 0.566 0.463 0.103
COETHNIC SOUTH EUROPEAN 0.235 0.149 0.086
COETHNIC EAST EUROPEAN 0.114 0.077 0.037
COETHNIC NORTH EUROPEAN 0.103 0.061 0.042
COETHNIC INDIAN 0.098 0.040 0.058
COETHNIC JEWISH 0.091 0.052 0.039
COETHNIC CHINESE 0.041 0.016 0.024
COETHNIC KOREAN 0.007 0.003 0.003
COETHNIC JAPANESE 0.004 0.002 0.002
COETHNIC OTHER 0.114 0.067 0.047
COETHNIC DISTINCT GROUPS(a) 0.466 0.311 0.155
COETHNIC INDISTINCT GROUPS(b) 0.955 0.914 0.041
COETHNIC ALL GROUPS 0.970 0.935 0.035
MAHALANOBIS ETHNIC DISTANCE 10.35 14.15 3.79
(a) For both actual and counterfactual pairs, “COETHNIC DISTINCT GROUPS” = 1 if any of (COETHNIC
SOUTH EUROPEAN, COETHNIC EAST EUROPEAN, COETHNIC NORTH EUROPEAN, COETHNIC
INDIAN, COETHNIC JEWISH, COETHNIC CHINESE, COETHNIC KOREAN, COETHNIC JAPANESE)
= 1
(b) For both actual and counterfactual pairs, “COETHNIC INDISTINCT GROUPS” = 1 if any of
(COETHNIC ANGLOCELTIC, COETHNIC WEST EUROPEAN, COETHNIC OTHER) = 1
37
Table 3: Relationship between ethnic proximity and probability of VCcompany match
Panel A displays Probit estimates and marginal effects derived from Probit estimates (dy/dx) of the relationship between ethnic distance and the
probability that a VC invested in the startup company with which it is paired. A VCcompany pair is the unit of analysis in the regressions. The dependent
variable is set to one for actual VCcompany pairs (i.e. pairs for which the VC invested in the company) and zero for counterfactual VCcompany pairs.
Robust standard errors clustered at the VC level are shown in brackets. Panel B presents the average differences in ethnic proximity between actual VC
company pairs and counterfactual VCcompany pairs after constructing the counterfactual sample through Propensity Score Matching (caliper matching).
We use **, *, and + denote p<0.01, p<0.05 and p<0.1 respectively.
Panel A: Probit regression results
Column # 1 2 3 4 5 6
D.V. = VCCompany match (0/1) Probit dy/dx Probit dy/dx Probit dy/dx
COETHNIC ANGLOCELTIC 0.008 0
[0.031] [0.000]
COETHNIC CHINESE 0.126** 0.0008**
[0.039] [0.000]
COETHNIC EAST EUROPEAN 0.03 0.0002
[0.030] [0.000]
COETHNIC INDIAN 0.148** 0.0009**
[0.030] [0.000]
COETHNIC JAPANESE 0.066 0.0004
[0.099] [0.001]
COETHNIC JEWISH 0.072* 0.0004*
[0.029] [0.000]
COETHNIC KOREAN 0.033 0.0002
[0.079] [0.000]
COETHNIC NORTH EUROPEAN 0.048+ 0.0003+
[0.027] [0.000]
COETHNIC SOUTH EUROPEAN 0.070** 0.0004**
[0.023] [0.000]
COETHNIC WEST EUROPEAN 0.025 0.0001
[0.021] [0.000]
COETHNIC OTHER 0.017 0.0001
[0.025] [0.000]
38
Table 3, Continued
COETHNIC DISTINCT GROUPS 0.074** 0.0004**
[0.017] [0.000]
COETHNIC INDISTINCT GROUPS 0.03 0.0001
[0.037] [0.000]
LOG ETHNIC DISTANCE 0.085** 0.0004**
[0.014] [0.000]
LOG GEOGRAPHIC DISTANCE 0.120** 0.0006** 0.120** 0.0006** 0.120** 0.0006**
[0.004] [0.000] [0.004] [0.000] [0.004] [0.000]
INDUSTRY DISTANCE 0.497** 0.0025** 0.501** 0.0025** 0.502** 0.0025**
[0.042] [0.000] [0.042] [0.000] [0.042] [0.000]
LOG N. OF CO EXECUTIVES 0.077** 0.0004** 0.092** 0.0005** 0.079** 0.0004**
[0.013] [0.000] [0.011] [0.000] [0.012] [0.000]
LOG N. OF VC PARTNERS 0.123** 0.0006** 0.136** 0.0007** 0.127** 0.0006**
[0.013] [0.000] [0.013] [0.000] [0.013] [0.000]
Constant 1.633 1.671 1.235
Co Year Fixed effects Y Y Y
VC Year Fixed effects Y Y Y
Industry Fixed effects Y Y Y
% ethnic personnel in VC & Co Y Y Y
Likelihood ratio chisquare 2936.4 2760.0 2764.0
Prob > Chi2 0 0 0
Observations 1,300,761 1,300,761 1,300,761
Panel B: Propensity Score Matching results
Variable Sample Treated Controls Difference S.E.
COETHNIC DISTINCT GROUPS Unmatched 0.468 0.311 0.157** 0.008
ATT 0.468 0.440 0.027** 0.012
LOG ETHNIC DISTANCE Unmatched 1.989 2.212 0.223** 0.016
ATT 1.989 2.042 0.053** 0.021
39
Table 4: Relationship between ethnic proximity and probability of VCcompany match by funding round
Panel A displays Probit estimates and marginal effects derived from Probit estimates (dy/dx) of the relationship between ethnic distance and the
probability that a VC invested in the startup company with which it is paired separately for the first four rounds of funding. A VCcompany pair is the
unit of analysis in the regressions. The dependent variable is set to one for actual VCcompany pairs (i.e. pairs for which the VC invested in the company)
and zero for counterfactual VCcompany pairs. Robust standard errors clustered at the VC level are shown in brackets. Panel B presents the average
differences in ethnic proximity between actual VCcompany pairs and counterfactual VCcompany pairs after constructing the counterfactual sample
through Propensity Score Matching (caliper matching). We use **, *, and + denote p<0.01, p<0.05 and p<0.1 respectively.
Panel A: Probit regression results
Column # 1 2 3 4 5 6 7 8
D.V. = VCCompany match (0/1) Probit dy/dx Probit dy/dx Probit dy/dx Probit dy/dx
COETHNIC DISTINCT GROUPS 0.082** 0.0003** 0.065** 0.0001** 0.050* 0.0001* 0.029 0.0001
[0.014] [0.000] [0.020] [0.000] [0.022] [0.000] [0.028] [0.000]
LOG GEOGRAPHIC DISTANCE 0.132** 0.0004** 0.095** 0.0002** 0.082** 0.0002** 0.075** 0.0001**
[0.003] [0.000] [0.004] [0.000] [0.005] [0.000] [0.006] [0.000]
INDUSTRY DISTANCE 0.893** 0.0026** 0.617** 0.0013** 0.509** 0.0010** 0.510** 0.0009**
[0.033] [0.000] [0.048] [0.000] [0.059] [0.000] [0.068] [0.000]
LOG N. OF CO EXECUTIVES 0.020* 0.0001* 0.043** 0.0001** 0.113** 0.0002** 0.091** 0.0002**
[0.009] [0.000] [0.013] [0.000] [0.018] [0.000] [0.023] [0.000]
LOG N. OF VC PARTNERS 0.133** 0.0004** 0.112** 0.0002** 0.101** 0.0002** 0.093** 0.0002**
[0.011] [0.000] [0.014] [0.000] [0.015] [0.000] [0.017] [0.000]
Constant 0.891 1.537 2.43 2.291
Co Year Fixed effects Y Y Y Y
VC Year Fixed effects Y Y Y Y
Industry Fixed effects Y Y Y Y
% ethnic personnel in VC & Co Y Y Y Y
Likelihood ratio chisquare 4217.0 1810.5 1080.6 814.7
Prob > Chi2 0 0 0 0
Observations 3,001,809 2,151,494 1,479,842 1,000,557
40
Table 4, Continued
Panel B: Propensity Score Matching (PSM) results
Variable = COETHNIC (DISTINCT
GROUPS) Sample Treated Controls Difference S.E.
Round #1 Unmatched 0.441 0.316 0.124** 0.007
ATT 0.441 0.411 0.029** 0.010
Round #2 Unmatched 0.499 0.352 0.147** 0.011
ATT 0.499 0.467 0.032* 0.016
Round #3 Unmatched 0.504 0.375 0.129** 0.014
ATT 0.504 0.488 0.015 0.020
Round #4 Unmatched 0.499 0.386 0.113** 0.018
ATT 0.499 0.486 0.013 0.026
41
Table 5: Relationship between ethnic proximity and probability of VCcompany match by company lifestage
Table displays Probit estimates and marginal effects derived from Probit estimates (dy/dx) of the relationship between coethnicity and the probability that
a VC invested in the startup company with which it is paired for companies at different life stages during the first round of funding. A VCcompany pair
is the unit of analysis in the regressions. The dependent variable is set to one for actual VCcompany pairs (i.e. pairs for which the VC invested in the
company) and zero for counterfactual VCcompany pairs. Robust standard errors clustered at the VC level are shown in brackets. We use **, *, and +
denote p<0.01, p<0.05 and p<0.1 respectively.
Column # 1 2 3 4 5 6 7 8 9 10
Lifecycle Stage Seed Stage Early Stage Expansion Stage Late Stage Buyout & Acq. Stage
D.V. = VCCompany match (0/1) Probit dy/dx Probit dy/dx Probit dy/dx Probit dy/dx Probit dy/dx
COETHNIC DISTINCT GROUPS 0.111** 0.0003** 0.087** 0.0003** 0.101** 0.0003** 0.074 0.0002 0.019 0.0001
[0.031] [0.000] [0.021] [0.000] [0.031] [0.000] [0.065] [0.000] [0.036] [0.000]
LOG GEOGRAPHIC DISTANCE 0.146** 0.0003** 0.134** 0.0004** 0.122** 0.0004** 0.142** 0.0003** 0.107** 0.0003**
[0.007] [0.000] [0.004] [0.000] [0.005] [0.000] [0.012] [0.000] [0.007] [0.000]
INDUSTRY DISTANCE 0.967** 0.0023** 0.865** 0.0025** 0.809** 0.0026** 0.844** 0.0019** 0.963** 0.0029**
[0.069] [0.000] [0.042] [0.000] [0.057] [0.000] [0.120] [0.000] [0.058] [0.000]
LOG N. OF CO EXECUTIVES 0.023 0.0001 0.019 0.0001 0.025 0.0001 0.084+ 0.0002+ 0.008 0
[0.022] [0.000] [0.012] [0.000] [0.020] [0.000] [0.044] [0.000] [0.016] [0.000]
LOG N. OF VC PARTNERS 0.147** 0.0003** 0.123** 0.0004** 0.108** 0.0003** 0.113** 0.0003** 0.199** 0.0006**
[0.021] [0.000] [0.014] [0.000] [0.018] [0.000] [0.032] [0.000] [0.018] [0.000]
Constant 1.14 0.109 1.085 1.176 1.201
Co Year Fixed effects (19) Y Y Y Y Y
VC Year Fixed effects (39) Y Y Y Y Y
Industry Fixed effects (18) Y Y Y Y Y
% ethnic personnel in VC & Co Y Y Y Y Y
Likelihood ratio chisquare 1573.46 2754.38 1250.45 481.49 1185.67
Prob > Chi2 0 0 0 0 0
Observations 620,085 1,338,428 500,846 101,377 422,198
42
Table 6: Company status and ethnic proximity
Status of companies % companies Mean COETHNIC
(DISTINCT GROUPS)
Mean Ethnic
Distance
Went Public (IPO) 22.24 0.61 7.50
Acquired or Pending acquisition 22.22 0.49 10.30
Private 48.63 0.41 11.95
Merger or LBO 5.88 0.38 12.54
Bankrupt 1.03 0.38 10.25
43
Table 7: Relationship between ethnic proximity and probability of successful exit
Table displays estimates of the relationship between ethnic proximity and the probability that the company exits through acquisitions and IPOs. The
estimation sample consists of actual VCcompany pairs, formed across different rounds of funding (each pair is represented once, regardless of whether
the VC funded the company in multiple rounds) and the dependent variable is set to one if the company exited through an IPO or acquisition, and zero
otherwise. Robust standard errors, clustered at the VC level, are shown in brackets. We use **, *, and + to denote p<0.01, p<0.05 and p<0.1, respectively.
Column # 1 2 3 4 5 6 7
D.V. = IPO+Acquired(0/1) Probit dy/dx Probit dy/dx OLS OLS OLS
COETHNIC ANGLOCELTIC 0.250** 0.099**
[0.054] [0.021]
COETHNIC CHINESE 0.047 0.018
[0.063] [0.025]
COETHNIC EAST EUROPEAN 0.02 0.008
[0.038] [0.015]
COETHNIC INDIAN 0.054 0.021
[0.044] [0.017]
COETHNIC JAPANESE 0.001 0.001
[0.147] [0.058]
COETHNIC JEWISH 0.133** 0.053**
[0.044] [0.017]
COETHNIC KOREAN 0.295** 0.111**
[0.114] [0.041]
COETHNIC NORTH EUROPEAN 0.081+ 0.032+
[0.043] [0.017]
COETHNIC SOUTH EUROPEAN 0.097** 0.038**
[0.032] [0.013]
COETHNIC WEST EUROPEAN 0.061+ 0.024+
[0.033] [0.013]
COETHNIC OTHER 0.027 0.011
[0.042] [0.017]
44
Table 7, Continued
COETHNIC DISTINCT GROUPS 0.079** 0.031** 0.025* 0.028** 0.037**
[0.027] [0.011] [0.010] [0.010] [0.010]
LOG GEOGRAPHIC DISTANCE 0.006 0.002 0.005 0.002 0.004* 0.004* 0.004*
[0.005] [0.002] [0.005] [0.002] [0.002] [0.002] [0.002]
INDUSTRY DISTANCE 0.311** 0.122** 0.310** 0.122** 0.070* 0.080* 0.079*
[0.069] [0.027] [0.068] [0.027] [0.035] [0.036] [0.036]
LOG N. OF CO EXECUTIVES 0.551** 0.216** 0.522** 0.205** 0.163** 0.175** 0.180**
[0.024] [0.010] [0.022] [0.009] [0.007] [0.007] [0.007]
LOG N. OF VC PARTNERS 0.068** 0.027** 0.046** 0.018**
[0.015] [0.006] [0.013] [0.005]
LOG TOTAL FUNDING 0.115** 0.045** 0.115** 0.045** 0.029** 0.029** 0.029**
[0.008] [0.003] [0.008] [0.003] [0.002] [0.002] [0.002]
VC BOARD MEMBER EXISTS 0.020+ 0.016
[0.010] [0.010]
FOUNDER EXISTS 0.092**
[0.009]
Constant 1.102 1.106 0.324 0.146 0.165
Co Year Fixed effects (19) Y Y Y Y Y
VC Year Fixed effects (39) Y Y Y Y Y
VC Fixed effects (2007) N N Y Y Y
Industry Fixed effects (18) Y Y Y Y Y
% ethnic personnel in VC & Co Y Y Y N N
Observations 17,418 17,418 17,418 17,418 17,418
Likelihood ratio chisquare 3362.19 3021.79
Prob > Chi2 0 0
Rsquared 0.289 0.292 0.297
45
Table 8: Relationship between ethnic proximity and probability of successful exit (two
stage estimates)
The table displays estimates of the relationship between ethnic proximity and the probability that the company
exits through acquisitions and IPOs. The estimation sample consists of actual VCcompany pairs, formed
across different rounds of funding and the dependent variable is set to one if the company exited through an
IPO or acquisition, and zero otherwise. Column 1 presents baseline OLS estimates. Column 2 displays 2
Stage Least Squares (2SLS) estimates obtained by using the average of the binary measure of “COETHNIC
DISTINCT GROUPS” for each focal company’s stateindustryfunding year as an instrument for
COETHNIC DISTINCT GROUPS. Column 3 displays 2Stage Least Squares (2SLS) estimates obtained by
using fixed effects for the states, industries, and years, as well as fixed effects for the interactions of state
industry and industryfunding years as instruments for COETHNIC DISTINCT GROUPS. Column 4
presents the secondstage of the Heckman selectioncorrection model. The first stage is estimated with the full
set of explanatory variables and the instrument used for the estimations in Column 2 to satisfy the exclusion
restriction; it uses the 10% sample of possible VCcompany pairs employed in the matching regressions.
Robust standard errors are clustered at the stateindustryfunding year level for the estimates in Column 2 and
3 and at the VC level for the estimates in Columns 1 and 4 and are shown in brackets. We use **, *, and + to
denote p<0.01, p<0.05 and p<0.1, respectively.
Column # 1 2 3 4
D.V. = IPO+Acquired (0/1) OLS 2SLS 2SLS Heckman
COETHNIC DISTINCT GROUPS 0.025* 0.121** 0.169** 0.133**
[0.010] [0.046] [0.065] [0.041]
LOG GEOGRAPHIC DISTANCE 0.004* 0.004* 0.003+ 0.178**
[0.002] [0.002] [0.002] [0.058]
INDUSTRY DISTANCE 0.070* 0.071* 0.068* 0.810**
[0.035] [0.032] [0.031] [0.247]
LOG N. OF CO EXECUTIVES 0.163** 0.142** 0.131** 0.241**
[0.007] [0.012] [0.016] [0.048]
LOG TOTAL FUNDING 0.029** 0.029** 0.029** 0.014**
[0.002] [0.003] [0.003] [0.003]
LOG N. OF VC PARTNERS 0.210**
[0.067]
Inverse Mills Ratio 1.618**
[0.527]
Constant 0.324 0.791 0.958 4.031
Co Year Fixed effects Y Y Y Y
VC Year Fixed effects Y Y Y Y
VC Fixed effects Y Y Y N
Industry Fixed effects Y Y Y Y
% ethnic personnel in VC & Co Y Y Y Y
Observations 17,418 17,418 17,418 3,222
Rsquared 0.285 0.285 0.279 0.213
46
Table 9: Relationship between ethnic proximity and postIPO performance
Table displays Ordinary Least Squares estimates of the relationship between ethnic proximity and postIPO
performance. The estimation sample consists of 2,943 actual VCcompany pairs for companies with data on
market capitalization (dependent variable for the estimations in Column # 1 and 2, expressed in million $) and
1,316 actual VCcompany pairs for companies with data on net income one year after IPO (dependent variable
for the estimations in Column # 3 and 4, expressed in million $). Robust standard errors, clustered at the VC
level, are shown in brackets. We use **, *, and + to denote p<0.01, p<0.05 and p<0.1, respectively.
Column # 1 2 3 4
D.V. Market capitalization Net income
COETHNIC DISTINCT GROUPS 0.091* 0.111* 0.005* 0.009+
[0.041] [0.055] [0.002] [0.005]
LOG GEOGRAPHIC DISTANCE 0.012+ 0.011 0.272 0.897
[0.006] [0.010] [0.389] [0.812]
INDUSTRY DISTANCE 0.302** 0.370+ 9.174 0.658
[0.093] [0.191] [6.723] [20.550]
LOG N. OF CO EXECUTIVES 0.350** 0.311** 2.262 0.204
[0.036] [0.048] [2.066] [3.846]
LOG TOTAL FUNDING 0.113** 0.147** 1.758* 2.498+
[0.019] [0.029] [0.800] [1.384]
LOG N. OF VC PARTNERS 0.036+ 2.201+
[0.019] [1.229]
Constant 1.604 4.154 132.943 15.687
Co Year Fixed effects Y Y Y Y
VC Year Fixed effects Y Y Y Y
VC Fixed effects N Y N Y
Industry Fixed effects Y Y Y Y
% ethnic personnel in VC & Co Y Y Y Y
Observations 2,943 2,943 1,316 1,316
Rsquared 0.326 0.577 0.257 0.553
47
Figure 1: Screening and influence’s role on selecting investments
The figure depicts the densities of signals observed by a VC, conditional on a startup being of a particular
quality, scaled by the prior probability that a startup is of that quality. The upper pair of densities occurs when
startups are socially close to the VCs (i.e.
23
), and the lower density pair occurs when startups are
distant (i.e.
1). The darker densities are those conditional on the startup being high quality scaled by the
prior probability that startups are high quality (i.e. 14
), and the lighter densities are those conditional on the
startup being low quality scaled by the prior probability that startups are low quality (i.e. 134
). Assume
that the VC invests in startups if and only if the probability of success exceeds 13
. First, suppose that post
investment influence (of high quality startups) is 12
, regardless of social distance. Then, the intersection of
the dark and light densities determines the threshold signals. Observe that the VC accepts lower signals from
close startups, but the probability of success and the expected company quality at each threshold signal is
identical. Now assume that the probability of successful postinvestment influence (of high quality startups) is
100% for close startups but only 50% for distant ones. Vertical dashed lines denote the new thresholds—VCs
accept relatively even lower signals from near startups than when influence was independent of social proximity.
But now, although the overall probability of success of startups at these respective thresholds is identical, the
quality of marginally accepted close startups is lower, because the VC anticipates smoother postinvestment
influence for close companies.
48
Figure 2: Histogram of number of VCs funding each company
Figure 3: Histogram of number of funding rounds for each company
0.1 .2 .3 .4
Fraction
0 2 4 6 8 10 12 14 16
Number of VCs funding Company
0.1 .2 .3
Fraction
1 2 3 4 5 6 7 8 9 10
Total number of funding rounds
49
Figure 4: Kernel density estimates of ethnic proximity by company status
Figure 5: Kernel density estimates of actual – predicted performance
0.05 .1 .15 .2
Density
0 5 10 15 20
Mahalanobis Distance
Acquisitions
Mergers
Bankrupt
Private
IPO
kernel = epanechnikov, bandwidth = 0.6708
0.5 11.5
Density
1 .5 0.5 1
Exit performance exceeds expectations (Residuals from exit equation)
COETHNIC (DISTINCT) = 0 COETHNIC DISTINCT = 1
kernel = gaussian, bandwidth = 0.0651
50
Theory Appendix
Definition 1: 1⁄ is the Mill’s ratio.
Lemma 2: The function strictly increases for all .
Proof: Case 0:
′0⇔ 1
′
1
(9)
From Gordon (1941) 1
⁄. From Sampford (1953) 01⁄
′1. Thus,
1
1
and the inequalities in (9) hold.
Case 0 : Clearly ′0, for all 0, as the denominator () is increasing and the numerator
(1Φ) is decreasing. Thus,
′0
because 0, 0, and ′0. ∎
Corollary 1: The function 1 strictly increases for all .
Proof: Taking the derivative,
1
′0
because
0 from Lemma 2 and ′0 is wellknown.
Lemma 3: For all 0 1
11
1
Proof. Observe that
11
11
1′1
11
1
1
0
51
where
. From Sampford (1953) 01⁄
′1, such that the inequality holds when 0.
Furthermore, since hazard rate of the normal distribution is wellknown to increase (i.e. 1⁄
′
0,∀), the inequality must also hold for 0. Since the form of the lemma clearly holds with equality as
→∞, the RHS of the lemma approaches the LHS from above, everywhere. ∎
For notational purposes, define
∗1
and
∗
.
Lemma 4: The probability that an investment is high quality changes with distance according to
′
∗ (10)
where
1
1110
Proof: From the quotient rule
ϕ
1Φ1Φ1ϕ
ϕ1
1Φ1Φ1
1
111
To simplify calculate the derivatives
1
2ln1
′
′
1
2ln1
′
′
Substituting and simplifying yields the form of the lemma. The inequality on follows from ′0. ∎
Proposition 1: Closer investments are more likely to succeed (i.e.
⁄0 ).
Proof: From the product rule
′
52
Substituting
from Lemma 4 and rearranging we can write
⁄0 iff
′1
1
1
∗
Since is increasing (from Lemma 2) the LHS is always negative. Note that ∗, by
definition. Thus, the RHS is positive if 1
1
1
1
This follows immediately from Lemma 3. ∎
(Sub)Model of Effort
The purpose of this section is to show that explicitly modeling effort yields an equilibrium probability that
an undertaken investment succeeds, specified by , such that ′0.
A successful venture investment yields profit to a VC and to a company. An unsuccessful venture
yield profits normalized to 0 to both. Let the probability that the venture succeeds be given by a function of
VC and company effort, ,, increasing, concave and complementary (i.e. 0, or doing one’s
own job is more effective if the other party has done theirs) . The cost of effort for each is given by a
positive distance scalar times respective increasing, convex cost of effort functions and .
Increasing cost of effort in distance may reasonably be thought of as higher communication costs between
socially distant individuals or a host of other similar factors.
The VC and company solve ,∗
and ∗,
respectively. These yield the standard first order conditions (FOCs) for both VC and company, in which
marginal expected benefit equals marginal cost
∗,∗′∗
and ∗,∗′∗
respectively. Now we will show that effort for both parties decreases in distance .
53
Proposition 3: (i) ∗
0. (ii) ∗
0.
Proof: Recall from the Implicit Function Theorem that
∗
∗
∗
∗
∗
∗
∙
1
∆
∗
∗
∗
∗
(11)
where the determinant of the Jacobian matrix is given by
∆
∗
∗
∗
∗ (12)
and the elements of the matrices are given by
∗
∗
∗,∗′∗
∗,∗∗0 (13)
∗
∗
∗,∗′∗
∗,∗0 (14)
∗
∗
∗,∗′∗
∗,∗0 (15)
∗
∗
∗,∗′∗
∗,∗∗0 (16)
∗,∗′∗′∗0 (17)
∗,∗′∗′∗0 (18)
All inequalities follow, because is increasing, concave and complementary, and is increasing and
convex.
We claim that the determinant of the Jacobian is positive (i.e. ∆0). Expanding the terms in (12) according
to their definitions in equations (13) through (18), this is true if and only if
∗,∗∗∗,∗∗∗,∗
which can be rearranged
∗,∗∗,∗
∗,∗
∗,∗∗∗,∗∗∗∗
54
The LHS is positive due to the concavity of G (i.e.
0), and the RHS is negative due to the
concavity of G (i.e. 0 and 0) and the convexity of (i.e. 0). Thus, we have proved our
claim that ∆0.
Finally, then, the proposition holds if and only if the elements of the computed column vector in equation
(11) are positive. Both elements sign positively directly from applying the inequalities in equations (13)
through (18). ∎
Thus, smaller distances increase effort by both parties. Since G increases in both parameters, we know that
the probability of a successful venture decreases as distance grows. In other words, writing the equilibrium
efforts as functions of distance ∗,∗, where′0, as specified in the main model, is a
sufficient statistic for an embedded effort model.
One might model other unobserved features in which postinvestment success, such as monitoring
effectiveness or coordination, depend on (ethnic) distance to similar effect. The details of the various
extensions would differ slightly, but the comparative static on the equilibrium probability of success with
respect to distance would also be negative—our current formulation captures this essential feature
generally. The basic intuition is that if social distance interferes with one of the activities positively related
to the success of the venture or makes it more costly to engage in, the less of the activity will occur and the
probability of success will decrease.
55
Data and Methods Appendix (Additional descriptives and robustness checks)
Table A1: Most common ethnic family names for USbased company executives
Ethnic group Last name Frequency % of last names
within group Ethnic group Last name Frequency % of last names
within group
ANGLOCELTIC SMITH 91 0.88
JEWISH COHEN
25 3.41
ANGLOCELTIC MILLER 65 0.63
JEWISH FRIEDMAN
23 3.13
ANGLOCELTIC LEE 51 0.49
JEWISH GOLDSTEIN
20 2.72
ANGLOCELTIC DAVIS 50 0.48
JEWISH LEVY
17 2.32
ANGLOCELTIC JONES 49 0.47
JEWISH KATZ
14 1.91
ANGLOCELTIC BROWN 45 0.44
JEWISH GOLDBERG
14 1.91
ANGLOCELTIC ANDERSON 40 0.39
JEWISH ROSEN
14 1.91
ANGLOCELTIC WILLIAMS 39 0.38
JEWISH GOLDMAN
13 1.77
ANGLOCELTIC YOUNG 34 0.33
JEWISH SHAPIRO
11 1.50
ANGLOCELTIC CLARK 34 0.33
JEWISH KAPLAN
10 1.36
CHINESE WANG 41 6.67
KOREAN KIM
43 19.20
CHINESE CHEN 38 6.18
KOREAN LEE 10 4.46
CHINESE CHANG 27 4.39
KOREAN CHOI
9 4.02
CHINESE LIU 24 3.90
KOREAN HWANG
8 3.57
CHINESE WU 24 3.90
KOREAN HAN
8 3.57
CHINESE WONG 23 3.74
KOREAN CHIN
7 3.13
CHINESE LIN 22 3.58
KOREAN CHO
7 3.13
CHINESE LI 19 3.09
KOREAN SONG
7 3.13
CHINESE HSU 12 1.95
KOREAN YOON
7 3.13
CHINESE ZHANG 12 1.95
KOREAN YI 5 2.23
EAST EUROPEAN NOWAK 3 0.32
NORTH EUROPEAN JOHNSON 57 8.30
EAST EUROPEAN ABRAMSON 3 0.32
NORTH EUROPEAN JACOBSON 12 1.75
EAST EUROPEAN PECK 3 0.32
NORTH EUROPEAN FRANK 11 1.60
EAST EUROPEAN BRODSKY 3 0.32
NORTH EUROPEAN PETERSON 9 1.31
EAST EUROPEAN SKOK 3 0.32
NORTH EUROPEAN ROTH 9 1.31
EAST EUROPEAN ZAK 3 0.32
NORTH EUROPEAN HANSEN 9 1.31
EAST EUROPEAN KRASNOW 3 0.32
NORTH EUROPEAN JENSEN 8 1.16
EAST EUROPEAN HURWITZ 3 0.32
NORTH EUROPEAN LARSON 7 1.02
EAST EUROPEAN SAGAN 2 0.22
NORTH EUROPEAN PETERSEN 6 0.87
EAST EUROPEAN SCHIFTER 2 0.22
NORTH EUROPEAN WALL 5 0.73
56
Table A1, Continued
Ethnic group Last name Frequency % of last names
within group Ethnic group Last name Frequency % of last names
within group
INDIAN SHAH 28 3.58
SOUTH EUROPEAN MARINO 7 0.47
INDIAN PATEL
27 3.45
SOUTH EUROPEAN GARCIA 7 0.47
INDIAN GUPTA
23 2.94
SOUTH EUROPEAN FERNANDEZ 6 0.41
INDIAN DESAI 12 1.53
SOUTH EUROPEAN RODRIGUEZ 6 0.41
INDIAN SINGH 9 1.15
SOUTH EUROPEAN OLIVER 6 0.41
INDIAN MEHTA
9 1.15
SOUTH EUROPEAN FERRARI 5 0.34
INDIAN SHARMA
9 1.15
SOUTH EUROPEAN LOPEZ 5 0.34
INDIAN JAIN 8 1.02
SOUTH EUROPEAN RUIZ 5 0.34
INDIAN KHANNA
6 0.77
SOUTH EUROPEAN PEREZ 5 0.34
INDIAN MEHRA
6 0.77
SOUTH EUROPEAN ROSSI 4 0.27
JAPANESE WATANABE
6 2.94
WEST EUROPEAN WEISS 22 0.57
JAPANESE YAMAMOTO
3 1.47
WEST EUROPEAN SCHWARTZ 18 0.47
JAPANESE MATSUMOTO
3 1.47
WEST EUROPEAN BECKER 16 0.42
JAPANESE YANO 2 0.98
WEST EUROPEAN KLEIN 14 0.36
JAPANESE SASAKI 2 0.98
WEST EUROPEAN ROSE 14 0.36
JAPANESE SAKAI 2 0.98
WEST EUROPEAN WAGNER 13 0.34
JAPANESE YAMADA 2 0.98
WEST EUROPEAN MARKS 13 0.34
JAPANESE SAITO 2 0.98
WEST EUROPEAN MEYER 13 0.34
JAPANESE FUJII 2 0.98
WEST EUROPEAN SIEGEL 12 0.31
JAPANESE KIKUCHI 2 0.98
WEST EUROPEAN STERN 12 0.31
OTHER KHAN 9 1.00
OTHER NGUYEN
7 0.78
OTHER HERNANDEZ
7 0.78
OTHER AHMED
6 0.67
OTHER LAU 5 0.56
OTHER CAO 5 0.56
OTHER TORRES
5 0.56
OTHER SOTO 4 0.45
OTHER ORTIZ 4 0.45
OTHER AMIN 4 0.45
57
Table A2: Industry distribution of portfolio companies funded by USbased VCs
Table shows the industry distribution of the 11,235 USbased startup companies funded by USbased VCs and
started during the years 19912010 in our sample.
Industry category N of Cos % of Cos
Internet Specific 2,438 21.7
Computer Software 2,329 20.73
Medical/Health 1,430 12.73
Communications 919 8.18
Biotechnology 785 6.99
Semiconductor/Electronics 725 6.45
Industrial/Energy 568 5.06
Consumer Related 539 4.8
Computer Hardware 401 3.57
Financial Services 315 2.8
Business Serv. 310 2.76
Transportation 137 1.22
Other 129 1.15
Manufacturing 118 1.05
Construction 41 0.36
Computer Other 20 0.18
Utilities 16 0.14
Agriculture/Forestry/Fisheries 15 0.13
TOTAL 11,235
58
Table A3: Most common ethnic family names for USbased VC partners
Ethnic group Last name Frequency % of last names
within group Ethnic group Last name Frequency % of last names
within group
ANGLOCELTIC SMITH 91 0.88
JEWISH COHEN
25 3.41
ANGLOCELTIC MILLER 65 0.63
JEWISH FRIEDMAN
23 3.13
ANGLOCELTIC LEE 51 0.49
JEWISH GOLDSTEIN
20 2.72
ANGLOCELTIC DAVIS 50 0.48
JEWISH LEVY
17 2.32
ANGLOCELTIC JONES 49 0.47
JEWISH KATZ
14 1.91
ANGLOCELTIC BROWN 45 0.44
JEWISH GOLDBERG
14 1.91
ANGLOCELTIC ANDERSON 40 0.39
JEWISH ROSEN
14 1.91
ANGLOCELTIC WILLIAMS 39 0.38
JEWISH GOLDMAN
13 1.77
ANGLOCELTIC YOUNG 34 0.33
JEWISH SHAPIRO
11 1.50
ANGLOCELTIC CLARK 34 0.33
JEWISH KAPLAN
10 1.36
CHINESE WANG 41 6.67
KOREAN KIM
43 19.20
CHINESE CHEN 38 6.18
KOREAN LEE 10 4.46
CHINESE CHANG 27 4.39
KOREAN CHOI
9 4.02
CHINESE LIU 24 3.90
KOREAN HWANG
8 3.57
CHINESE WU 24 3.90
KOREAN HAN
8 3.57
CHINESE WONG 23 3.74
KOREAN CHIN
7 3.13
CHINESE LIN 22 3.58
KOREAN CHO
7 3.13
CHINESE LI 19 3.09
KOREAN SONG
7 3.13
CHINESE HSU 12 1.95
KOREAN YOON
7 3.13
CHINESE ZHANG 12 1.95
KOREAN YI 5 2.23
EAST EUROPEAN NOWAK 3 0.32
NORTH EUROPEAN JOHNSON 57 8.30
EAST EUROPEAN ABRAMSON 3 0.32
NORTH EUROPEAN JACOBSON 12 1.75
EAST EUROPEAN PECK 3 0.32
NORTH EUROPEAN FRANK 11 1.60
EAST EUROPEAN BRODSKY 3 0.32
NORTH EUROPEAN PETERSON 9 1.31
EAST EUROPEAN SKOK 3 0.32
NORTH EUROPEAN ROTH 9 1.31
EAST EUROPEAN ZAK 3 0.32
NORTH EUROPEAN HANSEN 9 1.31
EAST EUROPEAN KRASNOW 3 0.32
NORTH EUROPEAN JENSEN 8 1.16
EAST EUROPEAN HURWITZ 3 0.32
NORTH EUROPEAN LARSON 7 1.02
EAST EUROPEAN SAGAN 2 0.22
NORTH EUROPEAN PETERSEN 6 0.87
EAST EUROPEAN SCHIFTER 2 0.22
NORTH EUROPEAN WALL 5 0.73
59
Table A3, Continued
Ethnic group Last name Frequency % of last names
within group Ethnic group Last name Frequency % of last names
within group
INDIAN SHAH 28 3.58
SOUTH EUROPEAN MARINO 7 0.47
INDIAN PATEL
27 3.45
SOUTH EUROPEAN GARCIA 7 0.47
INDIAN GUPTA
23 2.94
SOUTH EUROPEAN FERNANDEZ 6 0.41
INDIAN DESAI 12 1.53
SOUTH EUROPEAN RODRIGUEZ 6 0.41
INDIAN SINGH 9 1.15
SOUTH EUROPEAN OLIVER 6 0.41
INDIAN MEHTA
9 1.15
SOUTH EUROPEAN FERRARI 5 0.34
INDIAN SHARMA
9 1.15
SOUTH EUROPEAN LOPEZ 5 0.34
INDIAN JAIN 8 1.02
SOUTH EUROPEAN RUIZ 5 0.34
INDIAN KHANNA
6 0.77
SOUTH EUROPEAN PEREZ 5 0.34
INDIAN MEHRA
6 0.77
SOUTH EUROPEAN ROSSI 4 0.27
JAPANESE WATANABE
6 2.94
WEST EUROPEAN WEISS 22 0.57
JAPANESE YAMAMOTO
3 1.47
WEST EUROPEAN SCHWARTZ 18 0.47
JAPANESE MATSUMOTO
3 1.47
WEST EUROPEAN BECKER 16 0.42
JAPANESE YANO 2 0.98
WEST EUROPEAN KLEIN 14 0.36
JAPANESE SASAKI 2 0.98
WEST EUROPEAN ROSE 14 0.36
JAPANESE SAKAI 2 0.98
WEST EUROPEAN WAGNER 13 0.34
JAPANESE YAMADA 2 0.98
WEST EUROPEAN MARKS 13 0.34
JAPANESE SAITO 2 0.98
WEST EUROPEAN MEYER 13 0.34
JAPANESE FUJII 2 0.98
WEST EUROPEAN SIEGEL 12 0.31
JAPANESE KIKUCHI 2 0.98
WEST EUROPEAN STERN 12 0.31
OTHER KHAN 9 1.00
OTHER NGUYEN
7 0.78
OTHER HERNANDEZ
7 0.78
OTHER AHMED
6 0.67
OTHER LAU 5 0.56
OTHER CAO 5 0.56
OTHER TORRES
5 0.56
OTHER SOTO 4 0.45
OTHER ORTIZ 4 0.45
OTHER AMIN 4 0.45
60
Table A4: Checking for Balancing, Propensity Score Matching
Table reports results of tests that check whether the covariates used to generate our propensity scores are, in fact, balanced across the treated and non
treated groups. We carried out the tests for the full set of variables used in our Logit regressions to calculate propensity scores, that is the set of
observable VC characteristics, company characteristics, and VCcompany pair characteristics (except ethnic proximity) described in Section 3.2.
However, we do not report the results for covariates such as VC and Company founding years, and industry dummies for the sake of brevity. We
find that the % bias after matching was less than 5% for all the covariates, and the corresponding tstatistics were unable to reject the null
hypothesis for similarity between the treated and control groups for all the covariates. These results confirm that our matching was quite effective
at constructing an effective control group.
Variable Sample
Mean
%reduction
in bias
ttest
Treated Control %bias t p>t
Log geographic distance Unmatched 5.18 6.61 69.6 52.12 0.00
Matched 5.18 5.16 1 98.6 0.32 0.75
Industry distance Unmatched 0.77 0.79 8.4 4.70 0.00
Matched 0.77 0.77 1.2 86.1 0.44 0.66
Log # Co executives Unmatched 2.01 1.83 25.5 13.85 0.00
Matched 2.01 2.00 2 92.1 0.86 0.39
Log # VC partners Unmatched 2.34 1.89 44.8 26.00 0.00
Matched 2.34 2.36 2.5 94.3 0.98 0.33
61
Table A5: Relationship between ethnic proximity and probability of successful exit through IPOs
Table displays estimates of the relationship between coethnicity and the probability that the company exits successfully either through IPOs and “good”
acquisitions or through IPOs alone. “Good acquisitions” are defined as those for which the transaction value of the acquisition (as reported by SDC
Platinum’s Merger and Acquisitions database) exceeded total VC investments. By this definition, 69% of the acquisitions in our sample were “good.” All
specifications are OLS VCfixed effects regressions and should be compared against the estimates in Columns 57 of Table 7. Robust standard errors,
clustered at the VC level, are shown in brackets. We use **, *, and + to denote p<0.01, p<0.05 and p<0.1, respectively.
Column # 1 2 3 4 5 6
D.V. = IPO+Acquired(0/1) IPO+"Good"