Venture Capital Communities1

Amit Bubna Sanjiv R. Das N. R. Prabhala

ISB, Hyderabad Santa Clara University University of Maryland

April 21, 2012

1 Preliminary draft, comments welcome. Please do not quote without permission. We thank
Pete Kyle, Jiekun Huang, Vladimir Ivanov, Laura Lindsey, Robert Marquez, Vojislav Maksimovic,
Manju Puri, Krishna Ramaswamy, Rajdeep Singh, Richard Smith, Anjan Thakor, Susan Woodward
and participants at the 2011 CAF conference, the 2011 FIRS conference, the 2011 Private Equity
conference, as well as seminar participants at ISB and the University of Maryland for helpful
comments. We especially thank Josh Lerner for extensive comments on a previous draft. The
authors may be reached at their respective email addresses: amit bubna@isb.edu, srdas@scu.edu,
and prabhala@umd.edu.
 Abstract

Venture Capital Communities

Syndication accounts for two-thirds of venture capital (VC) investment and is a robust
predictor of successful exit. We provide new evidence on the composition of VC syndicates
and report three main findings. First, we show that VCs neither pick partners at random nor
do they syndicate with a fixed set of partners. Rather, they have probabilistic preferences for
some partners over others, which leads to spatial clusters, “communities,” whose members
tend to syndicate with each other though not exclusively so. We characterize the number,
composition, and stability of VC communities. Second, we show that there are elements of
both assortative and disassortative matching in syndicate partner choice. Partners are similar
along dimensions of functional style but dissimilar in preferred partner size, influence, and
reach. Finally, firms sourcing capital from community VCs are more likely to exit successfully.
The results are consistent with models in which VCs learn by doing in which familiarity
with syndicate partners aids learning, or syndication models in which familiarity mitigates
incomplete contracting problems of hold up or free riding by facilitating trust, reciprocity,
and information flows.

Key words: Venture capital, syndication, boundaries of the firm, social networks
JEL classification: G20, G24
1 Introduction

Venture capitalists (VCs) raise capital from wealthy individuals and institutional investors
and invest in young firms that promise high upside. The venture capital industry has grown
significantly since its institutionalization in 1958. According to the National Venture Capital
Association, there are over 56,000 VC deals for $429 billion in the U.S. between 1995 and
2009. VC success stories include prominent firms such as Apple, Cisco, and Microsoft.

VC investing is risky. Firms receiving venture capital tend to be young and immature,
often with unproven business models. VC investing is also resource intensive, both in terms
of ex-ante screening and follow on support after investing. As Gompers and Lerner (2001,
2004) point out, VCs often contact over 100 references before making an investment. After
investing, VCs serve as directors, conduct site visits, provide strategic advice, help recruit
key personnel, or find partners for their portfolio companies.1

VCs use multiple strategies to manage the risks and resource demands of their investment
process. For instance, their investment contracts incorporate security design elements such
as priority or staging, and carefully lay out control rights over portfolio firms.2 A key
strategy used by VCs is syndication, or co-investment in portfolio firms together with other
venture capital investors. Syndication is economically important. In the U.S., syndicated
deals account for 44% of the number and 66% of VC investment proceeds. Only 5% of VCs
never syndicate and they are small, peripheral players.

Our study focuses on the composition of VC syndicates. To motivate our analysis, ob-
serve that any one syndication round involves partnership between a small number of venture
capitalists. As a VC does more deals, its list of syndicate partners grows. A venture capital-
ist often manages multiple, simultaneous, and only partially overlapping syndications with
several other VCs, much as in inter-firm alliances studied by Robinson and Stuart (2007)
and Robinson (2008). We examine the forces that drive the composition of syndicates. Do
1
See Lerner (1995) for directorships, Gorman and Sahlman (1989) for advising, Hellmann and Puri (2002)
for professionalization, and Lindsey (2008) for strategic partnering roles of VC firms. See Da Rin, Hellmann,
and Puri (2012) for a recent survey of VC research.
2
See Cornelli and Yosha (2003) or Neher (1999) on security design, or Kaplan and Stromberg (2003,
2004), Robinson and Stuart (2007) and Robinson and Sensoy (2011) for evidence on VC contracts.

1
venture capitalists select syndicate partners randomly? If not, what kind of VCs tend to syn-
dicate with each other and what are the economic motivations for these choices? Are there
performance consequences for portfolio firms? We present new evidence on these issues.

Briefly, we report three key findings. First, we find that VCs do not pick syndicate
partners at random. Rather, VCs have pronounced preferences for some partners over others,
resulting in the clustering of VCs into spatial agglomerates or “communities.” Community
members are more likely to syndicate with each other than with outsiders. We characterize
the number, size, and stability of VC communities. Second, we find that VC syndication
reflects both assortative and disassortative matching. On some dimensions, VCs prefer similar
VCs but on others they prefer dissimilar VCs as syndicate partners. Thus, we reconcile the
apparent contradiction between the like-prefer-like (homophily) findings of Cestone, Lerner,
and White (2006) and Du (2011) and the diversity evidence of Hochberg, Lindsey, and
Westerfield (2011). Finally, we find that firms sourcing capital from community VCs are more
likely to have successful exits. Thus, economic motivation rather than a passive preference for
the familiar drives VC preferences to syndicate with select partners. While we confirm prior
evidence that syndication is beneficial, we show that syndicate composition also matters.3

To illustrate our main point, Figure 1 depicts the frequency distribution of the syndicate
partners of J. P. Morgan Ventures. Between 1980 and 1999, the firm had 640 different
syndicate partners. The long and thin right tail in Figure 1 shows that the firm has many
partners. However, there is a thick left mass as well, indicating that the firm clearly prefers
some syndicate partners over others. Examples of J. P. Morgan’s preferred partners include
Kleiner Perkins, Oak Investment Partners, and the Mayfield Fund. Figure 2 shows similar
patterns for other venture capitalists such as Matrix Partners, Sequoia Capital, and Kleiner
Perkins. Each of these firms syndicates with several partners but exhibits preference for some
partners over others. This preference is the basic ingredient for forming “VC communities.”

Before discussing the results, we briefly consider why VCs may self-organize into clusters of
communities containing preferred syndicate partners. One motivation comes from the view
3
See Brander, Amit and Antweiler (2002), Lerner (1994), Cestone, Lerner and White (2006), Sorensen
(2007) and Das, Jo and Kim (2011).

2
that VCs acquire skills through learning by doing (Goldfarb, Kirsch, and Miller (2007),
Sorensen (2008)). Success in VC investing demands considerable skill. While some skills are
endowed, others can only be learned hands-on through investing because VC-funded firms
often have unfamiliar, emerging business models. Community formation is consistent with
the joint hypothesis that VCs learn from investing and that learning is enhanced with (pre-
ferred) familiar partners. The importance of familiarity is outlined by Gompers and Lerner
(2004), who write that VCs will not invest unless acceptable syndicate partners agree that a
project is desirable. Familiarity enhances information flows due to better understanding of
partners’ norms and processes (Gertler (1995); Porter (2000)).

Sociologists also emphasize a role for familiarity. In an influential article, Granovetter

(1985) points out that agents place far greater confidence in information flows from trusted
sources. Thus, VCs place more faith in assessments of the syndicate partners they are familiar
with. Finally, incomplete contracting theories also predict a preference for familiar partners.
In models such as Grossman and Hart (1986) or Hart and Moore (1990), the suspicion that
partners will free ride or hold up initial investors results in underinvestment. Syndicating
with familiar partners mitigates these problems by reducing information asymmetry between
syndicate partners, or by building trust and enhancing reciprocity (Guiso, Sapienza, and
Zingales (2004), Bottazzi, Da Rin and Hellmann (2011)). Familiarity can be temporal,
based on past dealings with a partner, or spatial, when the potential partner has transacted
with a VC’s network of former partners.4 Either force leads to clusters of VCs who tend to
syndicate with each other.

We analyze syndication data based on U.S. venture financing deals between 1980 and
1999. The sample comprises 33,924 unique VC investment rounds in 15,455 portfolio firms.
The beginning of our sample is after the enactment of the Employee Retirement Income
Security Act, which led to the institutionalization of the VC market. The sample ends in
1999, which allows sufficient time for judging the performance of VC-funded portfolio firms.
4
For example, the VC firm Matrix Partners writes in its website “... The best way to get
in touch with our team is through an introduction from someone you know in our network.”
http://matrixpartners.com/site/about partnering-with-matrix, accessed May 3, 2011. The founder of
LinkedIn, Reid Hoffman, makes a similar point in his autobiography.

3
As in Hochberg, Ljungqvist, and Lu (2007), we employ rolling windows of 5 years. Thus, a
first set of communities is based on syndications between 1980 and 1984, the next set uses
venture financings between 1981 and 1985, and so on. We obtain exit data from the SDC
Platinum M&A database for mergers and the new issues database for IPO exits.

Communities are spatial agglomerates of VCs who are more likely to syndicate with
each other than with others. Identifying these agglomerates is a cluster detection problem,
a computationally difficult (NP-hard) one. The computational complexity arises because
we do not fix the number of clusters, allow each cluster to have a different size, and permit
porous boundaries so members can syndicate both within and outside their preferred clusters.
Several partitions satisfy such a flexible definition. Communities are identified by optimizing
modularity, or the difference between the number of within-community syndications and the
expected number of syndications given the deal flow. Identifying communities is difficult
because of the enormous computational complexity involved in sifting through all possible
partitions. We exploit the walk trap approach recently developed in the physical sciences
(Pons and Latapy (2005), see Fortunato (2009) for a review) to identify communities.

We report three major sets of results. First, we detect several communities in each 5-year
sub-period in the sample. There are between 12 and 35 communities in different periods with
a median community size of 13. About 20% of venture capitalists in any one period cluster
into communities. These clusters are stable. The pairwise similarity between communities
in one period and the next far exceeds the similarity between bootstrapped communities at
the 1% level. The average VC in a community is 16 times as likely to syndicate within the
community than with outside VCs. Thus, VC communities are both stable and tight-knit.

We next consider the nature of a VC’s preferred partners. What type of firms prefer
to syndicate with each other? One view is that VC syndicates should reflect disassortative
matching. VC clusters may include firms with heterogeneous skill sets, making a broad set of
skills and resources readily available to VC syndicates. On the other hand, VCs may exhibit
a preference for the similar. This can reflect the greater reliance of VCs upon the judgments
of similar VCs, which leads to clusters of similar VCs, reflecting assortative matching (Lerner,
1994; Cestone, Lerner, and White, 2006). Alternatively, similarity could be driven by pure

4
behavioral preferences for the similar (McPherson, Smith-Lovin and Cook (2001)).

The received VC literature reports contradictory results on assortative versus disassorta-

tive matching. Lerner (1994), Cestone, Lerner, and White (2006), and Du (2011) support
the like-prefer-like view while Hochberg, Lindsey, and Westerfield (2011) argue that VCs
prefer dissimilar partners. We examine this issue through the lens of communities. By def-
inition, communities reflect agglomerates of VCs who prefer to syndicate with each other.
By comparing variation in a characteristic within communities to variation within random
bootstrapped communities, we can test whether VCs prefer to syndicate with similar or dis-
similar partners along the characteristic. We find rather subtle patterns in matching that
reflect both assortative and disassortative matching. While there is similarity along style
dimensions such as industry or stage focus, we find heterogeneity on dimensions such as VC
size, influence, and geographical focus.

To determine whether the preference for the familiar has economic motivation, we in-
vestigate the performance of firms funded by venture capitalists belonging to communities.
Following the VC literature (e.g., Brander, Amit and Antweiler (2002) or Sorensen (2008)),
we define success as exit via an IPO or merger. Firms funded by community VCs are more
successful in logit models predicting 10-year exits, Cox hazard models, or competing hazards
models. The results are also robust to several controls. These include measures of VC repu-
tation such as age or funds managed by a VC, a VC’s eigenvector centrality, industry fixed
effects, ownership structure of VCs, or their stage and industry focus. We emphasize that
we control for whether a deal is syndicated or not. We do find that syndication matters.
Our new point is not about syndication per se, but about syndicate composition, and the
additional beneficial effects from the interactions with familiar syndicate partners.

The rest of the paper is organized as follows. Section 2 formally develops the community
detection problem, reviews the related work, and places it in the context of the VC liter-
ature on networks. Section 3 discusses the data. Sections 4 and 5 present the results on
performance and community composition. Section 6 concludes.

5
2 Communities

2.1 Basic Idea and Applications

The essential idea underlying community formation dates back at least to Simon (1962). In
his view, complex systems comprising several entities often have coherent subsystems, or
communities, that serve specific functional purposes. Identifying communities embedded in
larger entities can help understand the functional forces underlying larger entities. To make
these ideas more concrete, we review applications of community detection in the physical
and social sciences.

In the life sciences, community structures help understand pathways in the metabolic
networks of cellular organisms (Ravasz et al., 2002; Guimera and Amaral, 2005). Community
structures also help understand the functioning of the human brain. For instance, Wu et
al. (2011) find that there are community structures in the human brain with predictable
changes in their interlinkages related to aging. Community structures are used to understand
how food chains are compartmentalized, which can predict the robustness of ecosystems to
shocks that endanger particular species (Girvan and Newman, 2002). Lusseau (2003) finds
that communities are evolutionary hedges that avoid isolation when a member is attacked
by predators. In political science, community structures discerned from voting patterns can
detect political preferences that transcend traditional party lines (Porter, Mucha, Newman,
and Friend, 2007).5

Fortunato (2009) presents a relatively recent and thorough survey of the research in com-
munity detection. Fortunato points out that while the computational issues are challenging,
there is sufficient progress to the point where many methods yield similar results in practice.
However, there are fewer insights on the functional roles of communities or their quantita-
tive effect on outcomes of interest. Fortunato suggests that this is a key challenge in the
literature. As he concludes “... What shall we do with communities? What can they tell
5
Other topics studied include social interactions and community formation (Zachary, 1977), word ad-
jacency in linguistics and cognitive sciences (Newman, 2006), collaborations between scientists (Newman,
2001), and industry structures from product descriptions (Hoberg and Phillips, 2010). For some community
detection datasets, see Mark Newman’s website http://www-personal.umich.edu/ mejn/netdata/.

6
us about a system? This is the main question beneath the whole endeavor.” This is an
area in which we offer progress. We show that community detection methods provide useful
insights into the economics of VC process. For instance, we gain insights into the process
by which VCs select syndicate partners, and find beneficial economic effects associated with
VC communities.

2.2 Defining and Detecting Communities

We turn to the formal definition of communities and discuss how we can detect them from VC
syndication data. As in Hochberg, Ljungqvist, and Lu (2007), we represent VC syndicated
deals as the square adjacency matrix A. The rows and columns represent venture capitalists.
Element A(i, j) equals the number of syndicate rounds in which VC i and j are partners,
so more frequent partnerships lead to greater weights. The diagonal element A(i, i) is zero.
While this representation is standard in the networks literature, it has economic content.
The implicit assumption is that venture capitalists get no benefit of learning from syndi-
cate partners when they deal with themselves. The matrix is undirected and symmetric,
effectively assuming that the benefits of syndicate interactions flow to all members.

Community detection methods partition VCs into clusters that tend to work together.
It is useful to point out the considerable flexibility and realism built into the definition
of our community clusters. We do not require all VCs to belong to communities. Nor
do we fix the number of communities that may exist at a time, and we also allow each
community to have different size. With this flexibility, the key computational challenge
is to find the “best” partition because the number of possible partitions of the VCs is
extremely large. Community detection methods attempt to determine a set of clusters that
are internally tight-knit. Mathematically, this is equivalent to finding a partition of clusters
to maximize the observed number of syndications between cluster members minus what is
expected conditional on the deal flow within the cluster, aggregated across all clusters. More
formally (see, e.g., Newman (2006)), we choose partitions with high modularity Q, where

" #
1 X di × dj
Q= Aij − · δ(i, j) (1)
2m i,j 2m

7
 In equation (1), Aij is the (i, j)-th entry in the adjacency matrix, i.e., the number of
P
syndicate transactions in which VC firm i and j jointly participated, di = j Aij is the total
1 P
number of syndications that VC firm i participated in (or, the degree of i) and m = 2 ij Aij

is the sum of all edge weights in matrix A. The function δ(i, j) is an indicator equal to 1.0 if
nodes i and j are from the same community, and zero otherwise. Q is bounded in [-1, +1].
If Q > 0, intra-community connections exceed the expected number given deal flow.

Community detection is an NP-hard problem for which there are no known exact solutions
beyond tiny systems (Fortunato, 2009). For larger datasets, one approach is to impose
numerical constraints. For example, graph partitioning imposes a uniform community size,
while partitional clustering presets the number of communities. As there is no reason to
believe that VC communities are of a particular size or that the number of VC community
clusters should be fixed, we do not use these approaches.

The less restrictive methods for community detection are called hierarchical partitioning
methods, which are “divisive,” or “agglomerative.” The former is a top-down approach that
assumes that the entire graph is one community and breaks it down into smaller units. It
often produces communities that are too large especially when there is not an extremely
strong community structure. Agglomerative algorithms, like the “walktrap” technique we
use, begin by assuming all nodes are separate communities and collect nodes into commu-
nities. The fast techniques are dynamic methods based on random walks, whose intuition
is that if a random walk enters a strong community, it is likely to spend a long time inside
before finding a way out (Pons and Latapy (2005)).6

2.3 Communities, Centrality, and Social Network Analysis

Community detection forms part of the literature on social network analysis. The starting
point for this work is a set of pairwise connections between individuals or firms, which
has received much attention in the recent finance literature. Cohen, Frazzini, and Malloy
(2008a, 2008b) analyze educational connections between sell-side analysts and managers.
6
See Girvan and Newman (2002), Leskovec, Kang and Mahoney (2010), or Fortunato (2009) and the
references therein for a discussion. Appendix A gives an example and code for modularity optimization.

8
Hwang and Kim (2009) and Chidambaran, Kedia, and Prabhala (2010) analyze educational,
employment, and other links between CEOs and directors. Pairwise inter-firm relations are
analyzed by Ishii and Xuan (2009) and Cai and Sevilir (2012) while VC firm connections
with founders and top executives are studied by Bengtsson and Hsu (2010) and Hegde and
Tumlinson (2011).

Our study is more closely tied to the finance work on the aggregate connectedness de-
rived from pairwise connections. These metrics are introduced to the finance literature by
Hochberg, Ljungqvist and Lu (2007), who study the aggregate connections of venture capi-
talists derived through syndications. They show that firms financed by well-connected VCs
are more likely to exit successfully. Engelberg, Gao and Parsons (2000) show that highly
connected CEOs are more highly compensated. While we control for the aggregate con-
nectedness metrics of VCs, our community measure is quite different both in terms of its
mathematical definition and more importantly, in the economic intuition it is intended to
capture. We expand on this point before getting to the data and results.

The simplest measure of aggregate connectedness, degree centrality, simply aggregates

the number of syndicate partners that a VC has worked with. A more subtle measure,
eigenvector centrality, aggregates connections but puts more weight on the connections of
VCs to more connected syndicate partners. Other related constructs are betweenness, which
reflects how many times a VC is on the shortest path between two other VCs, and closeness,
which measures a VC’s distance to all other VCs. The important point is that each of these
measures represents an attempt to capture a VC’s stature or influence as reflected in the
number of its own connections or from being connected to well-connected VCs.

Community membership, on the other hand, is a group attribute that reflects whether a
VC belongs to a spatial cluster of firms that tend to do business together. Community mem-
bership is a variable inherited by all members of a spatial agglomerate. However, centrality
is an individual-centered variable that captures a VC’s influence. Community membership
does not measure the reach or influence of a VC. Rather, it is a measure focused on interac-
tions between VCs, reflecting whether a VC deals with familiar partners. Neither community
membership nor centrality is a proper subset of the other.

9
 The differences between community and centrality are visually depicted in Figure 3, which
is reproduced from Burdick et al. (2011). The figure shows the high centrality of Citigroup,
J. P. Morgan, and Bank of America, well connected banks in co-lending networks. However,
none of these banks belong to communities, which are represented by banks in the left and
right nodes of the figure. In sum, community is a group attribute that measures whether a
VC belongs to a tight knit group. Centrality reflects the size and heft of a VC’s connections.7

3 Data

Data on venture capital syndications come from Thomson Financial’s Venture Economics
database. We analyze VC investments made from 1980 to 1999. Our starting point is 1980,
around the time the VC industry started growing rapidly (see, e.g, Figure 1 in Gompers and
Lerner, 2001). We end in 1999 to allow at least 10 years from investment to outcome. We
filter the raw data by dropping the cases in which the database does not disclose a VC firm
name, or lists the VC firm as an angel, individual or management and only include domestic
investments by U.S.-based VC funds in non-buyout deals.

VC firms can exit through mergers and acquisitions or through IPOs. We sample IPO
firms using data from Thomson Financial’s SDC Platinum. We match companies by their
CUSIP identifiers, cross-check the matches against actual names, and further hand-match
the names with those in the Venture Economics database. 1,470 ventures in our sample exit
via IPOs. We obtain M&A data from Thomson Financial’s SDC Mergers and Acquisitions
database. There are 3,545 exits via mergers in our sample.

Table 1 gives descriptive statistics for our sample at the level of the VC firm. Our sample
includes 1,962 unique VC firms. On average, a VC firm invests in 22 portfolio firms and
48 rounds. Each round involves investment of $19.48 million. Close to three-quarters of
the deals made by a VC are syndicated and about one-third of the rounds are classified as
early stage investments. The total total funds raised by a VC amount to about $128 million
(median = $17.50 million). The average age of each VC at the time of its last investment in

7
Newman (2010) brings out the distinctions further. See his Sections 7.1/7.2 on centrality and Section
11.6 on community detection.

10
our sample is a little less than 10 years. The VC headquarters are located in 127 Metropolitan
Statistical Areas (MSAs) in our sample. The two big VC clusters in California (CA) and
Massachusetts (MA) account for about 35% of the VC firms’ headquarters.8

4 Results: Communities

4.1 Community Detection

We detect community clusters using syndications in rolling 5-year windows. Thus, the first
community is based on VC investments from 1980 to 1984, the second community is based
on 1981-1985 investments, and so on. The 5-year window is similar to that in Hochberg,
Ljungqvist, and Lu (2007). As they write, it allows sufficient time to identify syndicate part-
nership choices while avoiding excessively long periods that may contain stale information.
We require a minimum community size of five members and require that the end-to-end
diameter not exceed one-fourth that of the entire network. This constraint is not binding
for our dataset.

We find several communities in the dataset. VCs cluster into between 12 and 35 clusters
of communities that prefer to syndicate with each other. There is considerable variation
in community membership status. Between 81 and 183 VC firms, about 20% of the VCs
active in a 5-year period, belong to communities. The median community has 13 members.
Figures 4–7 depict communities for four non-overlapping 5-year windows, viz., 1980–1984,
1985–1989, 1990–1994, and 1995-1999. We show the members of the largest three commu-
nities in different colors. The upper plots in each figure show the entire VC network. To
present a less cluttered view of the network, the lower figure plots the largest community
embedded within all communities of at least 5 members. We see greater density of connec-
tions within the largest community than its connections across communities. In Figure 4 all
large communities are connected to one another, but in Figures 5 and 6 there are satellite
communities that are large but disconnected from all other communities. Figure 7 shows
8
While VC firms may have satellite offices or may change offices over time, none of the VCs in our sample
moved between CA/MA and non- CA/MA during our sample period.

11
satellite communities in the upper plot, but the largest communities are well connected to
the rest of the communities. In the lower plot, all large communities are connected, but a
few peripheral ones at the edge of the network are relatively isolated.

4.2 Community Stability

We next consider whether the community status of VCs is stable. We then consider the
stability of community clusters. A VC who belongs to a community in one five-year period
need not necessarily belong to a community in another period. To assess the stability of a
VC’s community status, Table 2 reports whether a VC belonging to a community in period
t remains a member of a community in time period t + k where k equals 1, 3, or 5 years,
respectively. On average, 90% of community VCs continue to belong to a community in
the next period and 75% of community VCs remain community VCs five years later. These
statistics are quite stable across all sub periods, suggesting that community membership is
stable but not time invariant.

We next consider stability in the composition of communities. One difficulty in quantifying

stability comes from the flexibility of our approach. Specifically, we let the data determine
which VCs cluster into communities and place no restrictions on the size and composition
of their communities. Thus, a community comprising 9 VCs could stay unchanged in the
next period, break into two mutually exclusive communities whose union is the set of 9 VCs,
or break into more than two communities. In addition, new members can enter any of the
new communities or some VCs may not belong to any clusters in the next period. These
possibilities are illustrated in Table 3, which shows members of the community to which
Sequoia Capital, a well-known VC firm, belongs in two different 5-year windows, 1985-89
and 1990-94. Only two VCs, Mohr Davidow Ventures and Stanford University’s VC arm,
remained in the same community with Sequoia Capital. With a mixture of communities of
different sizes, it is hard to tell which community in period t + 1 was “formed” from which
community in a previous period. With a multitude of communities of varying sizes, these
possibilities multiply.

We quantify the stability of community composition using a Jaccard index. For any pair

12
of sets, the Jaccard index is defined as the number of members in the intersection of the two
sets divided by the number of members in the union of the sets. In other words, if A and B
|A∩B|
are a pair of VC communities, the Jaccard index for the two is given by J(A, B) = |A∪B|
. To

aggregate across communities, let At = A1 , A2 , ..., Am be the set of m communities at time

t, and B t+1 = B1 , B2 , ..., Bn be the set of n communities at time t + 1. For each community

Ai , we determine a composite Jaccard measure, JC(Ai , B) = J(Ai , Bj )j | J(Ai , Bj ) > 0 for

all j = 1, 2, ..., n. The average measure across all initial communities is a composite Jaccard
index. If there is one community both in periods t and t + 1, and all the period t members
remained in the same community next period, JC = 1. On the other hand, if the members
were split equally into two communities, JC = 0.5.

Table 4 presents the composite Jaccard measure averaged over all windows of time. To
assess significance levels, we use a bootstrap procedure. We simulate communities with
number and size distribution equal to what we find in every 5-year period in the data and
determine the empirical distribution of the composite Jaccard measure for the simulated
communities. We find that the Jaccard measure of our community is greater than that of
bootstrapped communities, at the 1% level of significance. Thus, community composition in
our sample is far more stable than would occur by chance.

4.3 Community VC Deal Characteristics

Table 5 reports descriptive statistics for financing rounds in our sample classified by whether
it was financed by a community VC or not. Syndication is common in VC investment.
In our sample, 15,220 out of 33,924 rounds (about 44%) are syndicated and these account
for 66% of proceeds. 10,056 out of 14,897 or 67% of syndications are community rounds.
Early stage rounds account for about a third of the sample. 45% of these are community
rounds. 16,270 deals or close to one-half of the investment rounds are in the geographical
clusters in California or Massachusetts. This reflects a concentration of VC investments
in these states and their representation in VC databases (Kaplan, Sensoy and Stromberg
(2002)). Community VC-based financing rounds account for about 60% of these rounds.

13
3,372 rounds have corporate VC participation and 7,586 rounds have a financial VC firm.
In each case, communities account for 58% of the financing rounds.

Venture Economics classifies VC portfolio firms into 10 industries. We report these data
in Panel B. The software industry with 20% accounts for the largest share of financing
rounds in our sample, followed by medical or health firms, communications and media, and
internet firms. Interestingly, community VC is more likely for the more risky and complex
business models characteristic of software businesses and less likely for consumer product or
industrial businesses. The finding indicates that VC firms rely more on familiar partners in
riskier industries.

Panel C in Table 5 describes key characteristics across rounds. There is greater invest-
ment in rounds with a community VC ($48 million) than in rounds with no community VCs
($29 million). Besides higher investment per round, community rounds tend to have more
VC firms than rounds with no community VC. This may reflect the greater representation
of community VC firms in syndicated rounds. However, even within the subsample of syndi-
cated rounds, community rounds have 4 VCs on average compared to 3 VCs in rounds with
no community VC. This pattern holds for early stage rounds and initial financing rounds.
Panel D describes exit statistics, to which we turn in Section 5.

5 Do Similar or Dissimilar VCs Co-Syndicate?

In principle, communities could comprise VCs of similar or diverse attributes. The case for
diversity is based on resource complementarity. VC investing requires resources and skills
along multiple dimensions. Specialist VCs could each provide different skill sets. It is easier
for a VC firm to draw upon resources and skills that it does not possess if a familiar partner
has the desired complementarity. This preference for what a VC does not already have could
result in disassortative matching and communities with heterogeneous members. From this
viewpoint, communities are soft-border conglomerates, with less rigid borders than firms but
better defined boundaries than arm’s length spot contracting with unfamiliar partners.

Alternatively, following Lerner (1994) and Cestone, Lerner, and White (2006), commu-

14
nities could also be composed of similar-attribute VC firms. For instance, Gompers and
Lerner (2004) find that VCs will often not syndicate unless the investment is vetted by a
partner they trust. If VCs have greater trust in judgments by VCs with similar caliber and
experience, communities will be homogeneous. The behavioral literature provides another
motivation for homogeneity. In a well-known article, McPherson, Smith-Lovin, and Cook
(2001) argue that “birds of a feather flock together,” which leads to assortative matching and
clusters of similar-attribute VCs. The learning hypothesis (Cestone et al., 2006; Sorensen,
2008)) makes more predictions about the specific dimensions on which there is similarity.
Underlying this view are the evaluation and screening abilities of partners, which generate a
useful “second opinion” for VCs. If this force drives the preference for similar syndicate part-
ners, the within-community similarity should primarily be along the dimension of functional
expertise rather than (say) geographical reach.

We test whether the preference for familiar partners is assortative or disassortative by

examining the within-community variation of measurable VC attributes. To benchmark
the results, we compare the variation in attributes for our communities with similar vari-
ation for bootstrapped communities. We simulate the bootstrapped communities so that
they are equal in number and size distribution to what we identify using actual syndica-
tion data. Lower within-community variance for a characteristic indicates a preference for
similar-attribute VCs as syndicate partners, while greater variance indicates preference for
heterogeneous partners.

Table 6 reports the results with p-values in the rightmost column. Panel A in Table 6
reports the average characteristics of VC community members. This panel sheds light on the
types of VCs that cluster into communities. We find that older, larger, prestigious (with high
centrality) VC firms concentrated across states tend to form communities. Communities are
not particularly focused in terms of stage of investment, though they are focused in terms
of industry and geography of investments. Panel B reports the variation in characteristics
within the community compared to the variation for bootstrapped communities. We find that
communities are more concentrated in terms of industry and stage focus, consistent with the
idea that focused firms are more likely to deal with and learn from each other. On the other

15
hand, communities have greater variation in assets under management, in VC influence, and
diversity in the portfolio company state. These results suggest that communities permit VCs
to extend their reach in generating new investments.

Our results partially reconcile contradictory findings in the VC literature. While Du

(2011) reports that venture capitalists are more likely to syndicate with similar partners,
a result resembling the Q-Q assortative matching result of Robinson and Rhodes-Kropf
(2008) in the M&A literature, Hochberg, Lindsey, and Westerfield (2011) argue that VCs
partner with dissimilar VCs. Our results suggest a somewhat nuanced view of the like-
prefer-like versus diversity debate, as we find self-similarity in syndicate partner choice in
some dimensions but not all of them. The industry and stage focus result suggests that
focused VCs prefer focused VCs as partners, consistent with greater trust and benefits of
learning from similar partners. The heterogeneity findings are consistent with the resource
complementarity view of Hochberg, Lindsey, and Westerfield (2011).

The results make a broader point about “diversity” as an empirical construct. Individuals
– or firms – can be characterized by a vector of attributes. There is no economic reason for
similarity or dissimilarity on all dimensions. Teams can be homogeneous on some dimensions
but diverse on others. Understanding the dimensions on which there is similarity and those
in which there is dissimilarity is perhaps more informative and useful than attempting to
force fit all attributes to generate a single diversity index (Harrison and Klein, 2007).

6 Performance

We examine the performance consequences of sourcing capital from a community VC. While
success would normally be measured as an excess rate of return or alpha, return data are
unavailable for broad VC investment samples. We follow the literature (e.g., Gompers and
Lerner, 2000; Lindsey, 2008; Sorensen, 2008; Das, Jo, and Kim, 2011) and define success as
exit via an IPO or a merger. Gompers and Lerner (2000) experiment with several empirical
definitions of success, including different ways of handling mergers, and find that the results
are similar. We report similar findings.

16
 We examine performance by relating exits to community membership constructed from
5-years before the calendar year of an investment. This approach follows the strategy of
Hochberg, Ljungqvist, and Lu (2007). As they point out, the lags result in conservative
models where past 5-year syndication patterns predict outcomes (exits) that happen many
years later. Panel D of Table 5 presents univariate performance results. We find that 12,622
(or 37%) of financing rounds exit through IPOs or M&As. IPOs account for 11%, or a
third of these. In community rounds, 14% exit through IPOs and 29% exit through mergers
compared to 9% and 24% for non-community VC rounds, respectively. We find similar
patterns when considering exits classified by the number of portfolio companies rather than
number of rounds of financing. 13% of companies sourcing funds from a community VC firm
at least once have IPO exits compared with 7% of companies who never have community
VC financing. Finally, 78% of all rounds with a community VC proceed to a next round of
financing compared to 65% of the rounds with no community VC.

6.1 Multivariate Models: Specification

We consider two specifications for investing success. Following, e.g., Hellmann and Puri
(2002), we model the time to exit using a Cox proportional hazards specification. The Cox
model allows a flexible non-parametric baseline hazard with covariates creating parallel shifts
in the hazard function. For easier economic interpretation, we report the Cox results in the
form of the exponentiated hazards ratio. A ratio greater than 1.0 for a variable indicates
that the variable increases the time to exit, while a ratio less than 1.0 indicates that the
variable lowers the time to exit. We also consider a probit model in which success equals
exit either through an M&A transaction or an IPO within 10 years of the investment round.

Our primary interest is how sourcing funds from a community VC is related to exit. The
key variable of interest is the Community Dummy, which takes value 1.0 if the round has
at least one community VC, and zero otherwise. Several papers dating back to at least
Brander, Antweiler, and Amit (2002) find that syndication is a key determinant of success.
Accordingly, we include a control for whether a round is syndicated or not. We include
several other controls suggested by the recent VC literature (e.g., Lindsey, 2008). Early

17
Stage equals 1.0 if the financing is in an early stage round, and zero otherwise. As early
stage companies are more risky and have more unproven models, they are perhaps less likely
to exit successfully compared to later stage firms. The literature in economics suggests that
there is geographical clustering or agglomeration that conveys economic benefits to firms
located in geographic clusters (e.g., Glaeser (2010)). In the context of VC financing, well-
known geographic clusters are in California (CA) and Massachusetts (MA). To control for
firm agglomeration effects, we include a dummy variable that takes the value 1.0 if a portfolio
company is located in either CA or MA, and zero otherwise.

We include controls for the characteristics of venture capital firms participating in a financ-
ing round. Following Hellmann, Lindsey, and Puri (2008), VC arms of financial institutions
may have systematically different success rates. Thus, we control for financial institution
venture capitalists. We control for whether a VC in a financing round is a corporate VC
arm or not. Several papers stress the role of VC experience and skill.9 For instance, Kaplan
and Schoar (2005) identify the importance of experience in VC fund performance. Sorensen
(2007) points to the greater likelihood of an IPO of a portfolio company that is funded by a
more experienced VC. We control for VC skill as IPO Rate, or the rate at which it is able
to take its portfolio companies public.10 Following Lindsey (2008), we define Experience as
the average age of the participating VCs as of the year before the financing round.11

We consider more measures of VC skill. Following Hochberg, Ljungqvist and Lu (2007),

we include the lead VC’s eigenvalue centrality based on investments from t − 1 through
t − 5. Given the particular challenges associated with early stage financing, a VC with
experience in early stage may be considered to be different in terms of value and skills than
an investor without such experience. Such differences in investment focus could also affect
company performance. We define Early Stage Focus as the proportion of companies that
9
For a recent review, see Krishnan and Masulis (2011).
10
In calculating the IPO rate, we follow Krishnan and Masulis (2011) who find strong evidence that the
number of completed IPOs in a VC’s portfolio over the prior 3 calendar years relative to the number of
companies it actively invested in is a predictor of portfolio company performance.
11
Our definition modifies Lindsey’s definition on two fronts. First, we consider age based on the VC firm’s
founding year rather than its entry into Venture Economics. Second, we consider a VC’s experience based
on time periods prior to the financing round in question.

18
the participating VCs invested at an early stage until the year prior to the financing round.
Similarly, each industry presents its own challenges. Skills and expertise for biotechnology
investing can vary from those needed in software. We define Industry Focus as the proportion
of companies funded by the participating VCs in the same industry as the portfolio company
until the year prior to the financing round.

Finally, we consider the possibility that there are benefits of geographic agglomeration not
only for portfolio firms but also for venture capitalists. We include a geographical cluster
control for the VCs, which takes the value 1.0 if at least one of the participating VCs is
located in California or Massachusetts, and zero otherwise. The agglomeration variable
captures soft and informal information flows through local social interactions that are more
likely in areas where VCs are geographically concentrated. All our specifications include year
and industry fixed effects.

6.2 Baseline Estimates

Table 7 reports the basic estimates. In specification (1), the hazard ratio for community VC
is 1.11 and exceeds 1.0 at the 1% level. Thus, community VC financing speeds exit. Among
the controls, the company-level variables are statistically significant. The coefficient for Early
Stage is less than one, suggesting that early stage deals take longer to exit. Companies in
geographical clusters of California and Massachusetts are likely to exit sooner, perhaps due
to improved resource flows and better decision-making arising out of agglomeration (Porter,
2000; Glaeser (2010)). Ownership matters. Financing rounds with corporate or financial
institution VCs are likely to exit sooner.

We find that syndicated ventures tend to exit faster. A VC firms’ reputation for taking
its companies public, measured by the IPO Rate, is not statistically significant in explaining
speed of exit. As in Hochberg, Ljungqvist, and Lu (2007), we find that a more centrally
networked VC facilitates faster exit at the 10% level of significance. We show that this result
is centered in subsamples of later round financing. VC experience, in terms of age at the
time of financing, early stage focus or specific industry focus, is not statistically significant.
However, funding from VCs in the California-Massachusetts cluster predicts quicker exit.

19
The important finding in Table 7 is that VC community is significant even after including
these controls.

Specification (2) in Table 7 reports probit estimates that model the probability of exit
within 10 years. Most results are similar to the Cox results. One difference is that the
IPO rate is significant at the 10% level in the probit model but not in the hazards model.
However, the community variable continues to remain significant and is associated with a
greater likelihood of exit. We also estimate but do not report univariate specifications with
community dummy alone and partial specifications that include it with subsets of controls.
We note that the VC community variable is significant in these models as well with a similar
or higher exponentiated hazards ratio.

6.3 Next Round Financing

In this section, we consider follow-on financing as a measure of success. Follow-on rounds of

financing involve reassessment of the portfolio company. New investors are often brought in,
incumbent VCs increase their investment, and both sets of investors have the opportunity
to re-evaluate and reconsider the progress of the portfolio firm. Thus, attracting follow-on
funding can be viewed as an alternative metric of success. Cochrane (2005) suggests that
round-by-round financing data can be used to construct VC performance metrics.

We rely on Venture Economics codes to specify the round number. These data are not
without noise. In some instances, the first round of financing available in the database may
not be round number one and round numbers may be missing between rounds. We take a
conservative approach. We only consider those rounds that are identifiably numbered and
do not have missing data for subsequent rounds when one exists. These criteria reduce the
sample of first three rounds from 22,683 to 22,271 rounds.12

Table 8 shows the round-by-round results with both the Cox and the probit specifications.
Community VC accelerates the progression to a future round of financing in the earlier rounds
but matter less in later rounds. Among the control variables, the coefficient on the early
12
The 412 rounds we lose due to missing sequential round numbers are spread evenly through the sample
period and in both early and non-early stages.

20
stage dummy variable is positive and statistically significant. One interpretation of this
finding is that staged financing is more prevalent at the early stage firms given the greater
informational issues with these firms (Cornelli and Yosha (2003)). Thus, VC firms manage
early stage financing through more frequent injections of smaller amounts of capital.

Neither a corporate VC nor a financial institution VC helps with subsequent financing

rounds in the initial stages. In fact, financial institution VCs have a negative and significant
effect, perhaps because VC arms of financial firms are structurally different (Hellmann,
Lindsey and Puri (2008)). As before, syndicated rounds have a higher chance of obtaining
future funding and do so sooner. VCs’ reputation for taking portfolio companies public
(IPO rate) seems to have an adverse effect on subsequent funding after round 2. Eigenvector
centrality is significant in round 1 in the Cox specification at 10% significant level but not in
the probit specification. However, it matters in round 3 under both specifications, suggesting
that centrality and community play complementary roles. Perhaps thick rolodexes are more
critical in later stages when it provides firms access to a broader set of resources such as
personnel or strategic contacts.

Portfolio companies located in California or Massachusetts experience a higher likelihood

of next-round financing except in round 2. VC firms belonging to the geographical clusters in
California and Massachusetts have a positive effect in earlier round but not in later rounds.
VC experience, in terms of the participating VCs’ age, does not help and even impedes
progress in initial rounds. Early stage focus is associated with a greater likelihood of and
faster follow-on financing or exit. Thus, firms that declare specialization in early stage
ventures appear to accelerate a firm’s progress to a next round of financing. In each of the
specifications, the industry focus on the participating VCs is statistically insignificant. In
any event, the key result is that community VC remains significant in facilitating follow-on
financing, particularly in the initial rounds.

6.4 Competing Hazards

The Cox and probit models treat IPO and M&A exits symmetrically. While an IPO is a
reasonable proxy for an exit, an M&A exit is a more noisy proxy for success. As Gompers

21
and Lerner (2000) point out, some exits via M&A are successful while others may reflect
failure. One possibility is to focus on IPO exits alone as a measure of success. However,
firms can exit via IPOs only if a merger has not already occurred. To accommodate this
possibility, we consider a competing hazards model. Here, we consider the effect of multiple
hazards by modeling intervening subhazards.

Table 9 presents the results of the competing risks model where we treat IPO as the event
of interest and M&A as the competing risk event. Specification (1) repeats the analysis of
Table 7 while specifications (2) - (4) repeat the analysis of Table 8 with competing hazards
model instead of the Cox model. Our main results are essentially unchanged. In specification
(1), community VC in a financing round shortens the exit time to IPOs significantly. The
economic magnitude, 13%, is similar to the 11% we find in the Cox model. As before,
community VC financing does not matter in later rounds but is significant in round 1. In
specification (2), community VCs speed exit to future financing by about 15% compared to
a point estimate of 16% in the Cox model estimated in Table 8.

7 Conclusion

Syndication is a pervasive feature of venture capital financing. About two-thirds of the

venture capital financing raised in the U.S. market is syndicated. Syndication is a robust
determinant of successful VC exit. Over a period of time, a venture capital firm is likely to
form syndicates several times and do so with different partners.

While there is extensive evidence that VC syndication matters, less is known about the
process and economic consequence of picking syndicate partners. Our study provides new
evidence on these questions. We find that VCs do not pick syndicate partners at random. Nor
do they associate with a fixed set of partners. Rather, VCs exhibit associative preferences in
which they are probabilistically more likely to syndicate with some VCs than others. This
preference leads VCs to cluster into spatial agglomerates that we term “communities.”

Identifying clusters of VC communities is computationally challenging when we attempt

to incorporate real world features of VC syndications. We impose no a priori restrictions

22
on the number of communities nor any requirements that communities have the same size.
In addition, we let communities have porous boundaries to allow for the fact that outside-
community syndications are less likely but not altogether precluded. We employ recent
developments in the physical sciences literature to identify communities. Using 20 years of
venture financing data, we find robust evidence of VC community formation. About 20% of
all VCs cluster into communities. We find that communities are both stable and tight-knit.

The existence of communities shows that VCs exhibit a strong revealed preference for
familiar syndicate partners. This result adds texture to the view in the VC literature that
syndication is beneficial. As the literature points out, syndication permits risk-sharing,
access to broader resources, or better vetting of portfolio firms. Our findings suggest that
familiarity with partners enhances these benefits. Another interpretation of our results is
that while syndication is beneficial, it can also pose a fresh set of problems for venture
capitalists. Suspicions of ex-post hold up and free riding by partners can lead to insufficient
effort and undo the benefits of syndication. Syndicating with familiar partners can mitigate
these problems by reducing information asymmetry, building trust, and enhancing reciprocity
between partners. Alternatively or additionally, VCs may learn through doing and learning
is enhanced when partners are familiar. The propensity to pick preferred syndicate partners
can be interpreted as an outcome of these forces.

We also analyze the attributes of VCs within communities. What kinds of VCs tend to
syndicate with each other? At one end of the spectrum is the view that like prefers like,
due to a behavioral preference for the similar or because VCs rely more on the judgments of
similarly endowed partners, consistent with Lerner (1994) and Cestone, Lerner, and White
(2006). On the flip side, there may be a preference for heterogeneity in partner attributes
that can also help by easing access to broader skills and resources. The received evidence
is mixed. While Du (2011) finds evidence of similarity, Hochberg, Lindsey, and Westerfield
(2011) find that heterogeneity matters. We find subtle patterns consistent with both views.
While there is similarity along style style dimensions such as industry or stage focus, we find
heterogeneity on dimensions such as VC size, influence, and geographical focus.

Finally, we show that firms that source capital from community VCs are more likely

23
to experience successful exit. The finding suggests that economic motivations, rather than
passive behavioral preferences or second order transaction costs such as reduced paperwork,
motivate VCs to syndicate from a small short list of preferred syndication partners. We
emphasize that we do control for whether a deal is syndicated or not and find that it matters.
Our point is that syndication has benefits beyond its value for a specific deal. There is
additional value from the social interactions from syndication that benefit VCs in their
future deals. These benefits are not localized to the few VCs who derive thick rolodexes from
syndication but flow to members of communities who tend to cluster together in syndicates.

Our approach has immediate applications in other areas of finance such as loan syndica-
tions and underwriting where financial intermediaries form syndicates. As Robinson (2008)
points out, inter-firm collaborations also involve partially overlapping, non-exclusive, and si-
multaneous collaborations with multiple partners. Our study provides a practical method for
characterizing such collaborations and detecting clusters of preferred partners within these
networks. Communities can be viewed as organizational forms that help resolve the infor-
mation and incentive problems of these settings, intermediate organizational forms between
hard-boundary conglomerates that internalize transactions and arms-length spot contracting
with random partners. An interesting question is whether there is agglomeration into com-
munities in these other areas. If so, it is interesting to understand if this is associated with
better economic outcomes or it is driven by economically neutral phenomenon such as pure
behavioral preferences for the familiar or transaction cost motives to lower administrative
or operational costs. More generally, it is interesting to understand whether evolutionary
processes in biology and the natural sciences that give rise to communities are also mimicked
in settings where agents interact for economic benefits.

24
A Calculating Modularity

In order to offer the reader a better sense of how modularity is computed in different settings,
we provide a simple example here, and discuss the different interpretations of modularity
that are possible. The calculations here are based on the measure developed in Newman
(2006). Since we used the igraph package in R, we will present the code that may be used
with the package to compute modularity.

Consider a network of five nodes {A, B, C, D, E}, where the edge weights are as follows:
A : B = 6, A : C = 5, B : C = 2, C : D = 2, and D : E = 10. Assume that a community
detection algorithm assigns {A, B, C} to one community and {D, E} to another, i.e., only
two communities. The adjacency matrix for this graph is

 
0 6 5 0 0
6 0 2 0 0 
 

 
{Aij } = 
 5 2 0 2 0 

 0 0 2 0 10 

0 0 0 10 0

The Kronecker delta matrix that delineates the communities will be

 
1 1 1 0 0
1 1 1 0 0
 
 
 
{δij } = 
 1 1 1 0 0 


 0 0 0 1 1 

0 0 0 1 1

The modularity score is

" #
1 X di × dj
Q= Aij − · δij (2)
2m i,j 2m

1 P 1 P
where m = 2 ij Aij = 2 i di is the sum of edge weights in the graph, Aij is the (i, j)-
th entry in the adjacency matrix, i.e., the weight of the edge between nodes i and j, and
P
di = j Aij is the degree of node i. The function δij is Kronecker’s delta and takes value
1 when the nodes i and j are from the same community, else takes value zero. The core of
h i
di ×dj
the formula comprises the modularity matrix Aij − 2m
which gives a score that increases

25
when the number of connections within a community exceeds the expected proportion of
connections if they are assigned at random depending on the degree of each node. The
score takes a value ranging from −1 to +1 as it is normalized by dividing by 2m. When
Q > 0 it means that the number of connections within communities exceeds that between
communities. The program code that takes in the adjacency matrix and delta matrix is as
follows:

#MODULARITY
Amodularity = function(A,delta) {
n = length(A[1,])
d = matrix(0,n,1)
for (j in 1:n) { d[j] = sum(A[j,]) }
m = 0.5*sum(d)
Q = 0
for (i in 1:n) {
for (j in 1:n) {
Q = Q + (A[i,j] - d[i]*d[j]/(2*m))*delta[i,j]
}
}
Q = Q/(2*m)
}

We use the R programming language to compute modularity using a canned function,

and we will show that we get the same result as the formula provided in the function above.
First, we enter the two matrices and then call the function shown above:

> A = matrix(c(0,6,5,0,0,6,0,2,0,0,5,2,0,2,0,0,0,2,0,10,0,0,0,10,0),5,5)
> delta = matrix(c(1,1,1,0,0,1,1,1,0,0,1,1,1,0,0,0,0,0,1,1,0,0,0,1,1),5,5)
> print(Amodularity(A,delta))
[1] 0.4128

We now repeat the same analysis using the R package. Our exposition here will also show
how the walktrap algorithm is used to detect communities, and then using these communities,

26
how modularity is computed. Our first step is to convert the adjacency matrix into a graph
for use by the community detection algorithm.

> g = graph.adjacency(A,mode="undirected",weighted=TRUE,diag=FALSE)

We then pass this graph to the walktrap algorithm:

> wtc=walktrap.community(g,modularity=TRUE,weights=E(g)$weight)
> res=community.to.membership(g,wtc$merges,steps=3)
> print(res)
$membership
[1] 0 0 0 1 1

$csize
[1] 3 2

We see that the algorithm has assigned the first three nodes to one community and the
next two to another (look at the membership variable above). The sizes of the communities
are shown in the size variable above. We now proceed to compute the modularity

> print(modularity(g,res$membership,weights=E(g)$weight))
[1] 0.4128

This confirms the value we obtained from the calculation using our implementation of the
formula.

Modularity can also be computed using a graph where edge weights are unweighted. In
this case, we have the following adjacency matrix

> A
[,1] [,2] [,3] [,4] [,5]
[1,] 0 1 1 0 0
[2,] 1 0 1 0 0
[3,] 1 1 0 1 0

27
[4,] 0 0 1 0 1
[5,] 0 0 0 1 0

Using our function, we get

> print(Amodularity(A,delta))
[1] 0.22

We can generate the same result using R:

> g = graph.adjacency(A,mode="undirected",diag=FALSE)
> wtc = walktrap.community(g)
> res=community.to.membership(g,wtc$merges,steps=3)
> print(res)
$membership
[1] 1 1 1 0 0

$csize
[1] 2 3

> print(modularity(g,res$membership))
[1] 0.22

A final variation on these modularity calculations is to use a Kronecker delta matrix that
has diagonal elements of zero. In the paper we use the first approach presented in this
Appendix.

28
B Variable Definitions

Variable Description

Dummy Variables
Community Equals 1.0 if there is at least one community VC in the fi-
nancing round and zero otherwise
Early Stage Equals 1.0 if the round is an early stage financing and zero
otherwise.
Company Geographical Equals 1.0 if the portfolio company funded by the VC is in
Cluster the state of California or Massachusetts and zero otherwise.
Corporate VC Equals 1.0 if there is at least one venture capitalist who is
the corporate VC arm of a firm.
FI VC Equals 1.0 if there is at least one financial institution VC in
the round
Syndicated Equals 1.0 if the round is syndicated, zero otherwise
VC Geographical Cluster Equals 1.0 if at least one participating VC is in the state of
CA or MA
Other Variables
IPO Rate natural log of one plus the average of each participating VC’s
ratio of IPOs to number of portfolio companies in the last
three years prior to the financing round
Centrality lead VC’s eigenvector centrality, normalized for the sample
in each specification
Experience natural log of one plus the average age, in years, of the par-
ticipating VCs from their founding until the year prior to the
financing round
Early Stage Focus natural log of one plus the proportion of companies that the
participating VCs invested at an early stage until the year
prior to the financing round
Industry Focus natural log of one plus the proportion of companies funded by
the participating VCs in the same industry as the portfolio
company until the year prior to the financing round.

29
References
Bengtsson, O., Hsu, D. (2010). How do venture capital partners match with startup founders?
Working Paper.

Bottazzi, L., Da Rin, M., Hellmann, T. (2011). The importance of trust for investment:
Evidence from venture capital. Working Paper.

Brander, J. A., Amit, R., Antweiler, W. (2002). Venture-capital syndication: Improved

venture selection vs. the value-added hypothesis, Journal of Economics and Management
Strategy, v11, 423-452.

Burdick, D., Hernandez, M., Ho, H., Koutrika, G., Krishnamurthy, R., Popa, L., Stanoi,
I.R., Vaithyanathan, S., Das, S.R. (2011). Extracting, linking and integrating data from
public sources: A financial case study, IEEE Data Engineering Bulletin, 34(3), 60-67.

Cai, Y., and Sevilir, M. (2012). Board connections and M&A Transactions, Journal of Fi-
nancial Economics 103(2), 327-349.

Cestone, Giacinta., Lerner, Josh, White, Lucy (2006). The design of syndicates in venture
capital, Harvard Business School Working Paper.

Chidambaran, N. K., Kedia, S., Prabhala, N.R. (2010). CEO-Director connections and fraud,
University of Maryland Working Paper.

Cochrane, John (2005). The risk and return of venture capital. Journal of Financial Eco-
nomics 75, 3-52.

Cohen, Lauren, Frazzini, Andrea, Malloy, Christopher (2008a). The small world of investing:
Board connections and mutual fund returns, Journal of Political Economy 116, 951–979.

Cohen, Lauren, Frazzini, Andrea, Malloy, Christopher (2008b). Sell-Side school ties, forth-
coming, Journal of Finance.

Cornelli, F., Yosha, O. (2003). Stage financing and the role of convertible securities, Review
of Economic Studies 70, 1-32.

30
Da Rin, Marco, Hellmann, Thomas, Puri, Manju (2012). A survey of venture capital research,
Duke University Working Paper.

Das, Sanjiv., Jo, Hoje, Kim, Yongtae (2011). Polishing diamonds in the rough: The sources
of syndicated venture performance, Journal of Financial Intermediation, 20(2), 199-230.

Du, Qianqian (2011). Birds of a feather or celebrating differences? The formation and impact
of venture capital syndication, University of British Columbia Working Paper.

Engelberg, Joseph., Gao, Pengjie, Parsons, Christopher (2000). The value of a rolodex: CEO
pay and personal networks, Working Paper, University of North Carolina at Chapel Hill.

Fortunato, S. (2009). Community detection in graphs, arXiv :0906.0612v1 [physics.soc-ph].

Gertler, M.S. (1995). Being there: proximity, organization and culture in the development
and adoption of advanced manufacturing technologies, Economic Geography 7(1), 1-26.

Girvan, M., Newman, M. (2002). Community structure in social and biological networks,
Proc. of the National Academy of Science 99(12), 7821–7826.

Glaeser, E., ed., (2010). Agglomeration Economics, University of Chicago Press.

Goldfarb, Brent, Kirsch, David, Miller, David, (2007). Was there too little entry in the dot
com era?, Journal of Financial Economics 86(1), 100-144.

Gompers, P., Lerner, J. (2000). Money chasing deals? The impact of fund inflows on private
equity valuations, Journal of Financial Economics 55(2), 281-325.

Gompers, P., Lerner, J. (2001). The venture capital revolution, Journal of Economic Per-
spectives 15(2), 45-62.

Gompers, P., Lerner, J., (2004). The Venture Capital Cycle, MIT Press.

Gorman, M., Sahlman, W. (1989). What do venture capitalists do? Journal of Business
Venturing 4, 231-248.

31
Granovetter, M. (1985). Economic action and social structure: The problem of embedded-
ness, American Journal of Sociology 91(3), 481-510.

Grossman, S., Hart, O. (1986). The costs and benefits of ownership: A theory of vertical
and lateral integration, Journal of Political Economy 94(4), 691-719.

Guimera, R., Amaral, L.A.N. (2005). Functional cartography of complex metabolic networks,
Nature 433, 895-900.

Guiso, L., Sapienza, P., Zingales, L. (2004). The role of social capital in financial develop-
ment, American Economic Review 94, 526-556.

Harrison, D., Klein, K. (2007). What’s the difference? Diversity constructs as separation,
variety, or disparity in organization, Academy of Management Review 32(4), 1199-1228.

Hart, O., Moore, J. (1990). Property rights and the nature of the firm, Journal of Political
Economy 98(6), 1119-1158.

Hegde, D., Tumlinson, J. (2011). Can birds of a feather fly together? Evidence for the
economic payoffs of ethnic homophily, Working Paper.

Hellmann, T. J., Lindsey, L., Puri, M. (2008). Building relationships early: Banks in venture
capital, Review of Financial Studies 21(2), 513-541.

Hellmann, T. J., Puri, M. (2002). Venture capital and the professionalization of start-up
firms: Empirical evidence, Journal of Finance 57, 169-197.

Hoberg, G., Phillips, G. (2010). Product Market Synergies and Competition in Mergers and
Acquisitions: A Text-Based Analysis, Review of Financial Studies 23(10), 3773–3811.

Hochberg, Y., Ljungqvist, A., Lu, Y. (2007). Whom you know matters: Venture capital
networks and investment performance, Journal of Finance 62(1), 251-301.

Hochberg, Y., Lindsey, L., Westerfield, M. (2011). Inter-firm Economic Ties: Evidence from
Venture Capital, Northwestern University Working Paper.

32
Hwang, B., Kim, S. (2009). It pays to have friends, Journal of Financial Economics 93,
138-158.

Ishii, J.L., Xuan, Y. (2009). Acquirer-Target social ties and merger outcomes, Working
Paper, SSRN: http://ssrn.com/abstract=1361106.

Kaplan, S. N., Schoar, A. (2005). Private equity performance: Returns, persistence and
capital flows, Journal of Finance 60, 1791-1823.

Kaplan, S. N., Sensoy, B., Stromberg, P. (2002). How well do venture capital databases
reflect actual investments?, Working paper, University of Chicago.

Kaplan, S. N., Stromberg, P. (2003). Financial contracting theory meets the real world:
Evidence from venture capital contracts, Review of Economic Studies 70, 281-316.

Kaplan, S. N., Stromberg, P. (2004). Characteristics, contracts and actions: Evidence from
venture capital analyses, Journal of Finance 59, 2177-2210.

Krishnan, C. N. V., Masulis, R. W. (2011). Venture capital reputation, in Douglas J. Cum-

mings, ed., Handbook on Entrepreneurial Finance, Venture Capital and Private Equity,
Oxford University Press.

Lerner, J. (1994). The syndication of venture capital investments. Financial Management

23, 1627.

Lerner, J. (1995). Venture capitalists and the oversight of private firms, Journal of Finance
50 (1), 302-318

Leskovec, J., Kang, K.J., Mahoney, M.W. (2010). Empirical comparison of algorithms for
network community detection, ACM WWW International Conference on World Wide
Web.

Lindsey, L. A. (2008). Blurring boundaries: The role of venture capital in strategic alliances,
Journal of Finance 63(3), 1137-1168.

33
Lusseau, D. (2003). The emergent properties of a dolphin social network, Proceedings of the
Royal Society of London B 271 S6: 477–481.

McPherson, M., Smith-Lovin, L., Cook, J. (2001). Birds of a feather: Homophily in social
networks, Annual Review of Sociology 27, 415-444.

Neher, D. V. (1999). Staged financing: An agency perspective, Review of Economic Studies

66, 255-274.

Newman, M. (2001). Scientific collaboration networks: II. Shortest paths, weighted networks,
and centrality, Physical Review E 64, 016132.

Newman, M. (2006). Modularity and community structure in networks, Proc. of the National
Academy of Science 103 (23), 8577-8582.

Newman, M. (2010). Networks: An introduction, Oxford University Press.

Pons, Pascal, Latapy, Matthieu, (2005). Computing communities in large networks using
random walks, Journal of Graph Algorithms and Applications 10, 191–218.

Porter, M.E. (2000). Location, competition and economic development: Local clusters in a
global economy, Economic Development Quarterly 14(1), 15-34.

Porter, Mason., Mucha, Peter, Newman, Mark, Friend, A. J. (2007). Community structure
in the United States House of Representatives, Physica A: Statistical Mechanics and its
Applications 386(1), 413–438.

Ravasz, E., Somera, A.L., Mongru, D.A., Oltvai, N., Barabasi, A.L. (2002). Hierarchical
organization of modularity in metabolic networks, Science 297 (5586), 1551.

Robinson, D. (2008). Strategic alliances and the boundaries of the firm, Review of Financial
Studies 21(2), 649-681.

Robinson, D., Sensoy, B. (2011). Private equity in the 21st century: cash flows, performance,
and contract terms from 1984-2010, Working Paper, Ohio State University.

34
Robinson, D., Stuart, T. (2007). Financial contracting in biotech strategic alliances, Journal
of Law and Economics 50(3), 559-596.

Simon, Herbert (1962). The architecture of complexity, Proceedings of the American Philo-
sophical Society 106 (6), 467–482.

Sorensen, Morten (2007). How smart is smart money? A Two-sided matching model of
venture capital, Journal of Finance 62, 2725-2762.

Sorensen, Morten (2008). Learning by investing: evidence from venture capital, Columbia
University Working Paper.

Wu, K., Taki, Y., Sato, K., Kinomura, S., Goto, R., Okada, K., Kawashima, R., He,
Y., Evans, A. C. and Fukuda, H. (2011). Age-related changes in topological organiza-
tion of structural brain networks in healthy individuals, Human Brain Mapping 32, doi:
10.1002/hbm.21232.

Zachary, Wayne (1977). An information flow model for conflict and fission in small groups,
Journal of Anthropological Research 33, 452–473.

35
Figure 1: Frequency distribution of J.P. Morgan’s partners.

36
 180  

160  

140  

120  
No  of  interac,ons  

100  

80  

60  

40  

20  
Matrix  
0  
Sequoia  
1   2   3   4   5   6   7   8   J.  P.  Morgan  
9  
10   11   12  
13   14   Kleiner  
15   16  
17   18  
19   20  
Top  20  VC  partners  

Figure 2: Distribution of the number of interactions of four top firms with their top 20 collabora-
tors.

37
 FigureFigure
3: Communities and centrality
1: Co-lending in bank co-lending
network for 2005. networks.

PN
tion: xi = j=1 Lij xj , ∀i. This may be compactly repre-
sented as x = L · x, where x = [x1 , x2 , . . . , xN ]! ∈ RN×1 Figure
and L ∈ RN×N . We pre-multiply the left-hand-side of the
equation above by a scalar λ to get λ x = L · x, i.e., an
eigensystem. The principal eigenvector in this system gives previously, an
the loadings of each bank on the main eigenvalue and rep- presented in T
resents the influence of each bank on the lending network. trality are see
This is known as the “centrality” vector in the sociology lit- J.P. Morgan
erature [6] and delivers a measure of the systemic effect a Citigroup (no
single bank may have on the lending
38 system. Federal regula- and contribut
tors may use the centrality scores of all banks to rank banks Figure 2 sh
in terms of their risk contribution to the entire system and after 2005. C
 1980_1984

●
●
●
●●
● ●
●
●●
●
●●
● ●
●
●●●
●
●●●
●
●
●
●●●
●●
●●

1980_1984

● ●

●● ●
●●
●
● ●
●
●
● ●
●

Figure 4: Network graph for connected VCs (1980–84). The upper plot shows the network of all
VCs in communities (1180 in all), and blue, green, and red nodes in the center of the network
are the VCs in the top three largest communities, respectively. The lower plot shows the network
comprised only of the 134 VCs who are members of the 18 communities that have at least five VCs.
The darker nodes in the lower plot show the VCs in the largest community.

39
 1985_1989

●
● ●●
●●●
● ●●
●●●
● ●●
●
●
●●
●●
●
●●●
●
●
●● ●●●
●
●●
●
●
●
● ●

1985_1989

● ●
● ●
●
●
●
● ●
● ● ●
● ●
● ● ●
● ● ● ● ●
●
●
●

Figure 5: Network graph for connected VCs (1985–89). The upper plot shows the network of all
VCs in communities (1295 in all), and blue, green, and red nodes in the center of the network
are the VCs in the top three largest communities, respectively. The lower plot shows the network
comprised only of the 180 VCs who are members of the 18 communities that have at least five
VCs. The darker nodes in the lower plot show the VCs in the largest community. Note the single
satellite community at the bottom of the lower plot. Such a community has low centrality.

40
 1990_1994

● ●●
●
●
● ●●●●
●●●
●
●
●●
●
●●●
●
●●
●
●●●
● ●
●
●●●
●

1990_1994

●
●

● ●
●
● ●
●
●●
●
●

●
●
●

Figure 6: Network graph for connected VCs (1990–94). The upper plot shows the network of
all VCs in communities (953 in all), and blue, green, and red nodes in the center of the network
are the VCs in the top three largest communities, respectively. The lower plot shows the network
comprised only of the 114 VCs who are members of the 14 communities that have at least five VCs.
The darker nodes in the lower plot show the VCs in the largest community. Note the two satellite
communities above the main one in the lower plot. Such communities have low centraity.

41
 1995_1999

● ●●
● ●●●●
●●
●●●
●●
● ●●●●●●●●
●
●●●●
●
●
●
●
●
●
●
●
●●
●●
●
●●●
●
●
●●●●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●●● ●
●
●●● ● ●
● ●

1995_1999

●
●
●
● ●
●

●
● ●●
● ● ●
●
● ●
● ●● ● ●
● ● ●
●
● ● ● ●●●
●●
● ● ●● ●● ●
● ● ● ●● ●
● ● ● ● ●
● ● ●● ●
●
●
● ●
● ●
●
●

Figure 7: Network graph for connected VCs (1995–99). The upper plot shows the network of all
VCs in communities (2772 in all), and blue, green, and red nodes in the center of the network
are the VCs in the top three largest communities, respectively. The lower plot shows the network
comprised only of the 379 VCs who are members of the 35 communities that have at least five VCs.
The darker nodes in the lower plot show the VCs in the largest community.

42
Table 1: Venture Capitalists in our sample. This table provides descriptive statistics of the 1,962
unique U.S.-based VCs in our database over the entire 20-year period, from 1980 to 1999. Data are
from Venture Economics and exclude non-US investments, angel investors, and VC firms focusing
on buyouts. Size is the sum of the capital under management in all funds that were active during
1980-1999. Total investment is the sum of a VC’s investments over this time period. % Deals
Syndicated is the fraction of all rounds that a VC invested in that were syndicated. % early stage
deals is the fraction of a VC’s deals that are in the early stage. Age is defined as the difference in
the year of the VC’s last investment in the sample period and the VC firm’s founding date.

Variables: Mean Median # Observations

# Rounds 47.98 9.00 1,962
# Companies 21.64 7.00 1,962
Investment per round ($ mm) 19.47 10.56 1,945
% Deals Syndicated 73.62 80.90 1,962
% Early Stage Deals 35.95 33.33 1,962
VC size ($ mm) 128.01 17.50 1,552
Total Investment ($ mm) 595.14 110.47 1,945
Age 9.59 6.00 1,950
# VC firms per MSA 14.24 3.00 127
VC firm in CA/MA (dummy=1) 0.35 0.00 1,962
Table 2: Stability of community status. The table provides data on the number of VCs who belong
to community clusters in each 5-year window and the fraction of these that remain in a community
after 1, 3, and 5 years from the initial window.

Window # Community VCs After 1 year After 3 years After 5 years

1980-1984 134 0.90 0.85 0.77

1981-1985 153 0.96 0.90 0.80
1982-1986 180 0.93 0.80 0.72
1983-1987 177 0.96 0.87 0.77
1984-1988 205 0.87 0.78 0.67
1985-1989 180 0.92 0.83 0.71
1986-1990 169 0.88 0.76 0.69
1987-1991 125 0.88 0.79 0.77
1988-1992 130 0.93 0.78 0.75
1989-1993 111 0.86 0.77 0.71
1990-1994 114 0.89 0.80 0.77
1991-1995 112 0.82 0.80
1992-1996 146 0.93 0.89
1993-1997 173 0.90
1994-1998 246 0.94
1995-1999 379
Table 3: Sample Communities. This table details venture capitalists that belong to two sample
communities, one each for 1985-1989 and 1990-1994. We chose the communities that had Sequoia
Capital as a member in both time periods.

Sample community from the period 1985–89:

(1) Arthur Rock & Co., (2) Asset Management Company Venture Cap-
ital, (3) Associated Venture Investors (AKA: AVI Capital), (4) Bryan
& Edwards, (5) Draper Fisher Jurvetson (FKA: Draper Associates),
(6) GT Technology Fund, (7) MedVenture Associates (AKA: MVA),
(8) Mohr Davidow Ventures, (9) New Zealand Insurance, (10) Nippon
Investment & Finance Co Ltd., (11) OSCCO Ventures, (12) Pacific
Venture Partners, (13) Partech International, (14) Sequoia Capital,
(15) Stanford University, (16) Suez Ventures (FKA: Indosuez Ventures),
(17) Technology Venture Investors.

Sample community from the period 1990–94:

(1) Avalon Ventures, (2) Berkeley International Capital Corp., (3) Del-
phi Ventures, (4) Frazier Healthcare and Technology Ventures(FKA
Frazier & Co), (5) Integral Capital Partners, (6) Kleiner Perkins Cau-
field & Byers, (7) Mayfield Fund, (8) Mohr Davidow Ventures, (9) Se-
quoia Capital, (10) Silicon Graphics, Inc., (11) Stanford University,
(12) Technology Investment Fund, Inc., (13) Trinity Capital Partners,
(14) Vertex Management Pte, Ltd. (AKA: Vertex Venture Holdings),
(15) W.S. Investments.
Table 4: Stability of community composition. We identify community clusters in a 5-year window
and examine whether the communities in the next five year window are similar to the ones in the
previous period. We use the Jaccard similarity index which measures the similarity between every
pair of communities in the adjacent period, and average it across all non-empty intersections. The
Jaccard index is defined as the ratio of the size of the intersection set to the size of the union
set. To benchmark the index, we generate a similar index for bootstrapped communities simulated
using the same community sizes and number of communities in each 5-year rolling window as in
our sample. The last column shows the p-values testing the equality of the composite measure for
the community and bootstrapped community. ∗∗∗ denotes significance at the 1% level.

Window 1 Window 2 Community Bootstrapped p-value

Community

1980-1984 1981-1985 0.188 0.064 0.01∗∗∗

1981-1985 1982-1986 0.175 0.060 0.01∗∗∗
1982-1986 1983-1987 0.182 0.056 0.01∗∗∗
1983-1987 1984-1988 0.217 0.058 0.01∗∗∗
1984-1988 1985-1989 0.141 0.055 0.01∗∗∗
1985-1989 1986-1990 0.177 0.052 0.01∗∗∗
1986-1990 1987-1991 0.155 0.052 0.01∗∗∗
1987-1991 1988-1992 0.155 0.050 0.01∗∗∗
1988-1992 1989-1993 0.252 0.055 0.01∗∗∗
1989-1993 1990-1994 0.123 0.062 0.01∗∗∗
1990-1994 1991-1995 0.246 0.065 0.01∗∗∗
1991-1995 1992-1996 0.143 0.055 0.01∗∗∗
1992-1996 1993-1997 0.128 0.042 0.01∗∗∗
1993-1997 1994-1998 0.135 0.041 0.01∗∗∗
1994-1998 1995-1999 0.109 0.042 0.01∗∗∗
Table 5: Descriptive statistics for 33,924 rounds in 13,541 unique portfolio companies from 1985-
1999. A round is a community round if at least one VC firm participating in it comes from a VC
community. Communities are detected using a walk trap algorithm applied to syndicated deals over
five year windows rolled forward one year at a time. The sample comprises VC deals obtained from
Venture Economics excluding buyouts, angel investments and non-US deals. Industry classifications
are as per Venture Economics. Exit data are obtained by matching with Thomson Financial IPO
and M&A databases.

Variable Total Community Round Not Community Round

Panel A: Counts By Round
# Deals 33,924 15,220 18,704
—Round 1 11,018 3,581 7,437
—Round 2 6,881 3,015 3,866
—Round 3 4,784 2,410 2,374
Syndicated 14,897 10,056 4,841
Early stage 12,118 5,472 6,646
Geographical Cluster 16,270 9,607 6,663
Rounds with
—Geographical Cluster VC 19,678 12,140 7,538
—Corporate VC 3,372 1,923 1,449
—FI VC 7,586 4,415 3,171

Panel B: Percentage By Venture Economics Industry

—Biotech 6.8 7.3 6.3
—Commu&Media 12.1 13.3 11.1
—Hardware 7.3 9.0 6.0
—Software 19.8 22.7 17.5
—Semiconductor, Electricals 7.0 7.9 6.3
—Consumer Products 7.8 5.3 9.9
—Industrial, Energy 5.9 3.4 8.0
—Internet 11.0 11.9 10.3
—Medical 13.7 15.0 12.7
—Others 8.5 4.4 11.9
Panel C: Round Statistics
Proceeds ($ million) 38 (13) 48 (21) 29 (9)
# VCs 2.08 (1) 2.89 (2) 1.42 (1)
—in syndicated rounds 3.46 (3) 3.85 (3) 2.64 (2)
—in early stage rounds 1.93 (1) 2.53 (2) 1.43 (1)
—in round 1 1.54 (1) 2.03 (2) 1.31 (1)
—in round 2 2.00 (1) 2.70 (2) 1.45 (1)
—in round 3 2.38 (2) 3.23 (3) 1.52 (1)

PANEL D: Exit
Rounds with
—IPO exits 3,828 2,071 1,757
—M&A exits 8,794 4,363 4,431
—Follow-on funding 23,972 11,903 12,069
Table 6: Similarity of Same-Community VCs. The table compares key community characteristics
with those of random communities generated through bootstrapping based on matching commu-
nity sizes and number of communities in each 5-year rolling window. For each community (and
bootstrapped community), we generate the mean and standard deviation of the characteristic. In
Panel A and Panel B, the table presents the average value of these measures across communities
(and random communities). Age uses the number of years between a VC’s last investment in a
5-year window and the founding year of the VC firm. Assets under management (AUM), in ($
million), uses the sum of a VC’s active funds during each 5-year period. Centrality is based on
each VC’s eigenvector centrality determined for each 5-year rolling window. Ownership HHI is the
Herfindahl index based on the different types of VC ownership in a community. VC State HHI is
the Herfindahl index based on the number of VCs in each state. Industry HHI is the Herfindahl
index based on the amount invested in each industry, while Stage HHI is the Herfindahl index based
on the amount invested in each stage of investment. Company State HHI is the Herfindahl index
based on the amount invested in each state by VCs. The last column shows the p-values testing
the equality of the means of the community and bootstrapped community characteristics. ∗∗∗ , ∗∗ ,
and ∗ denote 1%, 5% and 10% significance, respectively.

Community Bootstrapped p-value

Community

Panel A: Mean
Age 9.51 8.60 0.01∗∗∗
AUM 138.50 82.24 0.01∗∗∗
Centrality 0.09 0.03 0.01∗∗∗
Ownership HHI 0.44 0.43 0.25
VC State HHI, # 0.43 0.20 0.01∗∗∗
Stage HHI 0.34 0.34 0.50
Industry HHI 0.25 0.21 0.01∗∗∗
Company State HHI 0.37 0.27 0.01∗∗∗

Panel B: Standard Deviation

Age 7.78 8.17 0.15
AUM 165.06 131.51 0.05∗∗
Centrality 0.09 0.05 0.01∗∗∗
Stage HHI 0.24 0.27 0.01∗∗∗
Industry HHI 0.25 0.31 0.01∗∗∗
Company State HHI 0.50 0.30 0.01∗∗∗
Table 7: Time to exit and probability of exit. Specification (1) reports the estimates of a Cox
proportional hazards model. The dependent variable is the number of days from financing to
the earlier of exit or April 30, 2010. Specification (2) reports the estimates of a probit model in
which the dependent variable is 1.0 if there is a successful exit (IPO or merger) within 10 years
of the investment round and 0 otherwise. See Appendix B for a description of the independent
variables. The sample comprises VC deals obtained from Venture Economics excluding buyouts,
angel investments and non-US deals. All specifications include year and industry fixed effects,
which are not reported for brevity. Both the specifications are overall significant at 1%. t-statistics
based on robust standard errors are in parentheses. ∗∗∗ , ∗∗ , and ∗ denote significance at the 1%,
5% and 10% levels, respectively.

Cox Probit
(1) (2)
Community 1.110 0.068
(3.96)∗∗∗ (3.44)∗∗∗

Early Stage 0.907 -0.047

(-4.63)∗∗∗ (-2.98)∗∗∗

Company Geographical Cluster 1.057 0.033

(2.54)∗∗ (2.05)∗∗

Corporate VC 1.327 0.193

(8.71)∗∗∗ (7.65)∗∗∗

FI VC 1.078 0.050
(3.02)∗∗∗ (2.70)∗∗∗

Syndicated 1.319 0.219

(11.99)∗∗∗ (12.66)∗∗∗

IPO Rate 1.078 0.071

(1.34) (1.65)∗

Centrality 1.022 0.019

(1.90)∗ (2.19)∗∗

VC Geographic Cluster 1.044 0.030

(1.75)∗ (1.68)∗

Experience 0.985 -0.014

(-1.14) (-1.46)

Early Stage Focus 1.050 0.015

(0.57) (0.25)

Industry Focus 0.910 -0.049

(-1.15) (-0.80)

# Observations 30769 32362

Table 8: Success through next round financing or exit. Specifications (1)-(3) report the estimates
of a Cox proportional hazards model. The dependent variable is the number of days from financing
to the earliest of the next financing round, exit, or April 30, 2010. Specifications (4)-(6) report
the estimates of a probit model in which the dependent variable is 1.0 if there is a successful exit
(IPO or merger) or financing round within 10 years of the investment round and 0 otherwise. See
Appendix B for a description of the independent variables. All specifications include year and
industry fixed effects, which are not reported for brevity. The sample comprises VC deals obtained
from Venture Economics excluding buyouts, angel investments and non-US deals. t-statistics based
on robust standard errors are in parentheses. All specifications are overall significant at the 1%
level. ∗∗∗ , ∗∗ , and ∗ denote significance at the 1%, 5% and 10% levels, respectively.

Cox Probit
Round1 Round2 Round3 Round1 Round2 Round3
(1) (2) (3) (4) (5) (6)
Community 1.163 1.139 1.058 0.169 0.222 0.113
(4.53)∗∗∗ (3.40)∗∗∗ (1.29) (4.04)∗∗∗ (4.02)∗∗∗ (1.67)∗

Early Stage 1.332 1.263 1.194 0.261 0.229 0.233

(10.38)∗∗∗ (7.98)∗∗∗ (4.84)∗∗∗ (8.30)∗∗∗ (5.45)∗∗∗ (4.08)∗∗∗

Company Geographical Cluster 1.078 1.002 1.098 0.083 0.046 0.140

(2.68)∗∗∗ (0.08) (2.50)∗∗ (2.48)∗∗ (1.02) (2.59)∗∗∗

Corporate VC 0.891 0.974 0.992 -0.126 -0.023 0.106

(-2.32)∗∗ (-0.58) (-0.17) (-2.06)∗∗ (-0.31) (1.24)

FI VC 0.889 0.927 0.908 -0.142 -0.126 -0.067

(-3.64)∗∗∗ (-2.14)∗∗ (-2.47)∗∗ (-3.78)∗∗∗ (-2.49)∗∗ (-1.06)

Syndicated 1.468 1.293 1.384 0.528 0.505 0.581

(13.89)∗∗∗ (7.63)∗∗∗ (7.87)∗∗∗ (14.70)∗∗∗ (10.39)∗∗∗ (9.48)∗∗∗

IPO Rate 1.242 0.766 0.977 -0.098 -0.303 -0.051

(0.86) (-2.90)∗∗∗ (-0.20) (-1.34) (-2.77)∗∗∗ (-0.34)

Centrality 1.027 1.022 1.051 0.019 0.039 0.093

(1.82)∗ (1.41) (2.72)∗∗∗ (0.92) (1.39) (2.56)∗∗∗

VC Geographical Cluster 1.062 0.974 0.891 0.090 -0.009 -0.098

(2.06)∗∗ (-0.74) (-2.67)∗∗∗ (2.66)∗∗∗ (-0.18) (-1.64)∗

Experience 0.952 0.978 0.944 -0.064 -0.065 -0.045

(-3.32)∗∗∗ (-1.23) (-2.38)∗∗ (-3.88)∗∗∗ (-2.73)∗∗∗ (-1.46)

Early Stage Focus 1.386 1.867 2.005 0.183 0.639 0.552

(3.70)∗∗∗ (5.28)∗∗∗ (4.29)∗∗∗ (1.80)∗ (4.06)∗∗∗ (2.72)∗∗∗

Industry Focus 1.148 0.978 1.003 0.074 -0.007 0.036

(1.53) (-0.19) (0.02) (0.71) (-0.05) (0.18)

# Observations 9471 6218 4320 9867 6522 4563

Table 9: Competing Hazards - Time to IPO or M&A. In Specification (1), the dependent variable
is the number of days from financing to the earlier of exit or April 30, 2010. IPO within the first
10 years of the investment round is the event of interest while merger is the competing risk. In
Specifications (2) - (4), the dependent variable is the number of days from financing to the earliest
of the next financing round, exit, or April 30, 2010. IPO or next round financing is the event of
interest while merger is the competing risk. See Appendix B for a description of the independent
variables. The sample comprises VC deals obtained from Venture Economics excluding buyouts,
angel investments and non-US deals. All specifications include year and industry fixed effects,
which are not reported for brevity. Both the specifications are overall significant at 1%. t-statistics
based on robust standard errors are in parentheses. ∗∗∗ , ∗∗ , and ∗ denote significance at the 1%,
5% and 10% levels, respectively.
Exit Only Round1 Round2 Round3
(1) (2) (3) (4)
Community 1.133*** 1.154*** 1.076 0.992
(0.05) (0.05) (0.05) (0.06)
Early Stage 0.853*** 1.371*** 1.290*** 1.192***
(0.03) (0.05) (0.05) (0.06)
Company Geographical Cluster 1.065 1.052 0.972 1.037
(0.04) (0.04) (0.04) (0.05)
Corporate VC 1.577*** 0.798*** 0.926 0.952
(0.08) (0.05) (0.06) (0.06)
FI VC 1.185*** 0.883*** 0.882*** 0.895**
(0.05) (0.04) (0.04) (0.05)
Syndicated 1.409*** 1.410*** 1.153*** 1.250***
(0.06) (0.05) (0.05) (0.07)
IPO Rate 1.211** 0.786*** 0.817* 0.859
(0.11) (0.07) (0.10) (0.12)
Centrality 1.006 1.009 1.021 1.034
(0.02) (0.02) (0.02) (0.03)
VC Geographical Cluster 1.070 1.040 0.969 0.905*
(0.05) (0.04) (0.04) (0.05)
Experience 1.051** 0.947*** 0.965 0.935**
(0.03) (0.02) (0.02) (0.03)
Early Stage Focus 0.514 1.591*** 1.749*** 2.078***
(0.08) (0.18) (0.26) (0.43)
Industry Focus 1.278* 1.109 1.019 0.958
(0.19) (0.12) (0.15) (0.18)
# Observations 30,757 9,439 6,231 4,339