Manager Networks and Investment Syndication: Evidence from. Venture Capital. Vineet Bhagwat. December 6, 2011 JOB MARKET PAPER.

Transcription

1 Manager Networks and Investment Syndication: Evidence from Venture Capital Vineet Bhagwat December 6, 2011 JOB MARKET PAPER Abstract I explore whether the educational connections between managers of venture capital (VC) rms can alleviate the search frictions and coordination costs experienced by the rms when seeking to form economic ties with other organizations. Two given VC rms are almost three times as likely to syndicate an investment together if their managers share an educational connection. Investments syndicated by VC rms whose managers share educational connections are more likely to receive follow-on funding, more likely to achieve IPO exit, and have a shorter time to IPO. These eects are stronger in early stage investments and for those VC rm-pairs syndicating with each other for the rst time, and do not appear to be due to similarities in manager latent talent. I thank my thesis committee for endless support on this and other papers: Yael Hochberg (Chair), Mitchell Petersen, Joshua Rauh, Paola Sapienza, and Linda Vincent. Without their guidance, this paper would not have been possible. I gratefully acknowledge support from the Coller Institute of Private Equity at London Business School. This paper has beneted from conversations with the Kellogg Finance Baglunch attendees, Jonathan Brogaard, and Jingling Guan for support and feedback. Any errors are my own. Author Aliation: Finance PhD Student, Kellogg School of Management, Northwestern University. Contact v-bhagwat@kellogg.northwestern.edu

2 1 Introduction In a frictionless capital market, establishing and monitoring economic ties with other rms is costless. Search frictions and coordination costs, however, are pervasive, increasing the costs of contracting with other rms. Understanding the methods through which a rm overcomes such frictions is thus important to our knowledge of inter-rm economic interactions. I focus on the role of manager networks in potentially facilitating and adding value to economic ties between rms. Manager networks could impact the contracting environment between rms in several ways. As conduits for information acquisition, manager networks could reduce search costs related with nding potential rm partners (Cohen et al., 2008, 2010). After establishing economic ties, manager networks may also help reduce costs of coordinating eort amongst the parties. On the other hand, manager networks may foster nepotism or reduce the objectivity of the two parties, thereby reducing the economic benets of the tie (Fracassi and Tate, 2012). In this paper, I employ venture capital (VC) rms as the testing ground to assess the impact of manager networks on inter-rm economic ties. Specically, I analyze whether the educational network of managers can impact the syndication decisions of venture capital rms by asking two questions. First, do two VC rms syndicate investments more often if the managers at the two rms are part of the same educational network? Second, do syndicates formed with manager network connections outperform those syndicates where no such network connection exists? To answer the rst question, I analyze the probability and frequency with which two VC rms form a syndicate as a function of the educational connections between their managers, and other controls. I answer the second question by analyzing, all else equal, the probability of IPO and time to IPO of investments made by connected syndicates versus unconnected syndicates. After establishing the main results, I investigate various subsamples of the data to understand some of the key drivers of the results and to test alternate hypotheses. My ndings are most consistent with manager networks adding value by reducing search and coordination frictions between rms. My analysis focuses on VC rms for several reasons. First, VC rms frequently engage in inter-rm economic ties by forming syndication partnerships with other VC rms to co-invest in entrepreneurial ventures (Lerner, 1994). In addition, after forming the syndicate, rms actively manage and nurture the startup company, which requires substantial coordination of eort between 1

3 the syndicate members (Gorman and Sahlman, 1989; Lerner, 1995; Hellmann and Puri, 2002). The contracting environment in which VCs operate, however, is fraught with risk and uncertainty, raising the required search and coordination costs. VCs typically invest in extremely young, high growth rms, where there is much uncertainty and disagreement regarding the viability of the product, market, and management team (Gompers and Lerner, 1999). Furthermore, information regarding the investments is largely soft in nature, and thus dicult to easily verify and transfer among participants that do not have a common context or history (Petersen, 2004). All these characteristics likely heighten contracting frictions, but also make for fertile conditions in which the network of a manager can provide a crucial avenue to mitigate such frictions. I focus on a particular subset of a manager's network connections formed through an educational institution. This measure of a manager's network is also employed by Cohen et al. (2008); Engelberg et al. (2011); Fracassi (2011), among others. One main advantage to this approach is that educational connections are formed many years, even decades, prior to the syndication decisions of interest. Thus concerns of reverse causality are lessened in this context. The results indicate that manager networks facilitate the formation of syndicates and that syndicates with manager educational connections outperform those lacking these connections. After controlling for a variety of factors, I nd that two VC rms are almost three times as likely to syndicate an investment with each other if managers at the two rms are connected by virtue of receiving the same degree from the same educational institution. While the probability of two given rms syndicating in a given year is 0.5% if they lack an educational connection between their managers, the probability of syndicating is 1.4% for two rms with such a connection. In addition, indirect connections, formed through a connection of a connection for example, also have a positive (but smaller) eect on the probability of two rms forming a syndicate. These eects are signicant at the 1% level and hold even when controlling for other known determinants of syndication, such as the relative size, experience, and prior success of the VC rms. The increased syndication due to manager educational connections need not be value increasing, as syndicates formed through manager networks could be due to nepotism or lack of objectivity. The evidence in this paper, however, indicates that manager educational connections are associated with better investment outcomes. Syndicates where a manager-level education connection between the VC rms exists are more likely to receive follow-on funding (a measure of intermediate success), 2

4 exit via IPO (a measure of long-term success), and have a shorter duration between rst investment and successful exit. The eect is economically large. All else equal, as compared to syndicates without educational connections, the likelihood of a successful exit via IPO increases from 12.5% to 17% for syndicates with educational connections. In addition to going public more often, companies receiving money from syndicates with educational connections also go public almost a year sooner: 13 versus 16 quarters. These results are consistent with manager networks increasing inter-rm economic activity, potentially allowing the parties to make value-increasing economic decisions. To shed light on the mechanism for these eects, I examine a number of dierent cuts of the sample. First, I analyze the stages of the investment cycle where educational connections have the most impact. The earlier stages of an entrepreneurial venture is when the risk and uncertainty surrounding the investment is highest, thus requiring more monitoring and eort by the VC to ensure that the investment is on track. If manager networks ease the costs of coordinating eort or transmission of soft information amongst the partner rms, they could be more valuable in the earlier stages of the investment. The evidence provides some support for such a view. Educational connections between syndicate members have a positive marginal eect on the probability of follow-on funding for companies that are in the earlier stages of development and no statistical impact for companies that are in the later stages. However, as the standard errors for the later-stage subsample are large, it is not possible to statistically distinguish the marginal eect of educational connections between the two subsamples. Nevertheless, it is promising that the impact of educational connections on the probability of follow-on funding is statistically zero for later stage investments. Second, I investigate how prior syndication activity between rms impacts the role of educational connections on investment success. The intuition is that educational connections represent just one channel through which managers can be connected. A shared work history can also connect managers. Therefore, if manager connections facilitate and add value to syndicated investments, educational connections should be more valuable in this regard to those rm pairs who do not have a shared work history. The results conrm this intuition. The marginal eect of an educational connection on the probability of IPO is twice as large within the subsample of rm pairs syndicating with each other for the rst time. Among the rm pairs syndicating with each other for a second or more time, educational connections do not have a statistically signicant eect on the probability 3

5 of IPO. 1 Essentially, prior syndication experience with another rm seems to be a substitute for educational connections with that rm. This indicates that both may add value to inter-rm economic ties through similar mechanisms, such as the ability to reliably convey soft information or better coordinate eort. Third, I explore whether the investment benets of educational connections come from a personal knowledge of the other party or from sharing a common alumni network. To do this, I rene my denition of manager connections such that two managers are connected if they received the same degree from the same educational institution and they are of similar age. As those managers of similar age are more likely to have overlapped at the educational institution, this measure of educational connections may capture aspects of a manager's educational network related with personal connections. In unreported ndings, the results are stronger for this rened denition of manager educational connections. The probability of syndicating an investment in a given year and the subsequent probability of IPO is signicantly higher for two VC rms whose managers share the same school/degree and are of similar age, as compared to two VC rms whose managers just share the same school/degree. This indicates that the benets of a manager connection may be stronger when a personal knowledge of the other party is potentially involved. Collectively, the results are consistent with the hypothesis that manager networks add value by opening channels through which rms search for and coordinate with potential partners. However, there could be other possible explanations for the increased syndication and investment success of two rms with educationally connected managers. Two VC rms with a common investment focus are potentially more likely to syndicate with each other and selectively hire managers from certain schools, introducing a spurious relationship between educational connections and syndication between two rms (focus hypothesis). The outperformance of investments by connected syndicates seems inconsistent with this explanation. If manager connections are spurious to the choice of syndication partners due to a common investment focus, they should have little eect on investment success, and I should thus expect no dierence between the performance of connected versus unconnected syndicates. The results discussed 1 Investments made by rm pairs syndicating with each other for a second or more time do perform better, on average, than investments by rm pairs syndicating with each other for the rst time. However, among the rm pairs syndicating with each other a second time, those with educational connections do not outperform those without educational connections. 4

6 earlier, however, are at odds with this prediction: investments by connected syndicates have a statistically higher probability of IPO and lower time to IPO than investments by unconnected syndicates. Moreover, if the investment focus is constant over time, it should be absorbed by the inclusion of a xed eect for each rm-pair, and the coecients of the xed eects model will re- ect within-pair changes in the variables over time. After the inclusion of a xed eect for each rm-pair, manager educational connections are still associated with an increase in the probability of syndication between two given rms. Another potential alternate hypothesis is that manager networks could proxy for latent manager ability (ability hypothesis). VC rms tend to hire disproportionately from top-ranked schools, and as a result, many managers in the sample may come from the right tail of the talent distribution. If this is the case, then it could be possible that syndicates where an educational connection exists are also the syndicates lled with more talented managers. Thus, these syndicates could be more successful not because of the manager's network, but because of the talent of the people involved. I address this concern in three ways. First, the results are qualitatively and quantitatively robust to excluding the top-ranking schools from my analysis. 2 In other words, the ndings are equally valid for managers from lower-ranking school and those from the top-ranking schools. Second, as a falsication exercise, I alter the denition of an educational connection such that two rms are connected, not if their managers attended the same school, but if their managers attended the same ranking category of school. This alternate measure of manager-level connections is not a measure of the manager's network, but a proxy for the manager's ability. If the ability story is true, I should still observe that two rms that are connected in this manner will have better investment outcomes than rms not connected in this manner. I nd that this is not the case: investments made by two rms whose managers attended the same ranking category school are no more successful than investments by two rms whose managers did not attend the same ranking category school. Third, if the ability hypothesis has a rst-order eect, then I would expect syndicates with a manager network connection to always outperform syndicates without any such connection. The results presented in the second part of the paper clearly show this is not the case. When the two rms involved have a shared investment history, syndicates with manager connections do not outperform syndicates without such connections. It is only when the two rms have never been in contact with 2 This nding is also robust to dening top-ranking schools by various methods. 5

7 each other (from a syndication perspective) that educational connections increase the probability of success. These ndings stand contrary to the predictions of an explanation based on managerial talent. As discussed in detail in Section 5, the results of this paper seem inconsistent with other competing explanations, such as homophily, reputational concerns, and selectively inviting connected rms to syndicate attractive deals. The ndings of this paper, however, do not strictly rule out the role of investment focus, managerial ability, or other competing mechanisms. Furthermore, the explored set of alternatives may not be fully exhaustive. However, I believe the preponderance of evidence indicates that manager networks can add value to inter-rm contracts by opening channels to share information and better coordinate resources. To my knowledge, this is the rst paper to analyze how manager networks between venture capital rms can aect the economic ties the rms form. 3 Unlike other economic environments studied in prior work on manager networks, coordination of resources between two parties is important in VC syndicates. A recent, growing literature in nance emphasizes the role of manager networks in facilitating information ow between two parties. Executive compensation, loan terms from banks, analyst stock recommendations, and investment decisions all seem to be aected by information transfers through the social network of managers (Engelberg et al., 2011, 2010; Fracassi, 2011; Cohen et al., 2008, 2010). While information is certainly important, VC represents a unique environment where coordination between syndicate members is critical as well. By analyzing syndication decisions of VC rms, I introduce coordination of eort as another potential mechanism through which manager networks can facilitate and add value to inter-rm economic activity. The nding that manager connections can add value by potentially reducing coordination frictions between VC rms stands in contrast to Fracassi and Tate (2012), who nd that CEO-director network connections weaken the intensity of board monitoring and reduce rm value. As the authors nd this negative relationship is stronger for rms with weaker shareholder rights, it is possible that the contrast in ndings is due to dierences between public rms and VCs in their governance measures. A second contribution of this paper relates with rm-level networks. Firm-level networks can be 3 Bengtsson and Hsu (2010) and Hegde and Tumlinson (2011) analyze the connections between VC managers and the entrepreneurs in which they invest. I analyze the manager connections between two given VC rms. 6

8 a crucial asset to a rm in the sourcing and success of deals (Hochberg et al., 2007, 2010; Sorenson and Stuart, 2001). Firm-level connections allow the rm to not only have superior access to deals, but also the ability to potentially shut out competitors. This paper digs one level deeper to shed light on one potential mechanism behind the formation of rm-level networks. Manager networks serve to increase the likelihood of syndication, thereby allowing the rm to establish economic ties with other rms. Manager networks may thus play a crucial role in facilitating the interplay between rms. The remainder of the paper is organized as follows. Section 2 describes the data and variable denitions. Section 3 analyzes the eect of manager educational connections on the probability and frequency of syndication between VC rm pairs. Section 4 analyzes the eect of manager educational connections on the success of syndicated investments. Section 5 explores alternate hypotheses for the results and Section 6 concludes. 2 Data and Variable Denitions 2.1 Data Venture capital investment data comes from Thomson Financial's Venture Economics database (VE). I concentrate on investments in U.S. companies made by U.S. based VC rms between 1980 and I begin in 1980 as venture capital as an asset class that attracts institutional investors has only existed since then. I obtain IPO information for the venture-backed companies from Securities Data Corporation's (SDC) Global New Issues database. Manager network data comes from Standard and Poor's Capital IQ database, which provides biographical data for top-level managers of around 500,000 public and private companies in the United States and Europe. I restrict my attention to the set of VC rms in the VE database. I am able to obtain employment and educational information on 956 managers in 390 VC rms from the years 1980 to The 390 VC rms participate in 31,501 investment rounds involving 10,314 companies. 46% of the investment rounds and 55% of the companies involve syndicated funding. The sample represents approximately 10% of the universe of rms in Venture Economics. This 4 Throughout the paper, I will use rm to denote the VC and company to denote the entrepreneur's start-up company. 7

9 is due to the fact that biographical information is not available in Capital IQ for managers of the rest of the rms. While the rms in the sample account for a relatively small portion of all venture rms in VE, they account for about 50% of all dollars invested by venture rms in VE. The rms in the sample, as compared to the rms in the entire VE database, have invested three times as much money, invested in four times as many companies, and have six times as much assets under management. In addition, the rms in the sample have raised on average three more funds and raise new funds more frequently. However, there is no statistical dierence between these two set of rms in terms of percent of investments that eventually go public or experience a trade sale. Collectively, the sample statistics indicate that the sample consists of the more prominent, mature VC rms in the database. While this certainly represents a selective sample of all venture rms, any bias created as a result should work against nding a relationship between manager networks and inter-rm economic ties. If manager networks play a role in facilitating inter-rm ties, they should help the less established rms more than the bigger, more established rms. The fact that the sample consists of the more established venture rms should, if at all, bias against nding any signicant relationship. 2.2 Variable Denitions Manager Connections The VE database provides a list of managers that are aliated with each fund within a VC rm and the year the fund was created. Using the fact that the vast majority of funds have a life of ten years, I code each manager as being employed at the VC rm for the ten years a fund with which the manager is associated is in existence. To determine if two managers are connected, I use the educational history of each manager from Capital IQ. While I would like to dene a connection between two managers only if they overlapped at an educational institution, unfortunately Capital IQ does not provide the year of graduation. Thus, I dene two managers as being connected if they received the same degree from the same educational institution. The assumption is that two managers who belong to the same school network are more likely to communicate with each other than two managers belonging to dierent school networks. 8

10 While I do not have the graduation year for the managers, their age is available. As age can provide a rough estimate of when the managers attended the educational institution, in robustness exercises, I rene my denition such that two managers are connected if they received the same degree from the same educational institution and they are of similar age. 5 As those managers of similar age are more likely to have overlapped at the educational institution, this measure of educational connections may capture aspects of a manager's network related with personal connections rather than just belonging to a common social network. The ndings are discussed in Section 5.1. From the manager-level educational connections, VC rm i and VC rm j are considered connected in year t if a manager employed at rm i went to the same school and received the same degree as a manager employed at rm j. Table 1 documents the most common educational institutions, degrees, and institution-degree combinations. Stanford and Harvard make up a large majority of the sample: 30% of the managers in the sample received a degree from one of these two schools. The two most common institution-degree combinations are Harvard MBA and Stanford MBA, accounting for 14% of the managers. Other schools like the University of Pennsylvania and University of California-Berkeley are also sizable, accounting for 9% of the manager educational degrees. Since top ranked schools make up a large portion of the sample, unobservable ability or quality of the managers could play a role in the syndication and success of the investments. I specically address this point with analysis in Section Syndication and Other VC Variables As in Hochberg et al. (2007) and Hochberg et al. (2010), I dene syndication at the investment round level. VC investments are typically staged into multiple investment rounds, allowing the VC to learn more information about the product and market as it is revealed over time, while retaining the option to re-invest or terminate (Gompers and Lerner, 1999). Thus I dene two rms as syndicating if they both invest in the same investment round for a given company. In order to isolate the eect of manager connections on syndication, I control for other known determinants as shown in the prior literature. Broadly speaking, venture capital rms syndicate investments in order to gather information on investment opportunities and get a second opinion 5 For Bachelor's degree, I dene similar age as those managers within 4 years of each other. For other degrees, I dene similar age as those manages within 6 years of each other. 9

11 on their initial investment (Lerner, 1994; Brander et al., 2002). Prior work shows that VCs tend to syndicate with other rms of similar size, past performance, and experience (Lerner, 1994). Therefore, I build measures of relative rm size, experience, and success. One measure of rm size is the total assets under management (Assets Under Management) by a rm in a given year, calculated as the sum of the size all its active funds. Assets Under Management thus captures the total amount of money available to the rm to make investments. Related to such a measure is the total dollars invested by a rm from inception to the current year (Total $ Invested). Firms that are more experienced or successful, are more likely to raise larger amounts of money in subsequent fund raising (Kaplan and Schoar, 2005), and therefore can invest more money in the coming years. Total $ Invested thus not only captures past size, but also aspects of rm experience and success. A more tangible measure of success in VC is whether the investment exited via IPO or sale. For each rm in a given year, I calculate the percentage of past investments that exited via IPO or sale up to and including the current year (%IPO or Mergers). 6 The relative investment experience of each rm is also a potential determinant of syndication. As investments tend to require substantial monitoring, some VCs specialize to a certain industries and invest in physically close companies (Sorenson and Stuart, 2001). As one motivation for syndication is to learn more information about an investment, rms that are experts in an industry or local area may be sought after as potential syndication partners. To capture this eect, I calculate a Herndahl measure of the concentration of a rm's past investments into the 50 U.S. states (State HHI ). For industry concentration, I calculate a similar measure (Industry HHI ) using the six broad industry categories dened by the VE database: Biotechnology, Communications and Media, Computer Related, Medical/Health/Life Science, Non High-Technology, and Semiconductors and Other Electronics. 7 Given the importance of industry and local knowledge, rms may seek potential partners for an investment within the set of rms that have invested in that particular industry and state in the past. 6 Additional measures of size and experience include number of companies invested, number of funds raised, and age of the rm. Results are robust to inclusion of these additional measures. In the interest of brevity, I present results using the three measures mentioned. 7 To give an idea of the industry classication, I document some popular companies that were rst nurtured and developed through the venture capital system. Tivo, Inc., a leading innovator and producer in digital video recording, is classied under Communications and Media. Electronic Arts (EA), a developer of popular computer and video game software, was funded through the venture capital industry in the 1980s and falls under Computer Related. Finally, classied under Non High-Technology, is Staples, Inc. a major retailer of oce supplies. 10

12 For example, if Omega Ventures is looking to nd a syndicate partner for an biotechnology startup in Illinois, they may rst seek out other venture rms that have invested in Illinois biotechnology startups in the past. In addition, search and coordination costs may be lower for two rms that have both invested in the same industry and state in the past. To capture the intuition that rms in the same market may be more likely to syndicate with each other, I dene an indicator variable I(Both Firms Invested in Same Market) that takes the value of 1 if in the prior ten years, rm i and rm j made investments in the same market. A market is dened as the combination of a particular industry and state. For example, investments in Texas semiconductor companies are a distinct market from Texas telecommunications companies or California semiconductor companies. I also include two further controls for manager networks. If manager networks indeed facilitate economic interactions between rms, then VCs that employ managers with more educational connections (to any rm) may have dierent syndication behavior than VCs with managers with fewer educational connections. I calculate for each rm in a given year, the total number of manager educational connections to all other managers of any rm in the sample (# of Firm Ties). For example, if rm i employs only one manager and she has 2 connections to rm j and 3 connections to rm k, then # of Firm Ties will equal 5 for rm i in that year. Additionally, a rm may be likely to syndicate if it has exhibited a propensity to syndicate in the past. I control for this by calculating the total number of investments a rm has syndicated in the past (# of Syndications). Table 2 gives summary statistics for the above variables. The unit of observation is a VC rmyear. As there are 390 rms and approximately 12 years of observations per rm, there are 4,730 rm-year observations. On average, a rm manages $178 million per year and has invested $951 million in its lifetime. Note that these values are highly skewed, as the median for the respective variables is $31 million and $260 million. A tiny fraction of VC rms, thus, manage and invest a substantial portion of the total VC investments. The average VC rm is relatively concentrated to a particular industry and region, as the mean industry and state Herndahls are 0.5. The majority of a VC's prots come from small amount of deals: 16% of a VC's investments exit successful via IPO or sale, while the remaining are written o and liquidated. Finally, the table shows that VCs frequently syndicate, as the average rm has syndicated 63 deals. Given that in the sample, an average rm invests in about 80 companies over its life, a signicant fraction of a VC's investments are syndicated. 11

13 Lastly, in addition to the above variables related to syndication, I use Venture Economics' classi- cation to categorize an investment as belonging to one of three stages: Seed/Early, Expansion, and Later. In Section 4.3, I examine the relative success of connected versus unconnected syndicates across the dierent stages of the investment cycle. The early/seed stage consists of the VC providing a small amount of capital used to build a prototype and prove the idea is a viable one and not a gment of the entrepreneur's imagination. If the concept is proven to the VCs liking, the company moves through to the expansion stage, where the company focuses on applying working capital to start and expand production, marketing, and development. Thus, the role of the VC shifts from support to strategy. The nal stage of investment is late stage, where the company has reached a fairly stable growth rate and is looking to go public in the near future. The company has a proven track record of sales and growth and is now shifting focus to how to best structure itself for a public oering (Metrick 2007). 3 Do Manager Connections Facilitate Syndication? 3.1 Empirical Setup In order to investigate the role of manager networks in facilitating inter-rm economic ties, I rst analyze the propensity of two VC rms to syndicate an investment in a given year as a function of their manager educational connections, and other known determinants of syndication. For each year of the sample, I construct unique pairs between all active rms. A rm is classied as active for all years after, and including, the year of its rst investment. 8 Thus, each observation is identied as a rm i, rm j, and year t combination. I dene two measures of syndication activity between rms. The variable, I(Pair Syndication), is an indicator that takes the value 1 if VC rm i and VC rm j syndicate an investment in year t, and 0 otherwise. Additionally, the variable # of Syndications measures the number of times VC rms i and j syndicate an investment with each other in year t. I will employ both these measures as dependent variables in analyzing whether educational connections facilitate syndications. For the independent variable of interest, I calculate three measures of manager-level connections 8 The results are robust to alternate classications, such as a rm is only active in the years where it makes at least one investment. 12

14 between two VC rms. Briey, they are the following: an indicator measuring whether an educational connection exists between two rms, a discrete variable measuring the number of educational connections, and a discrete variable measuring the network distance between two rms. I rst dene an indicator variable, I(Manager Tie), equal to 1 if in year t, a manager working in VC rm i received the same degree from the same educational institution as a manager working in VC rm j, and 0 otherwise. The rst independent variable of interest, I(Manager Tie Last 3 Years), takes on the value 1 if I(Manager Tie) equal 1 in any year between t 1 and t 3, inclusive, and 0 otherwise. The primary reason for using values over the last three years is that the advantage of a newly established manager-level connection may take a few years to fully develop. Two rms may not start working together and sharing information immediately, or even if they do, may take a while to nd a suitable investment to syndicate. All the analysis and results are robust to using longer or shorter windows to measure the manager-level connections between two rms (current year, past 1 year, and past 5 years). To measure the strength of manager connections between two rms, I dene Normalized # of Manager Ties as the number of connections between managers in VC rms i and j in year t, divided by the total number possible. For example, suppose rm i employs 2 managers and rm j employs 5 managers, making a total of 10 possible manager connections between the two rms. If one of the managers at rm i received the same degree from the same school as all ve managers at rm j, then the value for Normalized # of Manager Ties would be 1/2. The second independent variable of interest, ln(1 + Normalized # of Manager Ties Last 3 Years) is the natural log of one plus the average normalized connections between rms i and j over the years t 1 to t 3, inclusive. I take the natural log to allow for a decreasing marginal eect of each subsequent connection between rms and to remain scale-invariant. The third and nal measure of manager-level connections is a discrete variable for the network distance between two rms, Network Distance. The rst two main independent variables treat all managers who did not receive the same degree from the same school as completely unconnected. However, not all unconnected managers are equally unconnected. For example, two managers who are friends of a friend could be considered more close in connection than two randomly selected managers. If manager networks matter for investment decisions, indirect ties, or friends of friends, 13

15 should also have an impact. For example, suppose Adam and Bill received their Bachelor's degree from Stanford and Adam and Charles received their MBA degree from Kellogg. Bill and Charles are thus friends of a friend (through Adam) and have a network distance of 2 (Bill needs to take two steps to reach Charles). Bill and Charles are possibly more likely to communicate with each other than two randomly selected managers. To capture this eect, I dene Network Distance as the shortest path between two rms through their managers' educational ties, i.e. the minimum network distance between any two managers at the respective rms. Thus, if rm i employs Bill and rm j employs Charles, and no other employees of the two rms are connected, Network Distance would take the value 2 for the rm pair in year t (since the distance between Bill and Charles is 2). I take the average of Network Distance between rms i and j over the prior three years ([t 1, t 3]), and employ this average as my third independent variable of interest. 3.2 Univariate Analysis Table 3 gives summary statistics for the main dependent and independent variables. When interpreting the statistics, it is important to keep in mind that each unit of observation is a unique rm i, rm j, year t combination. Firms only appear in the dataset starting with the year of their rst investment. Given that there are 390 rms and on average, twelve years of data for each rm pair, there are 462,692 observations. 9 Approximately 1.9% of the VC rm pairs syndicate an investment in a given year of the sample. Additionally, a given VC rm pair syndicate 0.03 investments together on average per year. This does not mean that VC rms rarely syndicate. In fact, 45% of a VC rm's investments are syndicated (not reported in table). The fact that there are 390 rms in the sample means that the likelihood of syndicating with a given rm in a year is quite low. Approximately 24% of the VC rm pairs share a manager-level educational connection, and have 0.52 such connections between them over a given three year period in the sample. Furthermore, 11% of the possible manager educational connections between two rms are actually realized. Note that if rm i employs 2 managers and rm j employs 5 managers, there are a total of 10 possible manager connections between the two rms. If one of the managers at rm i received the same 9 In a given year, if there are N active rms, then there will be (N 1) + (N 2) = N (N 1) 2 unique rm pairs in that year. The total number of observations will depend on the number of years each rm is active, which is dierent for each rm. 14

16 degree from the same school as all ve managers at rm j, then the value for Normalized # of Manager Ties would be 1/2. Lastly, the minimum network distance between two given rms in a year is, on average, 1.9. This implies that on average, the managers of two rms are connected through a connection of a connection. Table 3 also presents a univariate t-test of the observed syndication between connected and unconnected VC rm pairs. Of the VC rm pairs that do not share a manager-level education tie over the past three years, 1.2% syndicate an investment together in the current year. This percentage jumps to 4.1% for those VC rm pairs that do share a manager-level education tie over the prior three years. This dierence is statistically signicant at beyond the 1% level. Furthermore, rm pairs that do not have a manager-level education tie syndicate an average of investments per year while those that do have manager connections syndicate investments per year. This dierence is also statistically signicant beyond the 1% level. Thus, in univariate analysis, the observed rate of syndication is signicantly higher in the years when two VC rms share a managerlevel education tie than in years when they do not. 3.3 Manager Connections and the Probability of Syndication Between Two Firms Univariate analysis indicates that rms with manager-level connections syndicate together more often. This section will analyze whether this observed relationship holds after controlling for other known factors that aect syndication. For each year in the sample, the unit of observation is a rm pair. Thus for the set of controls in the regressions, I employ the absolute value of the dierence in the variable between the pair. I estimate probit models where the dependent variable is an indicator for whether rms i and j have syndicated an investment in year t (I(Pair Syndication)). I estimate these models separately for the three independent variables of interest: an indicator for the presence of a manager-level connection between the two rms in the prior three years (I(Manager Tie Last 3 Years)), the natural log of one plus the normalized number of manager-level connections between two given rms (ln(1 + Norm. # of Manager Ties Last 3 Years)), and the network distance between two given rms (Network Distance). The results are reported in Table 4. Following Petersen (2008), all 15

17 columns cluster the standard errors two-way by each rm separately. 10 In addition, all regressions include year controls. Firms located geographically close to each other are more likely to syndicate with each other. The probability that two rms located in the same postal zip code syndicate an investment with each other in a given year is 1.3% versus 0.6% for two rms that are not in the same zip code. A similar magnitude eect is seen for those rms located within 100 miles of each other. In addition to geographic proximity, the results suggest that investment proximity is also an important predictor of syndication. All else equal, the probability that two rms that have invested in the same VC market (industry and state combination) in the past will syndicate with each other in a given year is 1%, whereas the probability is 0.1% for those rm pairs that have not. The coecients on the absolute dierence variables indicate that two rms are more likely to syndicate with each other if they have similar levels of industry and state investment concentration, and past investment success (as proxied by IPO and sales). However, rms seem to seek out partners that are dierent along the dimension of rm networks: syndication history and amount of manager-level ties to other rms. This would be consistent with a resource sharing motive for syndication, whereby rms seek out partners with complementary resources to their own that they lack, as suggested by Hochberg et al. (2011). In the rst column, the main independent variable of interest is an indicator for the presence of a manager-level education connection between the two rms over the prior 3 years (I(Manager Tie Last 3 Years)). The predicted probability of syndication is positively associated with the presence of a manager-level education tie at some point over the prior three years, and is statistically signicant at the 1% level. All else equal, the predicted probability of syndication between two rms is 0.5% if the two rms do not share a manager-level education connection, while it is 1.4% if the two rms do share such a connection. Eectively, a manager-level education tie almost triples the predicted probability of syndication between two rms in the current year. The second column investigates whether the strength of connections between the two rms impacts the syndication probability. The main independent variable is the natural log of one plus the 10 That is, standard errors are allowed to be correlated for all observations where either rm i or rm j is one of the syndicating rms. In unreported analysis, I also cluster by pair, where standard errors are allowed to be correlated only for observations where rm i and rm j are the two syndicating rms. As the two-way cluster results in generally higher standard errors, I take a conservative approach and report results using two-way clustering. 16

18 normalized number of manager-level connections between rm i and rm j over the past three years. The strength of manager-level education ties between two rms is associated with a signicantly higher probability of syndication. A one standard deviation increase in the log number of ties from the mean value results in an increase in the predicted probability of syndication between the two rms by 0.5%. Compared to the predicted probability of 0.7% at the mean value, this appears to be an economically signicant increase. The rst two columns of Table 4 indicate that syndication is positively associated with the presence and number of manager-level connections between two rms. However, not all unconnected managers are equally unconnected. For example, two managers who are friends of a friend could be considered more close in connection than two randomly selected managers. To investigate such a relationship, the main independent variable in the last column is the minimum network distance between managers at rm i and those at rm j (Network Distance). Higher minimum network distance between two rms is associated with a statistically and economically signicant lower predicted probability of syndications. At the mean values, the marginal eect of an decrease in network distance is approximately 0.5% and is signicant beyond the 1% level. All else equal, connected rm pairs (network distance of 1) have a predicted syndication probability of 1.3%, while the same for rm pairs whose managers are friends of a friend (network distance of 2) is 0.7%. Given that the mean syndication probability in the whole sample is about 2%, this represents an economically meaningful decline. It is important to note that even though manager educational ties are positive predictors of syndicating with a particular venture capital rm, the impact of other characteristics, like size or being located in the same zip code, have an equal or larger impact on the selection of syndication partner. While the impact of manager educational ties are relatively small compared to other characteristics, the marginal eects from Table 4 indicate that the economic impact is meaningful. It is possible that the results in Table 4 overstate the role of manager educational ties on the choice of syndication partner because the allowed set of potential candidate partner rms is too large. If some of the rm pairs within the large set of possible pairs are irrelevant alternatives, then the standard errors estimated in the rst three columns may be too small, thus overstating the eect of manager connections. I re-estimate the probit model and restrict the analysis to a plausible set of rm pairs in two ways. The rst method restricts the analysis to the set of rm pairs that have 17

19 invested in the same market (industry and state combination) at some point over the last ten years. The second method restricts the analysis to the set of rm pairs that syndicate at some point in the twenty-eight year sample period. While I do not report the results here for sake of brevity, all ndings are quantitatively and qualitatively robust to these restrictions. 3.4 Manager Connections and the Number of Syndications Between Two Firms If manager networks facilitate inter-rm economic ties, then I may observe not only an increase in the likelihood of syndications, but also in the number of syndications when two rms share a manager-level education connection. To investigate this, I run a negative binomial regression where the dependent variable is the number of times rm i and rm j syndicate any investment in year t, results reported in Table 5. Results are robust to a Poisson or Tobit specication. A Poisson specication assumes equal values for the conditional mean and variance of the dependent variable, while a negative binomial makes no such restriction. As the variance of the number of syndications per year, conditional on a manager-level connection, is twice as large compared to the conditional mean, I employ a negative binomial specication. 11 I estimate the negative binomial model separately for the three independent variables of interest: an indicator for the presence of a manager-level connection between the two rms (I(Manager Tie Last 3 Years)), the natural log of one plus the normalized number of manager-level connections between two given rms (ln(1 + Norm. # of Manager Ties Last 3 Years)), and the network distance between the two given rms (Network Distance). Standard errors are clustered two-way separately by each rm. In addition, all regressions include year controls. As in Table 4, rms in geographic proximity tend to syndicate with each other more often than those located farther apart. In addition, rms that invest in the same VC market also tend to syndicate with each other more often. Echoing the probit results in Table 4, rms syndicate more often with other similar rms on the dimensions of investment experience, while syndicating less often with rms of similar network characteristics. All else equal, the presence of a manager-level education connection is associated with a doubling of the numbers of syndications per year between two given rms. This increase is signicant at the 11 The mean and variance of the number of syndications is and 0.03 with no manager-level connection and and 0.14 with a manager-level connection. 18

20 1% level. The predicted number of syndications per year between two rms without a manager-level education connection is while the same for rms with such a connection is This indicates that manager educational connections are associated with not only an increase in the probability of syndication between two rms, but the frequency of syndication as well. While the presence of a manager connection increases the number of syndications, the second column investigates whether the strength of connections increases the number of syndications. The eect is strong both statistically and economically. A one unit log increase in the number of managerlevel education connections is associated with a 0.7 log unit higher syndications per year between the rm pair. This is economically signicant, as this increase represents a 1.97 times increase in the number of syndications in a year between two given rms. The last column investigates whether indirect connections also aect the number of syndications between two rms. A one unit increase in the network distance between two given rms is associated with a 0.6 log unit decline, or a 46% decline in the number of yearly syndications between the two rms, and is signicant at the 1% level. Thus the results suggest that the frequency with which two rms form syndicates is positively aected by the manager educational connections between the two rms. Similar to Table 4, it is possible that the results in Table 4 overstate the role of manager educational ties on the choice of syndication partner because the allowed set of potential candidate partner rms is too large. I re-estimate the negative binomial model and restrict the analysis to a plausible set of rm pairs in the two ways as described at the end of Section 3.3. While I do not report the results here for sake of brevity, all ndings are quantitatively and qualitatively robust to these restrictions. 3.5 Fixed Eects Specications to Account for Omitted Variables The results provide support for the hypothesis that both direct and indirect manager connections facilitate inter-rm contracts. It is possible, however, that the set of control variables do not account for all factors that aect syndication between two rms. If an omitted factor is correlated with the measure of personal connections, then the observed relationship between personal connections and syndications could be spurious. In short, the results presented could suer from an omitted variable problem. For example, a potential omitted variable could be the preferred investment style or thesis 19