Startup Accelerators and Ecosystems: Complements or Substitutes? 1

Transcription

1 Startup Accelerators and Ecosystems: Complements or Substitutes? 1 Daniel C. Fehder MIT Sloan September 22, 2015 In this paper, I explore how a startup's founding regional environment impacts its performance and whether admission to a startup accelerator changes that relationship. Entrepreneurs cobble together resources from numerous sources as they build their new firms, but they are constrained by their social and geographic proximity to these resources. I use this insight as a starting point to explore whether accelerators act as a complement to or substitute for prior access to resources. Using data from MassChallenge, a leading startup accelerator in Boston, this paper uses a regression discontinuity framework to evaluate both the overall impact of the program on its portfolio of startups and its heterogeneity based on: 1) the level of entrepreneurship resources in a startup s founding neighborhood and 2) their ability to access those resources. Startups birthed in neighborhoods with higher levels of entrepreneurial resources are associated with higher quality ideas, but not higher success conditional on the quality of their idea. Admission to the MassChallenge accelerator seems to change this relationship, differentially improving the ability of entrepreneurs from resource rich regions to reach key milestones. This finding suggests that accelerators change the way in which startup founders are able to access and leverage the resources in their home geographies. PRELIMINARY AND INCOMPLETE. NOT FOR CIRCULATION 1 I thank Fiona Murray, Scott Stern and Yael Hochberg and Ezra Zuckerman for their guidance. I also thank John Harthorne and Cory Bolotsky at MassChallenge for the opportunity to work with them and access their data. Financial Support from the Ewing Marion Kaufmann Foundation is gratefully acknowledged. The Faults are mine alone. Questions and comments are welcome: dfehder@mit.edu 1

2 1. Introduction At the beginning of a venture, startup founders build upon prior social and economic ties to gain critical information and resources. Indeed, strong advantages are conferred to founders with strong social networks and ties to critical resource holders (Aldrich, 2005; Shane and Cable, 2002; Hallen, 2008). These networked advantages not only provide individual advantage, but also generate inequality at the individual and group level (DiMaggio and Garip, 2012). In assessing the relative importance of social capital, it is difficult to fundamentally separate the quality of a venture s founder and their idea from the quality of their social capital because high quality individuals are more likely to form connections with other high quality individuals (Mouw, 2006). Such a question is of great significance to entrepreneurs, policy makers and scholars: If networked social capital serves such a critical purpose for early stage ventures, how might startup firms respond to an externally-provided increase in their networked social capital? Would certain startups benefit more from an increase in the number of experts and resource holders from whom they may create new ties? To what extent would the benefits of such interventions be mediated as a function of their existing social capital? While there is a generous literature on the importance of advantaged network positions within an industry for both individuals and firms (Burt, 1992; Stuart, 2000), there is less of an understanding of how these privileged positions are reached and the extent to which they can be engineered (Dhanasai and Parke, 2006). While recent research has explored the extent to which individuals deploy networked social capital instrumentally for economic gain (Obukova and Lan, 2014), less work has explored the extent to which networked social capital can be effectively created with the intention of near-term instrumental gain and even less that examine the impacts of purposefully designed programs intended to impact an actors social network (Srivastava, 2015). A program intended to improve a startup's stock of social capital will only have an impact if an externally provided "rolodex" of potential contacts can be transmuted into a valuable web of information and resource providers. Prior research has demonstrated that social capital is fixed in both social and geographic space. Because propinquity drives the efficiency of tie formation (Hallen and Eisenhardt 2012), the capacity of an entrepreneur to activate their individual social capital decays with geographic distance to alters, and distance has been demonstrated to drive the exchange of information in 2

3 economic settings beyond entrepreneurship (Almeida and Kogut 1999; Whittington, Owen- Smith, and Powell 2009). The micro-processes of strategic networking and activating social capital have been implicated as one of the key driving forces in the creation and maintenance of startup clusters (Sorenson and Stuart 2005; DiMaggio and Garip 2012; Dahl and Sorenson 2013). Thus, the resources and expertise available in a startups home region are likely to be key differentiators in determining startup performance. While frequently cited as the source of performance advantage of regions, there is little work that concretely explores the connection regional social capital, and startup performance. Even if they have abundant resources available, not all regions are created equal. Qualitative comparisons of regions have yielded observations of differences in the willingness of entrepreneurs and engineers in different regions for sharing technical and market information (Saxenian, 1992). While research in economics has tied this proliferation of knowledge to labor market institutions which allow the free circulation of workers between companies in California (Marx, et al., 2008; Rebitzer, 2005), recent work has also explored how the presence or absence of generalized social capital impacts the performance of regions in terms of entrepreneurial performance (Laursen, et al., 2012; Aldrich, et al., 2014; Samila and Sorenson, 2013). This growing literature connects more broadly to literature on how institutions in an industry or region shape the formal and informal channels through which firms gain access to cutting edge information (Murray, 2002; Owen-Smith and Powell, 2002) and shape that access into competitive advantage (Stuart, 2000). While variation in generalized social capital is broadly linked to performance at the regional level, we still have little knowledge of whether generalized social capital or can be mobilized by the catalyzing actions of individuals or organizations (Feldman, 2001; Fisman and Khana). Taken together, the literature suggests two questions: 1. What is the direct impact of an externally-provided increase in the potential networked social capital on the ability of a new venture to reach key milestones 2. How is the impact of such an external increase mediated by a new venture's founding region? Specifically its level of resources and its social capital. 3

4 This paper examines these questions through a novel network intervention of startup firms and startup ecosystems: Startup accelerators. Startup accelerators are fixed-term, cohortbased programs for startups that have emerged rapidly over the past decade. They provide startups with a time-delimited opportunity to access a rich set of contacts and potential social capital from an expanded set of mentors and experts. In this context, we can therefore ask: holding the ex-ante quality of the startups opportunity constant, to what extent do accelerator programs overcome ex ante differences in network quality and act as substitutes for entrepreneurs from regions with less quality networks. Alternatively, does time-delimited access to resources amplify existing differences amongst startups allowing firms with richer networked resources to grow larger (i.e. act as complements)? In this paper, we use detailed administrative data from one startup accelerator, MassChallenge, to explore these issues. Making use of the admission process of the program, described in detail below, this paper's core results make use of a fuzzy regression discontinuity framework extended to allow treatment heterogeneity at the border (Lee and Lemeuix, 2004; Becker, 2014). If, as is the case in our data, there is a discontinuous jump in the likelihood of being admitted to MassChallenge at the judge's score corresponding to the capacity limit of the program (the rank cut-off), yet there is continuity in the distribution of other key variables at that cut-off, we can use the regression discontinuity framework to explore not only the overall treatment effect of the program but its heterogeneity across a set of key inputs. Exploring treatment heterogeneity allows us to move beyond simple measurement of MassChallenge's impact on its portfolio startups and allows us to understand how this treatment effect is connected to broader social and economic processes in which the startups are enmeshed. History is littered with examples of failed initiatives to facilitate entrepreneurship or clusters (Lerner, 1999), yet researchers have begun to note that the broad and sustained experimentation in entrepreneurship programs at the national state and local level provide an opportunity to use natural variation and program design to explore the basic building blocks of early stage entrepreneurship (Chatterji, et al., 2013). In what ways can accelerators, as a new, entrepreneurled intervention into entrepreneurial firms and ecosystems help us understand how to improve the process whereby firms are born and grow? 4

5 Our results suggest that participant startups in the program experience significant variation in the impact of the program depending on the characteristics of the home region in which they were founded. Using a novel measure of entrepreneurial social capital, we find that startups that were initially founded in regions with strong generalized entrepreneurial social capital experience a larger benefit to the MassChallenge program in terms of reaching key early milestones. Similarly, we find that startups that were founded in regions with higher levels of early stage investors benefitted more from the program. Taken together, our results suggest that the MassChallenge program acts as a complement to a startup s prior advantages in terms of proximity to key entrepreneurship resources and access to generalized entrepreneurial social capital. In addition, we find evidence that startups founded in regions with higher access to early stage investors had on average higher quality ideas, but that their chances of success where not higher conditional on the quality of their idea. 2. Theoretical Framework Firms admitted to MassChallenge have to be at an early stage of their development, having reached neither substantial revenue nor fundraising. At this early stage of development, startups have not completely determined all the elements of their market offering and they have not built a reliable organization to address this opportunity. Indeed, the startups in our sample are almost entirely composed of founders. In this sense, these startups are susceptible to the liability of newness (Stinchcombe 1965). In just four months, MassChallenge changes the likelihood of the admitted startups reaching significant milestones compared to startups that are not admitted. In this section, we will consider various theoretical lenses through which to view this programperformance connection. In line with Stinchcombe s theory on the liability of newness, we would outline two broad classes of potential impact: 1) improvements to the internal capabilities of the startups and 2) improved capability for the startup to form meaningful external relationships. Of course, the internal/external divide is somewhat artificial startup performance derives from the interplay of the two. To date, however, much of the work on the impact of startup accelerators has interpreted the impact of startup accelerators as evidence of improvements in a startup s internal performance, especially learning (Hallen, Bingham, and 5

6 Cohen 2014). While the learning effects within the accelerator may be substantial, singular focus on the treatment within the accelerator might obscure the broader impact of accelerators: by substantially changing the way in which startups are able to make use of their existing social capital. Of course, one of the most important conditions for entrepreneurial performance is a good idea. While a potentially trite observation, studies of the impact of a program like accelerators are bedeviled by the potential that treated firms are simply better ex ante in terms idea, team, or both. Many accelerators and their portfolio firms were founded in areas (like Silicon Valley and Boston) that are rich in both entrepreneurial resources and new technological breakthroughs. Indeed, a substantial literature demonstrates that a founder s starting position in a social and industry structure will substantially influence their ability to recognize a new opportunity and their propensity to enter into a new venture conditional on recognizing that idea. For a given technological innovation (broad opportunity), there is substantial variation in the potential markets that can be addressed. Different choices can yield different outcomes of the venture, even if the differences among these choices can be difficult ex ante (Shane 2000). Broad differences have been observed in terms of the likelihood of employees from different firms to enter into entrepreneurship (Gompers, Lerner, and Scharfstein 2005). Thus, we argue: Hypothesis 1: Startups in regions with higher startup resources will have higher quality ideas on average After finding a worthy entrepreneurial opportunity and deciding to start a new firm, entrepreneurs still face a good deal of uncertainty about how to pursue the opportunity. At the very beginning of the venture, a startup is a collection of early founders, each with human capital to pursue (at least a portion of) the idea and a social network that can be tapped to draw in both market information, technical information about the startup process and also gain potential resources. As described above, startups can improve their life chances by either improving the internal workings of the company or improving their ability to interact with external partners. Accelerators programs address both of these channels. Very little evidence exists to support the notion that business acumen and education accelerates startup success. Huber, Sloof and Van Praag (2012) explore the effectiveness of early 6

7 entrepreneurship education, specifically, primary school entrepreneurship education in the Netherlands. They find a significant impact of entrepreneurship education on non-cognitive skill formation. Unfortunately, they cannot identify whether their measures of skill formation translate into future success. Oosterbeek, Van Praag, and Ijsselstein (2010) use an instrumental variables approach to examine an entrepreneurial education program for college students. They conclude that the education program did not increase self-assessed entrepreneurial skills. Of course, the structure of accelerator programs is substantially different from educational programs intended to impart general skills. While the details across accelerator programs can vary, it is most common for there to be a substantial period at the beginning of the program where startup founders disengage from the day to day operations of their business (Cohen 2013). During this period, their time is spent explaining their business idea and receiving constructive feedback about it. While early-stage startups can easily select incorrect routines and tactics, this liability is counterbalanced by the flexibility inherent in their organization to respond rapidly to new information to change the details of their business plan in ways that could substantially improve the life chances of their startup (Posen and Chen 2013). Thus, evidence of increased performance from startups admitted to accelerators has been interpreted as evidence of increased learning (Hallen, Bingham, and Cohen 2014). Of course, a substantial amount of a startup s time after graduation from the accelerator will be spent engaging with external parties, whether customers or resource providers. Through their open application process and publically available admissions data, accelerators also provide an important quality signal to external parties that can change future negotiations. Prior research has demonstrated that other endorsement signals, like the status of a venture investor, has an important impact on a startup s future performance (T. E. Stuart, Hoang, and Hybels 1999). Accelerators provide an endorsement for entrepreneurial ventures before venture investment where potential partners previously only had status signals of individual founders to infer the quality of the startup venture. Admission to a startup provides clear certification that experts have judged the quality of both the idea and the team to be high. It is not surprising, therefore, that startup founders frequently show their affiliation with top accelerator programs in a number of public forums like Linkedin, Angellist, and company websites. Based on these two channels, we thus argue: 7

8 Hypothesis 2: Admission to an accelerator program improves startup performance At the individual startup level, a good deal of work has been done to elucidate the mechanisms whereby individual startup founders are able to translate their individual social networks into ties for their new startup. Perhaps most unsurprisingly, entrepreneurs with ties to key resource providers like early stage investors are more likely to be able to meet early stage milestones. Across a number of studies utilizing both quantitative and qualitative evidence, research has demonstrated that founders with direct ties to investors prior to the establishment of a venture experienced lower rates of failure and were able to secure financing at higher rates than those without (Shane and Stuart 2002; Shane and Cable 2002; Hallen 2008). Evidence for the importance of indirect ties was more mixed with some studies showing that indirect ties were and important source of performance advantage (Shane and Stuart 2002) while others found no effect. At the core of these mixed results is the difficulty of converting an existing tie from a cofounder into direct support and facilitation of the new venture. High quality resource holders, whether they are investors or industry experts, face severe time constraints as they choose which startups to help and which to ignore. Prior research has demonstrated that participants in early stage ecosystems are often reticent to engage in too many relationships for fear of not only losing time but also diluting their reputation (Hallen and Eisenhardt 2012). At the heart of this phenomenon is the difficulty in evaluating early-stage ideas. Because the evaluation of early stage ideas is fraught with uncertainty, founder characteristics can act as proxies for quality in the resource allocation of key resource providers (Shane and Cable 2002). Formal membership in alumni networks, for instance, unlocks informal networks with key information as well as the capacity to generate other formal ties. It is also important to note that the majority of the studies focused on the impact of networks and networking strategies take place in regions with high levels of entrepreneurial resources. In places like Silicon Valley and Cambridge, starting networks and status characteristics are important attributes for founders seeking to form connections because there is a substantial amount of congestion. Thus status characteristics like a founder's alma mater serve two purposes: 1) they are a signal of quality and 2) they represent very real sources of potential relationships. In this capacity, local alumni networks, rather than alumni networks more broadly, have been shown to 8

9 significantly impact the likelihood of entrepreneurial entry (Kacperczyk 2013). Conditional on entry, the impact of geographically and socially proximate alters significantly improves entrepreneurial performance (Kalnins and Chung 2006). While an entrepreneur could theoretically seek information for any socially proximate alter, extreme time pressures that are at the heart of the liability of newness force them to seek help from nearby compatriots. The related but distinct mechanism of propinquity also reinforces the potential to activate and strengthen information exchange with nearby compatriots (Allen, et al. 1977). The informal exchange of information within a constrained region has been the posited mechanism in a broad range of theories in both innovation and entrepreneurship. Studies at the regional level assume have stressed the highly localized flow of technical and market information as a source of clustering of choice behavior ranging from patent citations to colocation of firms from the same industry (Jaffe, Trajtenberg, and Henderson 1993; Arzaghi and Henderson 2008). Similarly, studies noting the localization of venture capital investment have focused on monitoring and as the main driver of geographic choice of investments (Hellman 2007; Kovner, et al. 2011; Sorenson and Stuart 2001). Beyond personal networks, an entrepreneur is constrained by the broader openness of participants in a regional ecosystem to engage with them. Samila and Sorenson (2013) show that variations in the level of various measures of social capital in a region mediate the impact of variations in the supply of venture capital in terms of employment and firm founding rates. This work builds upon this work connecting general social capital at a regional level to the intensity of entrepreneurship. Other researchers have found a similar connection between regional social capital and the performance of entrepreneurs (Ruef and Kwon 2015). The institutional arrangements that facilitate (and sometimes hinder) the formation of useful connections can have a dramatic impact on the rate and direction of innovation and entrepreneurship. Substantial scholarship has traced out the performance implications of alliances in R&D heavy industries, but less attention has been paid to the nature of tie formation. Powell & Own-Smith (2004) explore the connections in the biotechnology industry as pipes or conduits to delimit the extremes of a continuum along which different types of ties can be arrayed, but the characteristics of ties are very much structured by a broader set of institutions and institutional logics ranging from the norms of profession to scientific ideas. These structure 9

10 the extent to which cumulative knowledge of past starts and ways of doing things are effectively shared and shared broadly in and entrepreneurial community (Murray and O'Mahoney, 2007). This is where accelerators help, they can provide an observable formal tie between a startup and a key institution (the accelerator) that facilitates the development of more informal networks key for the development of a startup through the mobilization of resources (Owen- Smith and Powell, 2004). This signaling effect of formal ties (Powell 1996) allows potential resource provides to give greater access to startups that have been admitted to an accelerator. Thus, we argue: Hypothesis 3: Startups from founding regions with higher levels of regional social capital will have a higher treatment effect from accelerator admission While forming a startup in a region with more openness, trusts and norms of reciprocity would be helpful in forming useful ties with external parties, higher social capital will not be as useful if there are not worthwhile parties with whom to form ties. Prior research has shown that the key actors in a startup ecosystem can have a dramatic impact on regional performance by acting as catalysts (Feldman 2004). In addition, largescale shifts in a local economy, like an IPO, can have downstream effects on the performance of subsequent startups in the region by freeing up resources (T. Stuart and Sorenson 2003). Similarly, regions without the right mixture of industries and industry size can often lack the localized expertise and amenities (lawyers, accountants, etc) necessary to successfully build a company (Chinitz 1961; Feldman 2003; Agrawal and Cockburn 2003). Therefore, we argue: Hypothesis 4: Startups from founding regions with higher levels of regional resources will have a higher treatment effect from accelerator admission While much economic activity, especially inventive activity is geographically localized, the underlying social and economic micro-mechanisms that drive this clustering of economic activity are still relatively unexplored. One of the main drivers for this lack of understanding is the difficulty in observing a slice of similar firms at relatively similar points in their development and observing how differences in their surrounding environment. 10

11 3. Institutional Setting This paper exploits rich administrative data from MassChallenge, the world's largest Startup Accelerator program by Cohort size. MassChallenge admits 128 startup firms each year in late June and provides them Office space, access to an educational program, and access to mentoring for four months. At the end of the program, a subset of the startups is selected to pitch their startup to a crowd of thousands of investors, corporate business development managers, and others. In this section, I will detail two aspects of the program in detail: the admission process, the resources provided. Each year, MassChallenge admits a cohort of startups for a four-month program which the call "Finalists". MassChallenge Finalists are admitted in a two-step selection process. In the first step, a small set of roughly 400 firms of credible prospective firms are culled from a broader list of more than 2,500 submissions. These "semi-finalists" are then reviewed and scored numerically by entrepreneurs, investors, and corporate partners drawn mainly from the Boston area. Examples judges include Joshua Boger, Founder and former CEO of Vertex Pharmaceuticals, Stefan Piesche, the CTO of Constant Contact, or Desh Deshpande, a prominent entrepreneur and investor. After this judging round, the applicant pool is culled to 128 startups 2, these MassChallenge 'Finalists' are given access to free office space and access to a wide network of mentors, drawn from the Boston entrepreneurship and innovation ecosystem. At the end of the program, a subset of the finalists will be given non-dilutive grants ranging from $25,000 to $150,000. MassChallenge is a non-profit corporation that draws its funding from corporate partnerships and government grants, so the program also focuses on generating connections between its startups and its corporate partners, often generating grants or advanced purchase agreements for its portfolio startups in addition to the final grants. The MassChallenge judging process provides an ex ante measure of the quality of the firms applying as well as a rich set of information about the applicant firms. This information will be used within a regression discontinuity framework to explore the impact of the 2 At the beginning of the program, there were fewer startups admitted but, critically for the empirical approach, the total number of admitted startups is determined before the beginning of the application process. That is, the amount of admitted startups does not vary in any given year based on the quality of the applicant pool.in that year. 11

12 MassChallenge program on applicants just above the admission cutoff. In addition, we will explore how this treatment varies with characteristics of the regions in which these startups were founded. MassChallenge evaluates each of its applicants in a two-stage process. In the first stage, each startup is independently evaluated by 3-5 startup experts. The backgrounds of these judges range from venture capitalists, to entrepreneurs to startup lawyers. In the first round of evaluation, the field of potential entrants is reduced from ~4,000 applicants to around 500 applicants, termed Semi-finalists, who are considered serious enough startups to be vetted in a more fulsome evaluation process. After each candidate is given the opportunity to make a pitch to a panel of judges (3-7 in number), they receive another score by each member of the panel. These scores are then aggregated by MassChallenge and used to array the applicants in a rankordered list. Each year there are a fixed number of slots. 3 While the MassChallenge program administrators use the composite score of each startup as the first and largest input into their admission process, there is some level of discretion amongst the administrators about who receives admission and who does not. Most importantly, though, the process of scoring and admission is not easily manipulated by either judges or applicants. The judges are randomly assigned to startups with which they have no ties. While in the judging panel, each judge fills out their own scoring sheet electronically without being able to view the scores of other judges. Furthermore, no feedback is given to the judges about where there panel or their startups fit in the wider distribution of scores. Thus, it is hard for any individual judge on any individual panel to manipulate the scoring to impact the outcomes of any team or group of teams. 4. Empirical Strategy Our objective is to identify the causal impact of admission to MassChallenge on key entrepreneurial milestones and variation in the impact of admission by characteristics of the startup s pre-treatment region. The heart of the empirical strategy is to exploit the numerical scores that generate admission decisions to compare startups that nearly received treatment to 3 In the first years of MassChallenge, the total number of slots was 125, but in 2012 they expanded the program by 3 slots to reach 128, a number of symbolic importance referring to the legendary Route 128 that encloses the Boston/Cambridge entrepreneurial ecosystem. 12

13 those that were close to not receiving treatment. To do so, we will use a regression discontinuity framework (RD) to explore first the overall treatment effect of MassChallenge and then the treatment heterogeneity with respect to the pre-treatment characteristics of the startups home region. Before beginning the RD analysis, however, we begin by a simple exercise of describing the variation in the ex-ante quality of ideas by the characteristics of a startup s home region. For each startup that applies to MassChallenge, we observe a set of characteristics at the time of application, the admission decision and how the startup performed over time in terms of reaching key performance objectives. Thus, the relationship between admission to the program and subsequent performance could be described by the following equation: Performance! = δ + γ accelerated + X! β + ε! 1) Here, γ measures the impact of acceleration conditional on the level of other covariates if we can assume that E acceleration, ε! = 0. Of course, it would be unreasonable to assume that selection into the accelerator is as good as randomly assigned. Startups selected into MassChallenge are more likely to have higher levels of capacity than startups that are not selected for treatment. Thus, a simple ordinary least-squared estimate of equation (1) will yield a biased estimate of the impact of MC on its portfolio firms. Instead, we begin with a simple analysis of the variation in the ex-ante quality of the startups applying to MassChallenge as function of key characteristics of their home region. Prior literature has suggested that startup clusters emerge because these regions have a higher proportion of opportunities (Audretsch and Feldman, 2004). If so, we should expect that entrepreneurs emerging from regions with a higher degree of entrepreneurial resources and activity will have higher quality ideas when applying to MassChallenge. We test this idea by estimating the following regression: JUDGE SCORE! = δ + X! β + ε! 2) Here X is a vector of characteristics describing the level of entrepreneurial activity in the startup s home region. β captures the association between the quality of the startups opportunity and the characteristics of their home region. Since judging is restricted to early stage startups that 13

14 have yet to achieve a significant amount of funding or revenue, the average quality of a startup applying to MC from a region with a given level of startup activity. After establishing a relationship between the ex-ante quality of startup and the characteristics of their home region, we then turn to an examination of the causal impact of MassChallenge on its startup firms. To do so, we employee a fuzzy regression discontinuity framework (RD) (Lee and Lemuix, 2010). Regression discontinuity uses information about the proximity of an individual startup to a threshold for treatment to estimate the impact of treatment on startups near the cutoff. In its most basic form when treatment is a deterministic function of receiving a score above the cutoff, a regression discontinuity design estimates the following model: Performance!! = δ + γ accelerated + f SCORE! cutoff! + ε!" 3) Here, accelerated is a dummy variable that measures the admission decision for a given startup and the function f provides a flexible control a startup s proximity to the treatment cutoff. Assuming that a startup is not able to fully manipulate their score, then γ provides a causal estimate of the impact of the program regardless of misspecification of the control function f as the number of startups approaches infinity. In our data, treatment is not a deterministic function of the score function. In Figure 1, we plot the probability of treatment as a function of the judging score. While there is a clear discontinuous jump in the probability of treatment at the capacity threshold, there is startups below the threshold that are admitted as well as startups above that are not admitted, requiring us to implement a fuzzy regression discontinuity approach. To do so, we estimate a two-stage regression where receiving a score above the capacity threshold for admission is used as a predictor for admission to MassChallenge. Specifically, our fuzzy regression discontinuity regressions estimate the following: accelerated!" = θ + δ above + f SCORE! cutoff! + η!" Performance!" = δ + γ accelerated + f SCORE! cutoff! + ε!" 4) We use of the judges scoring and the fixed capacity of the program to compare startups with scores nearer to the score cutoff for admission. The non-linearity displayed in Figure 1 and 14

15 measured in equation 4 identifies the local average treatment effect (LATE) of admission to MassChallenge as long as there is continuity in the potential treatment effect above and below the treatment threshold (McCrary, 2008). In simpler terms, startups need to be similar to one another just above and just below the probability discontinuity for treatment. In addition to visual inspection of the data in Figure 1, we perform the recommended McCrary test and find that we cannot reject the null hypothesis of local continuity. Thus, we interpret γ in this model as the causal impact of admission to the MassChallenge program. A visual preview of the treatment effect of MassChallenge is provided in the figures 2-3. While γ measures LATE of the MassChallenge program averaged over startups with very different backgrounds, we might ask if subsets of the population measured in equation (4) experience more or less impact from admission to MassChallenge. To do so, we extend our fuzzy regression discontinuity design to allow interaction effects with continuous variables. Assuming that these covariates are continuous at the treatment threshold, the regression discontinuity framework can not only measure the LATE but also the magnitude of the interaction of treatment assignment with the covariates of interest (Becker and Egger, 2013). Specifically, we are interested in estimating the following model: accelerated!" = θ + δ! above + f SCORE! cutoff! + η!" Performance!" = δ + γ accelerated + β! Region + β! Region accelerated + f SCORE! cutoff! + ε!" Here again γ measures the causal impact of the MassChallenge program overall, but now we have separately identified the impact of the characteristics of a startup s home region, β!, as well as how these home characteristics interact with the program s treatment effect, β!. 4) 4. Data 4.1. Sample Definition Our sample was composed of 1,152 startups that were semi-finalists or finalists in the MassChallenge Accelerator during the years This represents all cohorts from the beginning of the program for which at least two years of performance data could be collected at the time of data collection. In addition, we excluded startups that were pursuing social enterprise 15

16 startups that were non-for-profit and thus would not be pursuing outside financing from venture capitalists. [TABLE 1 ABOUT HERE] For each startup, we observe key characteristics of the startup from their application to MassChallenge. First, we observe their geography. For startups that did not provide their location at the zip code level, we supplemented the provided information with web searches. In addition, we received the identities of their founding team and links to personal websites and social media (LinkedIn, Twitter, etc.) that allowed us code the founding team for a range of covariates gender, age, and educational background. Lastly, we observed the aggregate judging score for each startup, JUDGING SCORE, that generates the admission decision for MC. Next, we collected performance measures for each of the startups in our sample. First, we collected employment levels in the startup on LinkedIn. For each of the startups in our sample, we counted the number of non-founder and non-intern workers listing employment in the startup. From these counts, we constructed two measures. First, we created a dichotomous measure of whether the startup achieved significant employment levels, SIGNIFICANT EMPLOYMENT, which we set to one if the startup had 15 or more employees within the first two years of application to MassChallenge and set to zero otherwise. Next we took logs of one plus the measured employment levels to construct the variable, LN EMPLOYMENT. Next, we downloaded early stage financing data from VentureXpert and Crunchbase. For each startup, we recorded tallied the total amount of financing raised in the two years after applying to MassChallenge to create two variables. The first was a dichotomous variable describing whether the startup achieve a significant amount of external financing, SIGNIFICANT FINANCING, which we set to one when a startup received more than $ 500,000 in early stage funding and set to zero otherwise. Next, we calculate the take the log of 1 + the total financing received by the startup to create the variable LN FINANCING. Lastly, we created a set of measures at the zip code level for the level of entrepreneurial activity in this region. First, we collected all data on entrepreneurial meetup groups in the United States from the website Meetup.com. Meetup provides a platform for informal interest-based communities to form and organize meetings. There are meetups for a range of activities from 16

17 book clubs to entrepreneurship support groups. We collected information on the number of entrepreneurship group meetings that were held in each zip code and near to each zip code using the distance weighting scheme most frequently used in the entrepreneurship literature (Sorenson and Audia, 2003; Sorenson and Stuart, 2003). We use this information to construct a measure of entrepreneurial social capital in the region, MEETINGS. Measures of club and affiliation group activity are frequently used to measure social capital in entrepreneurship (Sorenson and Samila, 2013; Kwon and Ruef, 2015) and more broadly in the innovation literature (Laursen, et al., 2012). Next, we create a similar measure of access to a key resource for entrepreneurs: local investors. We download information on all investment activity in the three years before an application year to MassChallenge from VentureXpert. We then create counts of distinct investors with headquarters in each zip code. We then created a weighted measure of investors at or near each zip code in our data using the same distance-based weighting scheme described above, REGIONAL INVESTORS. The dependent and independent variables for this study as well as the sources of the information are provided in TABLE Summary Statistics MassChallenge is one of the top startup accelerators in the country as well as a key player in the Boston/Cambridge Startup ecosystem. In Table 2, we provide summary statistics for our independent and dependent variables used in the study. The startups considered in this study are those that have reached semi-finalist status, as discussed above, meaning all startups in our sample that they were judged to be viable in an initial screening round. It is not surprising, therefore, that the unconditional mean for a startup to reach Significant Funding was 0.13 or 13% of our sample. This is far above the rate of funding for new startups, even in funding rich areas like Boston or Silicon Valley (Guzman and Stern, 2015). Similarly, the unconditional chance of a startup reaching significant employment was 0.19 in our sample or 19% of our total study population. While all startups face significant challenges, these startups are high quality regardless of the admission decision to MassChallenge. [TABLE 2 ABOUT HERE] 17

18 Our sample also represents four cohorts of the MassChallenge program beginning in the years 2010 and ending with the cohort in The mean cohort year of reflects the relatively even weighting of candidates across years with a slight shift towards later years that reflects a slight increase in the total number of admitted candidates from 120 in the year 2010 to 128 in later years. This increase in admitted candidates also corresponded to an increased number of semifinalist candidates in later years. Each semifinalist received an average score from a panel of judges. The highest theoretical score was 100 while the lowest theoretical score was 0. In our sample, we observe an average of 52.78, with a minimum score of 13 and a high score of 98. Thus, there is substantial variation in the perceived quality of the startups in our sample. Critically, the fuzzy RD design employed in this paper requires that there is a discrete change in the likelihood of admission for startups receiving a score that places them in a rank lower than the total class size. [FIGURE 1 ABOUT HERE] In Figure 1, we pool observations across all years and explore the relationship between admission to MassChallenge and the judging score. To do so, we center the scores for each year by subtracting by the cutoff score (e.g. the Judging Score of the startup in rank 128 for years where 128 startups were admitted). Figure 1 shows a dramatic increase in the likelihood of admission for startups that just above the rank threshold as compared to startups that were just below. For startups that are far away from the cut-off, the likelihood of admission is essentially zero, but there is a distinct increase in the likelihood of admission for startups closer to the rank threshold. It is this increase just before the rank threshold that requires us to pursue a Fuzzy RD design. In Figures 2 and 3, we explore visually the relationship between judging score and the likelihood of receiving significant funding and employing a significant number of people. Each figure is a binned scatter plot showing the average value for the dependent variable within each bin. Bins were constructed as 10 equally spaced bins across the support of the running variable. By providing a binned scatter plot rather than a simple scatter plot, our figure reveals more clearly the relationship between judging score and the conditional expected value of the dependent variable in question, the key empirical relationship explored in the results below. As 18

19 suggested in the broader regression discontinuity literature, providing these two plots allows the analyst to inspect the relationship between the running variable and the outcome variables of interest to assess the extent of model dependence in the estimates that follow (Lee and Imbens, 2010). In both figures, there is a discrete jump in the level of the dependent variable at the admission threshold suggesting that any findings in the results below are robust to model specification. [FIGURE 2 & 3 ABOUT HERE] Lastly, the results of our analysis of heterogeneity hinge on the continuity of our regional variables at the treatment threshold (as discussed above). To assess whether this assumption is valid in our measure of regional activity, we provide two binned scatterplots using the same methodology described above. Figure 4 plots the demeaned number of Meetups near a zip code as a function of the running variable. It shows that there seems to be a linear relationship between Meetup frequency and the running variable, but there is no discontinuity in this relationship at the rank threshold. Similarly, Figure 5 explores the relationship between the running variable and the number of active local investors near a startups zip code. Similar to the results in Figure 4, there seems to be a linear relationship between the running variable and the regional investors, but there is no discontinuity in this relationship at the rank threshold. Taken together, these two figures suggest that our assumption of continuity for our two key regional variables at the treatment threshold is well founded. 5. Results The empirical core of this paper explores two issues. First, we build upon the growing literature on the impact of startup accelerators on portfolio firms and measure the impact of the MassChallenge program on admitted firms. We use fuzzy regression discontinuity under a range of specifications to establish the causal impact of the MassChallenge program on the ability of startups to reach key startup milestones. Second, we move on to the heart of our analysis and examine how the characteristics of the founding region of the startup drive 5.1 Relationship between Regional Characteristics and Ex Ante Quality 19

20 Our regression results begin in Table 3 with a series of ordinary least squares regressions where JUDGING SCORE is the dependent variable. These models allow us to characterize the relationship between a region s level of entrepreneurial social capital and resources and its average level of quality for applicants to MassChallenge. Model 3-1 finds a positive and statistically significant relationship between the number of meetings proximate to a startups founding zip code and the average quality score of the startup. Each additional meeting held in a startup s zip code was associated with a increase in the quality of a startup s idea upon application to MassChallenge (the scores theoretically run from 0 100). Model 3-2 examines the impact of local early stage investors at or near a startups home zip code and finds a positive and statistically significant relationship. An increase of one early stage investor in a zip code is associated with a increase in the average score of a startup s application. [TABLE 3 ABOUT HERE] Taken together, these results suggest that startups that are founded in zip codes with higher levels of resources and entrepreneurial social capital are associated with higher quality opportunities. These results accord broadly with a spillover theory of entrepreneurship (Audretsch and Feldman, 2004) where entrepreneurial entry is driven by access to high quality opportunities that emerge from access to industrial and academic research that is locally available. Thus, it is critical to separate the quality of the opportunity as measured at the earliest stages of the venture from the ability of startup founders to mobilize resources to build upon the initial insight (Sorenson and Dahl, 2012). 5.2 Overall impact of the MassChallenge Program We now examine the causal impact of the MassChallenge Program. Because our results are driven by variation in the judging score of startups just above and below the treatment threshold, we are able to separate the regional variation in the quality of the startups driving the results in table 3 from the treatment effect of MassChallenge. The results of our analysis are presented in Table 4. Throughout the table, we use a range of specifications for our estimation of treatment effects to explore the sensitivity of our model estimates to specification assumptions as is suggested in the literature (Lee and Imbens, 2010). All regression have application year and industry fixed effects. Model 4-1 examines the overall 20

21 impact of MassChallenge on the capacity of a startup to find and hire a significant number of employees (>15). It uses a third-order polynomial model to model the relationship between the running variable and the outcome separately above and below the border. In this model, we find that admission to MassChallenge cases a positive and statistically significant increase of 0.33 percentage points in the likelihood of achieving significant employment. Model 4-4 examines that impact of MassChallenge again on the likelihood of achieving significant employment but using a local linear model with a Triangular kernel and IK bandwidth (IK CITE). In this model, we find a similar impact of the MassChallenge program on the likelihood of achieving significant employment showing that admission is associated with a percentage point increase in the likelihood of achieving significant employment. Taken together, these results show that MassChallenge has an impact on a startup s future employment that is robust to the specification of the RD model. These findings are broadly consistent with results from other studies on the impact of accelerators on their startup firms (Hallen, Bingham, Cohen, 2015). [TABLE 4 ABOUT HERE] Next, we examine the impact of MassChallenge admission on the ability of a startup to raise a significant amount of financing (> $500,000). Using a third order polynomial across all the data in our sample, Model 4-2 measures the change in the likelihood of raising significant financing when admitted to MassChallenge and shows that there is a positive and statistically significant increase of percentage points when a startup is admitted to MassChallenge. Similarly as above, Model 4-5 uses a local linear model to estimate the impact of MassChallenge. In it, we find that admission to MassChallenge is associated with a 0.5 percentage point increase in the likelihood of achieving significant funding. Here, there is a significant difference between the local linear and polynomial results, suggesting that startups just at the border may experience more of an impact of admission to MassChallenge than admitted startups of higher quality that are farther away from the cutoff. Next, we explore the impact of MassChallenge admission on the total number of dollars raised by startup firms. In model 4-3, we find that admission to MassChallenge is associated with a positive and statistically significant increase of in the total amount of funding received by startup firms. As before, the results increase when we turn to a local linear specification in model 4-6. Here, we find that admission to MassChallenge is associated with a 21