Business Accelerators: Evidence from Start-Up Chile - LSE

Business Accelerators: Evidence from Start-Up Chile1 Juanita Gonzalez-Uribe2 and Michael Leatherbee3

March 2015 This paper investigates whether government funded business accelerators create value for start-ups. We focus on the case of Start-Up Chile (SUP), an accelerator sponsored by the Chilean government, which provides participants with 40,000 USD (equity free) in seed capital, a work visa, and free office space, as well as the option to be selected into the mentoring arm of the programme where start-ups can access top mentors. Selection into the accelerator follows a rules-based approach: the top 100 applicants are selected every 4 months based on a ranking by external judges. We analyse start-up performance using webbased metrics for applicants that marginally rank above or below the 100th threshold. This analysis provides a clean causal estimate that deals with potential selection bias from heterogeneity in growth opportunities across start-ups. Our results do not allow us to rule out the possibility that participation in the accelerator has no impact on subsequent start-up performance. However, we find evidence, albeit weak, of differences in performance across participants in and out of the mentoring arm. These additional results provide new insights about the selection skills of government-sponsored programmes, and the potential value added role of mentoring for start-ups.

1

Su Wang provided excellent research assistance. Financial support from Abraaj Group is gratefully acknowledged. PRELIMINARY VERSION, DO NOT CITE WITHOUT PERMISSION. Corresponding author: Juanita Gonzalez-Uribe: [email protected]. 2 London School of Economics, [email protected] 3 Pontificia Universidad Católica de Chile, [email protected]

1

Encouraging entrepreneurship has become a major policy objective over the last decade. The rationale for public intervention is that while new businesses and ideas are crucial for economic growth and job creation, there is underinvestment in these areas, either because a funding gap exists (i.e., potential entrepreneurs with positive NPV projects do not get funded) or because stigma of failure is prevalent (i.e., risky but positive NPV entrepreneurial projects are forgone, because market’s beliefs on individual’s abilities negatively overweight failure). Government-sponsored programmes to spur entrepreneurship are now, as a consequence, common place. However, academic analysis about these programmes remains relatively scant, and the little existing evidence is quite glum (e.g., Brander et. al 2008, Lerner 2009). In this paper we take a careful look at whether government sponsored programmes to encourage entrepreneurship add-value to participants. We focus on a type of programme which has increasingly gained participation not only in the public, but also in the private sector: business accelerators. Accelerators are early stage financiers of high technology startups. In contrast to the investment practices of other early stage financiers, accelerators are structured as fixed-term and cohort-based programmes, which include mentorship and educational components, and offer shared-office space to participants (Cohen and Hochberg, 2014). From only one business acelerator in 2005, Y Combinator in Silicon Valley, there are now thousands worldwide, including Techstars which operates in several cities in the U.S., and Seedcamp, originally London-based and currently pan-European (e.g., Cohen and Hochberg, 2014). In spite of their prominence, business accelerators remain understudied in the economics literature due to data- and methodology-related challenges (see Fehder and Hochberg, 2014). Because participants are early-stage start-ups, they are often not legally incorporated, and are thus missing from standard business data sources. In addition, the

2

probability that these early stage start-ups “pivot” is so large, that is challenging to even define, let alone adequately measure, post-application performance. Finally, because simple comparisons across participants and non-participants confound the effect of the programme with the higher growth opportunities of teams that succeed in the competitive application processes, researchers are prevented from more in-depth analysis, unless selection processes are random or rules-based. Establishing empirically how business accelerators affect start-up value and what type of accelerator services have greater effects, thus, while essential for welfare and policy design is challenging. Our analysis overcomes several of these data and measurement challenges and provides the first formal analysis of an accelerator programme: Start-up Chile (SUP), an accelerator promoted by the Chilean government since late 2010. Participants in SUP receive a grant for U$40,000 (equity free), a one year work visa (i.e., the programme is open to Chilean and non-Chilean teams), shared office space for six months in Santiago de Chile, and the option to be selected into the Highway: the mentoring arm of the programme where participants are given additional access to top mentors. Every four months approximately 650 start-ups compete for the 100 coveted spots in SUP.4 One of the advantages of focusing on the case of SUP is that selection follows a rulesbased approach. Each round, applications are scored and subsequently ranked by external judges using three criteria: the quality of the founding team, the merits of the project, and the expected impact of the project on Chile’s entrepreneurial environment. Chilean government officials then use this external ranking to select from the pool of applicants (circa 650 every four months) the participants: roughly, the first 100 ranking start-ups.5 We argue that a

4

The 100 scale of the programme is based on budgetary restrictions. SUP has an annual budget of 15 million dollars and 100 start-ups every four months is the capacity of that budget. 5 Except in generation 2 were the SUP decided before the application round was opened to accept 150 participants.

3

regression discontinuity design on start-up performance based on this exogenous selection rule allows overcoming the endogeneity limitations of simple comparisons across participants and non-participants. The empirical strategy essentially compares performance of start-ups that rank marginally above and marginally below the 100th company threshold. For these close-call applicants, selection is akin to an independent random event (it is “locally” exogenous) and therefore uncorrelated to start-up growth opportunities. Intuitively, the average growth opportunities of start-ups that rank 97 are similar to those that rank 103. However, this small difference in rank leads to a discrete change in the probability that the start-up is accelerated: start-ups ranking below 100th are 14.5% more likely to participate in the accelerator. Our estimate captures the effect of this discrete change in selection at the 100 th ranked company threshold, and this estimate does not incorporate any observed or unobserved confounding factors as long as their effects are continuous around the threshold. We show that indeed, for start-ups that ranked closed to the 100th company threshold, selection is uncorrelated with observed start-up and founder characteristics. Hence, by focusing on these start-ups, we can plausibly estimate a casual effect. We present an analytical framework that shows how start-up performance should be affected by acceleration and how one can recover the value of acceleration from the outcomes of start-ups ranking near the 100th company threshold. Our analysis exploits hand-collected data at the applicant level for start-ups that applied to the accelerator during the 2010-2013 period. The accelerator provided us access to confidential records of the companies that applied to the programme, the evaluation scores from the panel of judges, and the selection decisions made. Based on these records, we collected information on subsequent start-up performance using surveys and extensive web-

4

searches on the businesses and the teams’ leaders in fund raising sites such as AngeList, Techcrunch, social media sites like Facebook, Linkedin, and in web-page tracking sites like Google Insights. The results do not allow us to reject the null hypothesis that the basic services offered by the government-sponsored accelerator, i.e., cash infusions and shared office space add no value to participant start-ups. However, the effect that we identify pertains, by definition, only to participants that have observations around the discontinuity, which affects the degree to which one can extrapolate the results of our analysis to others. In future versions of the paper we plan to explore this point further, by comparing the observable quality of applicants close and far from the 100th company threshold. We then exploit detailed data on the mentoring arm of SUP (i.e., The Highway) to analyse the importance of mentoring as part of the services traditionally offered by business accelerators. Two months into the accelerator, participants have the choice to apply for participation into the mentoring arm. The application process consists of a “pitch-day” in which start-ups do a formal presentation of their businesses to judges, both external (i.e., staff at other private accelerators in Chile such as Telefonica’s Wayra) and internal (i.e. staff at SUP). The judges independently score the start-ups, and then based on that score the staff at the accelerator selects roughly 20% of the participants into the mentoring arm. The accelerator provided us access to the additional confidential records detailing the scores of participants during the pitch-day and the selection decisions made. Using this information, and exploiting a data-driven selection rule—probability of acceptance increase by 40% if participants score more than 3.6 (over 5) in the pitch-day—we show that start-ups in the mentoring arm outperform their peers in the accelerator programme. Following a regression discontinuity approach in the participant’s sample that take part of the pitch-day, we compare

5

start-up performance for applicants near the 3.6 score threshold, and find evidence, albeit weak, that mentoring has a positive causal impact of start-up performance. In future versions of the paper we plan to: 1. explore the real effects of acceleration beyond start-up performance, by focusing on the potential effects on founders, 2. include results from a detailed survey on applicants regarding their experience in SUP, and their opinion on the most useful aspects of the programme. Finally, we will also present suggestive evidence of the more general impact of SUP on the Chilean entrepreneurial ecosystem, by comparing registering rates of start-ups in Chile across industries targeted and not targeted by SUP. Our paper contributes to the more general literature assessing the impact of early stage financiers on firms (e.g., Hellman and Puri (2000); Sorensen (2007) Kortum and Lerner (2000)) in two ways. First, we focus on a neglected type of investor: business accelerators. Second, our methodology allows us to uncover casual estimates; in contrast, the estimates in most pre-existing studies may be biased because of the non-random nature of selection process by early stage financers. For example, if better start-ups (e.g., those with unobservable better growth opportunities) are also more likely to be selected by an early stage financier, this would cause regression coefficients to be biased upwards. Our paper also contributes to our understanding on what types of services to start-ups appear to add more value, especially when imparted by government-sponsored programmes. Our results point to an important role of mentorship which complements studies in other fields such as subsistence businesses in developed economies (McKenzie and Woodruff (2008), De Mel et al. (2014)). Our paper has policy implications, in particular regarding design of policies to sponsor entrepreneurship. Our results suggest both that government-funded programmes can 6

develop valuable selection skills, and also that mentoring is potentially an important policy lever. Both these results are important. Business accelerators are becoming more frequent as policy tools to sponsor entrepreneurship: the SUP model has already been adapted in several countries such as the U.S., Brazil and Peru, and is in the process of being adapted to several other countries (e.g., Canada, Denmark and Spain, among others). Our findings can help policy makers understand how to adapt more successfully this type of programs to the idiosyncrasies of their countries, by understanding which is the crucial policy element. The rest of this paper is as follows. In Section 1 we describe the accelerator programme and the data provided, and detail the selection process. In Section 2 we explain the analytical framework and in Section 3 the identification strategy. In Section 3 we also present results, which we interpret in Section 4. We conclude in Section 5. 1.

INSTITUTIONAL SETTING: START-UP CHILE

SUP is a government-sponsored program launched in August 2010 to attract early-stage, high-potential entrepreneurs to bootstrap their ventures in Chile.6 The programme is run by the Ministry of Economy and is executed by the Chilean Economic Development Agency (CORFO), the leading organization for promoting innovation and entrepreneurship in the country. Its main long-term goal is to convert Chile into an innovation and entrepreneurial hub in Latin America not only by bringing in more entrepreneurs, but also by creating a much better-developed ecosystem of supporting institutions—including venture capital firms and angel investors. SUP offers four main benefits to participants. First, SUP provides selected start-ups with $40,000 equity-free seed capital. The capital is staged: 50% is delivered at the beginning of the programme, and the remaining 50%, 3 months after. The second instalment is 6

For more details on SUP see Applegate et al., (2012) and Gonzalez-Uribe (2014).

7

conditional on pre-determined performance milestones.7 The staging of capital provides incentives to entrepreneurs to provide effort, and accountability of participants’ expenditures. Second, SUP sponsors a temporary one-year work visa for accepted participants in order to attract foreign entrepreneurs. The programme also helps participants settle in Chile through a “buddy system”. The buddy-system pairs entrepreneurs with local members of the Santiago business community based on background interests and language. Local buddies advice participants on opening Chilean bank accounts, registering with the police, obtaining a local ID, and securing housing and mobile phones, in addition to checking in with participants once or twice a month throughout the entrepreneurs’ stay in the country. Third, SUP provides free, shared office space in downtown Santiago, equipped with WiFi, for all start-ups. Workshops on think-tanking and pitch-training based on peer to- peer teaching are held on-site. Start-ups also have access to SUP’s network of mentors. Starting in 2012, SUP expanded its programme to include more accelerator-type activities such as national and international pitch competitions. It created a mentoring arm within the accelerator known as the Highway, which provides additional resources to participants including access to the most renowned mentors and frequent monitoring by the SUP staff. Participants are carefully selected into the Highway after a pitch competition, in which external and internal judges rank participants. Roughly 20% of participants in each generation have classified into the Highway since SUP’s fourth generation. The SUP program, in turn, requires accepted entrepreneurs to stay in Chile for the six-month duration of the program, and contribute to the building of an entrepreneurial culture in Chile. During their stay, entrepreneurs have to accumulate 4,000 in “Return Value

7

In the inception of the programme, capital disbursements were neither pre-expense nor staged. This system was implemented in the first semester of 2013.

8

Agenda” (RVA) points, a system to measure the social contribution of participants in the Chilean entrepreneurial ecosystem. Participants have the option to attend, organize or innovate in social-related activities. Attendance refers to participation in local events, such as meetings and conferences at which entrepreneurs make themselves available to share knowledge and to network with locals. Organization can include giving a talk at a school, presenting a pitch to a local investor, or mentoring a local entrepreneur or student. Innovation refers to initiatives that actively engage the Chilean business community, such as starting a new business with a Chilean partner or patenting a product in Chile. 1.1.

DATA

We were given access to applicant records for seven generations of SUP. In total we have information on 3,258 applicants, 616 and 2,642 participants and non-participants, respectively. Panel A of Table 1 displays the number of applications judged per generation (i.e., not all applications are judged by YouNoodle as some are incomplete), the number of applications selected (e.g., and offer is extended by the accelerator to the start-up) and the number of applications that are formalized (e.g., the start-up accepts offer and reallocates to Chile for the 6 month duration of the programme).8 Panels B through D, and E through G, describe the composition of the sample by start-up and lead founder characteristics, respectively. For the empirical analysis, we bundle together all generations. While the average quality of start-ups on the accelerator is likely to change over time (e.g., as the accelerator gains recognition better start-ups may apply), we are unable to analyse

8

Results from the Global Entrepreneurship Monitor (GEM) report provide a basis for comparison between the entrepreneurs that apply to SUP, and the average Chilean entrepreneur. According to the latest GEM (2012), the average Chilean entrepreneur is 37.5 years old, is twice as likely to be male than female, has studies beyond those that are compulsory, and has a business that serves the consumer sector. The survey on micro entrepreneurship (EME) also provides a basis of comparison for the composition of Chilean SUP entrepreneurs. According to the EME of 2012 the average Chilean micro entrepreneur is male (69%), has between 45 and 59 years of age (39%), is responsible for a home (74%), has basic to mid-level education (67%) and its business belongs to the sectors: retail, restaurant and hotel (34%), agriculture and fishing (24%) and manufacturing (13%).

9

generations separately due to power considerations. We address this concern in our empirical strategy including generation fixed effects throughout. [INSERT TABLE 1 HERE] For the 3,258 start-ups that constitute our sample we hand-collect performance measures using extensive web-searches during the second semester of 2013. Table A1. in the Appendix has a list of the performance measures and their sources. Table 2 displays the summary statistics of these web-based performance measures. [INSERT TABLE 2 HERE] 1.2.

SELECTION INTO THE ACCELERATOR

Selection into SUP is a two-part process that takes place every four months. First, entrepreneurs apply to the programme and their applications are ranked by external judges. SUP outsources this first part to Younoodle, a consulting start-up in California, which provides and objective evaluation of the merit of the start-ups outside the particular context of the Chilean economy. Entrepreneurs fill in their applications through an open survey, and then Younoodle resorts to Silicon Valley experts (3-4 judges per application) who evaluate applications using three criteria: the quality of the founding team, the merits of the project, and the impact that it is likely to have on Chile’s entrepreneurial environment. Using the experts’ judging sheets, applicants are ranked. No ties are permitted; if companies tie in their judges’ score they are randomly ranked. The second part of the selection process is handled by CORFO, which makes the final decision based on Younnodle’s ranking. A threshold is pre-specified each round (normally

10

100), and only companies that rank above the threshold are meant to be selected.9 The threshold corresponds to the pre-determined size of the program and is decided on by the government before the application process begins as a function of its budget. The start-ups cannot precisely manipulate their ranking. Because start-ups do not know the judges’ scoring rules, and are unlikely to learn about these rules from past SUP participants, it is improbable that start-ups have room for manipulating their scores around the 100-th company cut-off. In addition, the judges are unlikely to manipulate the scores, as no judge evaluates all applications and only observe the very few he/she is asked to score. As it is common in government-sponsored programmes, however, the selection committee at CORFO does not strictly follow the selection rule and thus not all participants who rank above the 100th company threshold end up participating in the programme. Indeed, of the top 100 ranked applicants, about 75% of them are selected into SUP. The remaining 25% are selected by a committee among applicants ranked between 101 and 300 based on qualitative attributes of the applications. Although there is no 100% compliance of the selection rule, there is nonetheless a discrete jump in the probability of selection around the rank cutoff as shown in Figure 1. [INSERT FIGURE 1 HERE] Figure 1 plots the fraction of participating applicants against the normalized rank, as defined by the ranking of the start-up minus the predetermined size of the program of its generation (i.e., 100 for all generations except 150 for generation 2). The figure includes average participation rates by 10 applicant bins and the fitted value and 90% confidence interval from the regression

9

The threshold has been 100 in every generation, except the second generation where the threshold was set at 150.

11

(1)

𝑎𝑐𝑐𝑒𝑙𝑒𝑟𝑎𝑡𝑖𝑜𝑛𝑠 = 𝛿 + 𝛾𝑎𝑏𝑜𝑣𝑒𝑠 + 𝑓(𝑅𝑎𝑛𝑘𝑠 − 𝑐𝑢𝑡𝑜𝑓𝑓 𝑔 ) + 𝜀𝑠 ,

where the outcome variable acceleration is an indicator variable that equals 1 if the applicant participated in the accelerator, and 𝑓(𝑅𝑎𝑛𝑘𝑠 − 𝑐𝑢𝑡𝑜𝑓𝑓 𝑔 ) is a 4th degree polynomial of the normalized rank.10 The vertical line represents the ranking cutoff normalized at 0. The coefficient 𝛾, which in the plot corresponds to the difference in the vertical axis between the points where the left and right polynomials intersect in the cutoff, is a measure of discontinuity. As per visual inspection, there is a discontinuity in the probability of participation around the cutoff, which is sizable and significant. [INSERT TABLE 3 HERE] Table 3 presents the coefficient on the constant, 𝛿, and the coefficient on 𝑎𝑏𝑜𝑣𝑒𝑠 , 𝛾, using different specifications of equation (1): including generation fixed effects (column (2)), covariates (column (3)), and restricting the sample to a window of 48 observations around the cutoff (as calculated using the optimal bandwidth procedure of Calonico et al., 2014) and differentially weighting observations using a triangular kernel (column (4)). Across all specifications there is a significant jump in the probability of participation around the 100th company threshold. The estimate in column (3) implies that ranking above the cutoff increases probability of acceleration by 19%. The coefficient 𝛾 is significant at the 1% level and is stable across the first three columns. The estimate in column (4) implies that ranking above the cutoff increases the probability of acceleration by 30%. The coefficient is much larger as the observations included only correspond to those in a window of 48 ranks around The polynomial is evaluated on Rank s − cutoff g so that the coefficient on aboves corresponds to the effect of the selection rule on participation at the cut-off. In unreported regressions we verified that alternative parametrizations of 𝑓(. ) give similar results, including linear, quadratic and cubic polynomials. Different subsamples were also used, and different weights for the observations near the cut-off. Standard errors are robust. The standard RD implementation pools the data but allows the polynomials to differ on either side of the cut-off by interacting the normalized rank with aboves . However, I potentially have too few points to the left of the cut-off to estimate a control function separately on both sides. I verify, nonetheless, that results in Table 3 are robust to including different polynomials of 4 and 5 degrees, respectively, at the left and right side of the cut-off. 10

12

the cutoff, which excludes start-ups ranking between 150-160 which were abnormally likely to participate as can been seen in the plot (i.e., the fit of the polynomial around the normalized rank of 50 is relatively poor).11 This selection rule based on a “size-of-the-program” cutoff is useful in evaluating the causal effect of acceleration on start-up performance, because this cutoff dramatically changes the probability of acceleration but is likely continuously related to performance. For government-based programs this alternative evaluation method is important as these agencies are often unwilling to randomize based on ethical considerations. In the next section we present a simple analytical framework that shows how to recover the value of acceleration by focusing on applicants close to the cutoff. In section 3 we explain in detail how we exploit this selection rule in practice to identify the causal effect of SUP. 2.

ANALYTICAL FRAMEWORK

In this section, we present an analytical framework that shows how to recover the value of acceleration by focusing on applicants ranking close to the cutoff. We show that a discontinuity analysis is a simple way to deal with heterogeneity in unobserved growth opportunities across applicants. Denote as 𝑟 the ranking of the applicant and 𝑉(𝑟) the added-value of government-funded accelerator services. For simplicity, we assume throughout that the outcome of the selection process is binding, that the threshold for selection is 𝑟 ≤ 100, and that the value of acceleration to the start-up is fixed (i.e. is independent of 𝑟), such that 𝑉(𝑟) = 𝑉̅ if 𝑟 ≤ 100 and 0 otherwise. The objective of the empirical analysis is to estimate 𝑉̅ , the value of acceleration, which is not directly observable. Further assume that the underlying growth

11

One potential explanation is that judges check start-ups ranking between 150 and 160 as a final check on the sample. Interviews CORFO officials mentioned that their perceived “checking threshold” was closer to 200.

13

opportunities of the applicants can be represented by a function of the ranking 𝑟, 𝐺(𝑟), that is continuous around the 100th company threshold. For highly ranked applicants, growth opportunities are likely very high. Around the threshold, growth opportunities may not be as high, but most importantly, are comparable across participants in either side of the threshold. Since 𝐺(𝑟) is a continuous function of 𝑟, but 𝑉(𝑟) is discontinuous at the 100th company threshold, the performance of the applicant that one observes after acceleration is also discontinuous at the 100th company threshold. This implies that the difference in the performance at the 100th company threshold, VA, between a start-up that barely ranks above the 100th company and one that barely ranks below is exactly the value added of acceleration. Under the assumptions outlined before, 𝑉𝐴 = (𝑉̅ − 𝐺(𝑟)) − (0 − 𝐺(𝑟)) = 𝑉̅ . Therefore, one can recover the value of acceleration form the difference in performance across start-ups that rank close to the discontinuity. The only two crucial identification assumptions are that the distribution of start-up characteristics and growth opportunities is similar on both sides of the discontinuity, and that the probability of selection changes discretely when the company ranks below 100. We made a number of additional assumptions in our example, some of which do not necessarily hold in reality but are not crucial for identification. For example, as explained, the government committee sometimes decides to accept start-ups that fail to rank below 100. Hence, in this case one should expect 𝑉(𝑟) to be slightly positive to the right of the threshold and thus, the average performance to the right of the threshold will be less negative than if the selection rule were strictly binding. At the same time, start-ups may decide last minute to reject the offer; thus, 𝑉(𝑟) will be below the effective value of acceleration to the left of the threshold, and the average performance of start-ups in the left will be less positive than if selection were binding. Still, provided that 𝐺(𝑟) is continuous and the probability of selection

14

is discontinuous around the threshold,

𝑉𝐴 can be used as a measure of the value of

acceleration to the start-up. In this case, the value estimated at the discontinuity, 𝑉𝐴, is not equal to 𝑉̅ , as in the previous example. However, as Lee and Lemieux (2010) discuss the identification strategy is still valid as long as there is a discrete jump in the probability of selection at the 100th company threshold (this is the fuzzy regression discontinuity setting). The estimate recovered is the average effect of acceleration for start-ups ranking close to the threshold. An important issue, thus, is that the degree to which we can make generalizations based on our results, will depend on how different are the applicants ranking close to the threshold from other applicants. We return to this point in the next section when we discuss the results. Other important questions that arise when trying to infer the value of government-funded acceleration from differences in performance at the discontinuity are whether we should expect any effect of acceleration on start-ups that barely rank below or above the threshold, and whether these differences appropriately evaluate the policy performance. The issues here are (i) alternative sources of funding for rejected applicants (ii) objective function of the government-funded accelerator. On the one hand, if we take at face value the assumption behind government intervention that there is underinvestment in entrepreneurship, then we can assume that rejected applicants will most likely not have access to alternative sources of finance. In that case, differences in performance across start-ups ranking barely below and above the threshold will reflect the effects of alleviating financial constraints, as well as the potential value of additional services provided by the accelerators (e.g., knowledge spillovers from other accelerated projects). However, no differences may necessarily be detected if the policy objective is not to fund positive private NPV projects. The government-funded accelerator may pick negative private 15

NPV projects if their social return is positive, e.g., these investments can spark a cultural change that will help address future underinvestment in entrepreneurship. Sizable differences in performance across similar participants in either side of the threshold may thus constitute only a partial metric of welfare consequences. Finally, if the assumption of underinvestment in entrepreneurship is not valid and the objective function of the accelerator is to fund positive NPV projects, then rejected applicants ranking barely below the cut-off will likely find funding elsewhere (as we have assumed that start-ups in either side of the threshold have the same distribution of growth opportunities). Thus, we should expect to see differences in performance across start-ups closely ranking in either side of the cut-off, only if added value from early stage investors is not constant across investors. Under these assumptions welfare analysis is also nuanced due to potential crowding-out of private investment by the government funded-accelerator (e.g., Wallsten, 2000). If absent the government-funded accelerator accepted applicants would have been funded by the private sector, and there is heterogeneity in growth opportunities of start-ups such that those ranking far above the cut-off are on average better than those below, then private equity investors are negatively affected by the public programme (i.e., the better startups are funded with public funds). Sizable differences in performance across closely ranked start-ups in either side of the cut-off, do not necessarily imply then that the public accelerator is welfare improving. We come back to this discussion in Section 4 where we focus on interpretation of results. 3.

METHODOLOGY AND IDENTIFICATION STRATEGY

We now describe the empirical approach to measure the causal effect of acceleration on startup performance. Suppose start-up 𝑠 applies to the accelerator and is ranked at 𝑟𝑠 relative to all

16

other start-ups in its generation. We code the indicator for participation in the accelerator as 𝑎𝑐𝑐𝑒𝑙𝑒𝑟𝑎𝑡𝑖𝑜𝑛𝑠 =1. We are interested in the effect of acceleration on the performance of start-up 𝑠, 𝑜𝑢𝑡𝑐𝑜𝑚𝑒𝑠 . We can write (2)

𝑜𝑢𝑡𝑐𝑜𝑚𝑒𝑠 = 𝜋 + 𝛽𝑎𝑐𝑐𝑒𝑙𝑒𝑟𝑎𝑡𝑖𝑜𝑛𝑠 + 𝜀𝑠 ,

where the coefficient 𝛽 that we are interested in is the effect of acceleration on the performance measure, for example, survival, and 𝜀𝑠 represents all other determinants of performance (𝐸(𝜀𝑠 ) = 0). The problem with estimating a regression such as (2) directly is that acceptance into the accelerator is a highly endogenous outcome, and 𝑎𝑐𝑐𝑒𝑙𝑒𝑟𝑎𝑡𝑖𝑜𝑛𝑠 is unlikely to be independent of the error term (𝐸(𝑎𝑐𝑐𝑒𝑙𝑒𝑟𝑎𝑡𝑖𝑜𝑛𝑠 , 𝜀𝑠 ) ≠ 0), in which case the estimate of 𝛽 will be biased. To get a consistent estimate, we would ideally want participation in the accelerator to be a randomly assigned variable. The regression discontinuity framework that exploits the ranking by external judges helps us approximate this ideal setup because ranking in an arbitrarily small interval around the cut-off, is random, however, the probability of acceptance is dramatically different in either side of the cut-off in that small window. Intuitively, the idea is to compare the outcome of start-ups that almost participated in the accelerator as they barely ranked below the cut-off, with those that barely ranked above and almost didn’t participate. We implement this comparison using a fuzzy RD design (Imbens and Lenieux, 2007; Roberts and Whited, 2013). In order to conclude that any difference between start-ups ranking closely in either side of the cut-off is caused by participation in the accelerator, we assume that these two groups are statistically indistinguishable during the application stage. In the fuzzy RD setting, this assumption is equivalent to a continuous distribution of the unobserved residual at the ranking cut-off.

17

Following Lee and Lemieux (2010), we test whether the data rejects the identification assumption by inspecting the cross-sectional distribution of predetermined variables at the cutoff. We remark first that the distribution of applicants is by construction smooth at the cutoff because the selection mechanism is based on ranking. A visual test as suggested by McCrary (2008) is not very informative in this case. Moreover, as we argued, the start-ups are unlikely to precisely manipulate their ranking, because they do not know the judges’ scoring rules, and are unlikely to learn about these rules from past SUP participants. Judges are also unlikely to manipulate the scores, as no judge evaluates all applications and only observe the very few he/she is asked to score. We focus instead in testing whether at the time of application there were any systematic differences in characteristics of start-ups of founders in either side of the cut-off. In figure 2 we plot the averages for applicant’s characteristics at the time of application grouped in bins of 10 applicants. Five plots are shown for the variables Age, Chilean (i.e., a variable that equals one if the applicant leader is Chilean), Gender (i.e., a variable that equals one if the applicant is a man), Money Raised (i.e., an indicator variable that equals one if the start-up has raised external finance application), and Prototype (i.e., a variable that equals one if the project already has a prototype). The plots also show the fitted values form the applicant level regression of each of these variables on the polynomial 𝑓(. ) and the 𝑎𝑏𝑜𝑣𝑒𝑠 variable. Visual inspection suggests that there are no statistical discontinuities in the cross-sectional distributions of any of these variables around the cut-off. This result provides support for the identification assumption. [INSERT FIGURE 2 HERE] By substituting equation (1) into the regression model (2) and relabeling coefficients and the functional form 𝑓̆(. ), we obtain the fuzzy RD reduced form: (3)

𝑜𝑢𝑡𝑐𝑜𝑚𝑒𝑠 = 𝛼 + 𝛾 × 𝛽𝑎𝑏𝑜𝑣𝑒𝑠 + 𝑓̆(𝑅𝑎𝑛𝑘𝑠 − 𝑐𝑢𝑡𝑜𝑓𝑓 𝑔 ) + 𝜀 18

As noted by Hahn et al., 2001, in its simplest form, the fuzzy RD setting implements a Wald estimator for 𝛽. This estimator is equal to the coefficient of 𝑎𝑏𝑜𝑣𝑒𝑠 on regression (3), 𝛾 × 𝛽, divided by the coefficient of 𝑎𝑏𝑜𝑣𝑒𝑠 on regression (1), 𝛾. Thus the fuzzy RD procedure is akin to a setting where, conditional on 𝑓̆(𝑅𝑎𝑛𝑘𝑠 − 𝑐𝑢𝑡𝑜𝑓𝑓 𝑔 ), 𝑎𝑏𝑜𝑣𝑒𝑠 is an instrumental variable for 𝑎𝑐𝑐𝑒𝑙𝑒𝑟𝑎𝑡𝑖𝑜𝑛𝑠 . We estimate this regression using a two-stage least squares (2SLS) procedure where (1) is the first stage and (2) (including an additional term for the polynomial) is the second stage.12 An applicant ranking above the cut-off is more likely to participate in the accelerator, but applicants are still endogenously chosen by the accelerator, and also endougenously selfselect into accepting an invitation to participate. Our main assumption is that crossing the cutoff does not affect performance other tan through the change in the probability of acceleration, and that all applicants ranking below the cut-off that are chosen by the accelerator, would also be selected if they ranked above the cut-off (monotonicity). Then 𝛽 estimates the causal effect of acceleration on “compliers” of this instrument, that is, those applicants that end up accelerated because they ranked above the cut-off, and thus the estimate corresponds to a local average treatment effect (LATE). We remark that 𝛽 is precisely the estimate of VA from section 2. 3.1.

IMPLEMENTATION OF THE FUZZY RD

Implementations of RD inference vary considerably in the literature. Many researchers control for high-degree polynomials of the underlying continuous forcing variable, where the shape of the polynomial is often allowed to vary across the threshold, provided there are enough observations in either side (e.g., Lee and Lemieux 2010). This method borrows strength from observations far from the cut-off to estimate the average outcome for observations near it. In practice, cubic or higher order polynomials are used, often based on 12

To clarify the excluded instrument is only 𝑎𝑏𝑜𝑣𝑒𝑠 , the polynomial on the normalized cut-off is a control in the second stage.

19

statistical information criteria or cross-validation to determine the degree of the polynomial. Lee and Card (008) suggest a goodness-of-fit test to choose the polynomial degree. The polynomial function is estimated including a full set of bin dummies. Additional polynomial terms are added until the null hypothesis that the bin dummies are zero can no longer be rejected. The drawback in using this first approach is the potential bias introduced by using observations far from the cut-off. Others prefer a local analysis, which discards observations above some bandwidth h away from the cut-off, and estimates low-degree polynomial regressions on the remaining observations (e.g., Gelman and Imbens, 2014). Several methods to choose the bandwidth exist (e.g., Calonico et al., 2014). The drawback in using this second approach is the loss of efficiency due to discarding observations. In this paper we do not take a stand on this methodological discussion. Instead, we present results using both approaches. 3.2.

RESULTS: ACCELERATION AND START-UP PERFORMANCE

Figure 3 shows the average start-up survival as measured by a listing in AngeList by 2013 in bins of 10 applicants, and the fitted values of the reduced form regression (3). Visual inspection reveals a discontinuity in survival: an AngeList listing is more likely for applicants ranking above the cut-off relative to those ranking below. Given the identification assumption, the discontinuity in this survival metric is attributed to acceleration. [INSERT FIGURE 3 HERE] We formalize the intuition conveyed by the figure with regression tests as summarized in panel A of Table 4. Reported standard errors are heterosedasticity robust.13 Column (1) reports estimates from a simple OLS estimation of regression (2). The coefficient is positive and statistically significant at the 1% level: applicants that are accelerated are 43% more likely to survive. 13

In unreported regressions we repeat the analysis clustering standard errors by generation and results continue to hold. Consistent with potential small cluster bias (there are only 7 generations) we find that standard errors are most conservative without clustering.

20

Column (2) in Table 4 reports results from the fuzzy RD regression (3), the estimates for the polynomial terms are not included in the table to conserve space. The coefficient equals 0.49% and is statistically significant at the 1% level. The RD design does not require conditioning on baseline covariates, but doing so can reduce sampling variability. In Column (3) we present results after conditioning for selected covariates and generation fixed effects. Results continue to hold. Column (4) presents estimates allowing the polynomials to differ on either side of the threshold. The point estimate remains similar but is no longer significant, likely because we have too few data points to the left of the cut-off to estimate a control function separately on both sides. Finally, column (5) presents estimates using a local linear regression approach using a bandwidth of 47, which was optimally estimated following the procedure suggested by Calonico et al., (2014) (CCT). The coefficient remains positive, but more than halves in magnitude and is no longer significant, likely because 708 observations do not provide enough statistical power to distinguish an effect of size 0.12 (i.e., a power estimation indicates that at least 750 observations are needed to distinguish an effect of that magnitude if the mean is 0.2, the standard deviation 0.4, and the ratio between treated and control observations is 0.2). [INSERT TABLE 4 HERE] Interpretation of the results in Figure 3 and Table 4 as evidence that the accelerator affects survival is nuanced since these type of investors often encourage participants to initiate a listing in AngeList as a means of internal communication (this is certainly the case in Start-up Chile, which itself has a listing in AngeList that works as a networking platform with alumni (see: Gonzalez-Uribe, 2014)). If participants do not cancel their listings in the event of failure in order to maintain the contact with the program, our measure of survival may be biased. We thus repeat the analysis using two other survival metrics: listing in Crunchbase and listing in Linkedin, two other well-known platforms recording fundraising

21

activities and employee recruiting, respectively. Consistent with the AngeList indicator variable providing a potentially inaccurate measure of survival, Panel B in Table 4 shows that results are dramatically different when using these alternative survival measures. In particular, while the OLS estimates continue to be positive and significant at the 1% level, the Fuzzy RD estimates are not significant, and the point estimates are very small and negative. In unreported regressions, we find similar results when using other RD implementation methods including lower degree polynomials and local linear regressions, and other webbased performance proxies start-up growth, employment and fundraising. Results are similar across the different specifications, i.e., OLS estimates are positive and significant and fuzzy RD estimates are not significant, quantitatively smaller than the OLS estimates, and often negative. 4.

INTERPRETATION OF RESULTS

The results in Section 3 suggest that we cannot rule out the null hypothesis of no difference in performance across accelerated and rejected applicants near the cut-off. We now explore in depth the interpretation of these results. 4.1.

THE ADDED-VALUE OF CASH AND MENTORING

One potential interpretation is that the basic services offered by the accelerator (i.e., cash infusion and shared office space) are not enough or adequate to differentially affect subsequent performance on average. This interpretation is based on the assumption that complier start-ups ranking near the cut-off (i.e., recall that the fuzzy RD approach only allows us to estimate a local effect) are representative of average start-ups. To explore this interpretation in more detail, we explore dissimilarities in the services provided by the accelerator across participants. In particular, we look for differential performance across accelerated start-ups (i.e., participants) with different access to internal mentors. We focus on mentorship because provision of this service varies based on a

22

selection rule that can be exploited to identify the causal effect, and because mentorship is often an important part—usually not as easily scalable as monetary resources—of these programmes. Understanding whether it adds value is important for policy design, and possibly for our understanding of how accelerators affect performance more generally. 4.1.1. SELECTION INTO THE MENTOR ARM Participants in SUP have the option to join the Highway, the mentoring arm of the programme, which provides access to top mentors. Two months into the accelerator the application process for the mentor arm begins. It consists of a “pitch-day” in which start-ups do a formal presentation of their businesses to judges, both external (i.e., staff at other private accelerators in Chile such as Telefonica’s Wayra) and internal (i.e. staff at SUP) and a final decision by staff at the accelerator in the days following the pitch competition. The judges independently score the start-ups, and then based on that score the staff at the accelerator selects roughly 20% of the participants. While in each generation the number of accepted participants into the mentor arm is not strictly capped (in contrast to participation in the accelerator), an implicit selection rule is evident in the data: there is a discrete jump in the probability of selection into the mentoring arm of 34% if the start-up scores at least 3.6/5 during the pitch-day. Figure 4 shows the fraction of applicants participating in the pitch-day that are selected into the mentor arm. Visual inspection reveals this fraction is discontinuously higher for those participants that scored above 3.6 in the pitch-day. The figure also shows the ordinary least squares (OLS) fitted values and 90% confidence interval of the regression (4)

𝑚𝑒𝑛𝑡𝑜𝑟 = 𝜏 + 𝜇𝐴𝑏𝑜𝑣𝑒3.6 + 𝑔(𝑃𝑖𝑡𝑐ℎ_𝐷𝑎𝑦 𝑆𝑐𝑜𝑟𝑒) + 𝜀

where the outcome variable mentor is an indicator variable that equals 1 if the participant was mentored, 𝐴𝑏𝑜𝑣𝑒3.6 is an indicator variable that equals one if the participant scored above 3.6 23

during the pitch-day, and 𝑔(𝑃𝑖𝑡𝑐ℎ_𝐷𝑎𝑦 𝑆𝑐𝑜𝑟𝑒) is a 4th degree polynomial of the pitch-day score. The polynomial 𝑔(. ) depicted graphically as the smooth line on both sides of the cutoff controls for any underlying relationship between the fraction of participants that are mentored and the score of pitch-day. The coefficient 𝜇, which in the plot corresponds to the difference in the vertical axis between the points where the left and right polynomials intersect in the cut-off, is a measure of the size of the discontinuity. As per visual inspection, and as confirmed in the unreported regressions, the discontinuity is large—34%—and significant—1% level. [INSERT FIGURE 4 HERE] We implement a fuzzy RD design to identify the value added of mentoring by first confirming that the identification assumption of balanced covariates around the 3.6 score is satisfied. Figure 5 shows that participants in either side of the 3.6 score threshold are indistinguishable. The only significant difference is for the covariate Money Raised, i.e., selected participants into the mentor arm, which scored close to 3.6 in the pitch-day, are significantly more likely to have secured external financing prior to joining the accelerator. We deal with this issue by controlling for this covariate in the fuzzy RD regressions. Table 5 documents differences in performance across start-ups in and out the mentor arm using the OLS and the fuzzy RD approach based on the 3.6 score cut-off of the pitchday. The results in the table provide evidence, albeit weak, that mentoring has a positive causal impact of start-up performance. There are no significant differences in survival as measured by having a listing in AngeList, which is expected, as both, participants that are mentored and those that are not, are encouraged to list their companies in AngeList by the accelerator. However, there is evidence of impact on growth as measured by number of Facebook likes and employment as measured by company size in Linkedin. 24

[INSERT FIGURE 5 HERE] One interpretation of these additional findings is that there is heterogeneity in impact across services offered by the accelerator. While basic services such as cash infusion and shared office space appear not to add value, mentoring appears to contribute more to start-up growth. 4.2.

POLICY OBJECTIVE AND THE RIGHT UNIT OF ANALYSIS

One alternative interpretation of the findings is that focusing on start-up performance presents only an incomplete picture of the policy’s impact. Perhaps the correct unit of analysis should instead be the entrepreneur (founder) and not the firm. Start-ups pivot a lot and failing is part of growing. What matters most is that entrepreneur learns, tries again and is likely more to succeed after he is schooled in the accelerator. The main challenge here is measuring outcomes at individual level. We overcome this challenge by collecting information at the founder level from Linkedin. In future versions of the paper we will summarize the results from analysis at the founder level. The policy objective was not to help accelerated start-ups. In fact, the founder of the programme Nicolas Shea argued that the objective was to attract talent (not to be retained) but instead to inspire Chileans to become entrepreneurs and start an internal mentality revolution. This appears to be more general. Fehder and Hochberg (2014) find that there are subsequent local developments after accelerator programs. Perhaps, then, the correct test for the success of the policy is not participants’ outcomes but instead local development. To explore this question in detail we retrieve data on new business formation across the neighbourhoods (comunas) in Santiago de Chile. The comunas that are contiguous to Santiago Centro where SUP is located are: Independencia, 25

Providencia, Recoleta, Renca, Nunoa, San Joaquin, San Miguel, Pedro Aguirre, Estacion Central and Quinta Normal. We then explore whether business formation increases more in those near SUP’s headquarters and for those sector most related to SUP. In future versions of this paper we will summarize results from this analysis. 4.3.

WEB-BASED MEASURES ARE INACCURATE PROXIES OF

REAL

EFFECTS An alternative interpretation of the findings is that our measures of performance are not capturing real effects. This should not be such a crucial concern as these web-based metrics are the metrics used by investors in start-ups, so they are relevant for this type of company. One natural argument would be that because the program is Chilean perhaps we should focus on local networks. However, SUP has an international focus. As we have already argued an important of participants are foreign (see table 2) and the vast majority do not end up in Chile. This means that the relevant networks to analyse are the foreign ones. The official language in SUP is English and the focus is international. In one of the interviews the executives mentioned than an internal exit via a local accelerator such as Wayra was considered a failure. Another idea would be to use data form Chilean registry but this is not feasible: most projects are registered abroad. However, the more general concern that more real variables would be ideal is a relevant point. To address this issue we conducted a survey. An explanation of the surveys and list of questions, and results will be presented in future versions of the paper. 4.4.

GOOD AT SELECTION

The differences in performance calculated using the OLS methodologyindicate that the accelerator is good at selecting. We explore this point further by investigating the correlation

26

between success and subsequent start-up performance. In future versions we will summarize results from this analysis. 5.

CONCLUSIONS

In this paper we provide new evidence performance of government sponsored programmes that sponsor entrepreneurship. We focus on business accelerates a neglected yet increasingly popular type of early stage financiers both in the public and the private sectors. We quantify the causal impact of a government-funded accelerator in Chile, SUP, by simultaneously exploiting novel, rich micro-data and addressing concerns about unobserved heterogeneity. We find that we cannot rule out that the government-sponsored accelerator has an impact on start-up performance. Using additional data from the mentor arm of the accelerator, however, we find stronger evidence that accelerator services related to mentorship positively impact start-up performance. In future versions of the paper we plan to contemplate several explanations for the findings. Including potential value-added reflected in founders’ income or regional spillovers.

27

References Applegate, Lynda, William Kerr, Joshua Lerner, Dina D. Pomeranz, Gustavo herrero and Cintra Scott. Start-up Chile. Harvard Business Review. Cohen, Susan and Yael Hochberg, 2014, Accelerating Startups: The Seed Accelerator Phenomenon, available at: http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2418000 Fehder, Daniel and Yael Hochberg, 2014, Accelerators and the Regional Supply of Venture Capital Investment, available at: http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2518668 Gonzalez-Uribe, Juanita, 2014, El caso de Start-up Chile. Programa de atracción de talento para fomentar el emprendimiento, CAF, Development Bank of Latin America. http://scioteca.caf.com/handle/123456789/685#sthash.3uyjaDPp.dpuf Hahn, J., Todd, P., Van Der Klaauw, W., 2001. Identification and estimation of treatment effects with a regression discontinuity design. Econometrica 69, 201–209 Hellman, Thomas and Manju Puri, 2002, Veture Capital and the professionalization of StartUp Firms: Empirical Evidence, The Journal of Finance, pp. 169-198. Kerr, William, Joshua Lerner and Antoinette Schoar, 2014, The Consequences of Entrepreneurial Finance: Evidence from Angel Financings, Review of Financial Studies, 27(1) pp. 20-55. Kortum, Samuel and Joshua Lerner, 2000, Assessing the Contribution of Venture Capital to Innovation, RAND Journal of Economics 31(4), 674-692. Lelarge, Claire, David Sraer and David Thesmar, 2013, Entrepreneurship and Credit Constraints: Evidence from a French Loan Guarantee Program in NBER volume on "International Differences in Entrepreneurship" edited by Joshua Lerner and Antoinette Schoar, University of Chicago Press. Lee, David and Thomas Lemieux, 2010. Regression Discontinuity Designs in Economics, Journal of Economic Literature, vol. 48(2), pages 281-355, June. Van Der Klaauw, W., 2002. Estimating the effect of financial aid offers on college enrollment: a regression-discontinuity approach. International Economic Review 43, 1249– 1287. Van Der Klaauw, W., 2008. Regression-Discontinuity Analysis: A Survey of Recent Developments in Economics, Labour: Review of Labour Economics and Industrial Relations, Vol. 22 (2), 2008, p.219-245.

28

0

.2

.4

.6

.8

Figure 1 – Fraction of accelerated applicants

-100

0

100

200

z=Rank-cutoff The figure shows the average fraction of accelerated applicants in bins of 10 transformed ranks (i.e., 𝑧) and the fitted values and 90% confidence interval from the regression mode: 𝑎𝑐𝑐𝑒𝑙𝑒𝑟𝑎𝑡𝑖𝑜𝑛 = 𝛿 + 𝛾𝑎𝑏𝑜𝑣𝑒 + 𝑓(𝑅𝑎𝑛𝑘 − 𝑐𝑢𝑡𝑜𝑓𝑓) + 𝜀, where the outcome variable acceleration is an indicator variable that equals 1 if the applicant participated in the accelerator, and 𝑓(𝑅𝑎𝑛𝑘 − 𝑐𝑢𝑡𝑜𝑓𝑓) is a 4th degree polynomial of the transformed rank. The vertical line represents the ranking cutoff normalized at 0 for the modified ranking.

29

Figure 2 – Cross-sectional Covariates Age

28

.1

.15

30

.2

32

.25

34

.3

36

.35

Chilean

0

100

200

-100

0

z=Rank-cutoff

z=Rank-cutoff

Gender

Money_Raised

100

200

100

200

.2

.15

.2

.3

.25

.4

.3

.5

.35

.6

.4

-100

-100

0

100

200

-100

0

z=Rank-cutoff

z=Rank-cutoff

.3

.4

.5

.6

.7

Prototype

-100

0

100

200

z=Rank-cutoff

The figure shows that predetermined variables are continuous at the cutoff for applicants. Five plots are shown for the variables Age, Chilean (i.e., a variable that equals one if the applicant leader is Chilean), Gender, Money Raised (pre application), and Prototype (i.e., a variable that equals one if the project already has a prototype). All variables as of the application date. Plots show averages grouped in bins of 10 applicants. The plots also show the fitted values and 90% confidence interval of a modified versions of the regression in equation (1), 𝑜𝑢𝑡𝑐𝑜𝑚𝑒 = 𝛼 + 𝛽𝑎𝑏𝑜𝑣𝑒 + 𝑓̆(𝑅𝑎𝑛𝑘 − 𝑐𝑢𝑡𝑜𝑓𝑓) + 𝜀, with each of these variables as outcomes, on 𝑎𝑏𝑜𝑣𝑒, a variable that equals 1 if the applicant ranks above 100 th in its generation and 0 otherwise, and 𝑓̆(𝑅𝑎𝑛𝑘 − 𝑐𝑢𝑡𝑜𝑓𝑓) a 4th degree polynomial of the modified rank (i.e., 𝑧 = 𝑅𝑎𝑛𝑘 − 𝑐𝑢𝑡𝑜𝑓𝑓). The vertical line represents the ranking cutoff normalized at 0 for the modified ranking.

30

.1

.2

.3

.4

.5

Figure 3 – Effect of participation in accelerator on subsequent performance

-100

0

100

200

z=Rank-cutoff The figure examines the effect of participation in the accelerator on performance for all applicants irrespective of whether they participated or not (i.e., the reduced form estimates). The plot shows the average value of AngeList (i.e., a variable that equals 1 if the project has a listing on AngeList by December 2013) in bins of 10 applicants. The plots also show the fitted values and 90% confidence interval of a modified version of the regression in equation (1), 𝑜𝑢𝑡𝑐𝑜𝑚𝑒 = 𝛼 + 𝛽𝑎𝑏𝑜𝑣𝑒 + 𝑓̆(𝑅𝑎𝑛𝑘 − 𝑐𝑢𝑡𝑜𝑓𝑓) + 𝜀, with AngeList as outcome, on 𝑎𝑏𝑜𝑣𝑒, a variable that equals 1 if the applicant ranks above 100th in its generation and 0 otherwise, and𝑓̆(𝑅𝑎𝑛𝑘 − 𝑐𝑢𝑡𝑜𝑓𝑓) a 4th degree polynomial of the modified rank (i.e., 𝑧 = 𝑅𝑎𝑛𝑘 − 𝑐𝑢𝑡𝑜𝑓𝑓). The vertical line represents the ranking cutoff normalized at 0 for the modified ranking.

31

0

.5

1

1.5

Figure 4 – Fraction of mentored applicants

0

1

2 3 Pitch-Day Score

4

5

The figure shows the average fraction of mentored participants in bins of 0.2 scores by judges’ pitch-day scores, and the fitted values and 90% confidence interval from the regression: 𝑚𝑒𝑛𝑡𝑜𝑟 = 𝜏 + 𝜇𝐴𝑏𝑜𝑣𝑒3.6 + 𝑓(𝑃𝑖𝑡𝑐ℎ_𝐷𝑎𝑦 𝑆𝑐𝑜𝑟𝑒) + 𝜀 , where the outcome variable mentor is an indicator variable that equals 1 if the participant was mentored, 𝐴𝑏𝑜𝑣𝑒3.6 is an indicator variable that equals one if the participant scored above 3.6 during the pitch-day, and 𝑓(𝑃𝑖𝑡𝑐ℎ_𝐷𝑎𝑦 𝑆𝑐𝑜𝑟𝑒) is a 4th degree polynomial of the pitch-day score. The vertical line represents the implicit score cut-off of 3.6.

32

Figure 5- Cross-sectional covariates mentor arm Chilean

20

0

.2

30

.4

.6

40

.8

1

50

Age

0

1

2 3 Pitch-Day Score

4

5

0

1

2 3 Pitch-Day Score

4

5

4

5

Money Raised

-.5

0

0

.5

.5

1

1

Gender

0

1

2 3 Pitch-Day Score

4

0

5

1

2 3 Pitch-Day Score

0

.5

1

Prototype

0

1

2 3 Pitch-Day Score

4

5

The figure shows that predetermined variables are continuous at the cutoff for applicants. Five plots are shown for the variables Age, Chilean (i.e., a variable that equals one if the applicant leader is Chilean), Gender, Money Raised (pre application), and Prototype (i.e., a variable that equals one if the project already has a prototype). All variables as of the application date. Plots show averages grouped in bins of 10 applicants. The plots also show the fitted values and 90% confidence interval of a modified versions of the regression in equation (1), 𝑚𝑒𝑛𝑡𝑜𝑟 = 𝜎 + 𝜔𝐴𝑏𝑜𝑣𝑒3.6 + 𝑓̌(𝑃𝑖𝑡𝑐ℎ_𝐷𝑎𝑦 𝑆𝑐𝑜𝑟𝑒) + 𝜖, with each of these variables as outcomes, , 𝐴𝑏𝑜𝑣𝑒3.6 is an indicator variable that equals one if the participant scored above 3.6 during the pitch-day, and 𝑓(𝑃𝑖𝑡𝑐ℎ_𝐷𝑎𝑦 𝑆𝑐𝑜𝑟𝑒) is a 4th degree polynomial of the pitch-day score. The vertical line represents the implicit score cut-off of 3.6.

33

Figure 5- Effect of mentoring on subsequent performance

Employees Linkedin

-20

-1

0

0

20

1

40

2

60

Facebook Likes

0

1

2 3 Pitch-Day Score

4

5

0

1

2 3 Pitch-Day Score

4

5

The figure shows that predetermined variables are continuous at the cutoff for applicants. Five plots are shown for the variables Age, Chilean (i.e., a variable that equals one if the applicant leader is Chilean), Gender, Money Raised (pre application), and Prototype (i.e., a variable that equals one if the project already has a prototype). All variables as of the application date. Plots show averages grouped in bins of 10 applicants. The plots also show the fitted values and 90% confidence interval of a modified versions of the regression in equation (1), 𝑚𝑒𝑛𝑡𝑜𝑟 = 𝜎 + 𝜔𝐴𝑏𝑜𝑣𝑒3.6 + 𝑓̌(𝑃𝑖𝑡𝑐ℎ_𝐷𝑎𝑦 𝑆𝑐𝑜𝑟𝑒) + 𝜖, with each of these variables as outcomes, , 𝐴𝑏𝑜𝑣𝑒3.6 is an indicator variable that equals one if the participant scored above 3.6 during the pitch-day, and 𝑓(𝑃𝑖𝑡𝑐ℎ_𝐷𝑎𝑦 𝑆𝑐𝑜𝑟𝑒) is a 4th degree polynomial of the pitch-day score. The vertical line represents the implicit score cut-off of 3.6.

34

Figure 5 – Comunas in Santiago de Chile

The figure plots the names of the different Comunas in Santiago de Chile. SUP is located in the Santiago Comuna.

35

Table 1 - Composition sample Panel A: Selected and formalized start-ups by generation Generation

Selected Obs.

Mean

Participated Std. Dev.

Obs.

Mean

Total Std. Dev.

1

86

0.68

0.47

64

0.51

0.50

126

2

150

0.32

0.47

125

0.26

0.44

474

3

99

0.25

0.43

85

0.22

0.41

394

4

98

0.21

0.41

74

0.16

0.36

472

5

101

0.15

0.36

90

0.14

0.34

655

6

105

0.18

0.39

95

0.16

0.37

581

7

100

0.18

0.38

83

0.15

0.36

556

Total

739

0.23

0.42

616

0.19

0.39

3,258

Panel B: Capital raised before application by generation Generation 1

2

3

4

5

6

7

Total

1

462

3

13

0

0

0

479

107

10

290

354

492

450

357

2,060

10

1

72

72

116

92

134

497

50 K to 100K

3

1

20

15

24

24

50

137

100K to 500K

5

0

0

0

0

0

0

5

500K to 1 M

0

0

7

13

19

11

14

64