Appendix A: Detailed Explanation of Observational Analyses 1. Empirical Context: Crowdfunding A significant impediment to commercializing new ideas

Transcription

1 Appendix A: Detailed Explanation of Observational Analyses 1. Empirical Context: Crowdfunding A significant impediment to commercializing new ideas is the availability of seed-capital or early-stage funding. New ventures often find it difficult to raise funding from traditional sources of entrepreneurial capital, such as angel investors, banks and venture capital funds. As a result, entrepreneurs are increasingly turning to large, online communities of consumer-investors (Agrawal et al. 2013; Mollick 2014). This relatively new form of informal venture financing called crowdfunding allows entrepreneurs to directly appeal to the general public for financial help in getting their innovative ideas off the ground (Belleflamme et al. 2014). Specifically, project founders to raise funds by asking for small donations from a large number of individuals over the internet, sometimes in return for the product or service pitched. Convincing individuals to contribute involves a substantial investment of time, money, and social capital. Moreover, it entails a long-term commitment from project founders to fulfill project requirements. Initiating a crowdfunding project is an entrepreneurial endeavor a recent survey indicated that 90% of large crowdfunding projects in technology, videogames, and product design turned into ongoing businesses (Mollick and Kuppuswamy 2016). In 2015, crowdfunding platforms raised $34 billion collectively (Massolution 2015) and the number and variety of sites continues to expand (Younkin and Kashkooli 2016). Kickstarter alone has raised over $2.6 billion in pledges from 11.7 million backers to fund almost 113,000 creative ideas (Kickstarter 2016). The scale of crowdfunding projects continues to increase, with recent successes raising tens of millions of dollars. For example, one of the largest Kickstarter crowdfunded projects to date is Eric Migicovsky s e-paper Pebble Watch that integrates with an Android or iphone even though its original goal was $100,000, the project eventually raised over $10.2 million in 37 days from well over 65,000 backers. In February 2015, the founders of the Pebble watch went back to the crowd and raised an additional $20.3 million from over 78,000 backers. In theory, the relative ease of posting projects on crowdfunding platforms should reduce entry barriers for potential entrepreneurs, increasing the diversity of the founder pool. Moreover, since anyone can view projects posted online, the diversity of the pool of potential supporters increases considerably as well. As a result, crowdfunding represents a promising funding alternative for entrepreneurs and ventures that may have struggled with more traditional financing channels. Consistent with this view, recent studies show that crowdfunding provides support to high-quality projects (Mollick and Nanda 2015) that traditional experts would have likely rejected, and is particularly useful in improving the performance of female founders (Greenberg and Mollick 2016; Marom et al. 2015). For minority entrepreneurs, there is therefore optimism that by replacing a small set of geographically isolated and ethnically homogenous investors with a diverse and dispersed crowd, the significance of a founder s race will decline. 1

2 2. Data The data used in the observation field study comes from Kickstarter, the world s largest crowdfunding platform. Projects on Kickstarter are grouped into fifteen broad categories: Art, Comics, Crafts, Dance, Design, Fashion, Film and Video, Food, Games, Journalism, Music, Photography, Publishing, Technology, and Theater. To use Kickstarter, an individual (called creator in Kickstarter) creates a webpage for the project on the platform explaining the purpose of their project and the specific deliverables that they aim to produce with the contributed funds. Along with an end date for the project funding cycle, the founder also indicates the funding goal of the project, i.e., the amount of money they require to execute the project as specified. Founders can communicate with their supporters by posting public updates on the project webpage that everyone can see, or they can choose to communicate through social media. When a potential funder (called backer in Kickstarter) visits an active project s webpage, they are presented with all the project information initially posted by the founder. Moreover, potential backers are shown the current funding status of the project (i.e., the funds raised thus far), the ultimate funding goal, and the number of days remaining until the project funding cycle expires. An important feature of Kickstarter is its all-or-nothing aspect of fundraising. A project must be fully funded before its funding cycle concludes or no money pledged by any backer is transferred to the project founder. An over-ambitious funding goal may thus result in the fundraising effort falling short and consequently, raising no funds whatsoever. However, if a project does reach its fundraising goal, it can continue to receive contributions until its deadline. As a result, funded projects can exceed their original funding goal. While supporters in Kickstarter receive no financial returns, founders offer nonfinancial rewards in the form of tokens of appreciation or finished products. Our initial sample of 12,087 projects consists of every Kickstarter project launched between January 2012 and March 2014, with a goal of at least $5000 and with Facebook data available for the project founder (Mollick 2014) 1. We further restricted our sample to those 10,647 projects with a pitch video, in order to better measure the quality of the project idea (which we describe in greater detail below) 2. To determine the perceived race of a project s founder, we solicited respondents using Amazon s Mechanical Turk service to look at the photo associated with a given project and identify the race and gender of the founder (where available). Participants were recruited via a script that asked them to 1 We restrict our analysis to those projects with Facebook data available for the project s founder due to the importance of controlling for social network size as highlighted in prior work on crowdfunding (Mollick 2014). However, our results are similar if exclude social network size from the analysis and use the entire sample of projects. 2 Our results are similar if we exclude video pitch quality from the analysis and include projects without a video pitch. 2

3 Identify information about the founders of crowdfunding projects and paid $0.10 for each project they identified. We only accepted respondents with a 95% approval rating, located within the US, and with a minimum of 500 accepted prior HITs. To ensure reliability, respondents were asked two questions about the site (e.g. What is the founder s last name? and What was the project for? ), if they failed either check their answer was discarded and the URL was sent to a new rater. As a result, each project was assessed by a minimum of two participants and only coded as black if both raters agreed. In cases of disagreement, the projects were coded as non-black. Removing these projects from our study (Table A3) did not change the results. As a test of potential false-positives or false-negatives affecting the study, we randomly selected 100 projects identified as non-black by the raters and 100 projects identified as black, and then had two trained research assistants blindly (with no knowledge of the respondent s coding) identify the founder s race and gender. The results confirm the validity of the initial ratings. In 99.5% of cases the founders identified as black were verified as black, and in 98.5% of cases the non-black founders were also coded non-black (in three cases, one RA coded the race uncertain for a founder while the other RA and both mturkers coded the race non-black ). Of the 10,647 projects in our sample, 7,617 projects had photos that could be used to determine founder race and gender (others may have displayed a company logo or some other image). While these 7,617 projects represent our main sample, in additional robustness analysis we include projects without a founder picture in the sample to explore whether projects that conceal founder race fare differently than those with founder race visible. 2.1 Dependent Variables We use several project performance outcomes as dependent variables in our analysis. Our primary dependent variable, Project Funded, is a binary indicator that takes the value 1 if the project succeeded in reaching its fundraising goal, and 0 otherwise. Another project outcome, Raised, measures the total dollar amount raised by the project. We split Raised into two additional dependent variables Backers and Avg. Amount. While Backers measures the number of separate contributions received by the crowdfunding project, Avg. Amount measures the average size of a project s contributions, rounded to the nearest dollar (= Raised / Backers). Summary statistics of our main variables are displayed in Table 1. From Table 1, we see the mean value of Project Funded is 0.36 indicating that 36% of projects succeeded in achieving their fundraising goal. Our sample s success rate is very close to the overall success rate of 40% reported on the Kickstarter website. 2.2 Independent Variables 3

4 The main independent variable in our analysis is Black Founder, a binary indicator for whether the race of the founder is black. Black Founder takes the value 1 only if all raters agreed that the founder s race was black. This approach captures the perceived identity of the founder, irrespective of their self-identification, and allows us to estimate the influence of being seen as black by prospective supporters. We prefer a conservative measure, which requires full agreement, to better isolate the influence of founder race. Of the 7,617 projects in our sample, 556 projects (7.30%) were from black founders (Black Founder = 1). There are 201 additional projects (2.64%) where there was disagreement on the race of the founder, but at least one coder considered the founder black (Black Founder = 0). As seen in Table A3, the results are unchanged if we drop these projects from the analysis, rather than classify them as Black Founder = 0. In addition, given the central role played by videos in project pitches (Mollick 2014) and the need to control for potential differences in quality across projects, we created a novel measure of Video Pitch Quality. Specifically, we extracted the main video for each of the projects in our sample and showed it to subjects recruited through Amazon s mturk using the same protocol described above and a script offering $0.20 to watch and rate a crowdfunding video. The videos were shown in isolation, without additional details visible from the project s webpage, including any indication of the project s ultimate performance. Respondents were asked to rate (on a 5pt scale) the persuasiveness, professionalism, speaker s enthusiasm, and overall quality of the video (Chen, Yao, and Kotha 2009). To ensure the reliability of these ratings, we separately had two trained research assistants rate a random sample of 100 projects each, and their ratings were consistent with those given by the respondents (Cronbach's alpha > 0.8). Given the high degree of collinearity between these four measures, we combined them into a single scale measure of Video Pitch Quality (Cronbach's alpha > 0.9). Finally, we also control for a range of important variables used in prior studies of crowdfunding. First, we include an indicator, Female Founder, for the whether the founder of the project is female (Greenberg and Mollick 2016, Marom, Robb and Sade 2015). The gender of the founder was coded using an approach identical to the coding for the race of the founder. Second, consistent with recent work that has explored how project descriptions influence project outcomes (Gorbatai and Nelson 2015; Uparna and Bingham 2016), we use text-analysis software (LIWC) to measure key attributes related to the written pitch (Pennebaker et al. 2001). More specifically, we include measures for the number of words associated with negative emotions (Negative Words), positive emotions (Positive Words), and authenticity (Authenticity Words). Moreover, we control for linguistic content typically used by women (Female Words) and by men (Male Words). In addition, we control for the fundraising goal of the project in dollars (Project Goal), the median reward level associated with the project (Median Reward), the total number of words used to describe the project (Total Words), as well as fixed effects for the category associated with 4

5 the project and the month in which it launched. Finally, we also control for the number of Facebook Friends of the project founder (Facebook Friends). Prior research has shown the size of a founder s social network to be a strong predictor of crowdfunding success (Mollick 2014). Moreover, geographic differences in crowdfunding patterns (particularly early on in a campaign) largely stem from the strong support of a founder s friends and family, who tend to be more local than distant (Agrawal et al. 2015). As a result, controlling for social network size (Facebook Friends) also helps account for the magnitude of early (local) support founders may receive from their social network 3. As is clear in Table 1, many of our continuous variables are highly skewed to the right. As we expect diminishing returns to the effect of these variables over their range of values, we follow recent advice on modeling curvilinear relationships and log transform these variables when we include them in our empirical models (Haans et al., 2015). [Insert Table 1 Here] 2.3 Estimation Method In order to estimate the causal effect of founder race on crowdfunding performance, we must first address the possibility that projects from black and non-black founders differ in systematic ways. Underlying differences in project characteristics can lead to significant selection biases when examining the effect of race on performance (Heckman 1979). To reduce selection bias, we employ a nonparametric matching approach called coarsened exact matching (CEM) (Blackwell et al. 2009; Iacus et al. 2011). CEM involves coarsening a set of observed covariates, performing exact matching on the coarsened data, pruning observations so that strata have at least one treatment and one control unit, then running estimations using the original (but pruned) uncoarsened data (Aggarwal and Hsu 2014). The advantage of CEM (compared to other matching approaches like propensity score matching) is the ability to specify the degree of covariate balance ex-ante. However, as noted by Azoulay et al. (2013), the more finegrained the partition of the support for the joint distribution (i.e., the higher the number of strata incorporated into the analysis), the larger the number of unmatched, treated observations. As a result, there is a trade off between the quality of the matches and external validity the more covariates used in CEM, or the finer the partitions specified, the more difficult it is to identify a control match for each treatment observation. Consequently, the matched sample produced by CEM may be significantly smaller than the original sample (Bettis et al. 2014). In order to determine which covariates to match on, we identified the variables that differed significantly between projects by black and non-black founders. Model (1) of Table 2 displays the result 3 Agrawal et al. (2015) show that spatial differences in crowdfunding support are proxied by the support of a founder s offline friends and family, who tends to be local. We acknowledge the possibility that certain founders may have many offline friends or family, with few Facebook friends online (and vice-versa). Nevertheless, we assume a positive correlation between the size of founders offline friends and family network and the size of their online Facebook network. 5

6 of a logit model with Black Founder as the dependent variable. We find that the log-transformed variables Female Words, Facebook Friends, Project Goal, Total Words, and Median Reward, as well as certain category fixed effects, are significantly correlated with Black Founder (at the 5% level or lower). Moreover, Female Founder was marginally significant at the 10% level. As a result, to remove these underlying differences in projects characteristics between black and non-black founders, we matched on all these covariates. More specifically, we constructed a CEM matched sample of black and non-black founder projects by requiring exact matches for project category and Female Founder. Furthermore, we matched black and non-black founder projects on Female Words, Facebook Friends, Project Goal, Total Words, and Median Reward, using coarse buckets defined by the 5 th, 25 th, 50 th, 75 th, and 95 th percentiles of each log-transformed variable (Singh and Agrawal 2011). Our matched sample consisted of 663 projects 213 projects from black founders and 450 projects from non-black founders. In order to check whether the matched sample consisted of a more comparable set of projects from black and non-black founders, we re-estimated the logit model with Black Founder as the outcome using the matched sample (Model (2) in Table 2). As expected, none of the covariates (including the category fixed effects) remain significantly correlated with Black Founder in the matched sample. [Insert Tables 2 & 3 Here] 2.4 Results Using our matched sample of black and non-black founder projects, we turn to the effect of founder race on project performance. In Model (4) of Table 2, we use our matched sample to estimate a logit model of Project Funded as a function of Black Founder and our control variables (for reference purposes, Model (3) of Table 2 displays the logit results estimated using the full sample). We note that several control variables that are significant predictors of Project Funded in the full sample lose their significance in the model limited to the matched sample (e.g., Video Pitch Quality) a result due in part to the much smaller size of the matched sample. However, Black Founder has a negative coefficient that remains significant at the 1% level. As a result, we have strong evidence that black founders are significantly less likely to succeed compared to non-black founders. To evaluate the size of the negative effect, we compute the marginal effect of Black Founder (using the results of Model 4). We find that the probability of success falls from 0.40 for non-black founders, to 0.18 for black founders a decrease of 55%. The negative effect of Black Founder is observed again when we model Raised using a log-linear specification in Model (5). Interpreting the coefficient of Black Founder, we find that projects from black founders raise 86.1% less than comparable projects by non-black founders. 6

7 The simplest explanation for these findings is that the networks of black founders are not as wealthy as those of non-black founders. If this assumption were true, given that Kickstarter permits contributions as small as $1, we would expect black founders to receive the same number of contributions as non-black founders (with comparable network size), but that these contributions would be smaller in size. We therefore test whether the performance of black founders reflects a decrease in the number of contributions, the average size of contributions, or both. To this end, we model Backers and Avg. Amount as a function of Black Founder, in Models (6) and (7) respectively (again using a log-linear specification). We see that Black Founder has a negative and significant coefficient (at the 1% level) in both Models (6) and (7), indicating that black founders receive both fewer contributions and smaller contributions, the former of which is hard to attribute to differences in network wealth. These results are robust to alternative specifications where Raised, Backers, and Avg. Amount are estimated using count models for both the entire sample of projects (Table A1) and for the sample CEM sample (Table A2). Additional tests of robustness are outlined in detail below. 3. Tests of Alternate Specifications 3.1. Errors in coding founder race or gender As noted earlier, we pursued a conservative approach where assigned Black Founder a value of 1 only when both coders agreed that the race of the founder was black, and 0 otherwise. Similarly, Female Founder was assigned a value of 1 only when both coders agreed that the founder was female. However, to ensure that disagreement between coders did not bias our results, we dropped the 452 projects (5.9% of the sample; 3.6% projects had disagreement on founder gender, while 2.6% of projects had disagreement on founder race) where there was disagreement in race or gender assessments. We then re-generated a CEM matched sample and re-estimated our models of project success. The results are displayed in Appendix Table A3 and are very similar to the results presented earlier in Table Alternative approaches to estimating the treatment effect of Black Founder While CEM has significant advantages over traditional matching approaches that rely on the use of propensity scores (Iacus et al. 2011), we show that our results are robust to using the latter, more traditional solution for estimating treatment effects. Propensity score matching (PSM) entails a two-step approach (Rosenbaum and Rubin, 1983). In the first step, we use a logit to estimate the likelihood that a project is from a black founder versus a non-black founder. Using the model estimates, we obtain propensity scores (i.e., predicted probabilities) that the project is from a black founder. With these propensity scores, we use nearest neighbor matching to generate a matched sample of projects from black and non-black founders. Following recommended matching guidelines, we enforce a caliper of 0.01 to 7

8 ensure high quality matches and seek three control matches (non-black founder projects) for each treatment observation (black founder project) to gain more precision in our estimates and to reduce sampling bias (Caliendo and Kopeinig 2008, Stuart and Rubin 2008). The results of this matching process are displayed in Appendix Table A4. Model (1) shows the first stage logit used to generate propensity scores and Model (2) shows that the matched sample (PSM ) no longer appears to exhibit significant covariate differences between control and treatment groups. Using this matched sample, we estimate a logit model of Project Funded as a function of Black Founder and the other control variables in Model (3) (Models 4-6 show the results of log-linear models with alternative project outcomes). We again find strong evidence that the black founders experience significantly less success than their non-black peers. Moreover, we estimate the sensitivity of the estimated treatment effect to unobserved factors (Becker and Caliendo 2007) 4. We find that unobserved factors must increase the odds of treatment (a project from a black founder) by a factor of at least 2.75 before the treatment effect is sensitive to these factors (critical value of Γ = 2.75). An alternative treatment-effects estimator that relies on propensity scores but not the selection of nearest neighbors is inverse-probability-weighted regression adjustment (IPWRA) (Wooldridge 2007). The first stage of the IPWRA estimation entails the calculation of propensity scores as in the case of propensity score matching. However, instead of subsequently matching treatment and control observations on these propensity scores, treatment effects are calculated using regression adjustment with propensity scores as weights. This IPWRA approach enables one to consistently estimate the treatment effect parameters as long as we correctly specify at least one of the two models (either the outcome or treatment). As a result, IPWRA estimators are also known as Wooldridge s double-robust estimators (Wooldridge 2007). The results of the IPWRA estimator are presented in Appendix Table A5. Consistent with our earlier results, we again find that projects from black founders are significantly less likely to succeed compared to those from non-black founders Moderating effects of project category While the analysis above indicates that black founders are less likely to achieve crowdfunding success compared to non-black founders, it remains to be seen whether the negative effect is moderated by the category of the project. For instance, the increased success of women on crowdfunding platforms is largely the result of their disproportionate success in project categories where they are under-represented 4 We use the mhbounds command in STATA to determine the sensitivity of the estimated treatment effect. mhbounds calculates Rosenbaum bounds for average treatment effects on the treated in the presence of unobserved heterogeneity (hidden bias) between treatment and control cases, where both treatment and response variables are binary (Becker and Caliendo 2007). 8

9 (e.g., Technology) (Greenberg and Mollick 2016). To explore whether the negative effect of Black Founder varies by project category, we estimated a model of Project Funded where we interacted Black Founder with the category fixed effects and included them as additional covariates 5. Furthermore, due to the difficulty interpreting interaction terms in non-linear models (Ai and Norton 2003), we used a linear probability model to identify potential moderating effects. We find no significant interaction terms between Black Founder and the category fixed effects (see Appendix Table A6). As a result, in contrast to founder gender, the effect of founder race on project success does not appear to vary by project category. 3.4 Consideration of projects without a founder picture The preceding analysis has only focused on the performance of black founders relative to nonblack founders, using projects where founder race was discernable from the profile picture. However, as noted in the initial discussion of the dataset, a significant number of projects choose to forgo a picture of the founder in favor of some other image (e.g. company logo, personal avatar, etc.). Specifically, 3,030 projects, or 28.5% of the initial sample of 10,647 projects, lacked a picture where founder race was visible. This raises the question of how such projects fare relative to those where the race of the founder is visible. We created an indicator variable, No Picture, to denote projects without a profile picture that displays the founder s race. As before, underlying differences in project attributes between projects from black founders, non-black founders, and those without a personal picture must be accounted for before we can estimate treatment effects. We again use CEM to generate a matched sample of projects, now across three categories of treatment (black founder, non-black founder, founder without a picture). A multinomial logit in Model (1) of Table 3 shows the differences in project characteristics between the three treatment categories (to be consistent with prior analyses, the reference category is a project from a non-black founder). We constructed a CEM matched sample of projects by requiring exact matches for project category, and matching projects on Video Pitch Quality, log-transformed Female Words, Facebook Friends, Project Goal, Total Words, and Median Reward, using coarse buckets defined by the 25 th, 50 th, and 75 th percentiles of each variable. Due to an additional treatment level, our matched sample consists of even fewer projects than before 389 projects, with 80 projects from black founders, 194 projects from non-black founders, and 115 projects from founders without a picture. When we reestimated the multinomial logit from Model (1) using this matched sample (Model (2) in Table 3), we find that significant differences in projects characteristics across the 3 treatment categories have been eliminated. 5 We used the full sample rather than the matched sample for this analysis because the matched sample had several categories with few or zero projects. 9

10 In Models (3) and (4) of Table 3, we use a logit to model Project Funded as a function of Black Founder and No Picture, using the full sample and the matched sample, respectively. We again find that Black Founder has a negative and significant coefficient across both models (at the 1% level). Therefore, consistent with the earlier analyses, projects from black founders succeed at a significantly lower rate than do projects from non-black founders. No Picture has a positive and significant coefficient in Model (3), but an insignificant coefficient in Model (4), when the analysis is limited to the matched sample. As a result, we find no significant difference in performance between projects without a founder picture and those from non-black founders. Importantly, in Model (4), the coefficient of Black Founder is significantly different from that of No Picture (p < 0.01). Therefore, it is more beneficial for founders to conceal their picture, than to reveal themselves as black. For robustness, analyses with the matched sample using alternative project performance outcomes and model specifications are shown in Appendix Table A7. The results are consistent with those in Table Conclusion The field study above provides strong evidence that minority entrepreneurs experience less success on crowdfunding platforms. Further, ex-post analysis of the performance of successful projects indicates that black founders are no more likely to delay rewards or fail to deliver than non-black founders (Table 4), suggesting that the funding disparity does not reflect variations in ability. However, despite our attempts to reduce selection bias, through independent measures of project quality (using the video pitches) and coarsened exact matching (Iacus et al. 2011), some other unobserved confounding variable may nevertheless vary systematically between projects from black and non-black founders. Second, it remains unclear whether discrimination, net of founder and project quality, is conscious or unconscious, and whether it reflects statistical assumptions or dislike. Because of the known challenges in distinguishing between taste-based and statistical discrimination (Heckman 1998) we follow the recommendation of List (2004) and use a multi-stage experimental design to determine: (1) is there evidence of racial bias when founder and project quality is constant? And if so, (2) is the bias reflective of statistical assumptions, dislike, or unconscious associations? References V.A. Aggarwal, Hsu, D.H Entrepreneurial exits and innovation. Management Science, 60(4) A. Agrawal, Catalini, C. Goldfarb, A Some simple economics of crowdfunding (No. w19133). National Bureau of Economic Research. A. Agrawal, Catalini, C. and Goldfarb, A Crowdfunding: Geography, social networks, and the timing of investment decisions. Journal of Economics & Management Strategy, 24(2), pp C. Ai, Norton, E.C Interaction terms in logit and probit models. Economics Letters, 80(1) P. Azoulay, Stuart, T., Wang, Y Matthew: Effect or fable? Management Science, 60(1)

11 S.O. Becker, Caliendo, M Mhbounds-sensitivity analysis for average treatment effects. P. Belleflamme, Lambert, T. and Schwienbacher, A Crowdfunding: Tapping the right crowd. Journal of Business Venturing, 29(5) R. Bettis, Gambardella, A., Helfat, C., Mitchell, W Quantitative empirical analysis in strategic management. Strategic Management Journal 35(7) M. Blackwell, Iacus, S., King, G. Porro, G cem: Coarsened exact matching in Stata. Stata Journal, 9(4) 524. M. Caliendo, Kopeinig, S Some practical guidance for the implementation of propensity score matching. Journal of Economic Surveys 22(1) X.P. Chen, Yao, X., Kotha, S Entrepreneur passion and preparedness in business plan presentations: a persuasion analysis of venture capitalists' funding decisions. Academy of Management Journal. 52(1) A.D. Gorbatai, Nelson, L Gender and the Language of Crowdfunding. Academy of Management Proceedings (Vol. 2015, No. 1, p ). Academy of Management. J. Greenberg, Mollick, E Leaning in or leaning on? Gender, homophily, and activism in crowdfunding. Administrative Science Quarterly. forthcoming. R.F. Haans, Pieters, C., He, Z.L Thinking about U: theorizing and testing U and inverted U shaped relationships in strategy research. Strategic Management Journal. J.J. Heckman selection bias as a specification error. Econometrica: Journal of the econometric society J.J. Heckman Detecting Discrimination. The Journal of Economic Perspectives. 12(2) S.M. Iacus, King, G., Porro, G Causal inference without balance checking: Coarsened exact matching. Political analysis (August 23, 2011). Kickstarter Stats. Available at J.A. List The Nature and Extent of Discrimination in the Marketplace: Evidence from the Field. The Quarterly Journal of Economics. 119(1) D. Marom, Robb, A., Sade, O Gender Dynamics in Crowdfunding (Kickstarter). Available at SSRN: or Massolution, CF The Crowdfunding Industry Report. E. Mollick The dynamics of crowdfunding: An exploratory study. Journal of Business Venturing, 29(1), pp E. Mollick, Kuppuswamy, V Crowdfunding: Evidence on the Democratization of Startup Funding. In Revolutionizing Innovation: Users, Communities, and Open Innovation, edited by D. Harhoff, K. Lakhani. MIT Press. E. Mollick, Nanda, R Wisdom or madness? Comparing crowds with expert evaluation in funding the arts. Management Science, 62(6) J. Pennebaker, Francis, M., Booth, R Linguistic inquiry and word count [computer software]. Mahwah, NJ: Erlbaum Publishers. P.R. Rosenbaum, Rubin, D.B The central role of the propensity score in observational studies for causal effects. Biometrika, 70(1) J. Singh, Agrawal, A Recruiting for ideas: How firms exploit the prior inventions of new hires. Management Science, 57(1) E.A. Stuart, Rubin, D.B Best practices in quasi-experimental designs. Best practices in quantitative methods, pp J.M. Wooldridge Inverse probability weighted estimation for general missing data problems. Journal of Econometrics, 141(2) J. Uparna, Bingham, C The Positive of Negative: How Individuals Acquire Resources. Working Paper. P. Younkin, Kashkooli, K What Problems Does Crowfunding Solve? California Management Review. 58(2)

12 Appendix Table A1: Robustness analysis for full sample with alternative project performance outcomes and model specifications Log-Linear Poisson Log Log Log (Avg. (Raised) (Backers) Amount) Raised Backers Avg. Amount VARIABLES (1) (2) (3) (4) (5) (6) Black Creator *** *** *** *** *** *** (0.132) (0.0729) (0.0463) (0.123) (0.149) (0.0505) Video Pitch Quality 0.138*** 0.118*** ** 0.112*** (0.0295) (0.0188) (0.0101) (0.0299) (0.0513) (0.0137) Female Creator 0.339*** 0.227*** (0.0666) (0.0417) (0.0224) (0.0635) (0.0741) (0.0262) Log (Positive Words) 0.453*** 0.244*** 0.120*** 0.241* ** (0.104) (0.0620) (0.0352) (0.104) (0.280) (0.0453) Log (Negative Words) *** *** ** ** * ** (0.0888) (0.0553) (0.0296) (0.0762) (0.112) (0.0367) Log (Female Words) (0.0858) (0.0524) (0.0292) (0.0808) (0.104) (0.0385) Log (Male Words) (0.0782) (0.0473) (0.0269) (0.0729) (0.0875) (0.0437) Log (Authenticity Words) (0.0395) (0.0243) (0.0132) (0.0423) (0.0849) (0.0174) Log (Facebook Friends) 0.595*** 0.446*** *** 0.412*** 0.541*** (0.0272) (0.0164) ( ) (0.0308) (0.0423) (0.0121) Log (Project Goal) *** 0.420*** 0.265*** 0.187*** (0.0392) (0.0227) (0.0159) (0.0286) (0.0283) (0.0211) Log (Total Words) 0.372*** 0.248*** *** 0.245*** 0.304*** (0.0277) (0.0176) ( ) (0.0421) (0.0453) (0.0101) Log (Median Reward) 0.306*** 0.107*** 0.180*** ** *** (0.0426) (0.0236) (0.0150) (0.0271) (0.0277) (0.0174) Category Fixed Effects Yes Yes Yes Yes Yes Yes Month Fixed Effects Yes Yes Yes Yes Yes Yes Constant * *** 1.517*** *** 1.502*** (0.496) (0.298) (0.184) (0.502) (0.730) (0.260) N 7,617 7,617 7,617 7,617 7,617 7,617 R-Squared Pseudo R-Squared Robust standard errors in parentheses *** p<0.001, ** p<0.01, * p<0.05

13 Table A1 (continued) Negative Binomial Raised Backers Avg. Amount VARIABLES (7) (8) (9) Black Creator *** *** *** (0.0956) (0.118) (0.0486) Video Pitch Quality *** ** * (0.0216) (0.0265) (0.0116) Female Creator 0.147** 0.105* (0.0447) (0.0527) (0.0248) Log (Positive Words) ** (0.0721) (0.0947) (0.0372) Log (Negative Words) *** *** * (0.0613) (0.0656) (0.0354) Log (Female Words) (0.0620) (0.0711) (0.0345) Log (Male Words) (0.0560) (0.0637) (0.0355) Log (Authenticity Words) (0.0258) (0.0323) (0.0143) Log (Facebook Friends) 0.346*** 0.427*** (0.0250) (0.0242) (0.0102) Log (Project Goal) 0.505*** 0.285*** 0.183*** (0.0307) (0.0328) (0.0155) Log (Total Words) 0.150*** 0.168*** (0.0163) (0.0177) (0.0101) Log (Median Reward) 0.221*** * 0.199*** (0.0273) (0.0290) (0.0158) Category Fixed Effects Yes Yes Yes Month Fixed Effects Yes Yes Yes Constant *** 1.488*** (0.399) (0.428) (0.189) N 7,617 7,617 7,617 R-Squared Pseudo R-Squared Robust standard errors in parentheses *** p<0.001, ** p<0.01, * p<0.05

14 Table A2: Robustness analysis with CEM sample using alternative project performance outcomes and count model specifications Poisson Negative Binomial Avg. Avg. Raised Backers Raised Backers Amount Amount VARIABLES (1) (2) (3) (4) (5) (6) Black Founder *** ** *** *** *** *** (0.154) (0.206) (0.0876) (0.159) (0.155) (0.0780) Video Pitch Quality (0.0731) (0.0898) (0.0420) (0.0675) (0.0735) (0.0350) Female Founder * (0.152) (0.200) (0.106) (0.196) (0.212) (0.0966) Log (Positive Words) 0.520* 0.623* * 0.605* (0.217) (0.245) (0.119) (0.255) (0.263) (0.115) Log (Negative Words) ** ** * (0.176) (0.205) (0.142) (0.230) (0.241) (0.118) Log (Female Words) (0.221) (0.456) (0.176) (0.234) (0.265) (0.125) Log (Male Words) (0.180) (0.291) (0.134) (0.194) (0.217) (0.111) Log (Authenticity Words) ** ** ** (0.0889) (0.119) (0.0486) (0.0965) (0.100) (0.0446) Log (Facebook Friends) 0.551*** 0.812*** *** 0.649*** (0.0947) (0.115) (0.0479) (0.0805) (0.0957) (0.0381) Log (Project Goal) 0.306** * 0.320* (0.105) (0.133) (0.0578) (0.135) (0.139) (0.0565) Log (Total Words) 0.349*** 0.437*** *** 0.256*** (0.0960) (0.132) (0.0328) (0.0649) (0.0686) (0.0315) Log (Median Reward) 0.214* *** 0.882*** 0.588*** 0.304*** (0.0883) (0.135) (0.0618) (0.149) (0.142) (0.0572) Category Fixed Effects Yes Yes Yes Yes Yes Yes Month Fixed Effects Yes Yes Yes Yes Yes Yes Constant ** 1.980** ** 1.757** (1.644) (2.010) (0.685) (1.583) (1.683) (0.659) N Pseudo R-Squared Robust standard errors in parentheses *** p<0.001, ** p<0.01, * p<0.05

15 Table A3: Robustness analysis where projects with inconsistent coder assessments of race or gender are dropped Black Founder Full Black Founder CEM Project Funded Full Project Funded CEM Log (Raised) CEM Log (Backers) CEM Log (Avg. Amount) CEM VARIABLES (1) (2) (3) (4) (5) (6) (7) Black Founder *** *** *** *** *** (0.140) (0.285) (0.258) (0.153) (0.0850) Video Pitch Quality *** (0.0478) (0.0957) (0.0302) (0.122) (0.133) (0.0859) (0.0381) Female Founder *** ** 0.569* (0.108) (0.299) (0.0684) (0.373) (0.356) (0.225) (0.107) Log (Positive Words) ** * 0.661** (0.150) (0.343) (0.0959) (0.382) (0.404) (0.242) (0.120) Log (Negative Words) *** * (0.133) (0.304) (0.0910) (0.390) (0.366) (0.230) (0.113) Log (Female Words) 0.306* (0.119) (0.323) (0.0864) (0.390) (0.369) (0.218) (0.129) Log (Male Words) (0.122) (0.252) (0.0755) (0.333) (0.336) (0.205) (0.110) Log (Authenticity Words) ** (0.0630) (0.137) (0.0383) (0.161) (0.165) (0.0980) (0.0459) Log (Facebook Friends) 0.172*** *** 0.855*** 0.675*** 0.564*** (0.0515) (0.102) (0.0345) (0.171) (0.155) (0.0963) (0.0486) Log (Project Goal) 0.240*** *** *** (0.0488) (0.152) (0.0411) (0.223) (0.213) (0.126) (0.0648) Log (Total Words) *** *** 0.564* 0.604*** 0.410*** (0.0286) (0.104) (0.0340) (0.241) (0.162) (0.109) (0.0339) Log (Median Reward) *** *** 0.376* 0.892*** 0.426*** 0.316*** (0.0661) (0.141) (0.0353) (0.175) (0.235) (0.128) (0.0566) Category Fixed Effects Yes Yes Yes Yes Yes Yes Yes Month Fixed Effects Yes Yes Yes Yes Yes Yes Yes Constant *** ** * *** (0.728) (1.787) (0.496) (2.525) (2.609) (1.593) (0.814) N 7, , ψ R-Squared Pseudo R-Squared Robust standard errors in parentheses *** p<0.001, ** p<0.01, * p<0.05; ψ: 22 projects dropped from logit sample because 2 category indicators perfectly predict outcome.

16 Table A4: Robustness analysis with Propensity Score Matching (3 nearest neighbors; caliper = 0.01) Black Founder Full Black Founder PSM Project Funded PSM Log (Raised) PSM Log (Backers) PSM Log (Avg. Amount) PSM VARIABLES (1) (2) (3) (4) (5) (6) Black Founder *** *** *** *** (0.151) (0.143) (0.0810) (0.0498) Video Pitch Quality *** * (0.0470) (0.0528) (0.0633) (0.0642) (0.0395) (0.0213) Female Founder ** 0.357** 0.243** (0.107) (0.120) (0.129) (0.137) (0.0814) (0.0465) Log (Positive Words) ** (0.149) (0.175) (0.184) (0.220) (0.127) (0.0819) Log (Negative Words) ** *** *** (0.133) (0.147) (0.175) (0.177) (0.109) (0.0580) Log (Female Words) 0.317** (0.116) (0.130) (0.146) (0.163) (0.0953) (0.0547) Log (Male Words) * 0.191* (0.120) (0.134) (0.155) (0.160) (0.0961) (0.0530) Log (Authenticity Words) * (0.0613) (0.0677) (0.0741) (0.0770) (0.0454) (0.0250) Log (Facebook Friends) 0.165*** *** 0.676*** 0.492*** *** (0.0500) (0.0498) (0.0614) (0.0535) (0.0313) (0.0188) Log (Project Goal) 0.227*** *** * (0.0478) (0.0540) (0.0725) (0.0784) (0.0424) (0.0327) Log (Total Words) *** ** 0.236*** 0.145*** (0.0281) (0.0350) (0.0496) (0.0463) (0.0292) (0.0145) Log (Median Reward) *** *** 0.499*** 0.209*** 0.244*** (0.0644) (0.0653) (0.0651) (0.0837) (0.0449) (0.0310) Category Fixed Effects Yes Yes Yes Yes Yes Yes Month Fixed Effects Yes Yes Yes Yes Yes Yes Constant *** ** 1.534*** (0.710) (0.808) (0.937) (1.021) (0.585) (0.377) N 7,617 1,935 1,933 ψ 1,935 1,935 1,776 R-Squared Pseudo R-Squared Robust standard errors in parentheses *** p<0.001, ** p<0.01, * p<0.05; ψ: 2 projects dropped from logit sample because one category indicator perfectly predicts outcome.

17 Table A5: Robustness analysis with treatment effects estimated using Inverse Probability Weighting with Regression Adjustment (IPWRA) Group Comparison Average Treatment Effect (ATE) Project Funded Log (Raised) Log (Backers) Log (Avg. Amount) Black Founder = 1 vs. Black Founder = *** *** *** *** (0.0141) (0.127) (0.0697) (0.0468) Robust standard errors in parentheses *** p<0.001, ** p<0.01, * p<0.05

18 Table A6: Linear probability model of Project Funded to explore whether project category moderates the effect of Black Founder Project Funded Project Funded VARIABLES (1) (2) Black Founder *** ** (0.0164) (0.0690) Video Pitch Quality *** *** ( ) ( ) Female Founder *** *** (0.0127) (0.0127) Log (Positive Words) * * (0.0162) (0.0162) Log (Negative Words) *** *** (0.0150) (0.0150) Log (Female Words) (0.0151) (0.0151) Log (Male Words) (0.0131) (0.0131) Log (Authenticity Words) ( ) ( ) Log (Facebook Friends) *** *** ( ) ( ) Log (Project Goal) *** *** ( ) ( ) Log (Total Words) *** *** ( ) ( ) Log (Median Reward) *** *** ( ) ( ) Black Founder X Category[= Comics] (0.125) Black Founder X Category[= Crafts] (0.310) Black Founder X Category[= Dance] (0.249) Black Founder X Category[= Design] (0.0974) Black Founder X Category[= Fashion] (0.0836) Black Founder X Category[= Film & Video] (0.0754) Black Founder X Category[= Food] (0.0934) Black Founder X Category[= Games] (0.0855)

19 Table A6 (continued): Linear probability model of Project Funded to explore whether project category moderates the effect of Black Founder Black Founder X Category[= Journalism] (0.102) Black Founder X Category[= Music] (0.0780) Black Founder X Category[= Photography] (0.0934) Black Founder X Category[= Publishing] (0.0817) Black Founder X Category[= Technology] (0.0899) Black Founder X Category[= Theater] (0.116) Category Fixed Effects Month Fixed Effects Yes Yes Constant 0.242** 0.240** (0.0764) (0.0767) N 7,617 7,617 R-Squared Robust standard errors in parentheses *** p<0.001, ** p<0.01, * p<0.05

20 Table A7: Analysis of alternative measures of project success using CEM sample including projects without a founder picture Log-Linear Poisson Log Log Log (Avg. (Raised) (Backers) Amount) Raised Backers Avg. Amount VARIABLES (1) (2) (3) (4) (5) (6) Black Founder *** *** *** * * (0.406) (0.242) (0.141) (0.312) (0.300) (0.155) No Picture (0.375) (0.247) (0.108) (0.189) (0.236) (0.0950) Video Pitch Quality ** * * (0.209) (0.130) (0.0650) (0.123) (0.130) (0.0479) Log (Positive Words) 1.128* * (0.540) (0.350) (0.192) (0.437) (0.447) (0.152) Log (Negative Words) * ** (0.457) (0.301) (0.164) (0.319) (0.302) (0.147) Log (Female Words) (1.162) (0.733) (0.294) (0.501) (0.534) (0.238) Log (Male Words) * (0.458) (0.297) (0.141) (0.227) (0.231) (0.107) Log (Authenticity Words) * (0.232) (0.148) (0.0705) (0.127) (0.119) (0.0603) Log (Facebook Friends) 0.497** 0.320** *** 0.395*** (0.177) (0.0978) (0.0849) (0.113) (0.0982) (0.0830) Log (Project Goal) * 0.232* (0.188) (0.124) (0.0822) (0.106) (0.102) (0.0708) Log (Total Words) (0.137) (0.0949) (0.0323) (0.0680) (0.0902) (0.0317) Log (Median Reward) * ** (0.256) (0.137) (0.105) (0.103) (0.102) (0.0856) Category Fixed Effects Yes Yes Yes Yes Yes Yes Month Fixed Effects Yes Yes Yes Yes Yes Yes Constant ** (3.430) (2.065) (1.279) (1.801) (1.637) (1.035) N R-Squared Pseudo R-Squared Robust standard errors in parentheses *** p<0.001, ** p<0.01, * p<0.05

21 Table A7 (continued): Analysis of alternative measures of project success using CEM sample including projects without a founder picture Negative Binomial Raised Backers Avg. Amount VARIABLES (7) (8) (9) Black Founder *** *** ** (0.242) (0.218) (0.121) No Picture (0.236) (0.220) (0.0965) Video Pitch Quality * (0.104) (0.0951) (0.0549) Log (Positive Words) (0.374) (0.351) (0.174) Log (Negative Words) ** ** (0.255) (0.258) (0.133) Log (Female Words) (0.543) (0.507) (0.242) Log (Male Words) * (0.263) (0.265) (0.122) Log (Authenticity Words) * (0.118) (0.118) (0.0634) Log (Facebook Friends) 0.242* 0.273** (0.0952) (0.0833) (0.0632) Log (Project Goal) (0.203) (0.184) (0.0681) Log (Total Words) (0.0541) (0.0563) (0.0286) Log (Median Reward) 0.585* ** (0.241) (0.234) (0.0862) Category Fixed Effects Yes Yes Yes Month Fixed Effects Yes Yes Yes Constant ** (1.755) (1.676) (1.120) N R-Squared Pseudo R-Squared Robust standard errors in parentheses *** p<0.001, ** p<0.01, * p<0.05