Governments and donors around the world spend billions of dollars subsidizing entrepreneurship training programs. Unbiased evaluation estimates are paramount to understanding whether subsidies and investments in these programs are justified. The goal of this paper is to compare nonexperimental to experimental methods for evaluating the effectiveness of entrepreneurship training programs that provide training in management, marketing, accounting, legal, and other aspects of starting and running businesses. For the comparison, I take advantage of an unprecedented setting in which experimental and nonexperimental estimates are derived from the same underlying population of study participants. The Growing America through Entrepreneurship field experiment provides a uniquely large sample allowing for a separate nonexperimental analysis using only the control group. Experimental estimates indicate null effects of entrepreneurship training on business outcomes: business ownership, sales, and employment. In contrast, nonexperimental estimates using an extremely rich set of controls, including typically unobservable characteristics, such as previous family business experience, credit problems, and personality traits, find large, positive, and statistically significant effects. The nonexperimental estimates range from 21 to 22 percentage points (pp) for business ownership, $1300–2000 for average monthly sales, and 4–6 pp for hiring any employees at 1.5- and 5-year follow-ups. Nearest-neighbor and propensity-score matching models using detailed individual characteristics provide similarly large, positive, and statistically significant effects of entrepreneurship training on business outcomes. The findings have implications for choosing evaluation methods of government programs and provide evidence of positive selection bias which has more general implications on the scalability of entrepreneurship training programs to broader populations.

中文翻译：

---

评估创业培训：现场实验对于评估影响有多重要？

世界各地的政府和捐助者花费数十亿美元资助创业培训项目。公正的评估估计对于了解这些计划的补贴和投资是否合理至关重要。本文的目的是比较非实验性方法和实验性方法，以评估创业培训项目的有效性，这些项目提供管理、营销、会计、法律以及创业和经营企业其他方面的培训。为了进行比较，我利用了前所未有的环境，其中实验和非实验估计来自相同的基础研究参与者群体。“通过创业发展美国”实地实验提供了一个独特的大样本，允许仅使用对照组进行单独的非实验分析。实验估计表明创业培训对企业成果（企业所有权、销售和就业）没有影响。相比之下，使用极其丰富的控制组（包括通常不可观察的特征，例如以前的家族企业经验、信用问题和个性特征）进行的非实验估计发现了巨大的、积极的、统计上显着的影响。非实验性估计范围为：企业所有权为 21 至 22 个百分点 (pp)，平均月销售额为 1300-2000 美元，在 1.5 年和 5 年的后续行动中雇用任何员工为 4-6 个百分点。使用详细的个人特征的最近邻和倾向得分匹配模型提供了创业培训对业务成果同样巨大的、积极的、统计上显着的影响。研究结果对选择政府计划的评估方法具有影响，并提供了积极选择偏见的证据，这对创业培训计划向更广泛人群的可扩展性具有更普遍的影响。