Public funding for startups in Argentina: an impact evaluation

See Appendix A for further description of the selection process.

The response rate was lowered by the several monitoring surveys that BAE had previously implemented; the burden of replying progressively reduced participation and response rates. Appendix B provides details on the survey and questionnaire.

While the overall process remains similar in every program edition, both the scoring and the cutoff values change slightly. See Appendix A for details.

This methodology has been applied by, among others, Thistlethwaite and Campbell (1960) to analyze the impact of merit awards on future academic outcomes, Hoxby (2000) to evaluate the impact of class size on educational outcome, and Pettersson-Lidbom (2008) to estimate the impact of political parties on fiscal policies. See van der Klaauw (2008) for a survey.

As an example, consider the applicants to the 2011 edition. The application with the lowest score was a Web application for medical records and appointment management. The application with the highest score was a Web platform for sharing and managing audiovisual files. The projects close to the cutoff point (with scores around 55) were a Webpage for selling office supplies (a beneficiary) and a Webpage for managing and promoting student exchanges. These four ideas are not substantially different on the surface.

In Appendix E, we present the results of these regressions. Each table refers to a different outcome; the first two columns of the tables refer to the simple differences between beneficiaries and non-beneficiaries with no additional controls (column 1) or simply with program year dummies (column 2). The following columns report the regression discontinuity estimations. We tried different functions of the score, such as quadratic (columns 3, 5, and 7), cubic (column 4), and a Chebyshev polynomial of degree 3 (column 6). Additionally, we considered a subsample of those with scores between 45 and 65 (column 7), and we included additional controls (column 5) comprising a quadratic function of age and dummies for gender, industry, and education level. The results presented in the main text correspond to columns 3, 5, and 7.

As a robustness check, we alternatively defined duration as the number of months since the start of the relevant BAE edition. We also imposed censoring at different durations (e.g., at 48, 36, or 24 months). We found similar or stronger effects.

This is a proportional hazard model in which hazard h at duration t is estimated through a partial likelihood of the form \(\log h(t|X) = \alpha _{t}+\theta 'X\) and where the underlying hazard, \(\alpha _{t}\), is unspecified and identified after estimating the coefficients \(\theta\). Note that the hazard refers to the probability that the firm ceases its operations at a given duration. Thus, a negative coefficient on the “treated” coefficient implies that being a beneficiary increases the survival rate (i.e., the firm’s longevity).

A caveat to our exercise is that applications could represent more than one entrepreneur and the effects on employment could be driven by a preference among BAE selectors for larger groups of entrepreneurs. However, we find no correlation between the score and number of entrepreneurs listed on the application. Furthermore, when we compute initial employment figures minus the number of entrepreneurs listed on the application (i.e., non-founders), we find even higher and more significant effects, both in the regressions and in the local analysis. We thank an anonymous referee for pointing out this issue.

Throughout the paper, we assume that there is no displacement effect of BAE. In other words, the program is not diverting any other source of financing or harming the performance of other competing startups or firms. Given the small scale of the program, this “partial equilibrium” approach seems a reasonable assumption.

Appendix C provides more details about this exercise.

The average formal monthly wage was about US$ 1300 in 2011. Thus, the cost per job equates to about 3 months of wages.

Only three projects, one in the treatment group and two in the control group, received additional support from other programs, and these projects were not close to the cutoff of the running variable.