Long-Term Productivity Effects of Public Support to Innovation in Colombia

ABSTRACT

In this study, we evaluate the effect of innovation promotion programs administrated by the Colombian Innovation Agency (COLCIENCIAS). The evaluation focuses on programs that provide financial incentives for research and development (R&D)—matching grants and contingent loans—and encourage the formation of linkages among firms, universities, and other public research organizations. We use longitudinal firm-level data and adopt a fixed effects identification strategy to control for potential selection biases. The findings show that COLCIENCIAS financial incentives have increased labor productivity as a result of gains in total factor productivity (TFP) due to product diversification and, to a lesser extent, of capital intensification.

KEY WORDS:

• COLCIENCIAS
• Colombia
• fixed effects
• innovation
• policy evaluation
• productivity

Acknowledgments

The authors are grateful to Edgar Castro and Laura García for data management. The authors also thank two anonymous referees for their helpful suggestions and comments. The opinions expressed in this publication are those of the authors and do not necessarily reflect the views of the Inter-American Development Bank. The usual disclaimer applies.

ORCID

Alessandro Maffioli

http://orcid.org/0000-0003-1576-1817

Notes

^1. For a review of the effects of fiscal incentives in developed countries, see Hall and Van Reenen (2000). For an evaluation of the Colombian case, see Mercer-Blackman (2008).

^2. For a complete discussion on this topic, see Hall and Maffioli (2008).

^3. These studies include Czarnitzki and Fier (2002), Czarnitzki and Hussinger (2004), and Hussinger (2008).

^4. Even regarding the dynamics of investments, evidence for Israel shows that companies tend to use first the grant component of a supported R&D project and then, when they are certain of the effects, tend to increase, sometimes substantially, the private-funding component. Overall, there is crowding-in, but the timing of the evaluation also matters for the study of this effect (see Lach 1999).

^5. On the Argentina and Chile cases, see also Benavente et al. (2012) and Binelli and Maffioli (2007).

^6. Given the confidentiality of the data, the estimations were conducted following DANE’s microdata-access policy, which implies working in situ under the supervision of DANE’s staff and with blinded access to sensitive information.

^7. Created between 1995 and early 2000s, the CTDs include National Centers for Sector Development, Incubators of Enterprises of Technological Base, Regional Centers of Productivity, and Technological Parks.

^8. COLCIENCIAS also manages R&D Tax Credits programs; however, these resources are not an integral part of the COLCIENCIAS budget, and so they are not included in the previous figures. The tax incentives have been in place since 1992 and comprise a 75 percent deductible over expenditures in science and technology and VAT exemptions for assets purchases. Beneficiaries of these programs are mostly large firms.

^9. See, for example, Levin et al. (1987), Mansfield et al. (1981), and Martin and Scott (2000).

^10. Since the seminal works by Arrow (1962) and Nelson (1959), scientific and technological knowledge has been defined as a durable public good; that is, nonexcludable and nonrival. Furthermore, the nonrival character of new knowledge intensifies the need for creating incentives that can compensate for the nonappropriable benefits.

^11. The uncertainty could be both technical and commercial in nature. In the former case, it is not clear ex ante if the research projects will be able to achieve the technical solution to certain problems. In the latter case, the uncertainty is related to difficult assessment of the final users’ willingness to pay for a product or services that cannot be tested yet. Finally, the uncertainty and indivisibility of knowledge investments cause an even greater suboptimality in the allocation of resources.

^12. The regulation may allow and encourage firms to coordinate their R&D investment during the first stage of a project (e.g., the basic research stage) and then force them to engage in Cournot or Bertrand-type competition in the second stage (e.g., prototype development). On this topic, see, among others, Martin and Scott (2000).

^13. See Dosi and Nelson (1994), Metcalfe (1994), and Nelson and Winter (1982).

^14. According to the “network of learning” (Powell et al. 1996) and to the “interactive learning” approaches (Lundvall 1992; Morgan 1996), networks facilitate organizational learning and act as a locus of innovation. Thus, “organizational learning is both a function of access to new knowledge and the capabilities of utilizing and building on such knowledge” (Powell et al. 1996, 118).

^15. EAM firm-level data are protected by a statistical reserve regulation. We were able to work with them at DANE’s offices, thanks to a special cooperation agreement to facilitate academic research.

^16. Our panel is unbalanced. Because the Colombian Annual Manufacturing Survey data is a complete census of the manufacturing sector, any unbalance in the panel is only due to nonresponse and firm survival. The construction of a balanced panel including only those firms that survive and respond during the whole period of analysis could bias our estimates. For this reason, we prefer not to introduce any restrictions to the data in addition to those determined by our identification strategy (i.e., a common support and the attribution of weights).

^17. fully absorb any permanent heterogeneity at the firm, industry, and region level.

^18. For additional methodological discussion regarding pretreatment trends, please refer to Blundell and Costa Dias (2000), Dehejia and Wahba (1999), and Imbens et al. (2001).

^19. When estimating the probit, we exclude from the pool all posttreatment observations of beneficiary firms.

^20. We adopt a min-max criterion and eliminate control group firms that present a higher or lower average propensity score than the maximum or minimum propensity score of the treatment group, respectively.

^21. For a complete discussion on this topic, see section 5.2.1 of Angrist and Pischke (2008).

^22. This test is what validates our fixed effects identification strategy (see Angrist and Pischke 2009, 237; Heckman and Hotz 1989).

^23. Despite the fact that the final sample is just 25 percent of the full sample in terms of firms, the main results still hold. This is a reassuring sign that sample size is not affecting the results.

^24. The growth rate of the number of products is computed using the average treatment effect of 0.14 products per firm and the average baseline value of 5.14 products per firm (see Table 4).

^25. In the Appendix, we show the results of iteratively running the fixed effects model on a weighted and not weighted common support of firms constructed using different nearest neighbor matching algorithms (starting from fifty-five nearest neighbors down to five). Independently of the number of neighbors and weights used, results are similar in magnitude and significance to the ones discussed in this section.