Introduction

Innovation policy has emerged as a new area of economics during the last three decades (^{Fagerberg, 2017}). This is explained because the innovation is crucial for long-term economic growth and introduces novelty (variety) into the economy, but the innovation tends to cluster in certain sectors, which explains differences in performance between firms, regions, and countries (^{Fagerberg, 2006}). Moreover, innovation is a source of productivity for firms, but due to its public good characteristics, there is a suboptimal level of innovative efforts (^{d'Artis & Siliverstovs, 2016}, ^{Crespi, 2012}, among others). For the above reasons, the role of the state in innovation processes could be seen as correcting market and system failures that limit the development of knowledge, technology, and innovation (^{Borras & Edler, 2020}; ^{Edler & Fagerberg, 2017}; ^{Martin & Scott, 2000}).

In the case of Chile, in 1939 the Production Development Corporation (CORFO) was created, which in its beginnings helped in the industrialization of the country but changed its development strategy in the following decades due to the new development model based in the free market and openness to international trade. At the beginning of the 90's CORFO began to develop a set of programs of subsidies and credits to support the growth of the firms. At present, CORFO seeks to boost research and technological development with economic impact, encourage collaboration between firms, increase market competitiveness, encourage private investment and facilitate funding; then it has designed different programs with specific characteristics according to the desired objectives.

According to the data from the Longitudinal Survey of Firms, most programs used by firms and funded by CORFO are: a) The Guarantee Fund for Small Business Owners (FOGAPE) that allows access to funding for productive purposes, either working capital or investment; b) The Supplier Development Program (PDP) that seeks to support supplier firms or firms that have the potential to become direct suppliers of large firms, subsidizing part of the cost of diagnostic activities, design and implementation of a work plan to improve their skills and competences; c) The INNOVA program that includes different funding lines focused at promoting innovation, knowledge transfer, and technological capabilities; d) The Quality Development Program (FOCAL), which seeks to improve the competitiveness and productivity of firms by encouraging the implementation and certification of management system standards, protocols or products in their firms that allow access to more sophisticated or export markets; e) Technical Assistance Fund (FAT), which is a non-reimbursable contribution that covers part of the cost of a specialized consultancy to improve the firm management.

Despite the proliferation of many of these public funding programs, there is little evidence to support their effectiveness on innovative outputs in Chilean firms. For this reason, the present study seeks to complement previous studies¹ and evaluate the impact of CORFO on innovation in products and/or services, identifying if these programs have displaced private investment.

For this study, the information obtained from the three versions of the Longitudinal Survey of Firms (ELE) conducted in Chile covering the years 2007, 2009, and 2013 is used. The impact of the programs is estimated using econometric techniques for panel data using the subgroup of firms present in the three versions of the survey.

Literature review

Available data

The Longitudinal Survey of Firms (ELE) includes three waves conducted in the years 2007, 2009, and 2013, all drawn up by the Ministry of Economy and the National Statistics Institute. In general terms, the survey consists of five modules: accounting and finance, marketing and entrepreneurial environment, general management, human resources, and information and communication technologies.

In 2007 there were records of 10,213 firms surveyed, of which 152 received funding from a CORFO program, in 2009 there were records of 7,062 firms, of whom 218 received funding from a CORFO program, and in the year 2013 there are records of 7,267 firms of whom 292 received funding from some CORFO program. However, it is necessary to remark that the same firms have not been surveyed in all the ELE waves, some of them only in the first wave, others in the first and second (2,667 firms), others in the second and third (559 firms), and the same subgroup is in the first, second, and third (559 firms).

Table 1 presents the descriptive statistics of the variables used in this study. Firms that declare innovation for the introduction of new products and/or services correspond to 16.3% in 2007, 11.5% in 2009² and 15.6% in 2013. Most of the firms included in the survey are medium and large, and also belong to the industrial sector, construction and transport - communications. The average real sales of firms are approximately 52 billion of Chilean pesos in 2007, 49 billion of Chilean pesos in 2009 and 62 billion of of Chilean pesos in 2013. On the other hand, 92.3% of the firms are nationally owned, 6.7% foreign owned, and the rest are publicly owned firms. Most of the surveyed firms are limited liability firms (36.9%) and closed corporations (34.0%). In addition, 2.9% of the firms that are present in the three waves of ELE have obtained the funding from some program of CORFO in some waves. The PDP program has financed firms included in the three waves of the ELE, the INNOVA program has funded firms included only in the second and third waves of the ELE, the FOGAPE and FAT programs have funded firms included only in the first and second waves of the ELE, while other CORFO programs have funded firms included in the three follow-up waves. Finally, it should be mentioned that questions associated with innovative inputs, such as recruiting R&D personnel or having an R&D department, are only available for the second and third waves of the ELE survey.

Methodology

One of the main difficulties in identifying the impact of a public funding program on innovation is that it is not possible to observe the result variable (in this case innovation in products and/ or services) for the same firm with treatment (public funding program) or without treatment at the same instant of time.

This would not generate an econometric identification problem if the treatment was randomly assigned to the firms, but in practice it is not common that there is a random assignment due to the self-selection of the applicants to participate in this type of programs. However, it is possible to use a difference in differences estimator to estimate the average treatment effect in the treated group (assuming that the control group is a good estimate of the change experienced by the treated group if it had not received the treatment. Specifically, this method assumes that unobservable factors remain constant over time, that is, that there are common temporal effects across the treatment and control groups, and further, that it does not systematically change the composition within each group. The differences in differences estimator can also be implemented with panel data for longer periods through the following fixed-effects specification:

Where y_it is the dependent variable (innovation in products and/or services) of firm i in period t, x_it corresponds to the vector of explanatory variables (which includes a constant, characteristics of the firms and if they were awarded by the different programs of CORFO), β is the parameter vector of interest (which includes the effect of each of CORFO's funding programs), and in addition, the random error term is composed of a specific unobservable effect for each i (μ_i) a specific unobservable effect for each t(λt), and an effect that varies with i y t (ε_it).

Among the benefits of using panel data is the control of individual heterogeneity, which reduces the risk of biased results. In addition, panel data provides more information, more variability, less colinearity between variables, more freedom degrees, and greater efficiency. One of its main attractions is that it allows identifying and measuring effects that are not easily detectable with cross-sectional data or time series. However, there may be problems with the design and collection of information as firms could deliver incomplete data or simply refuse to respond over time, which can lead to distortions in measurement of variables and selectivity problems.

In the case that the dependent variable is binary (innovates or does not innovate in products and/or services) the structural model can be specified as:

As before, μ_i is the unobserved individual heterogeneity. In this case μ_i and β are the parameters to be estimated, but when N →∞ given a fixed T, the number of μ_i parameters is increased with N, so they can not be estimated consistently due to the problem of incidental parameters. In addition, the parameter β can not be consistently estimated with a model in differences when the dependent variable is binary.

Therefore, Chamberlain (1980) suggested to maximize a conditional distribution function to obtain estimates of β. In the case of a fixed-effects logit model only the observations of the individuals that change their state should be used, which implies to use a traditional logit that includes as explanatory variables to xit'-xit-1'. whereas the dependent variable takes the value 1 if y_it a changes from 1 to 0, and the value 0 if y_it changes from 0 to 1. This estimator is consistent and efficient only if there are no individual effects, otherwise they are inefficient because not all data are used. In addition, this approach does not generate computational simplifications that allow to estimate a probit model of fixed effects.

However, assuming that μ_i is not related to x_it the conditional distribution μ_i is not dependent on x_it, generating the so-called random effects binary model³. The probit specification is popular for estimating this model, whose fundamental assumption is that there is equicor-relation between successive errors belonging to the same unit. In this case u_i = μ_i + ε_it, it is also assumed that μi∼IIN0,σμ2 and εit∼IIN0,σε2 are independent of each other and independent of xit'. As Euit,uis=σμ2 for t ≠ s, the joint probability (y_1t, …, y_Nt) can not be written as the multiplication of the marginal probabilities of y_it, which complicates the estimation by maximum likelihood because it involves integrals with T dimensions. Fortunately, by conditioning for individual effects, the problem is reduced to a single integral involving the multiplication of a standard normal density function and T differences of two normal cumulative functions that are estimated with a Gaussian quadrature procedure.

Results

Different specifications of empirical models controlled by different groups of variables were generated to evaluate the impact of different CORFO's funding programs on the probability of innovations in products and/or services in Chilean firms. Specification 1 only includes funding programs without control variables. Specification 2 also includes control variables such as sales, number of employees, percentage of national private ownership, percentage of foreign private ownership, type of legal form of the firm (natural person, limited liability firm, individual liability firm, collective society, closed corporation, and open corporation) and if the firm exports. In specification 3 also includes as control variable the size of the firm (small 1, small 2, medium, and large). In addition, the economic sector (agricultural, mining, manufacturing, electricity - gas - water, construction, commerce, hotels - restaurants, transportation - communications, financial, real estate - business services, and services).

Table 2 shows the probit regressions with random effects that include the data from the three waves of the ELE survey. This table shows that the probability of performing innovation in products and/or services is affected positively and statistically significant by obtaining funding through the INNOVA program in all specifications. Also, it is observed that firms with higher sales, greater numbers of workers and exports are more likely to innovate in products and/or services. Finally, those firms that belong to the manufacturing sector and hotels - restaurants are less likely to innovate in products and/or services.

As mentioned before, the variables that characterize the firm's innovative inputs are only available in the second and third waves of the ELE survey. Therefore, Table 3 excludes the observations of the first wave of the survey in the four previously mentioned specifications, and also includes the specification 5 that also controls by the possession of an R&D department and by the recruitment of R&D personnel.

The results of the probit regressions with random-effects presented in Table 3 show that the impact of the INNOVA program on the probability of innovating in products and services is very similar for the same four specifications reported in Table 2 (including data for the three waves of the ELE survey). However, the most relevant aspect of this analysis is that by including the variables that are controlled by the innovative characteristics of the firms (which are available only for the last two waves of the ELE survey) it is possible to observe that the positive and statistically significant effect of the INNOVA program disappears. This is explained by the inclusion of the variable that reflects the possesion of an R&D department. Thus, it can be concluded that what actually increases the probability of innovations in products and/or services is this characteristic of firms and not obtaining funding from the INNOVA program. It could even be concluded that as the holding of an R&D department is an observable characteristic for CORFO when firms make their applications to this program it would be easy to choose as beneficiaries of the program firms that possess this characteristic, which are more likely to innovate even in the absence of the INNOVA program, this is known in the literature as picking winners.

Previous studies that estimated the aggregate impact of public instruments on innovation in Chile using cross-section data conclude that the firms that receive public incentives are more productive and/or innovative (^{Bravo-Ortega et al., 2014}; ^{Álvarez et al., 2011}; ^{Benavente, 2005}). However, studies that have used panel data to analyze the impact of specific programs (^{Benavente et al., 2007}; ^{Benavente & Crespi, 2003}) or pseudo-panel data to evaluate the aggregate impact of public instruments (^{Mardones & Zapata, 2019}) do not find effects of public funding on product innovation. The results of these latest studies agree with the findings obtained in the present research, which demonstrates the importance of controlling endogeneity problems and selection bias that can affect the impacts obtained with cross-sectional data. However, a distinctive aspect of this research compared to previous studies is that it evaluates different public instruments and controls for the innovative characteristics at the firm level.

Additionally, with data from the second and third waves of the ELE survey, it is possible to assess whether CORFO's public funding programs have any impact on the composition of the funding that firms choose to make their investments. Table 4 shows that the INNOVA program does not have a significant effect on the percentage of funding with own funds, factoring or leasing, bank loans, non-bank loans and capital increase, which shows that there is no crowding-out effect from this program on how to finance new investments. However, the FOGAPE program, which is a State guarantee for loans from banks, generates two opposing effects, reduces the need for factoring or leasing significantly, and increases the use of non-bank loans. In addition, the other programs included in the analysis do not have a significant effect on the funding structure of firms.

Conclusions

Previous studies have found that participating in CORFO funding programs generates positive impacts on sales, profits, labor productivity, total factor productivity, and number of hired workers. However, the present study complements these results by investigating the effect of funding programs on innovation in products and/or services, as one of the evaluated programs known as INNOVA that aims to promote innovation.

To carry out an ex-post evaluation of the most commonly used CORFO-funded programs, information was obtained from the ELE survey, which includes three follow-up waves for the same firms. However, it is only in the last two waves that variables are available that allow controlling for the innovative characteristics of firms such as the recruitment of R&D personnel and the holding of an R&D department.

When using the data from the three waves of the LES survey, there is a positive and significant effect of the INNOVA program on the probability of innovating in products and/or services. However, the rest of the programs do not have an impact on the innovative activity of firms. Nevertheless, when the probability of innovating in products and/or services is conditioned by the possession of an R&D department, the statistically significant effect of participating in the INNOVA program disappears. Therefore, it can be concluded that the INNOVA program is not fulfilling its objective of promoting innovation. In addition, since the holding of an R&D department is an observable characteristic before allocating the funds from the INNOVA program, it is possible to choose as beneficiaries firms that have generated innovation in products and/or services, even without having obtained this public funding, and therefore, the apparent positive effect of the program when it is not controlled by the innovative characteristics would be explained only by the bias in the selection of the beneficiaries.

Additionally, it is determined that the INNOVA program does not generate a substitution effect over other funding sources in the beneficiary firms, but the FOGAPE program helps reduce the need for funding through factoring or leasing, and also increases the use of non-bank loans.

As limitation of this study, it can be mentioned that the available panel data only have three time periods, and in the model specifications that control for innovative characteristics, only two time periods are used. Therefore, it is not possible to carry out placebo tests to simulate effects in periods prior to treatment. However, different model specifications are included to check the robustness of the impacts obtained.