Explaining differences in the returns to R&D in Argentina: the role of contextual factors

ABSTRACT

Argentinean firms’ investments in R&D are well below its regional peers. One potential explanation for this fact is the existence of low and heterogeneous returns for these investments. This paper uses novel microdata to estimate the returns to R&D and analyse the role of contextual factors in shaping its heterogeneity. The findings confirm that returns are indeed heterogeneous and depend on some important factors related to the market context, such as measures of uncertainty; and the knowledge context, such as knowledge spillovers. Acknowledging that heterogeneity of returns depends on firms’ context is crucial for designing innovation policies to boost private R&D returns.

KEYWORDS:

• Productivity
• R&D
• Argentina
• knowledge context
• market context

JEL CODES:

• L25
• O25
• O33
• O54

Acknowledgments

We thank the ENDEI team working at the National Directory of Scientific Information for their willingness to facilitate access to information and address our questions regarding the survey. We also thank Chad Syverson, Bernardo Diaz de Astarloa and Cecile Thioro Niang for comments on earlier versions of the paper and Wendy Brau for excellent research assistance.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 Benavente (2006) applied an adapted version of the CDM framework using Chilean cross-section data and found that neither R&D nor innovative results (share of innovative sales) have an effect on productivity (measured as value added per worker). Crespi and Zuniga (2012) applied the CDM framework on micro data for six Latin American countries. They find that greater investment in R&D leads to a higher probability of having at least one process or product innovation. Additionally, results show a positive impact of technological innovation on productivity (log of sales per employee) for all countries except Costa Rica. Moreover, they find that the magnitude of the results is very heterogeneous. Crespi, Tacsir, and Vargas (2016) used 2010 firm level data to analyse 17 Latin American countries. Results show that investment in R&D per worker increases the probability that the firm will innovate, and that this translated into a strong increase in labour productivity (measured as log of sales per employee). These results are robust to five different measures of innovation: innovation of product or process, product innovation, process innovation, innovative sales (share of sales of new products) and filing for intellectual property rights. In Argentina the CDM framework was first used by Chudnovsky, López, and Pupato (2004) and then by Arza and López (2010). Both papers showed that investment in R&D boosts firms’ labour productivity.

2 Maloney (2017) provided similar arguments regarding the need for innovation policy to complement the innovation system approach with other contributions from neoclassical economics, to capture often-underplayed barriers that prevent the accumulation and allocation of production factors other than knowledge and that could explain cross-country differences in innovation.

3 Pavitt (1984) classified industries according to the source of technology (e.g. inside/outside the firm, government/private-financed), users’ needs (e.g. price/performance) and methods used to appropriate benefits from innovation (e.g. secrecy, patents, time lags, unique knowledge, etc.). Similarly, Malerba and Orsenigo (1995) explored the dynamics of technological change and defined two different groups, labelled technological regimes, which were characterized by a specific combination of conditions of technological opportunity, appropriability of innovation, cumulativeness, and properties of the knowledge base defined at the sectoral level.

4 Another important related literature summarized in Griliches (1994) and more recently in Bloom et al. (2017) focus on the productivity of new ideas. This literature is more macro but has sector specific implications, since in some sectors ideas ‘are harder to get’ and this will affect their returns to R&D investments.

5 These elements gave birth to the new economic geography, and the study of economics of agglomeration, led by Krugman (1991).

6 Many empirical studies have attempted to find economic mechanisms to explain what was considered a statistical regularity of a highly concentrated market structure in a highly R&D-intensive sector. Since Scherer (1967) the relation between market competition was modelled in an inverted-U shape. However, since these variables are related in both directions the methodological challenge in empirical studies has been how to account for such endogeneity (e.g. Aghion et al. 2005; Davies and Lyons 1996; Sutton 1998).

7 The papers by Caballero and Pindyck (1996) and Pindyck and Solimano (1993) show that the threshold of the marginal return on capital that triggers investment increases with the volatility of the marginal return, and therefore investment decreases with volatility. Caballero and Pindyck analyse US manufacturing industries, while Pindyck and Solimano’s contribution is a cross-country study (indeed they found that the impact is larger for developing countries).

8 Data source is the same as in Figure 3. Firms include public and private enterprises. In Brazil data is only available until 2017.

9 A detailed explanation on how this variable is measured can be found in Table A.1 of Appendix A.

10 Small firms are those with 10–25 employees; medium firms, those with 26–99 employees and large, those with 100 or more employees. Sectors included can be seen in Table 3. For some sectors of special interest, information was disaggregated at 4 digits. The five regions were: Patagonia (including provinces of Chubut, Neuquén, Rio Negro, Santa Cruz and Tierra del Fuego); Cuyo (including provinces of Mendoza, San Juan and San Luis); the Northern region (including Chaco, Corrientes, Formosa, Misiones, Catamarca Jujuy, La Rioja, Salta, Santiago del Estero and Tucuman); Pampeana (including Buenos Aires, Cordoba, Entre Rios, Santa Fe and La Pampa) and the region of the Capital city and suburbs. Information was disaggregated at the province level.

11 Sampling methods changed between both waves. In this respect, only ENDEI 2 is relevant for our exercise of assessing the role of contextual factors since ENDEI 1 was not representative at regional level. Since waves cannot be matched at the firm level, we only use ENDEI 1 to build pseudo-panels at the size-sector-regional levels to construct some specific variables. This is the case of the firms’ capital stock (see footnote 12) and the instrumental variable we use for some exercises (see Appendix B and in particular footnote 35 within this).

12 Results are robust when we use sales as a dependent variable and include expenditure in intermediate goods in the regression.

13 We did not have information on capital stock. Thus, we proxied it using information on firms’ expenditure on energy, gas and fuel and investment in machinery. The former proxies the initial stock of capital since it accounts for energy costs on existing machinery (Frank 1959) while the latter accounts for gross fixed capital formation during the period and is also measured in units of energy expenses, which makes the addition possible. We used the following equation: (3) $\begin{array}{rl}{C}_{it}& ={\displaystyle \frac{TotalEnergyCost{s}_{sector-size}}{TotalWageCost{s}_{sector-size}}\ast WageCost{s}_{it}}\\ & +{\displaystyle \frac{\mathrm{\Delta }EnergyCost{s}_{2010-12Sector}}{Investmentinmachiner{y}_{2010-11Sector}}\ast {\begin{array}{c}Investmentin\\ machineryforInnov\end{array}}_{it}}\end{array}$(3) The first term of the sum aims at measuring initial capital stocks at the beginning of each year, while the second is an estimate of each firm’s capital formation during that year. The first factor in each term of equation (3) was calculated using information from ENDEI 1 due to lack of such data in the second wave of the survey.

14 Griliches (1967) estimated that firm’s R&D and firm’s productivity were connected in a bell-shaped lag structure and since then several strategies have been followed normally using R&D and its lags.

15 Bond and Guceri (2017) estimate the impact of R&D on productivity on a sample of UK establishments using both the estimated stock and the flow of investments with similar results.

16 In ENDEI, the questionnaire reads: ‘Industrial Design and Engineering Activities: they are those activities carried out within the firm: technical functions for production and distribution not included in R&D, drawings and graphics for establishing procedures, technical specifications and operational characteristics; installation of machinery; industrial engineer; and production start-up. These activities can be difficult to differentiate from R&D activities; for this it can be useful to check if it is a new knowledge or a technical solution. If the activity is framed in the resolution of a technical problem, it will be considered within the Engineering and Industrial Design activities. It should include the annual salary of the staff devoted to these activities according to the time dedicated’.

17 In addition, we also summed 1 to all reported observations in the RD&D variable to avoid missing information when applying natural logarithms to zero values.

18 More details on this selection procedure could be found in Arza et al. (2020).

19 We are aware of the limitations of using the Herfindahl index to measure market competition, but due to data limitation we could not build alternative measures such as Lerner or Boone indexes (Boone 2008).

20 We included the quadratic term of the Herfindahl index in the regression but we later dropped it to save degrees of freedom since it was not significant.

21 These sectors fall in the ‘other’ category because of the survey confidentiality requirements, given that there are a few large firms in each of these sectors which would otherwise be easily identified.

22 As firms construct their knowledge stock by investing in innovation activities, these decisions are very likely endogenous to firms’ productivity, and exogeneity of regressors cannot be assumed. Unobservable omitted variables such as workers’ know-how or managerial capabilities could affect both firms’ decisions regarding innovation, capital and labour, and firms’ productivity. Additionally, while larger knowledge stocks may increase firm productivity, more productive firms are more likely to be exposed and aware of innovation opportunities. Hence, these firms may be more prone to investing in innovation and increasing their knowledge stock than less productive firms, causing reverse causality issues.

23 The year of fieldwork for the first wave, which collected data for 2010–2012.

24 Given that we have only one valid instrumental variable, we decided not to instrument the interactions with all contextual factors. We would need to instrument eight variables (investment in RD&D and seven interactions) and we think this would lead to unreliable instruments and results.

25 In all estimations of Table 4, except for column (3), variables are expressed in per worker units. As our model is linear in logarithms, we can assume a Cobb–Douglas specification: $Y=A{L}^{\theta }{C}^{\gamma }{I}^{\delta }$ and dividing by L: ${\displaystyle \frac{Y}{L}=\left({\displaystyle \frac{A}{L}}\right)L^{\theta }{L}^{\gamma }{L}^{\delta }{\left({\displaystyle \frac{C}{L}}\right)}^{\gamma }{\left({\displaystyle \frac{I}{L}}\right)}^{\delta }=A{L}^{\text{β}}{\left({\displaystyle \frac{C}{L}}\right)}^{\gamma }{\left({\displaystyle \frac{I}{L}}\right)}^{\delta }}$ with $\text{β}=\theta +\gamma +\delta -1$. Hence, a negative coefficient for L implies that $\gamma +\delta +\theta <1$, i.e. decreasing returns to scale.

26 When knowledge stock is approximated by R&D we also find the significant and positive effect on productivity. However, when comparing results for R&D with and without FE (columns 4 and 5), the bias is negative, which means that if caused by omitted variables, they are correlated in opposite directions with R&D and with productivity. For example, there may be omitted innovative efforts which work as substitutes of R&D (but affect productivity positively), such as, possibly, design and engineering, since as we commented above it is very difficult for respondents to empirically discriminate them from R&D.

27 Firms with zero investment in RD&D are not included.

28 We estimated the short-run returns, but investment in RD&D may have higher returns in the longer term depending on the project nature.

29 F-test statistic for the null of equality of coefficients across sectors is rejected at 1% level of significance.

30 Sectoral and regional returns in monetary values were calculated as the mean value of individual firm returns within each sector and, in turn, region.

31 Patagonia region is a special case in terms of labour productivity since it is specialised on capital and natural-resources intensive industries and receives strong fiscal and economic support from the state for certain economic activities.

32 F-test for the null of equality of coefficients is not rejected at the usual significance levels in this case.

33 Low publication levels are sectors in the first quartile of the distribution of publications, while high publication levels are those in the fourth quartile.

34 We thank an anonymous reviewer for bringing this possible alternative explanation.

35 Calculated from Table 5 using the mean values of explanatory variables. We thank an anonymous referee for suggesting this calculation.

36 In the three cases depicted, the Kolgomorov–Smirnov test rejects the null hypothesis of equality of distributions, indicating that the distributions of returns are statistically different between groups.

37 Each of them represented around 13% of FONTAR budget in 2017. Data from FONARSEC and FONTAR available at ANPIDTYI-MINCYT (2020); data from COFECYT comes from the government webpage [https://www.argentina.gob.ar/ciencia/cofecyt/convocatorias/pfip-2017, accessed on November 2021].

38 This is possible given that samples were constructed to be representative of the Argentinean manufacturing sectors (by industry size). However, as ENDEI 1 was not constructed to be representative at the regional level, we could not divide the sample into the same regions as ENDEI 2. We therefore made an ad-hoc split of the sample considering that cases were relatively balanced: ‘Region’ is taken as a dichotomous variable signaling if the firm belongs to the Gran Buenos Aires region -comprising Ciudad de Buenos Aires and the main adjacent districts which belong to Buenos Aires Province- or to the rest of the country.

39 We assume that restrictions present during ENDEI 1 period (2010–2012) can represent restrictions in 2014, as the political administration was the same during both periods, as well as trade regulations.

40 The test rejects the null hypothesis that the Nagar bias of the second stage coefficient of RD&D exceeds a 20% of the ‘worst case bias’, i.e. the case in which instruments are completely uninformative and first and second stage errors are perfectly correlated.

Additional information

Funding

This work was supported by the World Bank [P168072] - Understanding and Promoting Firm-level Productivity and Growth in Argentina. The sponsor provided the funding, feedback and organised several workshops jointly with the Ministry of Production in Argentina to disseminate the results. The views of this paper are solely those of the authors.