1 Introduction

International flows of technological knowledge have an important effect on the ability of developing countries to learn and to innovate. Knowledge transmission from developed countries creates conditions for developing countries to catch up with the technological frontier, but, on the other hand, technological isolation slows down the development process and is conducive to technological and economic divergence. This paper studies the importance of patents and interpersonal links for technology diffusion across countries and asks to what extent international technology spillovers are mainly driven not only by the free flow of knowledge but also by interpersonal links and face-to-face contacts across countries.

This has important policy implications. If international interpersonal links and person-to-person contacts play a prominent role in fostering innovative domestic capacity, R&D subsidies could be effective only as long as they favor the international expansion of the network relations of local inventors. This has relevant consequences for the effectiveness of science and technology policies.

This paper is one of the first attempts to extend the economic analysis of R&D knowledge spillovers (at country and industry level) to developing countries and investigates empirically the determinants of international patent production in a selected number of Latin American countries (LACs). We ask whether foreign R&D activity affects the innovative performance of LACs at industry level via different channels of international knowledge flows. In particular, we focus on three mechanisms: foreign R&D, patent citation-related spillovers, and face-to-face contact spillovers based on co-inventorship relations. Of course, there are also other important channels of technological transmission that we do not deal with in this study, such as FDIs and bilateral trade. However, these channels affect, in particular, total factor productivity.Footnote 1

We are interested in studying whether the international patenting activity of LACs responds to international knowledge flows and we measure knowledge flows using patent citations and analyze the network of co-inventors from the patent documents. Assuming that inventors listed on the same patent know each other, if knowledge has at least a degree of tacitness, we expect a positive effect on the innovative activity of personal contacts. This in turn implies that the international mobility of inventors may play a crucial role in domestic innovative performance.

We use data for five big industrial sectors (Textiles and Food, Chemicals and Pharmaceuticals, Metals, Instruments Electronic and Non-Electrical Machinery, and Transportation), five Latin American countries (Argentina, Brazil, Chile, Colombia and Mexico) and the G-5 countries (France Germany, Japan, the UK and the USA) in the years between 1988 and 2003. We process the information contained in the US Patent and Trademark Office (USPTO) patent documents and their citations to build up the different indexes of R&D spillovers. Also, we match USPTO patent data with economic data at a sectoral level (domestic value added) and control for past innovative activity. In order to have a more complete picture of the patenting activities of the Latin American countries, we also provide some descriptive evidence on European Patent Office applications.

Overall, this paper provides a detailed account of the nature, sources and determinants of international patenting activity in Latin American countries. We show that a large part of the Latin American-invented patents belong to foreign companies with a foreign address or to a foreign subsidiary with a Latin American address, and top applicants at the USPTO and EPO are mainly US and German multinationals. The big Latin American patenters are active in a set of heterogeneous sectors that are not considered very R&D-intensive (e.g. Oil, Glass, Electric, Metals and Machinery). Second, econometric analysis shows that international knowledge spillovers from the G-5 countries are a significant determinant of inventive activity in the period considered. In particular, we find that, controlling for direct foreign R&D effects, both bilateral patent citations and face-to-face relationships between inventors are important additional mechanisms of knowledge transmission. Some of our results suggest that the latter is more important than the former.

In Section 2, we provide a short overview of the theoretical background of this study and justify the use of patent-based data to measure knowledge spillovers. In Section 3, we perform a descriptive analysis of the international patent activity in Latin American countries and network of knowledge relations across countries using patent citations and co-inventorship behavior. To have a clearer picture, we use data from different sources (i.e. the US and European Patent Offices). In Section 4, we construct our empirical model, and in Section 5, we describe the data we will use and our empirical strategy. More details are provided in the Appendix. Section 6 reports the main results from the estimation of different econometric specifications. In the final section, we conclude, discuss some important limitations and propose some directions for future work.

2 Background

This paper extends current studies on the economic impact of knowledge spillovers to developing countries and, in particular, to Latin American countries. We assess directly the determinants of innovative activity using a knowledge production function (KPF) (Pakes and Griliches 1984). The KPF is a methodological tool that tries to map research efforts into new knowledge. In the KPF baseline version, patent counts are used to approximate the production of new knowledge and R&D expenditure measures the R&D effort. However, in dealing with developing countries, external sources of knowledge—which originates spillover or is transferred to developing countries—are particularly important. Actually, much of the current debate about technology policy in developing countries is based on the assumption that a country’s innovative performance depends significantly on its relative technological capacities, its ability to absorb foreign (costly and specialized) knowledge, and its ability to learn how to adapt it to local needs (Cimoli and Dosi 1995; Cimoli et al. 2006).

R&D efforts either aim at lowering the costs of production (process-oriented R&D) or at producing new products or higher quality varieties of existing products (product-oriented R&D). Process-oriented R&D is often protected by secrecy (Mansfield 1986; Levin et al. 1987) and therefore it can be considered a minor source of spillovers, whereas product-oriented R&D generates spillovers through various channels such as the trade of goods which incorporate the innovation, and the generation of patent documents which allows other firms to collect relevant information.

Within the broad spectrum of product-oriented R&D, when new or improved goods are developed and traded, an increased price–quality ratio leads to a so-called rent spillover (Griliches 1979; van Meijl 1997); at the same time, when knowledge is mainly codified in publicly available sources such as scientific and technical literature (or also industrial espionage or reverse engineering), knowledge spills over between firms and countries, the so-called knowledge spillovers. Existing knowledge that is not perfectly protected may evoke new ideas which in turn lead to innovations (idea-creating spillovers) or may be simply absorbed and used to imitate (imitation-enhancing spillovers) (Los and Verspagen 2003). In the latter case, knowledge spillovers result in higher productivity, while in the former, they have a direct effect on innovative activity. Patent documents, as well as the mobility of R&D employees, are potential sources of idea-creating knowledge spillovers. Clearly, we focus on the role of foreign spillovers on innovative activity capturing the effect of foreign idea-creating knowledge spillovers.

It is worthwhile remarking that, in this paper, rent spillovers are not considered and patent citations and face-to-face interactions between inventors capture only a specific form of knowledge spillovers. For example, patent citations and inventors’ collaborations take place only if both source countries and receiving countries are actively engaged in R&D and apply for an international patent.

There is a vast literature that assesses international knowledge spillovers among developed countries.Footnote 2 Estimated international R&D spillover effects are typically significant and positive.Footnote 3 Recent empirical works show that extremely relevant sectoral knowledge flows cross national borders (Malerba et al. 2007). Bottazzi and Peri (2007) find that internationally generated ideas significantly affect innovation in a country. Branstetter (2006) uses a patent function to estimate firm-level spillovers. Based on a panel of 205 firms in five high R&D-sales ratio industries in the period 1985–1989, he provides strong evidence for Japanese intra-national knowledge spillovers and limited evidence that Japanese firms benefit from knowledge produced by American firms.

In the case of developing countries, there is a large literature on the microeconomic effects of FDI spillovers on total factor productivityFootnote 4 but there is still scant aggregate evidence of R&D spillovers on countries’ innovative outputs at sectoral and national levels. This paper focuses on two specific vehicles of knowledge spillovers: patent citations and collaboration via co-inventorship.

2.1 Patent citations as channels of knowledge flows

Patent citations are included in a patent document to delimit the scope of the property right and mention the relevant prior art. Citations are particularly reliable because they have a legal value. If patent A cites patent B, it can be reasonably assumed that B is a technological antecedent of A and that the knowledge embedded in B has been developed by A. Trajtenberg (1990) and Albert et al. (1991) are among the first scholars who empirically demonstrated that highly cited patents have higher economic and technological importance. If a patent is cited, it can also generate technological spillovers. Jaffe et al. (2000) tested this conjecture using USPTO patents and surveyed approximately 380 citing and cited inventors. Their results suggest that ‘communication between inventors is reasonably important, and that patent citations do provide an indication of communication, albeit one that also carries a fair amount of noise’ (p. 215). In addition, a consolidated stream of literature uses patent citations to track knowledge flows and spillovers (Jaffe et al. 1993; Jaffe and Trajtenberg 1996, 1999; Maurseth and Verspagen 2002; Malerba and Montobbio 2003; Peri 2005; Malerba et al. 2007).

Given that knowledge flows are inherently difficult to measure and that it is often problematic to assess the relevance of the source of knowledge and to evaluate the direction and the impact of the generated knowledge, patent citations have often been used to identify the direction of these knowledge spillovers between countries. If, for example, a patent with an inventor’s address from Argentina cites a patent with an inventor’s address in the USA, we could assume that some knowledge created in the USA has been used in Argentina and, as a result, patent citations could track the direction of knowledge spillovers between the two inventors and the two countries.

2.2 Patent co-inventors as channels of knowledge flows

The second major channel of knowledge transfer we consider in this paper passes through collaborations and face-to-face contacts. Processes of knowledge creation are importantly affected by the inventors’ community and network relationships (Breschi and Lissoni 2001). Similarly, research collaborations create fundamental social networks, in particular for developing countries: inventors who have studied or worked abroad not only benefit from the high standard of top international universities and companies, but also continue to rely on free information in subsequent research projects after the collaboration itself is finished. Therefore, research collaborations can indicate relational proximity and capture the spillover stemming from collaborative networks between regions and countries (Hoekman et al. 2008).

Singh (2005) has examined whether social networks of inventors are a significant mechanism for diffusion of knowledge and found that the existence of co-inventorship relations is associated with a higher probability of knowledge flows (measured in terms of citations): the probability of knowledge flows between inventions is a decreasing function of the social distance. Gonzalez-Brambilla et al. (2008) emphasized the relationship between social capital and knowledge creation, underlying the role of exchange and combination processes. In particular, using a database of international scientific publications and citations, they found that scientists in embedded networks have superior success because of better communication skills.

Citation patterns and co-inventor relations measure different kinds of disembodied knowledge flows. On the one hand, citations are able to measure flows of codified knowledge, that is, knowledge acquired by direct reading and comprehension of written and available documents such as publications and patents. On the other hand, if we assume that inventors listed on the same patent know each other, co-inventor relationships can be seen as a diffusion mechanism of non-codified knowledge (e.g. technical know-how, non-standardized production procedures, etc.). In fact, diffusion of non-codified knowledge requires face-to-face interactions, at least periodically, and is likely to have a great impact on the inventive activity.

Here, we apply this theoretical background to analyze international patenting in Latin America and the impact of international knowledge spillovers. We are aware that international patenting is a tiny portion of the innovative activity of these countries and, exactly for this reason, it is important to stress the peculiarities and specificities of international patenting before laying down the details of the empirical exercise. The next section is, therefore, dedicated to the precise understanding of the object of enquiry of this paper (see Montobbio 2007 for a broader discussion and comparison with other developing countries).

3 International patenting in Latin America

For this analysis, we use standard patent data sources from the European and US Patent Offices. Data sources and sectors of analysis are carefully explained in the Appendix. Table 1 shows the total number of Latin American granted patents at the USPTO by year (the country is assigned using the residence of the inventors). These numbers are small relative to the overall numbers in other countries. Top Latin American countries at the USPTO are Brazil and Mexico with, respectively, 2,155 and 2,102 patentsFootnote 5 in the period 1968 to 2003. Argentina and Venezuela follow with 1,037 and 704 patents, respectively. At the EPO, for the period 1978–2003, Brazil has the highest share with 1,688 patent applications, Mexico, Argentina and Venezuela follow with 678, 575 and 176 patent applications, respectively (see Table 2). It is important to note that, at the USPTO, Brazil and Mexico have almost the same number of patents, whereas at the EPO, Brazil has a total number of patents which is almost three times that for Mexico. This indicates that geographical proximity and economic agreements play a very important role and Mexican inventors tend to protect their innovations much more in the US market compared to the EU.Footnote 6 In recent years, no remarkable structural break is observable after the changes in domestic legislations due to the implementation of the TRIPs agreement in many countries.

Table 1 Patents at the USPTO by inventor’s country
Full size table
Table 2 Patents at the EPO by inventor’s country
Full size table

The rise in patent numbers documented in Tables 1 and 2 can be seen as the combined result of an increased propensity to patent world-wide and the increased use of international patents in LACs. Many authors have documented the explosion of patent activity world-wide and in the USA, in particular in semiconductors. This is related to a general reinforcement of IPR legislation (mainly, but not only, in the USA) and to institutional changes in the early 1980sFootnote 7 that favored changes in firm appropriability and IP strategies (e.g. Hall 2003). However, it can be noted that, on average, the growth of LAC patents is higher than the average growth of patents. This happens at both the USPTO and EPO during the 1990s (see Table 11 in the Appendix and Montobbio 2007).

It is important to underline the fact that an increasing share of the total Latin American-invented patents filed in the USA is the result of collaborative activity with foreign (in particular US, see Section 3.4) laboratories, companies and inventors (Fig. 1). It is worthwhile noting that these patents are mainly owned by US companies (such as Syntex USA, Delphi Technologies, Procter &Gamble, IBM, Hewlett-Packard and General Electric). Moreover, there is a non-negligible number of patents owned by US universities and research laboratories (e.g. Universities of Pennsylvania, California and Texas).

Fig. 1
figure 1

Share of international co-invented patents in the total Latin American patents by inventors (USPTO)

Full size image

3.1 Latin American-owned vs. Latin American-invented patents

The patent count, based on the inventor’s address, reflects more directly the inventive activity of laboratories and researchers in a given country. If a country’s patents are counted using the applicant’s address, results reflect ‘ownership’. Of course, this counts the inventive activity of a given country’s firms, even if their research facilities are located elsewhere. Typically, countries such as the United States or the Netherlands, where many multinational companies are located, have a relatively higher patent share when country is assigned on the basis of the applicant’s address (Dernis et al. 2001). The opposite occurs in most developing countries.

USPTO data do not report the applicant’s country, but it is possible to use EPO data on patent applications to understand the difference it makes to count patents using the applicant’s address.Footnote 8 As expected, counting patents with the applicant’s address reduces the number of patents in the main countries by approximately 41% (from 2,636 to 1,565, in the period 1977–2001, EPO data) with respect to patents with the inventor’s address. It is worthwhile noting that, out of 2,636 Latin American-invented patents, there are only 1,520 (56%) Latin American-owned patentsFootnote 9 (i.e. patents in which the applicant’s address is in a Latin American country). The rest are owned by foreign companies (1,213—44%)Footnote 10 (i.e. the company’s address is not in a Latin American country). Finally, it is important to note that, if we consider Latin American-‘owned’ patents, the share of patents with at least one foreign inventor is significantly lower (9%) than in the case of Latin American-‘invented’ patents. This indicates a low degree of internationalization of patentees resident in LACs.

Colombia, Mexico and Venezuela have the highest percentage difference between Latin American-owned and Latin American-invented patents. This means that, for these countries, a considerable part of the national inventors’ activity is performed in companies or institutions that do not have a legal address in the country. This asymmetry may partly reflect the internationalization of research and location of research and legal facilities by multinational firms and partly the fact that some Latin American inventors may be temporarily (or in some cases even permanently) active abroad but declare their address in Latin America.

3.2 Sectoral differences

Patents are classified according to very specific technological classes and, therefore, can be used to measure innovative activities in specific sectors of economic activity.Footnote 11 Table 12 in the Appendix shows the number and distribution of patents granted at the USPTO at a sectoral level. We observe that Chemicals and Pharmaceuticals, and Instruments, Electronics and non-Electrical Machinery are the two sectors that capture 80% of the total patents in Latin America, while, not surprisingly in traditional sectors such as Textiles and Food, the number of patents represents only 4% of the total. Table 12 in the Appendix also shows the number and distribution of patents by country: Chile seems to have a comparatively good production of patents in Metals, while Brazil displays a considerably high share of patents in Transportation.Footnote 12

3.3 Individual inventors

A more detailed look at these patents shows that many patent assignees are individual inventors. If we assign a patent to a country using the applicant’s address, 41.5% of Latin American patents at the EPO are owned by individual inventors. At the USPTO, 37.3% of the ‘Latin American-invented’ patents granted are ‘individually owned’.Footnote 13 These shares are considerably higher than average, considering that, for all patents at the USPTO and EPO, the shares of individually-owned patents are, respectively, 23% and 11%.Footnote 14 Typically, less developed countries and regions have a relatively higher share of individual inventors because firms, universities and research centers are less aware of the patent system and have relatively fewer resources to invest (relative to firms in advanced countries). Therefore, it is more likely that individuals decide to bear the expenses and file their own patents. Typically, these patents are considered less economically and technologically valuable because they are often the result of occasional activities and do not originate from well-funded R&D projects.

Some of these patents may actually belong to companies but are registered in the name of the owner as the applicant. This could be the case with micro companies, family companies or partly-informal companies. Given the great uncertainty of survival of small and medium companies—in a macro-economic context that is often unstable—companies prefer not to have the patent registered under the name of the company but rather under the name of the owner (for Argentina, see López et al. 2005). There may, however, be some exceptions to this negative interpretation. Some inventors, who are active abroad, keep the address of their home country. This inventive activity could be valuable, and these individual patents could signal cooperation with foreign countries and be an important vehicle of knowledge transferFootnote 15 as emphasized in previous sections.

3.4 Applicants

The concentration of assignees or applicants of international patents at the USPTO and EPO in Latin America is not very high. Many assignees or applicants are, in large number, different individual inventorsFootnote 16 and among the top applicants we find many US and German multinational companies. There are some big Latin American patentors, such as Petrobras, Embraco and INTEVEP-PDVSA, which are active in a set of heterogeneous sectors of activity that are not considered very R&D-intensive (e.g. Oil, Glass, Electric, Metals and Machinery). Almost no Latin American companies are active in the high tech and high growth sectors such as Electronics, Telecommunications or Pharmaceuticals. Table 3 shows the top 16 applicants at the USPTO and their patent numbers.

Table 3 Top 16 applicants at the USPTO (1978–2001) and relative numbers of patents
Full size table

The top ten patenting companies at the USPTO are (for the period 1978–2001, excluding ‘individually-owned patents’; in parenthesis there is the country of the inventors, not the address of the company which is not available in the USPTO database) INTEVEP (Venezuela), Petroleo Brasileiro s.a.—Petrobras (Brazil), Empresa Brasileira De Compressores (Brazil), Hylsa (Mexico), Carrier (Brazil), Syntex USA (Mexico), Vitro Tec Fideicomiso (Mexico), Hewlett-Packard (Mexico), Bayer (Brazil, Mexico and a few from Colombia and Argentina), Delphi Technologies (Mexico). The picture at the USPTO is quite similar to the EPO with a lower presence of German firms and a higher presence of US companies such as HP, IBM, Carrier or Colgate–Palmolive.

3.5 Citations

In order to address the issue of knowledge flows, in this section we track citation flows between Latin American countries and other geographical areas. Using USPTO citation data from the period 1975–2000, we build a matrix of citation flows across areas (CIT). Each element of this matrix {CIT kj} represents the number of patent citations flowing from country j into country k (i.e. the number of times patents with the inventors’ address in country j cite patents with an inventor’s address in country k). Note that CIT is squared and asymmetric and the elements on the main diagonal {CIT jj} are the number of citations that remain in the same specific country. Table 4 illustrates the matrix from the USPTO dataset. Each column represents the citing country and the rows are cited countriesFootnote 17 (e.g. Latin American patents cite Chinese patents ten times, equivalent to 3% of the total Latin American backward citations).

Table 4 Share of backward citations of different regions by destination country (USPTO data)
Full size table

Table 4 shows a very low share of citations among Latin American countries (4.29% of citations). This is similar to other countries such as China and India. Approximately 70% of the citations made and received are from US patents.Footnote 18 Finally, it can also be noted that knowledge flows from Latin American patents to patents invented in other regions are also extremely low. Our evidence shows that citations to Latin America from EU and US patents appear to be equal to 0.14% of the total outflow of their citations.

3.6 Co-inventors

Our second measure of knowledge flows is based on collaboration patterns between inventors. Table 5 shows the number of co-inventors and share by countries and sectors at the USPTO for five LACs (Argentina, Brazil, Chile, Colombia, and Mexico). In columns (1) and (2), we show the total number of inventors of USPTO patents that declare their residence, respectively, in a Latin American country and in a foreign country. In the other columns, the share of co-inventors resident in a foreign country is displayed. We consider only the co-inventors resident in the G-5 countries (USA, Japan, Germany, UK, and France).

Table 5 Number of co-inventors and share by countries and sectors
Full size table

Mexico has more international collaborations than the other LACs in terms of patenting activities: the G-5 co-inventors represent 31% of the total inventors of Mexican patents. At the opposite end, we find Argentina, where the G-5 co-inventors represent only 22% of the total number of inventors. Looking at the bilateral relationship, it is worth noting that the great majority of foreign inventors come from the USA: in all the LACs considered, the lowest share is for Brazilian patents with 56%. However, it is possible to distinguish different patterns of co-inventorship. Brazil has a higher co-inventors’ network with Germany (31%) and France (6%) with respect to other LACs, while Chile seems to have a significant collaboration with the UK (especially in Chemicals and Pharmaceuticals). Finally, if we consider sectoral differences, we find that more or less in all the countries, Chemicals and Pharmaceuticals and Instruments, Electronic and non-Electronic Machinery are the sectors with more international co-inventors.

4 The empirical model

This section outlines the empirical model we use to estimate international knowledge spillovers and, in particular, the effects of R&D performed in industrialized countries on the innovative activity of Latin American countries. Following Malerba et al. (2007), we start from the following KPF that relates R&D investments and the production of technological output:

$$ Q_{h,i,t} =f\left( {\mathop R\limits^-}_{h,i,t} ,\alpha ,\emph{v}_{h,i} \right)=\overline R_{h,i,t}^\alpha \emph{v}_{h,i} $$
(1)

where Qh,i,t is a latent measure of technological output in field i (i = 1, ...5), country h and period t. In addition, α represents the unknown technological parameter, and νh,i captures the country and technological field specific effects. We display a bar over the R&D variable because we assume that it is composed of domestic R&D efforts and international R&D efforts that produce usable knowledge at an international level. As emphasized in the previous section, we compare three different modes of knowledge flow. The first mode is pure spillover (IS1), the second is knowledge spillover through patent citations (IS2) and, finally, the third is knowledge spillover that is related to collaboration activities and face-to-face contacts (i.e. co-inventorship) (IS3):

$$ \overline{R}_{h,i,t}^\alpha =R_{h,i,t}^{\alpha 1} IS_{1h,i,t}^{\beta 1} IS_{2h,i,t}^{\beta 2} IS_{3h,i,t}^{\beta 3} $$
(2)

Moreover, we use patents as a noisy indicator of technological output:

$$ P_{h,i,t} =Q_{h,i,t} e^{\theta _t }u_{h,i} $$
(3)

We take into consideration possible common time effects in patenting (θ t) and differences in country-specific propensity to patent in each technological field (uh,i). Combining Eq. 3 with Eqs. 2 and 1 results in the following patent equation:

$$ P_{h,i,t} =R_{h,i,t}^{\alpha 1} IS_{1h,i,t}^{\beta 1} IS_{2h,i,t}^{\beta 2} IS_{3h,i,t}^{\beta 3} e^{\theta _t }\xi _{h,i} $$
(4)

We cannot directly estimate Eq. 4 because we do not have data on national R&D efforts at the sectoral level over time. However, even if we are interested in the effect of international spillovers on international patenting, we have to take into account some economic measures related with the trend in the size of the different industries in each country and national R&D investment in order to avoid omitted variable problems in the econometric approach. For this reason, we control national economic activity with data on value added (an additional specification includes the lagged dependent variable, see below), captured by the variable Xh,i,t:

$$ P_{h,i,t} =X_{h,i,t}^{\alpha 1} IS_{1h,i,t}^{\beta 1} IS_{2h,i,t}^{\beta 2} IS_{3h,i,t}^{\beta 3} e^{\theta _t }\xi _{h,i} $$
(5)

In general formulation, international knowledge spillovers are typically expressed as follows:

$$ IS_{h,i,t} =\prod\limits_f {R_{f,j,t}^{\lambda _{h,f,j,t} } } $$
(6)

where λ h,f,j,t weights the impact of R&D expenditures from foreign countries. R is the knowledge source and λ is the vehicle of knowledge spillovers. In our case, subscript f refers to the USA, the UK, Japan, France, and Germany, and h to Argentina, Brazil, Chile, Colombia, and Mexico. Our weights are sector-specific (sector j) and vary over time. Note that we adopt very large sectors and, therefore, we feel it legitimate to focus only on intra-sectoral R&D spillovers, neglecting inter-industry knowledge flows.

5 Data and methodology

Our econometric exercise uses different databases for the five Latin American countries (Argentina, Brazil, Chile, Colombia, and Mexico) and five industrial sectors (Textiles and Food, Chemicals and Pharmaceuticals, Metals, Instruments Electronic and Non-Electrical Machinery, and Transportation) in the period 1988–2003. We exclude Cuba and Uruguay from the econometric analysis and focus on the five countries with the highest number of patents. In particular, we use the USPTO-CESPRI database for patents and patent citations, the PADI-CEPAL database for value added and the OECD-ANBERD database for R&D data. We use USPTO data for the econometric exercise, as the US market is particularly relevant for Latin American countries, because there are more observations that can be used, and finally because, in USPTO data, there are many more patent citations.Footnote 19 Data sources and sectoral aggregations are thoroughly explained in the Appendix. Equation 5 captures the effect of the R&D effort performed in foreign countries in the production of USPTO patents by Latin American inventors. Taking logs of Eq. 5, we propose to estimate the following logarithmic specification:

$$ \ln P_{h,i,t} =\alpha _1 \ln X_{h,i,t} +\beta _1 \ln IS_1 +\beta _2 \ln IS_2 +\beta _3 \ln IS_3 +\theta _t +\varsigma _{h,i,t} $$
(7)

where the dependent variable is the log of the number of USPTO patents in county h (h = 1,...5), sector i (i = 1,...5), and time t (t = 1,...16 for the period 1988–2003). Note that our observational unit refers to industries (sectors) in different countries, for a total of 25 different groups.

The R&D stock in country f and sector i is calculated using the perpetual inventory method and, following the standard practice in the literature, we set the rate of depreciation δ at 0.12 (see Appendix).Footnote 20 Central to this paper is the calculation of international spillover variables. We measure three different channels of international knowledge spillovers. The first international spillover variable measures knowledge spillovers when knowledge is a public good and, once it is produced, it is freely available. Under this assumption, US$1 in R&D will have a direct impact on the knowledge production in other countries. We call this variable:

$$ \ln IS_1 =foreignR\& D\_tot_{h,j,t} =\sum\limits_f {\ln R\& D_{f,j,t} } $$
(8)

foreignR&D_tot is equal to the sum of the logarithm of R&D stocks in the main G-5 industrialized countries.Footnote 21 In this case, all weights λ h,f,j,t are set equal to 1. In addition, we have shown that the USPTO activity of Latin American countries is tightly linked to the activity of US companies and universities. Therefore, R&D expenditures in the USA are particularly important in terms of spillovers generated to Latin American countries. Then, in our regressions, we control for this aspect and also consider only the US R&D stock.

The second spillover effect is captured by patent citations. Patent citations are a paper trail that may signal that some knowledge flow occurs. Knowledge remains a public good but travels embedded in codified documents such as patents. We use USPTO citations to build a set of matrices that map citations between our five LAC countries and the G5 countries we considered. Each cell of the matrix is the number of citations in patents with at least one inventor resident in a LAC country to patents with at least one inventor resident in a specific G5 country. We build these matrices for each sector and for each year. Then we construct the weight λ h,f,j,t = cit h,f,j,t, which is the ratio of the number of citations flowing from country h to country f in sector j at time t over the total number of citations flowing from country h to all the G-5 countries in sector j at time t. As a result, our index of citation-based international knowledge spillovers (foreignR&D_cit) is calculated as follows:

$$ \ln IS_2 =foreignR\& D\_cit_{h,j,t} =\sum\limits_f {cit_{h,f,j,t} \ln R\&D_{f,j,t} } $$
(9)

The third spillover effect is related to interpersonal links and possibly face-to-face contacts. In this case, the fact that inventors work together on the same invention signals that some knowledge exchange takes place. We again use USPTO patent data to build up a second set of matrices. In this case, each cell (h,f) of the matrix is the number of patents with at least one inventor resident in country h and one inventor resident in country f. Again, we build up these matrices for each sector i and for each year t in the sample. Then we construct the weight \(\lambda _{\rm h,f,j,t} = coin\emph{v}_{\rm h,f,j,t}\), as the ratio of the number of patents with co-inventors in country h and country f in sector j at time t over the total number of patents with inventors in country h and all the G-5 industrialized countries in sector j at time t. As a result, our index of international knowledge spillover (foreignR&D_coinv), based on co-inventorship behaviour, is calculated as follows:

$$ \ln IS_3 =foreignR\& D\_coin\emph{v}_{h,j,t} =\sum\limits_f {coin\emph{v}_{h,f,j,t} \ln R\& D_{f,j,t} } $$
(10)

Table 6 displays summary statistics on the economic and patent data variables.

Table 6 Summary statistics for the regression variables
Full size table

6 Estimation results

Our estimation strategy follows three steps. First, we run simple fixed effect OLS regressions. We use fixed effects because they ensure consistency in the presence of correlation between the explanatory variables and the individual effects.Footnote 22 Therefore, we start with a set of static regressions using fixed effect models. Second, we control for possible spurious results due to common trends and test for the stationarity of the time series in the panel. Third, we use a lagged dependent variable to control for domestic innovative activity. In this last step, we estimate a dynamic panel using Within Group (Fixed Effect) estimation and GMM following Arellano and Bond (1991). Results are based on the assumption of stationarity consistent with the second step of this econometric exercise.

6.1 Static panel

We then start estimating Eq. 7 using Fixed Effects. Heteroscedasticity robust standard errors are applied. We take the log to have variables more closely distributed to normality and estimated coefficients expressed in terms of elasticity. In some cases, the number of patents is zero and the log of zero is not defined; in this case, we set zeroes equal to one and allow the corresponding observations to have a separate intercept (zero dummy) as in Pakes and Griliches (1984). In Section 6.2, we also perform a robustness check in this respect. In all specifications, we also include time dummies to control for common economic changes related to the calendar year.

Table 7 reports the robust Fixed Effect estimates of the parameters. All the specifications explain approximately 90% of the variation in international patenting. The first column includes only total foreign R&D stock (i.e. USA, Japan, Germany, UK, and France) as an input of the innovation function: an increase of 1% in total foreign R&D stock increases the innovative activity by 0.095% in terms of international patenting of our LACs. In Column 2, we assume that only R&D expenditures in the USA have a spillover effect on international patenting. Results show a strong positive spillover effect from the US R&D stock: the estimated coefficient is equal to 0.3 and statistically significant at the 1% level. Note that the size of this estimated coefficient is three times higher than in the case of total foreign R&D. This variable controls for foreign knowledge input effects as in Bottazzi and Peri (2007): US-generated ideas widen the basis of usable knowledge and generate further innovation based in LACs.

Table 7 Estimation results of Eq. 7
Full size table

Controlling for the effects of available ideas in a specific industry measured by US R&D stock, we proceed in columns (3), (4), and (5), adding as regressors the other ‘embedded’ international spillover mechanisms measured by the variables IS2 and IS3. These coefficients show that external R&D has a significant additional impact on patent production and, in particular, that citations and co-inventorship patterns are relevant channels of knowledge flows. The two estimated coefficients have similar sizes, being, respectively, 0.032 and 0.027, and are significant at the 1% level. Our results suggest that a significant portion of international knowledge spillover is embedded or in codified documents, such as patents that are publicly available, or in interpersonal links and contacts, such as cross-country collaborative efforts on specific innovations.

Finally, in column (6) we test the robustness of our results running a Fixed Effect Negative Binomial model in order to take into account that patents are a count variable and the results related to citation-based spillovers and co-inventorship-based spillovers do not change substantially. Conversely, the US R&D stock is smaller and not statistically significant. As we will see in the next paragraph, this variable is non-stationary and this may crucially affect the results.

6.2 First robustness check

We have 85 observations out of 400 in which the number of patents is zero: in this case, when the spillover effect passes through patent citations or patent co-inventors, the source of external R&D is zero by definition (it is not possible to have citations or co-inventors without patents). In order to check if the previous results are driven by this effect, we ran the fixed-effect model, dropping the observations where the number of patents is zero. Results do not change substantially. The coefficients associated with the spillover measured by citations and by co-inventors are significant and positive. In particular, a 1% increase in citation-weighted R&D generates a 0.029% increase in the domestic innovative output, while, for the co-inventors’-weighted R&D, we get a significant coefficient of 0.024%. The R&D performed in the USA has a greater impact, with an estimated elasticity of 0.24% (see Table 14 in the Appendix).

6.3 Stationarity tests

Our estimates rely on the assumption that our variables are stationary or co-integrated, and it is possible that serial correlation is spuriously driving the above results. We therefore perform the panel unit root test developed by Im et al. (2003). If it is assumed that the time series are independent across i, the null hypothesis is that all the series are non-stationary; if the contrary is assumed, some of the individual time series have unit roots. Table 8 shows the results. We find that the dependent variable and our measures of R&D spillovers weighted by citations and co-inventors are indeed stationary.Footnote 23 At the same time, the null hypothesis of unit root cannot be rejected for the other measures of foreign R&D we have used. Total foreign R&D stock and US R&D stock are therefore both non-stationary. For this reason, the estimations presented in Table 7 may be biased. In the following section, we check the robustness of our results excluding Total Foreign R&D and US R&D in order to obtain consistent estimates. In addition, we add a lagged dependent variable in order to estimate a dynamic version of our empirical model.

Table 8 Results for the IPS (2003) unit root test for panel data
Full size table

6.4 Dynamic panel

This section is devoted to control the robustness of our results. We control for an additional potential source of omitted variable bias, including a lagged dependent variable. This leads us to estimate a more general dynamic version of our empirical model. It is reasonable to think that international patenting is a cumulative and past-dependent process. Accordingly, we assume that the production of patents is an AR(1) process, and the number of patents at time t is also a function of the number of patents produced in the previous period, ceteris paribus. This helps controlling, together with value added, for domestic past innovative effort. Including a lagged dependent variable, we therefore have the following dynamic specification:

$$\begin{array}{rll} \ln P_{h,i,t} &=&\gamma \ln P_{h,i,t-1} +\alpha _1 \ln X_{h,i,t} +\beta _1 \ln IS_1\\ &&+\,\beta _2 \ln IS_2 +\beta _3 \ln IS_3 +\theta _t +\varsigma _{h,i,t} \end{array}$$
(11)

The errors ς h.i,t are decomposed into time invariant individual specific effects η h,i (in our case, 25 country-sector pairs), and the random noise ν h,i,t so that ς h,i,t = η h,i + ν h,i,t . One implication of model (11) is that the lagged dependent variable is correlated with the idiosyncratic disturbance—even if the disturbance is itself not serially correlated—because of a possible bias by the omitted individual specific effects (Greene 2003). The Ordinary Least Squares (OLS) estimates of γ in Eq. 11 are inconsistent, since the explanatory variable is positively correlated with the error term due to the presence of individual effects. The Within Group estimator eliminates this source of inconsistency by transforming the equation in order to eliminate the individual (country-sector) effect η h,i. Specifically, the mean values of the variables are calculated across the T − 1 observations for each unit, and the original observations are expressed as deviations from these means. Since the mean of the time invariant η h,i is itself η h,i, these individual effects are eliminated. Then we use OLS to estimate the transformed equation. Nevertheless, this transformation induces a possible correlation between the transformed lagged dependent variable and the transformed error term, especially in panels where the number of time periods available is small, so that the WITHIN estimator could also be inconsistent (Bond 2002).

Arellano and Bond (1991) propose an alternative estimation technique based on the GMM that corrects the bias introduced by the lagged dependent variable. In a dynamic panel model with unobserved individual heterogeneity, the idea is first-differencing Eq. 11 in order to eliminate the individual dummies (unobserved individual and time-invariant effects). However, this transformation implies that OLS estimates in the first-differenced model are inconsistent because of the dependence with the disturbance. So, sequential moment conditions are used, where lagged variables or lagged differences of the dependent variables are instruments for the endogenous differences, while the other variables can serve as their instruments. Instruments are required to be correlated with the instrumented variable and not correlated with the disturbance. In Arellano and Bond, estimators of the instruments are ‘internal’, that is, based on lags of the instrumented variables. In particular, in our case, the lags of the dependent variables or the lags of first differences must be correlated with the first difference and uncorrelated with the disturbance.Footnote 24

Table 9 shows the results. We compare WITHIN estimations with GMM estimations. Since GMM estimations are based on the assumption of stationarity, we cannot include foreign R&D stocks and US R&D stocks in the specification. This would return biased results. The Sargan test of over-identifying restrictions satisfies the underlying assumptions of the Arellano and Bond approach, suggesting that estimates reported are consistent and efficient.Footnote 25 Our results suggest that it is indeed important to control for a lagged dependent variable that is always statistically significant. International patenting is a cumulative and past-dependent process. Moreover, the estimated coefficients indicate that, on the one hand, the spillover effect measured by citations is still positive but not statistically significant, and, on the other hand, the estimated coefficient for international spillovers captured by co-inventors is still positive and significant. This result is important because it emphasizes the role played in international technological transmission by collaborations and person-to-person contacts.

Table 9 Dynamic panel, estimation of Eq. 11
Full size table

6.5 Differences across sectors

Finally, we look at the differences in terms of types of knowledge spillovers across sectors. We assume that parameters γ, α 1, β 1, β 2, β 3 in Eq. 11 are industry-specific. Table 10 shows the spillover estimates obtained from separate regressions on our five sectors. We run both a static fixed effect model and a dynamic model, using the GMM technique used in the previous section. Due to the limited number of observations, these results have to be handled with care. However, we show that the effects of international spillovers may differ across sectors. Focusing in particular on the more general dynamic specifications, our GMM results show that citation-based spillovers are positive and significant in all sectors. The values of the estimated coefficients range between 0.05 and 0.07. Second, knowledge flows measured through co-inventorship play a sensible and positive role mainly in the Chemicals and Pharmaceuticals sector, Instruments and Machinery, and Metals, with estimated elasticities equal, respectively, to 0.06, 0.04 and 0.03. It is worthwhile noting that value added has an important effect on international patenting only in Metals.

Table 10 Estimation of Eqs. 7 and 11 by sectors
Full size table
Table 11 Growth rate of patents by country and patent office
Full size table
Table 12 Number and distribution of USPTO patents by sector and country
Full size table
Table 13 Regression variables: correlation matrix
Full size table

7 Conclusions

A large body of literature emphasizes that international flows of technological knowledge have an important effect on countries’ ability to learn and innovate. This paper provides one of the first attempts to study different mechanisms of knowledge transmission from developed countries to developing countries at the industry level. In particular, we focus on the determinants of international patent production in a selected number of Latin American countries and explore the role of three channels of R&D spillovers: foreign R&D, patent citations-related spillovers and face-to-face contact spillovers based on co-inventorship relations. In the econometric analysis we use data for five big industrial sectors (Textiles and Food, Chemicals and Pharmaceuticals, Metals, Instruments Electronic and Non-Electrical Machinery, and Transportation) from five LACs (Argentina, Brazil, Chile, Colombia and Mexico) and the G-5 countries (France, Germany, Japan, UK and USA) in the years between 1988 and 2003.

Overall, this paper provides a detailed description of the nature and characteristics of international patenting (EPO and USPTO) in LACs. We show that a large part of the Latin American-invented patents belong to foreign companies with a foreign address or to a foreign subsidiary with a Latin American address, and top applicants at the USPTO and EPO are mainly US and German multinationals. The big Latin American patentees are active in a set of heterogeneous sectors of activity that are not considered very R&D-intensive (e.g. Oil, Glass, Electric, Metals and Machinery). We also show that individual inventors play a prominent role that is difficult to interpret but is linked to the fragile structure of many innovative activities in these countries.

Second, we apply GMM methods to estimate the effect of the three different types of knowledge spillovers. We find that international knowledge spillovers from the G-5 countries are a significant determinant of inventive activity during the period considered. In particular, the stock of ideas produced in the USA seems to have a strong impact on the international patenting activity of these countries. Moreover, controlling for these US-driven R&D effects, bilateral patent citations and face-to-face relationships between inventors are both important additional mechanisms of knowledge transmission. Some of our results suggest that the latter is more important than the former. Finally, we find some sectoral differences: knowledge flows measured through co-inventorship play a particularly important role mainly in the Chemicals and Pharmaceuticals sector, Instruments and Machinery and Metals.

These results have relevant policy implications. The relative weakness in many sectors of the LACs’ technological capabilities goes hand in hand with the lack of international integration of their inventive activities. The effectiveness of science and technology policies may depend upon the degree of internationalization of inventors’ activity and their international mobility. If international face-to-face contacts and collaborations display a positive marginal effect on domestic innovative activity, R&D subsidies and fiscal R&D policies should be complemented with policies oriented at the international expansion of network relationships of local inventors and companies.

However, these policy conclusions need to be handled with extreme care due to some important limitations of this study. First of all, we consider an extremely small portion of the LACs’ innovative activities. The absolute numbers displayed in Section 3 clearly indicate that few companies and individuals patent their technologies internationally. An alternative strategy could be to look at national patents at domestic patent offices. A first attempt to look at Brazilian data is provided in Laforgia et al. (2008). National patents are, however, heavily influenced by changes in national patent legislations.

A second important limitation of the paper, which is left to be addressed by future work, relates to the analysis of other important channels of technological transmission that we do not consider here, such as FDI and bilateral trade. Future work should be able to compare the relative importance of these different channels. Finally, this paper addresses only the R&D impact on international patenting. More evidence is needed to fully understand the final impact on fundamental economic variables such as labor or total factor productivity or patterns of trade. Montobbio and Rampa (2005) describe different types of relations between technological activity (using a similar set of USPTO patents) and export gains in nine large developing countries and show that they are heavily influenced by the sectoral structure of the economy.