Knowledge spillover effects: a patent inventor approach.

Aldieri, Luigi ; Vinci, Concetto Paolo

INTRODUCTION

The idea of knowledge spillover has been accepted in the theoretical and empirical literature on firm competitiveness and economic growth for the last 10 years. Knowledge spillovers exist if the social benefit from ideas is greater than the private returns to their inventors. Researchers have always considered the transmission of ideas, depending on space proximity and technological specialization. More precisely, they think that existing ideas generate new ones and only a proportion of the latter are codified in new patents. The non-codified share, embodied in the experience of inventors, may be available only through direct interaction, while the codified share operates as a public good as it is perfectly accessible to anybody who reads the patent.

The non-codified share of technological knowledge tends to assume a complex form transferable only through direct interaction (Feldman and Audretsch, 1999). The spatial proximity of inventors is helpful when direct interaction and knowledge flows between them; in turn direct interactions improve the stock of knowledge in the area where inventors are located. In other words, if knowledge is not simply available anywhere in space, as it is embodied in the inventors, both the location of the inventors and the characteristics of knowledge diffusion become crucial in order to comprehend the development of regions where inventors are located. This clarifies why the levels to which knowledge flows are restricted within geographic limits, and have received such particular consideration in the literature (Padmore and Gibson, 1998). This idea has inspired researchers to extend the innovation system framework to the regional dimension by studying knowledge flows directly within countries or specific sub-regions (Acs et al, 2002; Braczyk et al, 1998). Departing from the assumption that direct interactions spur the inventors of a region to habitually cite ideas from their neighbors (Jaffe et al, 1993), these studies evidence the local nature of innovative activity (eg Silicon Valley) (Audretsch and Feldman, 1996). For example, Jaffe et al (1993) have used this approach as a starting point for their empirical analysis of the geographic concentration of patent citations. According to the authors, citations are taken as paper trails of knowledge spillovers from the cited inventor to the citing inventor. Further studies of local systems of innovation (Audretsch and Feldman, 1996) confirm that citations are stronger at the local level.

On the other side, the codified share of technological knowledge is perfectly accessible everywhere; patents are used to generate new ideas and patent them. The patents ascribed to a firm denote the knowledge that a firm is recognized as having created by using knowledge generated elsewhere (Jaffe et al, 1993). In most cases a new patent from a firm includes a list of inventors from various counties, and thus the patent of a firm located in a country can be traced back to innovators located abroad. The geographic distance of inventors means the patent does not assume the construction company's nationality but has a cross-national dimension. This geographic distance may improve new patents from the knowledge and expertise of inventors coming from abroad and characterized by different cultures and technological orientations (Vagnani, 2015).

Other issues affecting knowledge spillovers involve technological specialization. Some researchers support the hypothesis that new ideas may be generated using existing ideas. The similarity of new ideas to previous ideas makes them easier to realize. Inventors are able to most simply recognize and engage external knowledge close to their existing knowledge base (Cohen and Levinthal, 1990). Even when they seek to embody new external knowledge in their new ideas-creation process, the resulting search procedures confine the external knowledge used to familiar and neighboring regions. Inventors are therefore expected to create ideas closely aligned with their knowledge (Aldieri, 2013).

Conversely, differences between the preceding ideas and the new ones generate the potential for non-overlapping knowledge bases. New knowledge comes from the combination of new components or new combinations of the existing components. As a consequence, it has been demonstrated that combining knowledge from different technological contexts increases the opportunity set of new components that can be utilized. Thus, inventors may use the same knowledge components in new ways or create breakthrough innovations by embodying knowledge very different from their own in realizing a new patent (Caiazza and Audretsch, 2013).

Considering the effects of both geographical and technological variables on firm innovation, the framework of our study may be summarized in Figure 1.

THEORETICAL FRAMEWORK

In this section, in order to better analyze the geographical extent of the knowledge interchange in the practice of the transmission of ideas that inventors generate during their innovative procedure, we contemplate a simple non-overlapping generation model where each generation of inventors from three different geographic areas, supposed to be the United States, European Union, and Japan, is assumed to consist of a continuum of two types of risk neutral agents, both of them normalized to one, and with an intertemporal preference rate equal to 0. Following Acemoglu (1996), the two groups of inventors are assumed to live for two periods, belong to different industries or different zones within the same geographic areas.

[FIGURE 1 OMITTED]

In the first period, at time t = 0, all types of inventors in the three different areas, in order to maximize their utility function, will determine the optimal number of patents, and their decisions are assumed to be irreversible. At t = 1 the number of mutual citations among the various inventors takes place both on a single-level and between countries, according to the following functional forms:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (1)

where [I.sup.US.sub.i,j,t], [I.sup.EU.sub.i,j,t], [I.sup.N.sub.i,j,t] are indicators capturing the number of citations by firm i(j) to a patent applied for by firm j[i) within each country, while [I.sub.i,j,t] is the index between countries, [p.sup.US.sub.i,t], ([p.sup.EU.sub.i,t]), ([p.sup.N.sub.i,t]) and [p.sup.US.sub.j,t], ([p.sup.EU.sub.j,t]), ([p.sup.N.sub.j,t]) are the number of patents of firm i and firm j belonging to the United States, EU and Japan. Moreover we assume what follows:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (2)

with 0 < [alpha], [beta] < 1

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (3)

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (4)

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (5)

with 0 < a, b, c < 1. Parameters [A.sup.US], [A.sup.EU], [A.sup.N] stand for the technological context and other firm's effects. (1) We may write:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII], (6)

where [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

As inventors have no ideas about any decisions concerning the patents of other inventors, foreign and not, in making their decisions about the optimal number of patents, the expected number of mutual citations will depend on the entire distribution of patents across all the entrepreneurs of the other groups. We define [[lambda].sub.i], [[lambda].sub.j], [[delta].sub.i], [[delta].sub.j], [[theta].sub.i] and [[theta].sub.j] as taste-positive parameters capturing disutility from investments made in order to obtain patents. As in Acemoglu (1996), the distributions of the taste parameters across inventors are common knowledge. The number of patents of any particular type of inventor will increase with the number of patents of all other inventors, both native and not. As result we can state what follows:

Proposition: Assuming [[theta].sub.i] = [[theta].sub.l], [[theta].sub.j] = [[theta].sub.2], [[lambda].sub.i] = [[lambda].sub.l], [[lambda].sub.j] = [[lambda].sub.2], [[delta].sub.i] = [[delta].sub.l], [[delta].sub.j], [[delta].sub.2]

* There exists a unique equilibrium given by: [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

* The equilibrium is Pareto inefficient and exhibits increasing pecuniary returns in the sense that a small increase in the number of patents of inventors will make everyone better off

* When a small group of inventors increases the number of patents by investing more in R&D, other inventors of the other type within the country and of any type abroad, will respond, and the equilibrium rate of return of all other inventors will improve

The above proposition states that with the incompleteness of contracts, there are increasing social returns a la Acemoglu (1996) in patents of the two types of inventors within the country and of any type abroad. Furthermore, there will be a stronger form of increasing social returns in the sense that when a small group of inventors of the first category in the United States (in the EU, or in Japan) decide to make an investment in order to increase the number of patents, native inventors of the second class, and foreign inventors of both types will respond by increasing their patents; as a result the rates of returns of inventors who have not invested more will improve. Indeed, in the empirical section, we show this same idea: the increase in patents, because of more investments in the innovative input of a firm, leads to an increase in the patents of other firms through the citation mechanism. This effect captures the core of the theoretical proposition, where we demonstrate the increasing returns of innovative activity of firms. Recall that, as opposed to the relative empirical literature, where firms are located on the basis of their geographical distribution, in this paper each firm represents a geographical vector, because we consider the geographical distribution of inventors. Thus, it is reasonable to expect that knowledge flows be between countries and just firms. In the empirical section, we measure the extent to which the knowledge flows may be affected by both technological and geographical dimensions.

DATA

We exploit two sources of data. First, we use information from the Hall et al. (2001) Patent Data file, (2) which is widely used in the empirical analysis of knowledge spillover. It refers to all patents taken out at the United States Patent and Trademarks Office (USPTO).

OECD, REGPAT database (2015), the patent database, from July 2013, (3) is the second source of information used in this analysis. This database covers patents from the European Patent Office (EPO)'s PATSTAT database, from April 2013. Two hundred forty international firms with patents both in the USPTO and EPO patent systems were selected for the analysis. The matching procedure had the same difficulties as in Aldieri (2013). The raw sample involves 35 European, (4) 83 Japanese and 122 US firms. Patents are attributed to the economic area where firms are located. The main contribution of this paper to the literature is that we consider the distribution of patent inventors to identify the firm's technological vectors. For example, 'SONY' is located in Japan and has 10,119 patents in USPTO data. Taking into account the distribution of patent inventors, there are 176 patents to be allocated to the European area and 493 to the United States. 'SONY' then has three technological vectors, one for every economic area. In Japan, the 'SONY' vector considers 10,119-176-493 = 9,450 patents. Following this procedure for all firms, we get 124 European, 122 Japanese and 201 US firm vectors in the USPTO data, and 49 European, 86 Japanese and 132 US firm vectors in the EPO data. We select particular firms because we need to identify firms whose patents are numerous both in USPTO and EPO data. We focus our attention on the European area, Japan, and the United States because most of innovation activity in the world happens there.

In Figure 2, we indicate the distribution of patent inventors by economic area both for USPTO and EPO data. As we may see, Germany is the leading country in the European area and this reveals its fundamental role in the worldwide innovation system after the United States and Japan.

In order to consider the intensity of knowledge flows between the source and the recipient of spillovers, we compute two proximity measures. On one hand, we follow the methodology developed by Jaffe (1986). This procedure rests on the construction of a technological vector for each firm based on the distribution of its patents across technology classes. These vectors allow firms to be located in a multi-dimensional technological space where technological proximity between firms is determined as the uncentered correlation coefficient between the corresponding technology vectors:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (7)

where [V.sub.i] is the technological vector of the firm [T.sub.ij] and is the technological proximity between firms i and j. In order to construct the technological proximity measures, we use the higher-level classification proposed by Hall et al. (2001), which consists of 36 two-digit technological categories. As far as the EPO data is concerned, 118 technological classes comprise the International Patent Classification at the two-digit level. In order to simplify the calculations, these 118 classes are grouped into 50 classes (Aldieri, 2011).

[FIGURE 2 OMITTED]

In order to consider the geographical proximity, we identify a dummy variable, which assumes the value of one if firms also have technological vectors in other economic areas and zero otherwise. In particular, we expect a negative effect of the defined geographical dimension on spillovers, because our proximity is constructed as a 'distance' variable: the more inventors are located in distant countries, the fewer spillover effects we should observe, as suggested in the literature (Bottazzi and Peri, 2003; Aldieri and Cincera, 2009; Aldieri, 2011).

In order to control the degree of concentration in the analyzed sectors, we include the Herfindahl index as an additional explanatory variable, as in Aldieri (2011):

[H.sub.jk] = [K.summation over (k=1)] [ms.sup.2.sub.jk] (8)

where [ms.sub.jk] are the percentage of patents of firm j in technological class k (K = 36 in USPTO data and K = 50 in EPO data) over the total patents of firm j.

In order to control firm size and absorptive capacity, we introduce the number of employees and the R&D capital stock for each firm in the estimation procedure. We obtained these variables from the EU R&D investment scoreboard (2013). In particular, the R&D capital stock is computed using the permanent inventory method (Griliches, 1979). We use the country and industry dummies to account for the characteristics of other geographical or technological firms.

THE MODEL

The model that is estimated is the following:

[p.sub.i][p.sub.j] = f([T.sub.ij], [G.sub.ij], [H.sub.i], [lk.sub.i], [le.sub.i] (9)

where Tij is the technological proximity, Gij is the geographical proximity, pi and pj are the number of patents of firm i and firm j, Hjk the Herfindahl index, [k.sub.i] is the R&D capital stock of firm i and [e.sub.i] is the number of employees of firm i. In particular, we compute a firm vector-by-vector matrix with (124 x 124) + (122 x 122) + (201 x 201) =70,661 observations for USPTO data and (49 x 49) + (86 x 86) + (132 x 132) =27,221 observations for EPO data. Our main dependent variable consists of the interaction between the number of patents of firm i and firm j. The idea is that if own innovation leads to a higher number of patents and this determines a higher number of patents in the other firms; this effect could be attributed to knowledge spillovers between firms. We expect a positive sign for the coefficients of technological proximity measure and a negative sign for the coefficients of geographical proximity, because a higher proximity leads to more knowledge flow between the firms, and then to more spillover, while the more distant localization of patent inventors might mean a spillover reduction. In Table 1, we provide the summary statistics of our sample.

As the dependent variable, the interaction between the patents of firms, is a count variable and it not normally distributed, OLS is not opportune (Greene 1994; Winkelmann and Zimmermann, 1995). For this reason, we should implement the Poisson model corrected for heteroskedasticity, as in Aldieri (2011), however, from the summary statistics table, there is first a significant proportion of zeros, the right tail of the distribution is very long. The overall mean is about 88, the maximum value is more than 64,000 for USPTO data, the overall mean is about 21 and the maximum value is more than 15,000 for EPO data. Second, there are some very large values that contribute substantially to overdispersion. These two features make it difficult to specify a model with a conditional mean and variance that captures the main features of the data. For this reason, we also estimate other two models: a negative binomial model with constant dispersion (NB1) and a negative binomial model with no constant dispersion (NB2). (5) Finally, we compare Poisson, NB1 and NB2 estimates using AIC (Akaike's information criterion) and BIC (Bayesian information criterion).

EMPIRICAL RESULTS

In Table 2 and Table 3, we report the results of the analysis based respectively on USPTO and EPO data. As explained in the previous section, we compute Poisson, NB1 and NB2 estimates. In order to identify the best model, we take into account the AIC and BIC information criteria in Table 4. On the basis of this procedure, the NB2 model is preferred, because of lower AIC and BIC. We may note that knowledge spillovers, proxied by the interaction between the patents of both firms, are sensitive to both technological and geographical measures. In general, the impact of the technological proximity is statistically positive and the effect of geographical proximity is negative. The results are in line with the relative literature (Aldieri and Cincera, 2009; Aldieri, 2011).

The result is also robust with respect to the patent system used in the analysis, even if we control for the concentration index, firm size, absorptive capacity, country dummies and industry dummies. In particular, the marginal effect relative to R&D capital stock is positive. This result seems to demonstrate that higher investment in R&D allows firm to benefit more from knowledge spillover, measured by patents. This feature is in line with the notion of absorptive capacity. The effect on spillovers confirms the core of propositions in the theoretical section.

The finding that the technological proximity and the geographical one are relevant in the mechanism of knowledge spillovers, and then in the diffusion of innovation ideas, has determinant repercussions on the policy implications in the industrial strategy of every country. Indeed, policymakers should prefer greater concentration in particular industries or mergers between firms in a particular geographic region so as to foster growth.

In order to investigate the role of the geographic dimension, we ran NB2 estimates by country in Table 5 and Table 6. As we may see, the impacts of technological and geographical proximities have the same sign as in the pooled estimation context. We may compare the results of each country with respect to those of world average, proxied by pooled coefficients. (6) We demonstrate that the marginal effects of European and Japanese firms are lower than world results while the marginal effects of American firms are close to (EPO data) or higher than (USPTO data) world results. The marginal effect relative to own R&D capital stock is negative for European and positive for Japanese and American firms. This result is robust with respect to the patent system used in the analysis. The result for the absorptive capacity could be caused by the technological gap between the countries (Table 7).

In order to also analyze the technological dimension, we ran the NB2 estimates by high-technology sector and by manufacturing sector. (7) The marginal effects of technological and geographical proximities are higher for high-technology firms. This result is robust with respect to patent office data. This seems to demonstrate the stronger relevance of spillover effects especially for more advanced technologically firms (Table 8).

DISCUSSION AND CONCLUDING REMARKS

In this paper we investigated the effects of technological and geographical proximity on knowledge flow in the United States, Japan, and Europe. In doing so, we introduced a patent inventor approach to measuring the proximity between the firms allocating patents on the basis of their inventors' distribution. This is a relevant contribution to the literature, because patents are usually attributed to the economic area where firms are located. We compared results between the United States, Japan, and Europe for the technological and geographical proximity effects on knowledge spillover. The empirical results indicate that there is a statistically significant impact from technological and geographical proximity on knowledge spillover, and that these results are robust with respect to the patent office data used in the analysis. These findings confirm the relevance of both technological and geographic dimensions in the literature, and also from a patent inventor's perspective (Lychagin et al, 2010).

In general, the impact of technological proximity is statistically positive and the effect of geographical proximity is negative. The finding that the effect of geographical proximity in terms of distance is negative seems to indicate a localization effect from spillover. Indeed, if a firm's technological vectors are located in its own economic area, this leads to greater knowledge spillover.

The industrial strategy of every country should thus consider both the technological and geographical aspects of the diffusion of innovative ideas.

There are some limitations in our analysis, however, which can be addressed in future research. First of all, we account for the correlation between spillovers and proximity measures, but it would be interesting to identify the causality of the innovative process, through the implementation of a methodological procedure able to deal with the endogeneity of relevant variables. To this aim it would be reasonable to assume a time lag between the variables in order to move one step closer to causality.

A second weakness of our analysis thus involves its static nature.

Finally, to improve our analysis, the results should be compared with those based on Japanese or Chinese patent data. In this way, we could also take into account the innovative ideas for the Asian economic area.

Acknowledgements

The authors are grateful to two referees whose comments greatly improved the quality of the paper. Results, conclusions, views or opinions expressed in this paper are only attributable to the authors.

REFERENCES

Acemoglu, D. 1996: Microfoundation for social increasing returns in human capital accumulation. The Quarterly Journal of Economics 111 (3): 779-804.

Acs, ZJ, Anselin, L and Varga, A. 2002: Patents and innovation counts as measures of regional production of new knowledge. Research Policy 31(7): 1069-1085.

Aldieri, L. 2011: Technological and geographical proximity effects on knowledge spillovers: Evidence from the US patent citations. Economics of Innovation and New Technology 20(6): 597-607.

Aldieri, L. 2013: Knowledge technological proximity: Evidence from US and European patents. Economics of Innovation and New Technology 22(8): 807-819.

Aldieri, L and Cincera, M. 2009: Geographic and technological R&D spillovers within the triad: Micro evidence from US patents. The Journal of Technology Transfer 34(2): 196-211.

Audretsch, DB and Feldman, MP. 1996: R&D spillovers and the geography of innovation and production. The American Economic Review 86(3): 630-640.

Bottazzi, L and Peri, G. 2003: Innovation and spillovers in regions: Evidence from European patent data. European Economic Review 47(4): 687-710.

Braczyk, HJ, Cooke, PN and Heidenreich, M. 1998: Regional innovation systems: The role of governances in a globalized world. London: UCL Press.

Caiazza, R and Audretsch, DB. 2013: A general framework for classifying spin-off. International Entrepreneurship Review 11(1): 15-30.

Cameron, AC and Trivedi, PK. 2013: Regression Analysis of Count Data, 2nd Edition, Cambridge University Press.

Cohen, WM and Levinthal, DA. 1990: Absorptive capacity: A new perspective on learning and innovation. Administrative Science Quarterly 35(1): 128-152.

European Commission. 2013: The 2013 EU Industrial R&D Investments Scoreboard, JRC Scientific and Technical Research series, http://iri.jrc.ec.europa.eu/scoreboard.html, accessed 7 January 2014.

Feldman, MP and Audretsch, DB. 1999: Innovation in cities: Science-based diversity, specialization and localized competition. European Economic Review 43 (2): 409-429.

Greene, WR. 1994: Accounting for excess zeros and sample selection in poisson and binomial regression models. Working paper 94-10, New York University Stern School of Business, New York.

Griliches, Z. 1979: Issues in assessing the contribution of research and development to productivity growth. Bell Journal of Economics 10(1): 92-116.

Hall, BH, Jaffe, AB and Trajtenberg, M. 2001: The NBER patent citations data file: lessons, insights and methodological tools. NBER Working Paper No. 8498, Cambridge, MA.

Jaffe, AB. 1986: Technological opportunity and spillovers of R&D: Evidence from firms' patents, profits and market value. American Economic Review 76(5): 984-1001.

Jaffe, AB, Trajtenberg, M and Henderson, R. 1993: Geographic localization of knowledge spillovers as evidenced by patent citations. The Quarterly Journal of Economics 108(3): 577-598.

Lychagin, S, Pinkse, J, Slade, ME and Van Reenen, J. 2010: Spillovers in space: does geography matter? NBER Working Paper No. 16188, Cambridge, MA.

OECD. REGPAT database, February 2015.

Padmore, T and Gibson, H. 1998: Modelling systems of innovation: II. A framework for industrial cluster analysis in regions. Research Policy 26(6): 625-641.

Vagnani, G. 2015: Exploration and long-run organizational performance: The moderating role of technological interdependence. Journal of Management 41(6): 1651-1676.

Winkelmann, R and Zimmermann, K. 1995: Recent developments in count data modelling: Theory and application. Journal of Economic Surveys 9(1): 1-24.

(1) It is only for simplicity that we don't introduce parameters capturing technological and geographical proximity.

(2) We use the updated patent data (1975-2002) downloaded from Hall's website: www.econ. berkeley.edu/ ~ bhhall/patents.html

(3) Please contact Helene.DERNIS@oecd.org to download REGPAT database

(4) The European economic group involves the following countries: Belgium, Denmark, Finland, France, Germany, Italy, the Netherlands, Sweden, and the United Kingdom

(5) See Cameron and Trivedi (2013) for a technical discussion of Poisson, NB1 and NB2 models.

(6) We thank the reviewer for their suggestion for this point.

(7) The industry sectors are: oil & gas, chemicals, basic resources, construction, manufacturing, automobiles, food & beverage, personal and household goods, health care, retail of food and drugs, media, travel & leisure, telecommunications, utilities, banks and high-technology.

LUIGI ALDIERI [1] & CONCETTO PAOLO VINCI [2]

[1] Department of Business and Economic Studies, Parthenope University of Naples, Via Generate Parisi 13, 80132 Naples, Italy. E-mail: aldieri@uniparthenope.it

[2] Department of Economic and Statistic Sciences, University of Fisciano, Via Giovanni Paolo II, 132, 84084 Fisciano, Salerno, Italy. E-mail: cpvinci@unisa.it

Table 1: Description statistics

                               Standard
Variable               Mean    Deviation   Minimum   Maximum

USPTO
  [p.sub.i][p.sub.j]   88.09    828.500       0      64,200
  [T.sub.ij]            0.23      0.249       0         1
  [G.sub.ij]            0.45      0.498       0         1
  [H.sub.jk]            0.01      0.017     0.000     0.112
  [Lk.sub.i]           14.53      2.106     5.989    21.048
  [Le.sub.i]           20.07      1.934     8.955    25.634

EPO
  [p.sub.i][p.sub.j]   21.91    529.824       9      15,300
  [T.sub.ij]            0.31      0.269       0         1
  [G.sub.ij]            0.11      0.307       0         1
  [H.sub.jk]            0.01      0.015     0.000     0.112
  [Lk.sub.i]           14.53      2.172     5.989    20.925
  [Le.sub.i]           19.96      2.018     8.955    25.634

Note: 70,661 observations.

Note: 27,221 observations.

Table 2: Full sample results (USPTO data)

                    Coefficient estimates      (b)Marginal effects

                                 Standard                  Standard
                                error (a,c)               error (a,c)

Poisson
  [T.sub.ij]          1.49 ***    (0.123)      12.75 ***    (0.987)
  [G.sub.ij]         -4.02 ***    (0.072)     -49.66 ***    (2.344)
  [H.sub.ik]        -14.02 ***    (2.628)    -119.99 ***   (22.066)
  [lk.sub.i]          0.38 ***    (0.027)       3.25 ***    (0.321)
  [le.sub.i]          0.41 ***    (0.037)       3.53 ***    (0.272)
  Pseudo [R.sup.2]    0.55

NB1
  [T.sub.ij]          0.32 ***    (0.011)      31.55 ***    (1.668)
  [G.sub.ij]         -0.56 ***    (0.006)     -54.65 ***    (2.014)
  [H.sub.ik]         -3.10 ***    (0.144)    -306.22 ***   (18.016)
  [lk.sub.i]          0.07 ***    (0.002)       7.16 ***    (0.345)
  [le.sub.i]          0.03 ***    (0.002)       3.28 ***    (0.247)

NB2
  [T.sub.ij]          1.10 ***    (0.112)      10.95 ***    (1.114)
  [G.sub.ij]         -4.08 ***    (0.043)     -59.67 ***    (1.776)
  [H.sub.ik]         -2.77 *      (1.665)     -27.67 *      (17.09)
  [lk.sub.i]          0.41 ***    (0.021)       4.11 ***    (0.201)
  [le.sub.i]          0.36 ***    (0.017)       3.55 ***    (0.212)

(a) *** Coefficient significant at 1%, * Coefficient significant at
10%

(b) Country dummies and industry dummies are included in the
estimation procedure. The European area is the reference country. The
oil & gas sector is the industry reference.

(c) Standard errors are corrected for heteroskedasticity.

Table 3: Full sample results (EPO data)

                     Coefficient estimates (b)    Marginal effects

                                   Standard                 Standard
                                  error (a,c)              error (a,c)

Poisson
  [T.sub.ij]          0.80 ***      (0.050)     10.75 ***    (0.693)
  [G.sub.ij]         -1.46 ***      (0.058)    -12.22 ***    (5.475)
  [H.sub.ik]         20.83 ***      (0.857)     27.91 ***   (11.999)
  [lk.sub.i]          0.10 ***      (0.011)      1.40 ***    (1.489)
  [le.sub.i]          0.11 ***      (0.013)      1.52 ***    (1.669)
  Pseudo [R.sup.2]    0.47

NB1
  [T.sub.ij]          0.44 ***      (0.023)     10.44 ***    (5.694)
  [G.sub.ij]         -0.67 ***      (0.019)    -10.77 ***    (2.421)
  [H.sub.ik]         11.61 ***      (0.495)     52.20 ***   (14.598)
  [lk.sub.i]          0.02 ***      (0.004)      6.89 ***    (1.201)
  [le.sub.i]          0.10 ***      (0.005)      1.32 ***    (1.212)

NB2
  [T.sub.ij]          0.78 ***      (0.042)     10.44 ***    (5.694)
  [G.sub.ij]         -1.20 ***      (0.036)    -10.77 ***    (2.421)
  [H.sub.ik]         38.99 ***      (1.021)     52.20 ***   (14.598)
  [lk.sub.i]          0.05 ***      (0.009)      6.89 ***    (1.201)
  [le.sub.i]          0.10 ***      (0.009)      1.32 ***    (1.212)

(a) *** Coefficient significant at 1%.

(b) Country dummies and industry dummies are included in the
estimation procedure. The European area is the reference country. The
oil & gas sector is the industry reference.

(c) Standard errors are corrected for heteroskedasticity.

Table 4: Comparison based on information criteria

                Poisson          NB1          NB2

USPTO Data
AIC           151000006656    1072995.8    1042269.8
BIC           151200006144    1073198.1    1042472.1

EPO Data
AIC           56780001280     614114.9     608521.1
BIC           56780001280     614296.9     608703.1

Table 5: Estimates by geographic area (USPTO data)

                        EU firms                   JP firms

                              Standard                   Standard
                             error (a,c)                error (a,c)

Coefficient estimate (b)
  [T.sub.ij]      0.65 *       (0.334)       0.43 ***     (0.105)
  [G.sub.ij]     -2.81 ***     (0.208)      -4.27 ***     (0.060)
  [H.sub.ik]    -22.32 ***     (2.092)      19.60         (2.298)
  [lk.sub.i]     -0.37 ***     (0.066)       0.21 ***     (0.030)
  [le.sub.i]      1.12 ***     (0.075)       0.56 ***     (0.025)

Marginal effects
  [T.sub.ij]      1.21 *       (0.639)       3.82 **      (0.915)
  [G.sub.ij]    -14.97 ***     (2.913)     -43.37 ***     (1.578)
  [H.sub.ik]     41.54 ***     (4.786)     174.64 ***    (19.939)
  [lk.sub.i]     -0.69 ***     (0.165)       1.83 ***     (0.285)
  [le.sub.i]      2.09 ***     (0.265)       5.01 ***     (0.263)

                          US firm

                                Standard
                               error (a,c)

Coefficient estimate (b)
  [T.sub.ij]       1.72 ***      (0.075)
  [G.sub.ij]      -4.46 ***      (0.039)
  [H.sub.ik]     -20.45 ***      (1.514)
  [lk.sub.i]       0.58 ***      (0.013)
  [le.sub.i]       0.26 ***      (0.013)

Marginal effects
  [T.sub.ij]      23.72 ***      (1.094)
  [G.sub.ij]     -62.342 ***     (1.276)
  [H.sub.ik]    -281.54 ***      (322.8)
  [lk.sub.i]       7.97 ***      (0.227)
  [le.sub.i]       3.53 ***      (0.190)

(a) *** Coefficient significant at 1%, ** Coefficient significant at
5%, * Coefficient significant at 10%.

(b) Industry dummies are included in the estimation procedure. The
oil & gas sector is the industry reference.

(c) Standard errors are corrected for heteroskedasticity.

Table 6: Estimates by geographic area (EPO data)

                        EU firms                   JP firms

                              Standard                   Standard
                             error (a,c)                error (a,c)

Coefficient estimate (b)
  [T.sub.ij]      0.70 ***     (0.133)       0.74 ***     (0.056)
  [G.sub.ij]     -1.72 ***     (0.088)      -2.07 ***     (0.068)
  [H.sub.ik]     14.97 ***     (1.664)      52.61 ***     (1.576)
  [lk.sub.i]     -0.08         (0.047)       0.31 ***     (0.021)
  [le.sub.i]      0.53 **      (0.049)       0.04 **      (0.019)

  Marginal effects
  [T.sub.ij]      2.27 ***     (4.491)       1.16 ***     (1.270)
  [G.sub.ij]    -50.37 ***     (3.121)     -21.78 ***     (0.453)
  [H.sub.ik]     48.71 ***     (56.53)     115.01 ***     (40.08)
  [lk.sub.i]     -2.52         (1.583)       6.85 ***     (4.716)
  [le.sub.i]      1.71 **      (1.807)       8.28 **      (4.179)

                         US firm

                              Standard
                             error (a,c)

Coefficient estimate (b)
  [T.sub.ij]      0.70 ***     (0.053)
  [G.sub.ij]     -0.96 ***     (0.040)
  [H.sub.ik]     39.62 ***     (1.421)
  [lk.sub.i]      0.01 ***     (0.009)
  [le.sub.i]      0.08 ***     (0.010)

  Marginal effects
  [T.sub.ij]      6.72 ***     (5.197)
  [G.sub.ij]    -65.56 ***     (2.147)
  [H.sub.ik]    381.01 ***    (143.69)
  [lk.sub.i]      0.60         (0.909)
  [le.sub.i]      7.35 ***     (0.949)

(a) *** Coefficient significant at 1%, ** Coefficient significant
at 5%.

(b) Industry dummies are included in the estimation procedure. The oil
& gas sector is the industry reference.

(c)Standard errors are corrected for heteroskedasticity.

Table 7: Estimates by industry sectors (USPTO data)

                        High technology            Manufacturing

                                 Standard                   Standard
                                error (a,c)                error (a,c)

Coefficient estimate (b)
  [T.sub.ij]         1.35 ***     (0.116)       1.22 ***     (0.158)
  [G.sub.ij]        -4.00 ***     (0.083)      -4.51 ***     (0.073)
  [H.sub.ik]        -0.72 ***     (3.323)      -4.04         (3.257)
  [lk.sub.i]         0.21 ***     (0.043)       0.48 ***     (0.033)
  [le.sub.i]         0.51 ***     (0.027)       0.25 ***     (0.032)

Marginal effects
  [T.sub.ij]        25.23 ***     (2.328)      10.93 ***     (1.360)
  [G.sub.ij]       -84.63 ***    (33.598)     -59.54 ***     (3.308)
  [H.sub.ik]       -13.54 ***    (874.37)     -36.24        (30.205)
  [lk.sub.i]         4.01 ***     (13.36)       4.28 ***     (0.314)
  [le.sub.i]         9.61 ***     (9.878)       2.22 ***     (0.320)

(a) *** Coefficient significant at 1%.

(b) Country dummies and industry dummies are included in the
estimation procedure. The European area is the reference country. The
oil & gas sector is the industry reference.

(c) Standard errors are corrected for heteroskedasticity.

Table 8: Estimates by industry sectors (EPO data)

                      High technology              Manufacturing

                                Standard                    Standard
                               error (a,c)                 error (a,c)

Coefficient estimate (b)
  [T.sub.ij]       0.42 ***      (0.080)       1.29 ***      (0.102)
  [G.sub.ij]      -1.40 ***      (0.063)      -1.369 ***     (0.083)
  [H.sub.ik]      40.11 ***      (2.238)      38.80 ***      (2.203)
  [lk.sub.i]      -0.06 ***      (0.017)       0.07 ***      (0.019)
  [le.sub.i]       0.26 ***      (0.018)       0.07 ***      (0.021)

Marginal effects
  [T.sub.ij]      44.49 ***      (8.544)      18.55 ***      (1.579)
  [G.sub.ij]     -96.47 ***      (3.519)     -12.39 ***      (0.551)
  [H.sub.ik]     425.01 ***     (259.30)     557.01 ***     (338.54)
  [lk.sub.i]      -6.403 ***     (1.761)       1.04 ***      (0.277)
  [le.sub.i]       2.70 ***      (1.838)       1.01 ***      (0.294)

(a) *** Coefficient significant at 1%, ** Coefficient significant at
5%, * Coefficient significant at 10%.

(b) Country dummies and industry dummies are included in the
estimation procedure. The European area is the reference country. The
oil & gas sector is the industry reference.

(c) Standard errors are corrected for heteroskedasticity.