Is innovation persistent at the firm Level? An econometric examination ...

18 downloads 1890 Views 295KB Size Report
Conference (Warwick, March 2002). 1 Université ... macroeconomic level can be supported by different firm-level behavioral assumptions. The aim of ..... played by the development of internal innovative know-how in Samsung's technological successes ..... what we call the relative innovation persistence in Table 1. We find ...
Is innovation persistent at the firm Level? An econometric examination comparing the propensity score and regression methods ∗

1

Emmanuel Duguet and Stéphanie Monjon

2

Revised, July 2004



We thank P. Aghion, T. Brodaty, B. Crépon, D. Encaoua, P. Hoonhout, C. Lebas and R. Veugelers for helpful suggestions and comments as well as the participants to the Spring Meeting of Young Economists (Copenhagen, April 2001), the CERAS-EUREQua-LEI seminar in Industrial Organization (Paris, May 2001), the CEPR-ECARES workshop (PMFM, Brussels, May 2001), the EEA annual meeting (Lausanne, August 2001), the EARIE meeting (Dublin, August 2001), the 50th AFSE annual congress (Paris, September 2001), the ADRES seminar (Paris, November 2001) and the Royal Economic Society Conference (Warwick, March 2002). 1 Université de Bretagne Occidentale and EUREQua - CNRS UMR 8594 – Université Paris I - 106-112, Boulevard de l’Hôpital, 75647 Paris Cedex 13. e-mail : [email protected] 2 This paper was written while visiting University College London. EUREQua - CNRS UMR 8594 Université Paris I, 106-112, Boulevard de l’Hôpital, 75647 Paris Cedex 13. e-mail : [email protected].

1

Is innovation persistent at the firm Level? An econometric examination comparing the propensity score and regression methods At the macroeconomic level, the persistence of technological change allows sustainable growth. But do the innovations come from the same set of firms or from a continuous renewal of innovators? On this point, the assumptions underlying the endogenous growth models differ and innovation persistence at the macroeconomic level can be supported by different firm-level behavioral assumptions. The aim of this article is threefold. Firstly, we evaluate a measure of the degree of innovation persistence at the firm level. Secondly, we analyze the factors underpinning the innovation persistence by testing the theoretical explanations that have been proposed in the literature. Lastly, we examine the robustness of the standard econometric methods used in innovation economics. We show that the persistence of innovation is strong at the firm level and that the right theoretical modeling depends on the size of the firm. While the small firms reveal strong learning-by-doing effects in the production of innovation, the persistence of innovation in the large firms relies on the persistence of formal research and development investments. Keywords: Community Innovation Surveys, Creative destruction, Innovation, Learning-by-doing, Matching, Persistence, Propensity score, Research and development. JEL Classification: C14, O31, O32.

L’innovation est-elle persistante dans les entreprises ? Un examen économétrique comparatif des méthodes de régression et d’appariement sélectif Au niveau macroéconomique, la persistance de l’innovation permet d’envisager une croissance soutenue. Mais les innovations proviennent-elles du même ensemble d’entreprises ou d’un renouvellement perpétuel des innovateurs ? Sur ce point, les hypothèses varient d’un modèle de croissance endogène à l’autre et la persistance macroéconomique de l’innovation peut se justifier par des hypothèses microéconomiques différentes. Cet article examine les trois questions suivantes. Premièrement, nous estimons un degré de persistance de l’innovation au niveau de l’entreprise. Deuxièmement, nous analysons les déterminants de la persistance de l’innovation en testant les différents modèles proposés dans la littérature. Troisièmement, nous examinons la robustesse des résultats trouvés par les méthodes économétriques habituelles. Nous montrons que la persistance de l’innovation est forte au niveau de l’entreprise et que la bonne représentation théorique dépend de la taille de l’entreprise. Alors que les petites entreprises font apparaître des effets d’apprentissage importants dans la production d’innovation, la persistance de l’innovation dans les grandes entreprises repose sur la persistance d’une recherche et développement formalisée. Mots-clés : Enquête Communautaire sur l’Innovation, Destruction créatrice, Innovation, Apprentissage par la pratique, Appariement sélectif, Recherche et développement. Classement JEL : C14, O31, O32.

2

I.

Introduction Innovation is a primary source of economic growth. But do innovations come

from the same set of firms or from a continuous renewal of innovators? On this point, the assumptions underlying the endogenous growth models are different. While Romer (1990) suggests a strong persistence of innovators, Aghion and Howitt (1992) take the opposite view and develop a neo-Schumpeterian model in which the process of 3

creative destruction leads to a perpetual renewal of innovators. More recently, Aghion, Harris and Vickers (1997), Stein (1997), Encaoua and Ulph (2000) and Aghion, Harris, Howitt and Vickers (2001) have shown that innovation persistence at the macroeconomic level can be supported by a rich number of firm-level behavioral assumptions. What is the empirical relevance of these different microeconomic foundations? This question relates to two central problems in economic theory. In the first place, the frequency with which firms introduce innovations plays a central role in the analysis of technical progress and economic growth (Romer, 1990; Aghion and Howitt, 1998). In the second place, understanding whether innovation is persistent or not at the firm level constitutes an important piece of evidence for finding and improving current theories of industrial dynamics, where some forms of dynamic increasing returns play a major role in determining degrees of concentration, the evolution of market shares and their stability over time (Baldwin and Scott, 1987; Geroski, 1989). Despite the theoretical importance of this topic, only few empirical works have examined the issue of innovation persistence at the firm-level data (Crépon and Duguet, 1997; Geroski, Van Reenen and Walters, 1997; Malerba and Orsenigo, 1999; Cefis and Orsenigo, 2001; Cabagnols, 2002). Overall, these works seem to conclude that innovation is persistent in a small number of firms only. Hence, as long as the majority of innovators would be involved in a creative destruction process, there would still remain an innovative core. However, several considerations led us to investigate the robustness of this conclusion. First, the data used to measure innovation mainly correspond to patent data which probably underestimate the number of innovative firms and therefore the

3

This difference of assumptions can be seen by comparing the innovation value equations of these models.

3

persistence of innovation. The main problem is that a patent involves both to innovate and to be the first to innovate. This means that patent data could measure the persistence of innovative leadership rather than the persistence of innovation. In the same way, the “major innovations” data involve some kind of leadership as well. Consequently, the evaluation of innovation persistence requires separating innovation from its commercial performance and from the intellectual protection strategies of the firms. In this article, we use detailed data coming from the Community Innovation Surveys (CIS) that provide information that satisfy these conditions. The second motivation of this paper is to illuminate the factors underpinning the innovation persistence by testing the theoretical explanations of the different models that exist. The previous studies, sometimes limited to sample statistics, seldom allow for the identification of the theoretical model that generates the data. This paper identifies the theoretical models that are compatible with the empirical regularities. Lastly we examine the robustness of the standard econometric methods. Ideally, we would like to measure the difference between, on the one hand, the innovative performance a firm makes today knowing that it has innovated in the past and, on the other hand, the innovative performance the same firm would have done if it 4

had not innovated in the past. There is no way to observe both quantities at the same 5

time for any firm. The standard econometric methods assume that the non-observable outcomes can be obtained from a regression model, which could be invalid. In this article, in order to solve this problem, we use the propensity score approach introduced by Rubin (1974), Rosenbaum and Rubin (1983) and developed by Heckman, Ichimura and Todd (1997) and use non-parametric methods of estimation (Härdle, 1990; Andrews and Buchinsky, 2000). Several interesting findings emerge from our research. First, we find a significant persistence of innovation in the lines of business that have significant technological opportunities. Second, we find that different factors are at the origin of this persistence. More interestingly, our results suggest that the factors that induce innovation persistence depend on the characteristics of the firms, especially on their

4

We evaluate the importance of innovation persistence of a firm through the impact of its past innovations on its current innovation. Geroski et al. (1997) apprehend the extent of innovation persistence of a firm by the number of consecutive years during which it obtains a patent or a major innovation. We discuss the differences between these two approaches below. 5 Either the firm innovated and we cannot observe what it would have done if it had not innovated, or the firm did not innovate and we cannot observe what it would have done if it had innovated.

4

size. Dynamic increasing returns in the production of innovations appear to play a major role on innovation persistence in the small firms. In the largest firms, this effect vanishes and the persistence of innovation originates from the persistence of the formal research and development investments. Lastly, we find that the standard regression methods provide correct results on average but conceal an interesting composition effect. The article starts with a short review of the literature that lays the theoretical foundations of our econometric model. In particular, we highlight the theoretical predictions about innovation persistence of different innovation models. Then we summarize the empirical results from previous studies. In the section 3, we describe the model and the data on which the study is based. Section 4 discusses some estimation issues and sets out the estimation methods. Section 5 outlines the empirical results and briefly discusses their implications. Section 6 concludes the paper.

II.

Theoretical framework and previous empirical studies 2.1 The theoretical framework Different types of models, which lead to different empirical predictions, can

explain the persistence of innovation at the firm level. The linear model of innovation establishes a simple relationship between the research and development (R&D) expenditures of a firm and its innovations. The firms that can support the sunk costs of R&D make inventions that lead to product or process innovations (Cohen and Klepper, 1996). In this model the successive innovations would originate from the continuity of R&D expenditures and are not directly connected between them. According to this vision of the innovative process, innovation is persistent only if R&D is. A second stream of the literature underlines the importance of the commercial success of past innovations to innovate again. Giving the difficulty of funding the R&D activities, the commercial success of past innovations helps to fund the current innovative activities: successful innovative firms make profits that can be allocated to future R&D investments. According to this theory, a firm that reached a commercial success in the past has more chance to innovate in the future merely because it reinvests its benefits in its research projects. Hence, “success breeds success” (Nelson and Winter, 1982): a past innovation that met commercial success becomes a necessary condition to finance the future research projects and to innovate persistently.

5

A third stream of literature emphasizes the strategic considerations in the innovation decision. A seminal article by Arrow (1962) suggested that firms in competitive markets have significantly greater incentive to invest in innovation than firms in markets characterized by a significant degree of monopoly power. Gilbert and Newberry (1982) showed that if there is free entry to the industry, Arrow's result does not hold. If there is no uncertainty on the innovation process (auction model) and if the innovation is non-drastic, the incumbents firms with monopoly power will rationally preempt potential entrant investment in innovation in order to continue to profit from the extension of existing market power to a new generation of technology, which should lead to a strong innovation persistence at the firm level. But Reinganum (1983) reestablished Arrow's result in the case of drastic innovations, by showing that under conditions of uncertainty, incumbent monopolists will rationally invest less in innovation than entrants will, for fear of cannibalizing the stream of rents from their existing 6

products. Successive innovations should come from different firms, which decrease the degree of innovation persistence. In these models the successive innovations are not directly connected but two firms with different market shares will not have the same incentive to remain innovative, which has an impact on the degree of innovation persistence at the firm level. A last stream of literature relies on the idea that innovations result from an accumulation of specific competencies (Rosenberg, 1976). More precisely, this literature considers that the innovative abilities that a firm develops when it invests in research projects do not necessarily depreciate rapidly over time. Therefore, the same knowledge or know-how may be applied to develop several innovations at successive 7

times. The competencies related to innovation do not only refer to the scientific knowledge available to the firm but also to an informal know-how in the production of innovations. According to that view, firms benefit from dynamic increasing returns in the form of learning-by-doing, learning-to-learn or scope economies in the production of innovations, which increases the innovation persistence at the firm level.

8

6 This result comes from the fact that an innovative monopoly replaces its own product so that it only earns the difference between its two innovative profits, while the incumbent earns a monopoly profit. 7 It is particularly true when the innovations are cumulative. See Scotchmer (1999) for a discussion of the concept of cumulative innovations. 8 There are many possible sources of dynamic increasing returns in the production of innovations. Chandler (1990) highlights the importance of such effects in the production of innovations.

6

Even if these different models do not explain the persistence of innovation in the same way, most of them suggest that we should observe a significant degree of 9

persistence. According to the first model, a weak persistence would mean that firms do not invest in R&D continuously. But, empirically, it is widely observed that there are large differences in R&D efforts across firms and that these differences in R&D effort are persistent over time (Hall, Griliches and Hausman, 1986). Only a part of the firms performs R&D but these firms generally invest in R&D activities continuously. According to the second model, weak innovation persistence would imply that firms would not reinvest their profits from past innovations in their current research projects. It is unlikely giving the importance of innovation for their competitiveness. In some industries, competition principally centers on innovation (Encaoua and Hollander, 2002). According to the last model, weak innovation persistence would imply that dynamic increasing returns in the production of innovations are not important. Some recent empirical analyses suggest the contrary. Kim (1997) showed the central role played by the development of internal innovative know-how in Samsung’s technological successes in the semi-conductors industry. Henderson and Cockburn (1996) and Nightingale (2000) studied the pharmaceutical industry. The first study shows that the firms of this industry have benefited from significant scope economies in the production of new drugs. Nightingale (2000) finds that the pharmaceutical firms have succeeded in lowering the production costs of new drugs by adopting new experimentation methods and organizational changes. Klette (1996) provides also some empirical evidence of the importance of scope economies in knowledge production for Norwegian manufacturing. If we combine these models, we should expect a significant degree of innovation persistence.

2.2 The previous empirical results The issue of innovation persistence at the firm level has been studied by Geroski, Van Reenen and Walters (1997). The authors use two different measures of innovation: annual patent data and annual major innovations data. They define the degree of innovation persistence of a firm as the number of consecutive years during

9

The model with strategic considerations is less clear on this point.

7

which it has a recorded innovative output (annual patent and annual major innovations) and note that “it is very hard to find any evidence at all that innovative activity can be self-sustaining over anything other than very short periods of time”. Cabagnols (2004) uses the same data to examine more precisely the impact of technological accumulation (measured by past patents) on their ability to persist in innovation. It is also another way of evaluating the innovation persistence in a firm. The author finds a positive and significant relationship. The same study reports comparable results for the French firms on similar data. The work of Crépon and Duguet (1997) is equally related to the issue of innovation persistence. They use a panel of R&D performers operating in France and 10

they also use patent data to measure innovation.

They estimate a dynamic count

data model that links the current number of patents to both the previous year number of patents and the amount invested in R&D. They also add an individual fixed effect that can represent differences in the firm’s propensity to patent, size or technological opportunities. They find that the effect of lagged patents on the current number of patents is significantly positive, which suggests a rather strong persistence in innovation among formal R&D performers. The descriptive statistics study by Malerba and Orsenigo (1999) also examined the issue of persistence by using patent data of six countries over the period 19781991.

11

This study shows that a large fraction of innovators is casual. Nevertheless,

there would still remain a stable group of innovators that apply for a large share of patenting. The results of this study are confirmed by Cefis and Orsenigo (2001) who find that both “great-innovators” and non-innovators have a high probability of staying in the same innovative state. Then, there would be a strong persistence of innovation in a core of firms. Globally, these studies suggest thus that the theoretical predictions would be applicable only to a restricted group of firms. But this conclusion could be conditioned by the specificity of data used in these studies. There is a difference between a patent and an innovation (Griliches, 1990). The most important difference for this study is that

10 R&D performers are defined according to the Frascatti criterion (at least one research working full time on R&D). 11 The six countries are: France, Germany, Italy, Japan, the U.K. and the U.S.A. Their patent data are available for different periods: 1978-1982, 1982-1985, 1986-1988 and 1988-1991.

8

a patent implies leadership in innovation.

12

Indeed, an innovative firm needs to be the

first to apply for a patent in order to be properly registered in the data set. Patent data do not only measure innovation but also the fact that a firm won an innovation race. Hence, a sample including firms that win the innovation race from time to time would show up a weak persistence of innovation, even if these firms innovate all the time. Moreover, the leadership itself could be poorly measured by patent data. This problem is linked to the intellectual property strategy of the firms. Firms can have an obvious interest in avoiding patenting an intermediate innovation in order to conceal the 13

knowledge that could be used by their competitors.

There is a strong empirical

evidence of this kind of behavior (Levin et al., 1987, and Cohen et al., 1997, on 14

American data; Duguet and Kabla, 1998, on French data).

Lastly, the firms tend to

patent more their product innovations than their process innovations (Arundel and Kabla, 1998). The patent data are thus biased in favor of product innovations. The second type of data used in Geroski, Van Reenen and Walters (1997) concerns innovations that met a commercial success. Here, the data avoid the biases associated to the patenting strategy of the firms but there remains one limitation. Since this definition involves a commercial success, the firms considered as innovative are likely to be either innovation leaders or commercial leaders. In the latter case, the data measure the ranking of the firms on the market, that is their ability to adapt the available innovations to consumers’ tastes. Our objective is to use a different measure of innovation and to develop a model in which it is possible to identify the determinants of innovation persistence with a more representative sample of firms.

12

For a theoretical analysis of the persistence of leadership, see Gruber (1992) or Denicolo (2001). Moreover, patents underestimate the innovation of small-sized firms that use fewer patents than large firms. See Acs and Audretsch (1988). 14 Duguet and Kabla (1988) show that on average only a third of the innovations are patented because of the information disclosure implied by the patent documents. Moreover, after controlling for the propensity to patent, R&D investments, market share and industry of the firm, the determinants “will to avoid litigation” and “technology negotiations” remain strongly significant in the explanation of the number of patent applications. Hence, strategic aspects are omnipresent in the patent numbers, which precludes from using it as a mere innovation indicator. This strategic dimension of patents is more and more present in the firms’ decisions to patent their innovations (Encaoua and Hollander, 2001). 13

9

III.

The model and the data 3.1 The model Our aim is twofold: evaluating a measure of the innovation persistence degree

at firm level and understanding the origin of the persistence then identified by confronting the different theoretical models. The measure of the innovation persistence degree we used is the difference between the probabilities to innovate depending on a firm innovated in the past or not. Once the persistence degree defined, it is possible to test the predictions of theoretical models. In order to determine which of the previous theoretical models is relevant, one should take into account the whole list of explanatory variables included in a regression, since the degree of significance found for each determinant obviously depends on all the explanatory variables included in the analysis. Crépon and Duguet (1997) take into account few explicative variables (through R&D and a fixed effect) and their sample is limited to the significant R&D performers only. Moreover, they use patent data so that the robustness of their results needs to be examined on other data sources. If the persistence of innovation comes from the persistence of R&D only, there should be no systematic innovation difference between firms once controlled for R&D expenditures. Hence, it is possible to test this by regressing a measure of innovation on a measure of past innovation and a measure of R&D. The issue is then whether there is an additional effect of past innovation on present innovation.15 If R&D and past innovation are both significant, other models should be investigated. In the same way, it is possible to control for the differences of commercial success of past innovations and for the differences of strategic considerations before examining the relation between past and current innovations. We use the share of firms’ sales due to recent innovations in order to measure the commercial success of past innovations and a variable of market share to capture the strategic considerations. Henderson (1993) used the same type of variable to control for strategic effects. It is more difficult to test the role of dynamic increasing returns in the innovation persistence. We examine whether past innovations still have an effect on present innovation once R&D, commercial success of past innovations and strategic

10

considerations have been controlled for.

16

If the dynamic increasing returns play an

important role on innovation persistence, past innovations should indeed remain significant.

3.2 The data17 The data used in this paper come from five data sets, including three about innovation. The first data set is the Innovation Survey “L’Innovation Technologique dans l’Industrie” conducted by the SESSI that was performed in order to prepare all the other innovation surveys in France. It was conducted in 1991 and provides information about the period 1986-1990. In order to identify innovation persistence, we also use two Community Innovation Surveys (henceforth CIS).

18

The CIS1, conducted in 1993,

covers the period 1990-1992, while the CIS2, conducted in 1997, refers to the period 1994-1996. These data sets give information about the implementation of innovation at the firm level, without any reference to their commercial success or their patenting status: whether the firm has innovated during the period considered, the inputs that it used to innovate, and so on. Data indicating if a firm has innovated during a period are used to measure the innovation. These data allow for the separation of leadership, either commercial or innovative, from innovation itself. Then we take into account different inputs used by firms to innovate. The CIS0 distinguishes formal R&D according to the Frascatti criterion implemented by the firm or its group, informal R&D identified as “technical studies” as well as R&D coming from external sources. The answers are provided on a four-point scale. The CIS1 distinguishes R&D implemented by the firm, its group, an external public source or an external private source. In this survey, firms answer on a five-point scale. Lastly the share of sales due to recent innovations is used as a commercial performance indicator of past innovations. These data are available in CIS0 and CIS1 on a four-point scale. The remaining data sets provide accounting information. First, the annual industry census 1985 (in French, “Enquête Annuelle d’Entreprises”) provides data

15

Once the R&D differences are controlled for, the impact of past innovation on current innovation reflects the innovation persistence that is not explained by R&D. 16 The size and industry-level differences are also accounted for. 17 See also the appendix II. 18 These surveys have been conducted in many European countries.

11

19

about the size of firms, their main line of business and their sales.

Second, the

annual line-of-business data set on domestic sales 1985 (in French, “Enquête Annuelle d’Entreprises par Fractions”) gives information on sales of firms in each line of business where they operate. This allows computing a domestic market share of each firm in 1985 that takes into account the fact that some firms are diversified.20 The market power variable is then a weighted average of the different market shares of a firm.

21

The merger of these data sets provides a sample of 621 firms operating in manufacturing sector and covers the period 1986-1996.

22

Year 1993 is missing since

no innovation survey covers this year. Table 1 provides the percentage of innovators by industry. These percentages are not directly comparable since the first survey spans 5 years while the two other innovation surveys cover 3 years. The highest innovation rates are reached in equipment goods, electrical components, chemicals and pharmaceuticals.23 Table 1 also presents some persistence statistics evaluated by comparing two consecutive surveys or the three surveys available. We find that the most innovative sectors are also the ones where innovation is the most persistent. Considering the percentage of firms that have declared an innovation in the three surveys, we find that globally 45% of industrial firms persistently innovate. However, this average conceals strong differences between industries. While equipment goods, transport equipment

19

EAE is the « Enquête Annuelle d’Entreprises » (the industry census). It is compulsory for all firms above 20 employees. 20 We computed a market share variable for other years. We discuss in the following why we have not kept these last variables in our regressions. 21 This measure is presented in detail in Crépon, Duguet and Kabla (1996). 22 We imposed the presence of firms in each innovation survey and in the 1985 E.A.E. survey. We thus examine the innovation persistence conditional to the existence of the firms in 1985 and to their survival up to 1997. The integration of entry and exit issues of innovative firms in our analysis would require another data that are not available. Nevertheless, the impact of new entrants on the date at which the incumbents decide to innovate is taken into account. Moreover, the role of new innovative firms, in particular of startups, could be relatively minor in France. In a recent study, Arora et al. (2000) show that in Europe 90% of research projects in biotechnological sectors are due to large installed firms, whereas in United States more than 50% of the research projects are accompanied by the creation of a new firm. Chesbrough (1999) finds similar results in the semi-conductors industry. 23 Globally, the first survey reports 79% of innovators and the following ones 63% of innovators. These figures seem high but there are two reasons for it: they cover a three-year or a five-year period and the definition of innovation includes new products and new processes as well as innovations of different heights. The advantages of this innovation definition are that it does not refer to the market performance of the firms or to their patenting strategy and that it allows for measuring innovation in a firm that persistently innovates but in a different manner over time, for instance by introducing a new product and then by improving on its production process. The case studies confirm the view of the innovative process according to which the persistence originates from the diversity of innovative behaviors (Kim, 1997).

12

and electrical components and chemicals have a strong persistence (over 60%), the oldest industries have a very small one (wearing apparel, textile, printing and publishing, foundry). These differences may express a difference of technological opportunities. One objection to this conclusion is that a weak persistence simply translates a weak percentage of innovators, since the latter is the maximum of the persistence percentage by definition. Therefore, we have corrected the persistence figures by dividing the persistence percentage by the innovation percentage. This is what we call the relative innovation persistence in Table 1. We find that with this correction, the persistence ranking remains almost unchanged. Equipment goods rank first, followed by electrical components, transportation equipment, chemicals, pharmaceuticals and houseware. Innovation and its persistence may partly originate in the persistence of the innovation inputs like the research and development (R&D) expenditures. Table 2 provides information on the innovation inputs from CIS0 and CIS1. The most quoted source is internal R&D, either formal or informal. Then come group and external R&D. The relationship may take the form of R&D cooperation or other means of technology transfers.24 Looking at the first column of Table 2, we also see that the size matters. The median size, which is not influenced by outliers, is the largest for the firms operating in transportation equipment, electrical components, chemicals and pharmaceuticals. The ability of the firms in these industries to spread the sunk cost of innovation may partly explain both their innovation rate and the persistence of innovation. A look at the median market share offers a similar result with one difference: the mining industries have both a large median market share and low innovation rate. These descriptive statistics are interesting but they do not allow for disentangling the importance of the determinants of innovation persistence that have been highlighted in the literature. For this, we need to use econometric methods.

IV.

Methodology Let yi denote the current innovative performance of firm i and Ti denote his past

innovation.

25

Each firm has two different innovative performances depending on

whether it innovated in the past (Ti = 1) or not (Ti = 0). We denote them y i (1) and y i (0 ) .

24 25

On this issue, see Duguet and Mac Garvie (2003). The current innovative performances also correspond to a binary variable.

13

The effect of past innovation on the whole population, called the average causal effect in the statistical literature, is defined as c = E[y i (1) − y i (0 )] . If it was possible to observe the performances of each firm in both states 0 and 1, the average effect of past innovation could be estimated by the difference of the corresponding sample averages. Since such data are not available, one needs to construct a counterfactual, that is an estimation of y i (1) for the firms that did not innovate in the past and an estimation of

y i (0 ) for the firms that innovated. The most widespread method is to use a parametric model y(.) that explains the performance y i by past innovation Ti and the characteristics of the firm (denoted Xi ): y i = y(Ti , Xi ) . The counterfactuals are simply obtained by y i (0 ) = y(0, X i ) and y i (1) = y (1, X i ) .

4.1 The Logit model When the performances are binary variables (e.g., to innovate or not), the evaluation can be obtained from a Logit model, that explains the probability to innovate today by past innovation and the characteristics of the firm (size, market power, industry, R&D expenditures, etc.). It is given by:

y (Ti , X i ) = F[X iβ + γTi ] , where F is the cumulative distribution function of the logistic distribution. The average effect of past innovation is therefore estimated by:

cˆ =

1 N 1 N ˆ ˆ { ( ) ( ) } y 1 , X y 0 , X − = ∑ ∑ F Xiβˆ + γˆ − F Xiβˆ , i i N i =1 N i =1

{(

) ( )}

where βˆ and γˆ denote the maximum likelihood estimates from the Logit model and N 26

is the number of firms in the sample.

One of the main limitations of the Logit method is that it assumes that there is always a perfect counterfactual given by the parametric model, in other words it assumes there can exist a firm that did not innovate in the past and that has exactly the same characteristics than the firm that has innovated (or more precisely the same Xiβ ). Therefore this method could be invalidated when innovative firms have different characteristics Xi from the non-innovative firms. The basic argument why the firms are

26

The same computation can be done for any binary explanatory variable. It is thus possible to compare the importance of past innovation with the importance of other explanative variables, especially with R&D.

14

not comparable may simply be that innovative firms do more R&D or have a size that differs from the size of non-innovative firms.

27

In this case, it is possible that the Logit

method provides inconsistent estimates of the effect of past innovations.

4.2 The Rubin method With the Rubin method, the counterfactuals are not obtained from a parametric model, but from the actual data. This methodology was first proposed by Rubin (1974, 1977) and developed by Rosenbaum and Rubin (1983) as well as Heckman, Ichimura and Todd (1997) among others. The intuition is as follows. If we had an experimental sample, a direct comparison of the percentage of innovation between the two sets of firms, defined by Ti, would provide a consistent estimate of c.

28

This is because the

performances difference between past innovative and non-innovative firms could only come from past innovation and the empirical average would be a consistent estimator of the expected causal effect. But past innovation is not allocated at random between firms. Rubin (1974) showed that it is nevertheless possible to evaluate the effect c if the following condition is fulfilled:

(y i (0), y i (1)) ⊥ Ti Xi

(H-1)

where ⊥ denotes statistical independence. This implies that one can evaluate the counterfactuals by:

E( y i (1) X i , Ti = 0) = E( y i (1) X i , Ti = 1) = E( y i (1) Xi ) E( y i (0) X i , Ti = 1) = E( y i (0) X i , Ti = 0) = E( y i (0) X i ) The only practical problem with this method is that it implies to match firms on large number of variables Xi . Fortunately, Rosenbaum and Rubin (1983) have shown that this condition can be simplified to conditional independence on the onedimensional propensity score defined as Pr (Ti = 1 X i ) . This selection probability conveys the following useful intuition. Suppose that we have a group of firms with the same probability to have innovated in the past. Inside this group, there are firms that innovated in the past and firms that did not. Hence, the allocation of past innovation between these firms can be considered as random. The comparison of the average

27

For evidence, see Cohen and Levin (1989) and Kleinknecht ed. (1996) on European data.

28

In this literature, T i is called the treatment and the firms that innovated in the past the “treated”.

15

performances inside homogeneous probability groups is therefore relevant. More precisely, Rosenbaum and Rubin (1983) have shown that:

(y i (0), y i (1)) ⊥ Ti Xi ⇒ (y i (0), y i (1)) ⊥ Ti Pr (Ti = 1 Xi ) . In practice, the propensity score is replaced by its estimate. Following the literature on discriminant analysis, the propensity score can be estimated by a Logit or a Probit model with Xi as explanatory variables (see Maddala, 1983).

29

But the

propensity score may also depend on a firm-level fixed effect, which represents any 30

time-invariant factors that influence the innovative performances of the firms.

For

instance, this effect may represent firms that have research teams with more successful researchers. We denote this fixed effect by α i . The identification condition becomes:

(y i (0), y i (1)) ⊥ Ti (Xi , α i ) ⇒ (y i (0), y i (1)) ⊥ Ti Pr (Ti = 1 Xi , α i ) . The fixed effect raises an issue because it is unobservable. There are two ways to deal with this problem. The first method is to estimate a fixed-effect Logit model on panel data and to use the predicted probability to match firms. But, unfortunately, this is not possible with the Community Innovation Surveys for the two following reasons: - In the fixed effect Logit model, the estimation proceeds by the conditional maximum likelihood method, where the conditioning variable is the number of times that a firm innovates (i.e., the sum of the innovation dummies). This implies that two kinds of events must be excluded from the regression (Hsiao, 1986). First, one should exclude the firms that have always innovated and, secondly, one should exclude the firms that have never innovated. The reason why the always-innovators are excluded by this method is that the conditional probability to innovate knowing that one has always innovated is equal to one, so that it does not contribute to the conditional likelihood (i.e., provides no relevant information for the estimation). The reason why neverinnovators are excluded is that the conditional probability to innovate knowing one has never innovated is equal to zero. Hence the higher the persistence of innovation is, the less there will be firms available for the fixed-effect Logit regression.

29

Notice that this is not the same as the parametric model. In the Rubin approach, the treatment is explained by a logit model and the outcome of the treatment remains non-parametric. In the parametric model, there is no matching and the outcome is directly explained by a logit model. 30 Taking into account an individual fixed effect implies that the selection equally comes from unobservable variables. See Heckman et al. (1998) on this issue.

16

- In a Logit model with a fixed effect and two years of data, one needs to do a regression on the difference of the explanative variables (Hsiao, 1986). Unfortunately the definition of research inputs changes between surveys over time. Hence, there is no way to take the difference. Therefore, the applied researcher has no other choice than to use another method. The second method is to find out observable variables that are strongly correlated to the fixed effects so that the permanent differences between firms can be controlled for. Thus, we need variables that, for example, give information on the competencies of the research team of a firm. The most obvious set of variables is the past innovative history of the firm. If a firm has a successful research team, its innovation history should score better than the one of another firm. Other time persistent variables related to innovation performance can also be used. The first innovation survey is especially useful for this purpose since it provides information on the innovation history of the firms over the five-year period 1986-1990. The long length of inquiry of this data set is an advantage when trying to control for individual effects. Firms that did not innovate over five years may not have successful research teams, while firms that innovated at least once in five years may have more competent research teams. Another time-persistent variable available in this survey is firm’s line-of-business which reflects a specific degree of technological opportunities. We denote these two variables by Z i . Replacing the fixed effect formulation by its observable counterpart, we use the following condition:

(y i (0), y i (1)) ⊥ Ti Wi ⇒ (y i (0), y i (1)) ⊥ Ti Pr (Ti = 1 Wi ) where Wi = ( X i , Z i ) . We perform the matching on the corresponding estimated propensity score. Whichever the method used, the first precaution to take is to check that the supports of the propensity score have a sufficient overlap between the two groups of firms (i.e. treated and non-treated firms) (Dehejia and Wahba, 1998). Some firms can be excluded from the sample because there exists no relevant counterfactual. It is therefore not possible to evaluate a causal effect for these latter firms. It is an important difference with the Logit method which uses all the observations. A simple way to evaluate the causal effect is to use the Nadaraya-Watson nonparametric estimator (Härdle, 1990). For each treated firm, we compute the difference between its performance and a local weighted average of the performances of its non-

17

treated neighbors, where the weights decrease with the difference of the propensity score. The kernel estimator of E[ y i (0 ) |T=1] is defined as:

[

] ∑ ω j ×y j

∀i ∈ I0 , Eˆ y i (0) T = 1 =

j∈I0

where ω j =

[

]

K (p i − p j ) / h , K (p i − p j ) / h

∑ j∈I1 [

]

where p i is the propensity score of the treated firm i, p j the propensity score of the nontreated firm j, K(x) is a Gaussian kernel, h the window and I0 the set of firms that have 31

not innovated in the past.

The average causal effect on the treated is obtained by: cˆ 1 =

∑{

]}

[

1 y i (1) − Eˆ y i (0) T = 1 , N1 = card(I1 ) N1 i∈I 1

The same method is used to compute the global effect: cˆ =

1 N 0 + N1

  Eˆ y i (1) T = 0 − y i (0) + y i (1) − Eˆ y i (0) T = 1  i∈I0 i∈I1

∑( [

[

]

] ∑ η j ×y j ,

with ∀i ∈ I1 , Eˆ y i (1) T = 0 =

) ∑(

and η j =

j∈I1



])

[

[



K (p i − p j ) / h

]

∑ j∈I0 K[(p i − p j ) / h]

.

These estimators are asymptotically normal and their variances can be obtained by the bootstrap. The optimal number of bootstrap repetitions is computed according to the method of Andrews and Buchinsky (2000). In this application, this gives between 787 and 1007 simulations depending on the estimation performed. The propensity score is re-estimated at each draw and all the estimations are performed under SASIML.

V.

The results 5.1 The Logit models Table 3 presents the results of Logit models. Innovation 1994-96 is first

explained by size, a full set of industry dummies, innovation 1986-90 and innovation

31

In order to make this estimation we took a Silverman window. For more information about this method, see Härdle (1990).

18

1990-92 (model 1). The sector dummies should control for the differences of technological opportunities. This first regression allows evaluating the degree of innovation persistence in our sample. Then we extend the model in order to try to understand the origin of innovation persistence identified (model 2). Our comments focus on the reduced forms. In model (1), we find that size increases the probability to innovate. For the industry dummies, the differences of coefficients are likely to be correlated to the degree of technological opportunities. The strongest industry effects are found in equipment goods, electrical components, chemicals and houseware. In these activities both the scientific base (or technology push) and the demand conditions (or demand pull) are high enough to sustain a high level of innovation. We also find that lagged innovations are significant. The coefficient of lagged innovations decreases with the importance of the lag (0.8 for two years, 0.6 for four years), which suggests that some kind of knowledge depreciation takes place. The advantage of previously innovating firms decays progressively over time. This result suggests the existence of a strong entry barrier to innovation. An immediate question can be put forward: do these lagged innovation variables control in fact for the lagged R&D expenditures and the other innovation inputs, commercial success differences of past innovations or market share differences? Therefore, this first regression should be augmented with other variables. This extension is also presented in Table 3. The model (2) explains innovation output (1994-96) by innovation inputs available in the CIS0 and CIS1 surveys, the share of sales due to recent innovations, market share, cross product of market share and size as well as lagged innovations, size and the industry dummies. This regression provides several interesting results. First, we find that lagged innovation is significant but only on the short run. Indeed, only the first lag is significant. It means that a part of the innovation persistence found in the previous model comes from innovation inputs, commercial success of past innovations or strategic considerations. Second, the effect of size is confirmed. The market share is significant as well and their cross product has a positive effect on innovation. This means that a large firm with a high market share has a stronger probability to innovate than a large firm with a low market share. Hence, size and market power have a complementary effect on innovation performance. This argument is reinforced if we recall that industries with high sunk costs tend to generate a market structure with large market share (i.e., few

19

competitors). These variables are thus capturing something else than just R&D expenditures (itself strongly correlated to size). Third, only some innovation inputs are significant in the innovation regressions. A strong formal R&D (1986-90) increases the probability to innovate, as well as a moderate use of group R&D (1986-90). Therefore, for formal R&D, there seems to be a threshold over which innovation takes place, while for group R&D there would be a bell-shaped relationship, since both its lowest and highest levels are not significant while the moderate level is. The reason may simply be that above a high threshold of group R&D the firm does not perform the innovation itself but rather adopts it from another firm in the group. This type of change is not recorded as an innovation internal to the firm. Moreover, the results suggest that informal knowledge related to informal R&D would have a higher depreciation rate than formal knowledge. One explanation is that formal R&D can be more easily codified and transmitted to new researchers or engineers or simply that the knowledge generated by formal R&D is relevant for a wider scope of application than the knowledge generated by informal research. Four, the share of sales due to recent innovations 1992 has positive impact on current innovation as well as past innovation (1990-92), but the share of sales due to recent innovations in 1990 is not significant; this suggests that the financial advantage coming from commercial success of past innovations decays quickly over time. Following the theoretical literature, we can attribute the effect of lagged innovation to dynamic increasing returns in the production of innovations. The fact that only a short lag is significant goes well in line with this conclusion since it involves a continuity of the innovation practice inside the firm. What is its importance? Table 4 compares the average effect of lagged innovation, R&D and technological opportunities. We find that the strongest effects are not at the firm level but at the industry level. Clearly, what matters the most is the technological opportunities of the line of business. Firms operating in electrical components, electrical machinery and cars have a 16% to 24% higher average probability to innovate in absolute terms than the other firms (for the same values of size, market share, innovation inputs and lagged innovation). The second important determinants are, with almost the same importance, R&D (+8% for internal R&D and +13% for a moderate group R&D) and past innovation which is assimilated to dynamic increasing returns (+10%). Hence, the effect of dynamic increasing returns seems to be comparable, for the average firm, to the one of a strong formal R&D.

20

5.2 The Rubin model and the evaluation of the causal effect Another estimate of the impact of dynamic increasing returns is presented in Table 5. We use an entirely different method, the Rubin propensity score approach, to evaluate the effect of lagged innovation (90-92) on current innovation (94-96). We find that, on average, the effect is +13% which is stronger but compatible with the previous estimate of +10%. Therefore our result is robust to the estimation method. But we have also investigated another issue linked to dynamic increasing returns. Basically, this kind of knowledge acquisition may be more important in the small firms. In order to test this, we have separated our sample in three sets of equal size according to the 33% and 66% percentiles of the sales distribution, and performed three separate evaluations of the effect of lagged innovation. The small-sized class presents the strongest effect with +25% (p-value 0.016), the medium-sized class +11% (p-value 0.194) and the large-sized class +9% (p-value 0.348). The effect of lagged innovation is not significant in the large firms where innovation does therefore rely on internal and group R&D, along with size, market power and technological opportunities. The source of innovation persistence may therefore depend on the size of the firms. Our result suggests that the innovation persistence of these firms may be due to their easier access to finance. These results confirm the ones of the Logit model. Moreover, the averages’ difference is close to the different estimators, so that we do not find evidence of a severe selection bias. The origin of innovation persistence thus depends on the size of the firm. Consequently, the relevance of the different theoretical models depends on the characteristics of the firms. For the largest firms, the linear model provides a good approximation whereas for the smallest firms, a relevant model should include dynamic increasing returns. This last conclusion is close to the results of Kleinknecht (1987) and Kleinknecht and Reijnen (1991) that emphasize the inadequacy of R&D data to evaluate the innovative competencies in the small-sized firms.

21

VI.

Conclusion Our first finding is that the innovation persistence is strong. Indeed, using a data

set that combines three Innovation Surveys, we show that, ceteris paribus, a firm that already innovated in the past has a stronger probability to innovate today. This persistence has several origins: persistence of research activities, commercial success of past innovations, technological opportunities and probably dynamic increasing returns. Our second finding is that the origin of the persistence depends on the size of the firm. Whereas the dynamic increasing returns hypothesis seems to play a major role in the small-sized firms, its weight decreases with the size of the firms. In the largest firms, the linear model is relevant, since we do not find any significant direct effect of past innovation on current innovation. The innovation persistence is due to the formal research of these firms. Consequently, the relevance of the innovation models depends on the size of the firm, so that the dynamic increasing returns model and the linear model are not conflicting but complementary. Finally, our results suggest the following functioning of innovation: the importance of dynamic increasing returns should decrease with the formalization of research and development activities. Clearly, the last class includes mostly firms with the highest formal R&D budgets, so that the persistence of innovation for these firms comes from the persistence of research. In order to evaluate the degree of innovation persistence at the firm level, both effects must be accounted for. The fact to omit the dynamic increasing returns lead to underestimate the innovation persistence, especially in small-sized firms. One direction may prove particularly fruitful for further research. It would be particularly interesting to pursue the analysis of the dynamic incentives of firms to innovate by distinguishing their technological position. More precisely, it would consist in examining whether a technological leader is more or less incited to innovate more often than a technological laggard. This extension would connect this work with previous empirical papers since the patent data measure the technological leadership of a firm.

22

References Acs Z.J. and Audretsch D.B. (1988). Innovation in Large and Small Firms: an Empirical Analysis. American Economic Review, vol. 78 (4), pp.678-690. Aghion P. and Howitt P. (1998). Endogeneous Growth Theory, MIT Press. Aghion P. and Howitt P. (1992). A Model of Growth through Creative Destruction. Econometrica, vol. 60, pp.323-351. Aghion P., Harris C. and Vickers J. (1997). Competition and Growth with Step-by-step Innovation: An Example. European Economic Review, vol. 41, pp.771-782. Aghion P., Harris C., Howitt P. and Vickers J. (2001). Competition, Imitation and Growth with Step-by-Step Innovation. Review of Economic Studies, vol. 68, pp.467-492. Andrews D. and M. Buchinsky (2000). A Tthree-step Method for Choosing the Number of Bootstrap Repetitions.Econometrica, vol. 68(1), pp. 23-51. Arora A., Gambardella A., Pammolli F. and Riccaboni M. (2000). The Nature and the Extent of the Market for Technology in Biopharmaceuticals, CNRS-CEPR International conference on « Technology Policy and Innovation : Economical and Historical Perspectives », Paris, November. Arrow K.J. (1962). Economic Welfare and the Allocation of Resources for Inventions", In R. Nelson (ed.), The Rate and Direction of Inventive Activity: Economic and Social Factors, Princeton, N.J.: Princeton University Press. Arundel A. and Kabla I. (1998). What Percentage of Innovations are Patented? Empirical Estimates for European Firms, Research Policy, 27, pp.127-141. Audretsch D. and Elston J. (2002). Does firm size matter? Evidence on the impact of liquidity constraints on firm investment behavior in Germany. International Journal of Industrial Organization, 20(1), pp.1-17. Baldwin W.L. and Scott J.T. (1987). Market Structure and Technological Change, Chur, Switzerland: Harwood Academic Publishers. Barlet C., Duguet E., Encaoua D. and Pradel J. (1998). The Commercial Sucess of Innovations : An Econometric Analysis at the Firm Level, Annales d’Economie et de Statistique, 49-50, pp.457478. Cabagnols A. (2004). “Technological learning and the persistence of innovation: A France-UK comparison based on a Cox model of duration”. Working Paper, Research center Auguste et Léon Walras UMR CNRS 5056. To appear in the Revue d’Economie Politique. Cefis E. and Orsenigo L. (2001). The Persistence of Innovative Activities: A Cross-countries and Cross-sectors Comparative Analysis. Research Policy, 30, pp.1139-1158. Chandler A.D. (1990). Scale and Scope : the Dynamics of Industrial Capitalism, The Belknap Press of Harvard University Press : Cambridge, MA. Chesbrough H. (1999). Arrested Development : The Experience of European Hard Disk Drive Firms in comparison with US and Japanese Firms, Journal of Evolutionary Economics, 9, pp.287329.

23

Cohen W. and Klepper S. (1996). A reprise of size and R&D. Economic Journal, vol. 106, pp.925951. Cohen W. and Levin R. (1989). Empirical Studies of Innovation and Market Structure. In Schmalensee et al. (eds), The Handbook of Industrial Organization, volume II. Cohen, W.M. and Levinthal, D.A., 1989, Innovation and Learning: The Two faces of R&D, Economic Journal, 99 (3), 569-596 Cohen W., Nelson R. and Walsh J. (2000). Protecting their Intellectual Assets: Appropriability conditions and why firm patent and why they do not in the American manufacturing sector. NBER Working Paper, n°7552. Crépon B. and Duguet E. (1997). Estimating the knowlege production function from patent numbers: GMM on count panel data with multiplicative errors . Journal of Applied Econometrics, vol. 12(3). Crépon B., Duguet E. and Kabla I. (1996). Schumpeterian Conjectures : A Moderate Support from Various Innovation Measures. in Kleinknecht A. (ed.), Determinants of Innovation: The Message from new Indicators, Mc Million. Crépon B. and Iung N. (1999). Innovation, emploi et performance. INSEE working paper, DESE, ref. G9904. Dehejia R.H. and Wahba S. (1998). Causal Effects in Non-experimental Studies : Re-evaluating the Evaluation of Training Programs. NBER Working Paper Series, 6586. Denicolo V. (2001). Growth with non-drastic innovations and the persistence of the leadership. European Economic Review, 45, 1399-1413. Duguet E. and Kabla I. (1998). Appropriation strategy and the motivations to use the patent system in France : an econometric analysis at the firm level. Annales d’Economie et de Statistique, 49-50 (Special issue on the economics and econometrics of innovation). Duguet, E. and M. MacGarvie (2003). “How well do patent citations measure flows of technology? evidence from French innovation surveys”. To appear in Economics of Innovation and New Technology. Working paper, Université de Paris I, Cahiers de la MSE n°2003(99). Available at ftp://mse.univ-paris1.fr/pub/mse/cahiers2003/V03099.pdf and at http://www.ssrn.com. Duguet E. and S. Monjon (2002). Creative destruction and the innovative core: is innovation persistent at the firm level?. University College London, Discussion Paper n°02-07. (http://www.econ.ucl.ac.uk/papers/working_paper_series/0207.pdf). Efron B. and Tibshirani R.J. (1993). An Introduction to the Bootstrap, Monographs on Statistics and Applied Probability 57, Chapman e& Hall, New York. Encaoua D. and Hollander A. (2002). Competition Policy and Innovation, Oxford Economic Review, forthcoming. Encaoua D. and Ulph D. (2000). Catching-up and Leapfrogging? The Effects of Competition on Innovation and Growth. Working Paper, EUREQua, University Paris 1. François J.-P. (1991). Une enquête sur l’innovation. Courrier des Statistiques, n°57. François J.-P and Favre F. (1998). L’innovation technologique progresse dans l’industrie. Le 4 Pages du SESSI, n°89.

24

Geroski P. (1989). Entry, Innovation and Productivity Growth, Review of Economics and Statistics, 71, pp.572-578. Geroski P., Van Reenen J. and Walters C. (1997). How Persistently do firms innovate? Research Policy, 26, pp.33-48. Gilbert R.J. and Newberry D.M. (1982). Preemptive Patenting and the Persistence of Monopoly", American Economic Review, 72, 514-526. Griliches Z. (1990). Patents Statistics as Economic Indicators : a Survey, Journal of Economic Litterature, 28, pp.1661-1707. Gruber H. (1992). Persistence of Leadership in Product Innovation. Journal of Industrial Economics, 40 (4), 359-375. Hall B. (1992). Investment and Research and Development at the Firm Level: Does the Source of Financing Matter?. NBER Working Paper, n°4096. Hall B., Griliches Z. and Hausman J. (1986). Patents and R and D: Is There a Lag?. International Economic Review, 27, pp.265-283. Hall B. and Van Reenen J. (2000). How effective are fiscal incentives for R&D ? A review of the evidence. Research Policy, vol. 29, pp.449-469. Härdle W. (1990). Applied non-parametric regression. Econometric Society Monograph, Cambridge University Press. Heckman J., LaLonde R. and Smith J. (1998). The Economics and Econometrics of Active Labor Market Programms. Working Paper prepared for the Handbook of Labor Economics, III, Ashenfelter O. and Card D. editors. Heckman J., Ichimura H. and Todd P. (1997). Matching as an Econometric Evaluation Estimator : Evidence from Evaluating a Job Training Programme. Review of Economic Studies, 39 (1), pp.33-38. Henderson R. (1993). Underinvestment and Incompetence as Responses to Radical Innovation: Evidence from the Photolithographic Alignment Equipment Industry. Rand Journal of Economics, 24(2), pp.248-270. Henderson R. and Cockburn I. (1996). Scale, Scope and Spillovers: The Determinants of Research Productivity in Drug Discovery. Rand Journal of Economics, 27(1), pp.32-59. Himmelberg C. and Petersen B. (1994). R&D and Internal Finance: A Panel Study of Small Firms in High-Tech Industries. Review of Economics and Statistics, 76, pp.984-1001. Hsiao C. (1986). Analysis of Patent Data, New York: Cambridge University Press. Kim L. (1997). The Dynamics of Samsung's Technological Learning in Semiconductors. California Management review, 39(3), spring, pp.86-100. Kleinknecht A. (1987). Measuring R&D in Small Firms : How Much are we Missing?, The Journal of Industrial Economics, 36 (2), december, pp.253-256. Kleinknecht A. (ed.) (1996). Determinants of Innovation: The Message from new Indicators, Mc Million. Kleinknecht, A. and J. Reijnen (1991). More evidence on the undercounting of small firm R&D. Research Policy, Vol. 20, pp. 579-587.

25

Klette T.J. (1996). R&D, Scope Economies, and Plant Performance. Rand Journal of Economics, 27 (3), pp.502-522. Levin R., Klevorick A., Nelson R. and S. Winter (1987). Appropriating the returns from research and development. Brookings Papers on Economic Activity, 3, pp.783-831. Lhuillery S. (1995). L’innovation technologique dans l’industrie. Le 4 Pages du SESSI, n°46. Maddala G. S. (1983). Limited dependent and qualitative variables in econometrics. Econometric Society Monograph n°3, Cambridge University Press. Malerba F. and Orsenigo L. (1999). Technological Entry, Exit and Survival : An Empirical Analysis of Patent Data. Research Policy, 28, pp.643-660. Nelson R. and Winter S. (1982). An Evolutionary Theory of Economic Change, The Bellknapp Press of Harvard University Press, Cambridge, MA. Nightingale P. (2000). Economies of Scale in Experimentation: Knowledge and Technology in Pharmaceutical R&D. Industrial and Corporate Change, 9(2), pp.315-359. Reinganum J. (1983). Uncertain Innovation and the Persistence of the Monopoly, American Economic Review, 73, 741-748. Romer P. M. (1990). Endogenous Technological Change, Journal of Political Economy, 98, pp.S71-S102. Rosenberg N. (1976). Perspectives on Technology, Cambridge UP. Rosenbaum P. and Rubin D. (1983). The Central Role of the Propensity Score in Observational Studies for Causal Effects. Biometrika, 70, pp.41-55. Rubin D. (1974). Estimating Causal Effects of Treatments in Randomized and Non-Randomized Studies. Journal of Educational Psychology, 66, pp.688-701. Rubin D. (1977). Assignment to a Treatment Group on the Basis of a Covariate. Journal of Educational Statistics, 2, pp.1-26. Scotchmer S. (1999). Cumulative Innovation in Theory and Practice. GSPP Working Paper 140, February (under revision for Journal of Economic Literature).

Stein J. C. (1997). Waves of Creative Destruction: Firm-Specific Learning-by-Doing and the Dynamics of Innovation. Review of Economic Studies,64, pp.265-288. White H. (1980). A Heteroskedasticity Consistent Covariance Matrix Estimator and a Direct Test of Heteroskedasticity, Econometrica, 48, pp.817-838.

26

Table 1: Innovation and persistence statistics Sample of 621 firms in French manufacturing, with 20 employees or more. Industry Number Innovation % Of firms CIS0 CIS1

CIS2

Innovation persistence % CIS0/CIS1 CIS1/CIS2 All CIS

Rank

Relative innovation persistence % CIS0/CIS1 CIS1/CIS2 All CIS Rank

C1: Wearing apparel and leather C2: Printing and publishing C3: Pharmaceuticals C4: Houseware D0+E1: Transport equipment E2: Non electrical machinery E3: Electrical machinery F1: Ore mining and glass F2: Textile F3: Wood and paper F4: Chemicals, rubber, plastic F5: Foundry and metalworking F6: Electrical components

46 35 34 45 42 91 33 43 34 39 65 83 31

41 40 91 84 90 85 97 79 71 79 92 73 94

33 34 65 90 79 80 88 58 53 44 77 53 84

26 29 71 71 83 75 88 56 50 51 77 52 90

24 14 62 53 76 76 88 51 47 41 74 46 77

23 14 62 53 76 76 88 51 47 41 74 46 77

7 0 50 44 67 66 82 37 29 33 60 30 70

12 13 6 7 3 4 1 8 11 9 5 10 2

59 36 68 63 84 90 91 65 67 52 80 62 83

40 17 77 81 88 85 93 72 61 76 82 68 92

16 0 55 53 74 78 84 47 42 42 65 41 76

12 13 6 7 4 2 1 8 10* 9* 5 11 3

Total

621

79

63

63

57

57

45

×

73

77

57

×

* The ranking is determined from all decimals.

27

Table 2: Innovation input statistics Sample of 621 firms in French manufacturing, with 20 employees or more. Sales*

Market share %

Innovative products** CIS0 %

Innovative products** CIS1 %

Research and Development CIS0:

Research and Development CIS1:

Percentage of firms declaring a strong use of the input indicated.

Percentage of firms declaring a strong use of the input indicated.

median

Median

mean

mean

Internal Formal

Group

Internal Informal

External

Internal

Group

External Public

External Private

C1: Wearing apparel, leather C2: Printing and publishing C3: Pharmaceuticals C4: Houseware D0+E1: transport equipment E2: Non electrical machinery E3: Electrical machinery F1: Ore mining and glass F2: Textile F3: Wood and paper F4: Chemicals, rubber, plastic F5: Foundry, metalworking F6: Electrical components

5 28 108 27 156 36 62 35 10 10 147 72 131

0.42 0.60 0.93 1.09 3.44 1.08 1.56 4.05 1.43 0.67 1.70 0.54 1.79

12.1 16.8 27.6 22.0 25.0 24.5 35.6 12.0 16.2 17.6 20.8 12.4 25.5

14.5 18.3 23.9 22.5 28.3 21.8 32.7 12.5 20.0 18.1 14.2 15.9 22.3

4 9 53 27 43 40 67 30 24 18 49 28 52

2 6 62 11 31 13 33 26 3 5 34 7 13

17 14 29 31 45 37 45 21 26 18 28 25 42

7 3 3 4 5 8 3 0 0 8 5 5 0

9 14 29 24 36 33 42 23 29 15 32 20 45

2 0 21 4 24 13 27 19 6 3 22 7 23

0 0 12 7 14 4 6 0 3 3 9 4 3

9 9 12 11 31 25 27 16 12 8 26 14 10

Total

34

1.09

20.9

20.6

34

18

29

4

27

13

5

17

*Millions of Euros ** The average is taken on innovative firms only.

28

Table 3: Logit Regressions Left-hand variable: Implementation of an innovation between 1994 and 1996 (0/1). Maximum likelihood estimation of the Logit model. In what follows, M indicates a moderate level and S a strong level. Model (1)

Coeff. Intercept Innovation 90-92 Innovation 86-90 CIS0:Formal RD-M CIS0:Formal RD-S CIS0:Group RD-M CIS0:Group RD-S CIS0:Informal RD-M CIS0:Informal RD-S CIS0:External RD-M CIS0:External RD-S CIS1:Internal RD-M CIS1:Internal RD-S CIS1:Group RD-M CIS1:Group RD-S CIS1:Public external RD-M CIS1:Public external RD-S CIS1:Private external RD-M CIS1:Private external RD-S %innovative products 1992 %innovative products 1990 ln(market share) ln(sales) ln(sales)×ln(market share) C2: printing and publishing C3: pharmaceuticals C4: houseware D0: cars E1: other transport equip. E2: non-electrical machin. E3: electrical machinery F1: ore mining and glass F2: textile F3: wood and paper F4: chemicals rubber plastic F5: foundry, metalworking F6: electrical components % Concordant predictions

Backward selection on (1)

Std error Coeff.

-1.0110 0.7914 0.3886

Std error Coeff.

0.4144 -0.6618 0.2168 0.8335 0.2591 0.5959

0.4755

0.0690 0.4656

-0.4339 0.6428 1.4896 1.3480 0.8838 1.3099 2.1506 0.5114 0.6861 0.6806 1.2398 0.7883 2.1039

0.5531 0.5630 0.5263 0.7007 0.8914 0.4635 0.6936 0.5196 0.5303 0.5226 0.4994 0.4515 0.7397

82.6

Model (2)

0.9450 0.7614 1.5722

0.6782 1.5335

81.7

29

Backward selection on (2)

Std error Coeff.

0.2300 -1.2304 0.2134 0.6793 0.2482 0.4044 0.1419 0.3900 0.7901 0.0498 -0.3492 -01468 -0.2193 -0.3691 -0.1047 -0.4730 -0.0528 0.1421 0.2976 0.1770 -0.0691 -0.0060 1.7870 0.2273 0.2120 0.0635 0.2895 0.0899 -0.1744 0.7461 0.3910 1.4438 1.4058 0.3102 0.2983 1.1491 0.5961 1.9171 0.1448 0.5038 0.6698 0.3472 1.0972 0.6765 0.6444 2.0625

84.1

Std error

0.4353 -0.6716 0.3576 0.5621 0.3255 0.3423 0.3102 0.4934 0.4048 0.8485 0.3762 0.3138 0.3051 0.3309 0.5458 0.4054 0.3424 0.3915 0.4408 0.3448 0.6068 0.3574 0.4051 0.9775 2.0418 0.7440 0.1157 0.1746 0.1260 0.3089 0.0408 0.3089 0.5725 0.6412 0.5390 1.0514 0.7413 1.0421 0.9448 0.4809 0.7592 0.7227 1.5095 0.5580 0.5536 0.5294 0.5248 0.7403 0.4582 0.7732 1.6979

83.2

0.1929 0.2449 0.2592 0.3771

0.9014 0.0988 0.1038 0.1038 0.3944 0.6201 0.3075 0.6109

0.3572 0.6655

Table 4: Average effects from the Logit model Difference of averages probability to innovate depending on the indicated dummy variable is set to 1 or to 0. We have kept the significant variables from the Logit regression only (see Table 3). In what follows, M indicates a moderate level and S a strong level.

Innovation 90-92 Innovation 86-90 CIS0:Formal RD-S CIS0:Group RD-M C4: houseware D0: cars E2: non-electrical machinery E3: electrical machinery F4: chemicals rubber plastic F6: electrical components

Backward selection on (1)

Backward selection on (2)

0.15142 0.10538

0.09609

0.14389 0.12098 0.22109 0.10740 0.21707

30

0.08115 0.13190 0.15813 0.15657 0.11958 0.21561 0.11572 0.23652

Table 5: Causal effect of innovation 90-92 on innovation 94-96 Nadaraya-Watson estimator of the causal effect, with a Gaussian kernel and a Silverman window. The optimal number of bootstrap repetitions is determined according to the method of Andrews and Buchinsky (2000) for standard errors. We use τ = 0.05 and pdb = 5 . The causal effect equals the difference between the probability that a firm that has innovated over 199092 innovates over 1994-96 and the probability that a firm that has not innovated over 1990-92 innovates over 1994-96. The size classes are determined by the 33rd and the 67th percentiles. They are computed on the sales of the firms in 1985. Model A includes lagged endogenous variables and model B is the reduced form of model A (see the appendix I). Propensity score (see the appendix)

Model A

Model B

Size classes :

All

Small

Medium

Large

All

Small

Medium

Large

Optimal number of bootstrap repetitions

897

824

787

1007

804

832

800

866

Effect on the treated

0.1270

0.2482

0.1145

0.0851

0.1604

0.2637

0.1754

0.0901

Std error p-value

0.0524 0.0153

0.1032 0.0162

0.0880 0.1936

0.0907 0.3478

0.0514 0.0018

0.1087 0.0153

0.0981 0.0739

0.0845 0.2867

Effect on the non-treated

0.1970

0.2808

0.1659

0.0533

0.2118

0.2431

0.2118

0.0773

Std error p-value

0.0578 0.0007

0.0964 0.0036

0.0914 0.0696

0.1155 0.6445

0.0550 0.0001

0.0940 0.0097

0.0859 0.0136

0.0886 0.3826

Global effect

0.1556

0.2675

0.1330

0.0777

0.1810

0.2519

0.1886

0.0871

Std error p-value

0.0487 0.0014

0.0895 0.0028

0.0832 0.1099

0.0896 0.3860

0.0473 0.0001

0.0910 0.0057

0.0856 0.0275

0.0831 0.2947

31

Appendix I: Estimation of the propensity score Left-hand variable: Implementation of an innovation between 1990 and 1992 (0/1). This variable is the “treatment” variable of the Rubin method. Maximum likelihood estimation of the Logit model. In what follows, M denotes a moderate level and S a strong level. Two variants of the model are estimated. The first one uses the lagged dependent variables (innovation and the share of sales made with new product): this is model A. The second variant (model B) is the reduced form of the first model and aims to avoid any simultaneity problem. These two representations of the propensity score lead to the same estimates of the causal effect so that the simultaneity problem can be ruled out.

Model A

Intercept ln(market share) ln(sales) ln(sales)´ln(market share) Innovation 86-90 %innovative products 1990 CIS0:Formal RD-M CIS0:Formal RD-S CIS0:Group RD-M CIS0:Group RD-S CIS0:Informal RD-M CIS0:Informal RD-S CIS0:External RD-M CIS0:External RD-S C2: printing and publishing C3: pharmaceuticals C4: houseware D0: cars E1: other transport equip. E2: non-electrical machinery E3: electrical machinery F1: ore mining and glass F2: textile F3: wood and paper F4: chemicals rubber plastic F5: foundry, metalworking F6: electrical components

Backward selection on model A

Model B

Backward selection on model B

Coeff

Std error

Coeff

Std error

Coeff

Std error

Coeff

Std error

-1.0303 -0.0610 0.4547 0.0282 0.6587 1.6358 0.3637 0.7339 -0.1834 -0.9268 0.1646 0.3747 0.3547 1.6636 -0.2598 0.1620 0.4474 0.5672 0.6136 1.3079 1.4577 0.5654 0.5290 -0.2843 0.9322 0.4714 0.9867

0.4231 0.1108 0.1239 0.0363 0.3202 0.7398 0.3295 0.2953 0.3688 0.3473 0.2916 0.2885 0.3292 0.8015 0.5586 0.6058 0.5218 0.6586 0.9403 0.4850 0.7071 0.5333 0.5436 0.5413 0.5165 0.4580 0.6633

-0.5198

0.2344

0.1620

0.0679

0.3966 0.1082 0.1217 0.0351

-0.3219

0.4213

-0.7304 -0.1145 0.5002 0.0180

0.3852

0.0672

1.0003 1.7039

0.2718 0.7164 0.5001 0.9578 0.0171 -0.6750 0.4641 0.7177 0.4458 1.7377 -0.3165 0.3086 0.7004 0.7434 0.6395 1.4657 1.7963 0.6919 0.7001 0.0409 1.1121 0.5591 1.1477

0.3229 0.2850 0.3652 0.3386 0.2705 0.2653 0.3310 0.8119 0.5477 0.5945 0.5062 0.6432 0.8998 0.4691 0.6971 0.5169 0.5281 0.5190 0.5017 0.4425 0.6478

0.5551 1.0788

0.3133 0.2775

-0.6199 0.6329 0.8760

0.3073 0.2586 0.2550

1.4500

0.7895

1.0644 1.2577

0.3155 0.5986

0.6877

0.3487

0.7523

0.2717

-0.9550

0.3120

1.5126 -0.7107

0.7737 0.4183

0.8882 1.0324

0.3209 0.5984

-0.8608

0.3999

32

Appendix II: Supplementary information on the data The innovation surveys The first innovation survey in France, namely “l'innovation technologique dans l'industrie”, was conducted in 1991. We refer to it as CIS0, since it was made to prepare the CIS surveys. The firms were asked to report retrospectively over the 1985-1990 period. Hence, the choice of the respondent was an important issue. Here intervenes the SESSI (Industrial Statistics Bureau of the Ministry of Industry) which is responsible of the Industry Census and of all innovation surveys in France (and more surveys). The basic organization is as follows: inside SESSI the same person always works with the same set of firms. A part of his (or her) job is to find the right interlocutor inside the firm. On each questionnaire appear the name and the phone number of the SESSI correspondent inside the firm. Here the correspondent (which is an employee of the firm) has to send the questionnaire to “a person responsible of innovation, development, strategy issues or to the boss himself” (literal translation). The name of the respondent, that can be different from the name of the correspondent, and its phone number, have to be systematically reported on the questionnaire. The respondent has a SESSI phone number he (or she) can use to have explanations on how to reply to the survey. The census file is used for the mailing that prints automatically the name of the correspondents on the questionnaire itself etc. In other words, this survey has been conducted by an administration that has for main purpose to collect data among industrial firms. The survey was presented as an appendix to the Industry Census, which is compulsory. While the Census was compulsory, the appendix was not, but it was not indicated on the questionnaire so that the firms could have believed that it was compulsory. This is likely to be the case since the response rate to the innovation survey is the same as the one of the industry census (85% for compulsory surveys in France). The possibility of a response bias is systematically studied by the specialists of SESSI, for all the surveys. They compute the response rate after the “first wave” of the survey for each size class and each industry in order to detect abnormal response rates (e.g., below 85%). When the questionnaire does not come back, they can launch a second wave. Last but not least, all French firms have a compulsory national identification number that is called the SIREN code. The use of this code is compulsory for all the relationship that a firm has with the administration (including taxes). Its main advantage is that it allows for matching all the surveys without loss for identification reasons.

33

The first innovation survey is linked to the Community Innovation Survey (CIS) since it was designed to prepare the future CIS. The information collected is made up of answers over the 1985-1990 period, so that no annual information is available. This survey provides information on the type of innovation that industrial firms have implemented, as well as which knowledge sources they have used to achieve these innovations. We have qualitative information (yes/no) about eight types of innovation including the five following types of innovation: Significant improvement on an existing product. Introduction of a product that is new for the firm and for the market; Introduction of a product that is new for the firm but not for the market; Significant improvement on an existing process; Process technological breakthrough (“Première de procédé technologique”). A firm that has performed at least one of these five types of innovation is considered as innovative. We take this definition because it corresponds to the one used in the CIS surveys (product or process innovation).

32

The questionnaire then turns to the sources of these

innovations. The questionnaire design clearly indicates the causality in the following way (literal translation) “Sources of innovations: in your firm, does the introduction of innovation result from:” and then comes the list of the innovation inputs. The importance of innovation sources is available on a four-point scale. The scale is: unimportant, weakly important, moderately important and very important. The two inputs used in this paper are formal and informal R&D: Formal R&D is defined as internal research and development with at least one full-time employee. This is the definition that is used for the R&D survey and it is the closest to the traditional R&D studies (i.e., the Frascatti criterion). Informal R&D is defined as “internal method and technical studies”. The interest of this measure is to avoid the undercounting of research in small firms, a common feature of most databases (Kleinknecht, 1987; Kleinknecht and Reijnen, 1991).

CIS 1 and CIS2 CIS 1 is the first international survey on innovation. It was also conducted by SESSI and reached the same response rate as the first innovation survey. Notice that we do not use the micro-aggregated version of the survey but the original French survey in which information is available at the firm level. Fewer firms were surveyed than in the original innovation survey of 1991 (see below). It provides information about the implementation of product and process

32

The remaining three types of the first survey are: organizational innovation, marketing innovation and packaging innovation.

34

innovation and on the innovation inputs. In order to keep comparable specifications across time, we have kept the formal and informal research inputs that are grouped together in only one question. The answer is on a five points scale: “not important”, “weakly important”, “moderately important”, “important” and “very important”. CIS2 is the second international survey on innovation. It was conducted in the same conditions as the two other surveys and reaches the same response rate. It includes much information but we just keep the implementation of a product or a process innovation. For more information about the data sources, see François (1991), Lhuillery (1995) and Favre and François (1998).

The sample Our sample results from the merger of these three innovation surveys and of the industry census in 1985. We impose the presence in 1985 because the first survey questionnaire refers to the period 1986-1990. The sample includes 621 firms. Even though merging is easily done with the SIREN code, we lose two types of firms by performing this operation. The first firms we lose are the firms that were not included in all the surveys. This allocation is random, so that it should not be a source of bias. The second firms we lose are the firms that did not survive the whole period. Here, we should check that our data are representative of the exit rates of the industry. Since there are both innovative and non-innovative firms in our sample, we should not expect that our firms survive longer. The appendix tables give the detail of the constitution of the sample and compare the entry and exit rates of the innovation survey with the ones of the whole industry. It clearly shows the innovation surveys are representative of the cohort of firms that were present in manufacturing in 1985. It is especially true of the first survey since all the firms in the CIS1 survey were respondents to the first innovation survey of 1991. The CIS2 survey, on the contrary, includes a large number of firms newly included in the survey. Globally, we can say that our sample is representative of the exit rate of the industry but that it exhibits a smaller entry rate. Therefore, we should interpret our results as valid for the cohort of manufacturing firms in 1985. The merger of all files gives 621 firms. This small figure comes mostly from the variation of the sampling of the two last innovation surveys (CIS1 and CIS2) since the exit rates are about the same in the census and in the CIS. Therefore, our sample is representative of the firms that belong to the cohort of 1985.

35

Table A.5 - Details on the Merger of the Innovation surveys (2 by 2)

File 1

CIS0

CIS0

CIS1

Number of firms in file 1 (A)

15498

15498

3342

File 2

CIS1

CIS2

CIS2

Remaining number of firms after the merger of file 1 and file 2

3329

3882

3881

0 2 2

821 558 1379

2840 284 3124

Explanation of the variation of the number of firms after the merger of file 1 and file 2: Entry in the survey (from file 1 to file 2) - old business included in file 2 - new business included in file 2 - total Exit from the survey (from file 1 to file 2) - by bankruptcy between the two surveys - by exclusion from the second survey (these firms are still in the census) - total

2626 9545

5330 7665

693 1892

12171

12995

2585

Entry in the survey – Exit from the survey (B)

-12169

-11616

+539

Remaining firms after the merger of file 1 and file 2 (A) + (B)

3329

3882

3881

Annualized Entry and exit rates (corrected for pure sampling movements): Exit rate from the census Exit rate from the CIS Entry rate in the census Entry rate in the CIS

10.3% 8.1% 6.6%