A
N
E
XAMINATION OF THE
E
FFECTS AND
C
OMPLEMENTARITIES OF
I
NTERNAL
AND
E
XTERNAL
R&D
ON
S
ALES
G
ROWTH IN
I
NNOVATING
EU
F
IRMS
MSC. INTERNATIONAL ECONOMICS AND BUSINESS
AUTHOR: H.M. TEN LOHUIS
T
ABLE OF
C
ONTENTS
I. Abstract ... 3
II. Introduction ... 4
III. Literature Review ... 4
III.1 Introduction to R&D ... 4
III.2 The Effectiveness of R&D ... 5
III.3 Outsourcing and R&D Outsourcing ... 7
III.4 Complexity of the Innovation Process ... 8
IV. Data and Methodology ... 10
IV.1 Methodology ... 10
IV.1.1 Primary Regression Equation ... 10
IV.1.2 Control Variables ... 12
IV.1.3 Regression Analyses Using Subsamples ... 12
IV.1.4 Quantile Regression ... 14
IV.1.5 Testing Robustness ... 17
IV.1.5.1 Median Regression ... 17
IV.1.5.2 Robust Regression ... 17
IV.1.5.3 Removing Outliers ... 17
IV.2 Data ... 18
IV.2.1 Variables ... 18
IV.2.2 Output ... 19
IV.2.3 R&D – Internal and External ... 20
IV.2.4 Interaction Term ... 20
IV.2.5 Control Variables ... 21
V. Data Analysis ... 22
V.1 Descriptive Statistics ... 22
V.2 Statistical Tests ... 26
V.3 Initial Regression Results ... 28
V.4 Adding Control Variables ... 29
V.5 More Advanced and Less Advanced Economies ... 29
V.6 High-tech and Low-tech Industries ... 30
V.7 Small, Medium and Large Firms ... 32
V.8 Quantile Regression ... 33
V.9 Robustness Analyses ... 34
VI. Conclusion & Evaluation ... 37
VII. Bibliography ... 38
Appendix A – Overview of Tables and Figures ... 42
I.
A
BSTRACT
This study examines the effects and the complementarities of investments in internal R&D and outsourced R&D on sales growth in more than 7,000 innovating firms in the manufacturing industry across 16 EU countries. The micro-level data is combined with industry-level control variables and examined using an array of empirical methods and subsamples. Past theoretical arguments and empirical studies suggest that investments in R&D promote sales growth and that internal and external knowledge sources are complementary. However, the data, combined with the methods used here do not provide consistent support for the hypotheses. Internal R&D has a negative effect on sales growth for almost all subsamples, whereas the results for external R&D are more ambiguous. Support is found for complementarity between internal R&D and external R&D in more advanced countries, medium-low-tech firms and in small firms. Finally, quantile regression shows that the generally negative effects on sales growth become positive (but insignificant) in top-performing firms.
II. I
NTRODUCTION
In much of the traditional outsourcing literature, it is often argued that firms can outsource segments of their value chain when there are low levels of asset specificity, risk and transaction frequency (Williamson, 1981; Riordan & Williamson, 1985). Furthermore, more recent literature on outsourcing, based on knowledge-based and resource-based theories on outsourcing, argue that firms will only outsource the less fundamental parts of their value chains (Kogut & Zander, 1992; Grant, 1996). However, as many authors have shown in recent years (Howells, 1999; Mol, 2005; EARTO, 2007), firms have decided to outsource R&D at increasing rates. Relational views on outsourcing (Dyer & Singh, 1998) argue that firms can work closely with suppliers, buyers and other partners to jointly innovate.
Beyond the various theories on outsourcing, the outsourcing of R&D appears to be playing an increasingly large role in innovation strategies. Several authors (Kleinknecht & Reijnen, 1991; Mol, 2005; Martínez-Noya & García-Canal, 2011) study what drives firms to outsource and even offshore R&D and under what conditions they may choose to do so. James Quinn (2000) even considers the outsourcing of innovation to be the “new engine of growth” and discusses many of the benefits of outsourcing R&D.
The effect of R&D on firm performance has been studied extensively since the 1960s. Many of these studies have found support for the relationship between R&D investments and sales (revenues) growth with increasingly large data sets and complex empirical methods ( Nolan, Oppenheim, & Withers, 1980; Lachenmaier &
Rottmann, 2011) However, as the complexity and the role of outsourced R&D in the innovation process increase, the following question is raised: How effective is outsourced R&D relative to internal R&D in promoting sales growth and to what extent are both sources complementary?
A relatively new dataset from the Community Innovation Statistics (Eurostat) presents micro-level data for an array of EU-member countries, which allows for the differentiation between internal and external investments in R&D in order to examine the question above for the manufacturing sector. This study uses a regression model with independent variables for internal and external R&D intensities, combined with an interaction term including internal and external R&D intensities, and industry control variables to estimate the effects on sales growth.
This study begins by introducing theoretical arguments for the effect of R&D on sales growth. Theoretical arguments are complemented by past empirical findings related to R&D, the outsourcing of R&D, and the complexities and complementarities of innovation in section II. Section IV presents the empirical model, data sources, variables and econometric assumptions. The empirical findings are discussed relative to the previous literature and the hypotheses in section V. Finally, section VI summarizes and concludes the findings of this study and suggests future areas for improvement.
III. L
ITERATURE
R
EVIEW
There are three main bodies of literature that are relevant to this paper. First, much of the original research on the effect of innovation or knowledge on performance began with studies relating R&D investments to sales growth. Other indicators of firm performance include productivity or employment growth, but this paper will focus on studies related to sales growth. The second body of literature studies the role of outsourcing. Outsourcing is a widely studied topic ranging throughout strategic management, business and economics. Finally, recent literature on R&D uses more detailed and larger data sets to study to study the complexities of R&D and the
complementarity of knowledge sources.
III.1
I
NTRODUCTION TO
R&D
virtually all research on the effect of technical knowledge uses R&D investments as a proxy for technical knowledge. Thus in the remainder of this paper, the term R&D is a synonym for technical knowledge and investments in technical knowledge. However, this is generally an imperfect proxy for knowledge, as R&D has many intrinsic uncertainties at both the technical and market levels in the process from initial idea to market success (Kafouros, 2008). R&D, for example, does not include the potential of firms learning from each other (spillover effects). Several fundamental matters of R&D will be introduced here.
First, there are various types of R&D, ranging throughout the ‘value chain of R&D’. The OECD has identified three main types (OECD, 2002):
Basic Research: general scientific research without a direct future purpose
Process R&D: R&D related to improving the production process (quality or efficiency)
Product R&D: R&D related to the improvement of existing products or the development of new products It is clear that all three types of R&D are different in nature. For example, basic research may take much longer to develop and may never make it to a final product. Alternatively, process R&D requires a firm’s innovation processes to be directed in a different direction than for product R&D. The complete “innovation value chain”, from basic R&D to technical success, is undoubtedly a complex process, without any guarantees. This is reflected in the innovation literature, which now includes studies on divergent innovation strategies (Srholec & Verspagen, 2008), reverse effects of internal R&D (Cohen & Levinthal, 1989) and complementarities of various R&D processes (Freitas, Clausen, Fontana, & Verspagen, 2010).
Even when technical success is achieved, commercial success is an entirely separate matter. This includes the difference between the private and public returns to R&D. This important differentiation was first introduced by Griliches (1979) and he argues that although a firm’s R&D department may be able to successfully develop a new product which replaces its competitors’ products on the market, the product may soon be reverse-engineered and copied. As such, other firms, and eventually the consumer, may gain from the spillovers of this technical knowledge. For example, a firm may invest heavily in the development of a new product with a vastly higher utility level; however, once the product is released all of its competitors immediately copy the new product such that there are no changes in the sales or profits of any of the firms. In this situation, the returns to R&D are captured publicly by the consumers, rather than privately by the firm that made the investment to develop the new product.
III.2
T
HE
E
FFECTIVENESS OF
R&D
Despite the complexity of the innovation process, a 2005 study by the consulting firm McKinsey & Co. states that the ability to innovate is widely considered to be the largest factor in promoting future firm growth by CEO’s of large companies. The following sections will introduce several theoretical models with arguments for R&D in promoting sales growth, followed by results from several empirical studies.
A traditional model of firm behavior (Hashi & Stojčić, 2012) suggests that competing firms will quickly copy all innovation. Thus, in the long run, all firms arrive at the same equilibrium, with equal profits in all firms (0 with perfect competition). However, because significant heterogeneity (size, profits, growth, etc.) between firms appears everywhere, other theories must be more suitable for explaining these observations.
For example, Joseph Schumpeter’s theory of creative destruction moved beyond the system of ‘circular flow’ above, and suggests that new products, production methods, organizational systems, and inputs replace older, inferior systems. Thus, his earlier theories (Schumpeter, 1934) suggest that the firm as an entrepreneur must continually innovate to promote growth. Hashi and Stojcic (2012) argue that Schumpeter’s later work
In contrast to neoclassical growth models, modern models from the 1980s attempt to internalize technical changes to the firm The evolutionary model of the firm (Nelson & Winter, 1982) introduces the idea that a firm’s routines determine its competitive advantage relative to other firms. In order to remain competitive, a firm must innovate to improve its routines relative to its competitors, which drives the economic environment. In contrast to the ‘steady state’ in classical economics, in which market competition has removed all inefficient firms and remaining firms arrive at equilibrium, Nelson and Winter suggest ‘steady change’, where innovation changes a firm’s routines and promotes growth. The evolutionary model suggests that changes occur in a manner that is difficult to predict with boosts and slowdowns during the process.
The endogenous growth literature (Grossman & Helpman, 1993) agrees that internal changes rather than external changes result in the differences between firms. Thus, investments in capital, innovation and human capital can promote firm growth. Furthermore, the theory also suggests that there is simultaneity between innovation and firm performance. Although an investment in innovation promotes sales growth, sales growth also affects the investment level in innovation. An institutional environment with patent laws and legal protection is also necessary to allow innovators to reap the benefits of their innovations. The main difference between the evolutionary and the endogenous model is that the endogenous model still suggests that technological
development follow a stylized, steady pattern, whereas the evolutionary model argues that development occurs in spurts. In the endogenous model, innovation occurs to return to a profit-maximizing equilibrium, whereas the evolutionary model does not include these assumptions.
Furthermore, Del and Papagni (2003) argue that private returns to R&D result mainly from competitive effects. Rather than increases in sales and/or profits from improved products or processes, R&D investments also create significant entry barriers for new entrants, sustaining lower levels of competition for incumbents. A secondary competitive effect of innovation is the business stealing effect, which occurs when a firm brings a new or improved product to the market, and “steals” a large portion of its competitors’ market share.
As stated earlier, many studies on the effectiveness of knowledge have taken place since the 1960s. Studies relating R&D to firm performance have taken many forms throughout the last 50 years, developing throughout the 1980s and 1990s, testing the effect of innovation on various measures of firm performance such as
productivity ( Mansfield, 1980; Clark & Griliches, 1984), employment growth ( Nolan et al., 1980; Lachenmaier & Rottmann, 2011) and sales growth ( Odagiri, 1983; Coad & Rao, 2008). These studies developed as data sets expanded, now including broader data and more controls, and as methods became more advanced. Recently, dedicated innovation surveys such as CIS for the EU and the BRDIS in the U.S.A. have allowed for more advanced types of studies in which the actual innovation results of R&D can be tested separately ( Klomp & Van Leeuwen, 2001; Hashi & Stojčić, 2012). The following paragraphs will introduce several of these studies, focusing on the effect of innovation on sales growth.
Studies on firm-level productivity focus on firm-level data, and add R&D inputs to a standard Cobb-Douglas production in addition to capital and labor inputs. On average, studies using this method found coefficients between 0.2 and 0.3 (Link, 1981; Goto & Suzuki, 1989; Mairesse & Hall, 1996). Nolan et al. (1980) find positive effects of R&D on employment for UK pharmaceuticals. In a more recent study, Lachenmaier and Rottmann (2011) find positive effects of both R&D inputs and innovation output on employment with a 20-year dataset for German manufacturing firms. The benefit of using productivity is that there is less of a causality issue as most authors will not argue that productivity greatly affects R&D investments.
In conclusion, the theoretical arguments and models introduced above are generally supported by the empirical results. Thus, in line with the findings summarized above, the following hypothesis is presented.
Hypothesis 1: Knowledge and innovation from internal R&D positively affect firm-level sales growth.
III.3
O
UTSOURCING AND
R&D
O
UTSOURCING
Outsourcing has been a hotly debated topic for many years that has only risen with media interest around offshoring. One of the older and more established views on outsourcing is the theory of transaction cost economics (Williamson, 1981; Riordan & Williamson, 1985). Williamson suggests that there are many types of transaction costs related to monitoring, risks, etc. and that a firm will attempt to minimize transaction costs while also taking into account the direct cost savings from outsourcing particular processes. Thus, Williamson argues that a firm will only decide to outsource when transaction costs are low; in cases of few transactions, minimal risks and low asset specificity. Later views on outsourcing such as knowledge-based and resource-based views (Kogut & Zander, 1992; Grant, 1996) argue that a firm will keep processes close to its core business internal, while outsourcing processes that are less fundamental. Leiblein and Miller (2003) extend this argument by saying that firms will not outsource portions of the value chain that are fundamental to future growth. Khanna and Phalepu (2000) also discuss the role of property rights protection in a country. In a country with weak property rights, a firm may be hesitant to outsource activities, as the risk of expropriation is much higher than in countries with more rigid property rights. Overall, the majority of traditional views on outsourcing described above suggest that R&D is a core activity in promoting future growth. Thus, when outsourcing R&D, transaction costs should be high because of the high levels of asset specificity and high risks of expropriation. Furthermore, firms that do decide to outsource R&D suffer because of these transaction costs and can be forced out of the market (Metcalfe, 1994).
However, in contrast to these traditional arguments, R&D is being outsourced at increasing rates. R&D
outsourcing can include everything from basic research at a university to entire long-term contracts to develop a new product (Howells, 1999; Quinn, 2000). An often quoted paper on R&D outsourcing is “Outsourcing Innovation: The New Engine for Growth” by James Quinn (2000). Quinn argues that in the current highly competitive environment, in which firms are consistently raising R&D expenditures, firms can gain from strategically outsourcing particular R&D processes. Certain parties may simply be better at particular R&D activities than others (i.e. Universities for basic R&D) and he gives several case study examples of companies that have successfully outsourced particular processes, which would not be supported by the traditional R&D literature. Several of his most salient arguments for R&D outsourcing are:
Resource Limits: (Smaller) firms have resource limits and do not have the resources to compete with the total resources of its network of suppliers, buyers, competitors, and dedicated R&D organizations
Specialist Talents: Particular companies and organizations have specialist talents that are often not contained within a single business unit
Multiple Risks: A dedicated research organization or supplier is able to spread the risks of R&D across its buyers
Attracting Talent: A firm in which R&D is not a core activity may not be able to attract the best talent Speed: Dedicated, small R&D companies are more flexible and may be able to bring innovations to the market quicker
Although these more recent arguments discuss ways in which firms may manage and benefit from outsourced R&D, there are also a few recent arguments against outsourcing such core processes. Windrum et al. (2008) study the effects of complete outsourcing. They find that firms may initially receive a boost in productivity and profits because of the cost savings. However, in the long run, productivity growth is hampered in comparison to firms that outsource less. These findings may show that the relational view of outsourcing is somewhat flawed as firms that outsource are not able to recoup the full vertically-integrated benefits.
Finally, several authors study the determinants that may motivate a firm to outsource R&D (Kleinknecht & Reijnen, 1991; Mol, 2005; Martínez-Noya & García-Canal, 2011). The authors make some surprising
discoveries, as the more traditional arguments for outsourcing R&D do not appear to affect firms’ decisions. Mol shows that R&D intensity has become less of an impediment to outsourcing. He suggests that knowledge and innovation now come from many sources, rather than a firms internal R&D department and firms can effectively outsource processes by cooperating with various knowledge sources. Martinez-Noya and Garcia-Canal presents three arguments for why firms may outsource or even offshore R&D activities. The level of technological capabilities, institutional property protection, and local responsiveness attitude positively affect the outsourcing and offshoring of R&D. However, the strongest argument they make is that a firm’s technological capabilities do not automatically translate to the required governance capabilities to successfully control outsourced R&D. Kleinknecht and Reijnen show that firms do not cooperate on R&D because of costs, risks or small firm size. Instead, the size of research institutions and a formal R&D department in the firm promote cooperation. Furthermore, international firms are also more likely to cooperate with firms outside of the home country. Ultimately, considering the recent arguments, studies and empirical data, the following hypothesis is tested in this study.
Hypothesis 2: Knowledge and innovation from outsourced R&D positively affect firm-level sales growth.
III.4
C
OMPLEXITY OF THE
I
NNOVATION
P
ROCESS
In contrast to much of the older literature on R&D, thus extending beyond the hypotheses above, recent literature has pursued interest in the complexity of the R&D process. Firms’ diverse innovation strategies suggest that there are different ways to ensure successful innovations and improve performance. Furthermore,
complementarities may exist between various knowledge sources when the joint effect of two different R&D investments on performance is larger than both effects in isolation. The following sections will introduce several studies on the complexities of R&D, followed by recent examinations of complementarities of R&D sources. Many authors (Griliches, 1979) studied direct relationships between R&D and productivity although there may be many more sources of knowledge and innovation, which also do not occur in a gradual fashion. Various authors (Srholec & Verspagen, 2008; Freitas et al., 2010) show that there are many sources of knowledge and innovation that firms may utilize to improve their products and productivity. Furthermore, the European Association of Research and Technology Organizations (EARTO), albeit an interest group, argues for the importance of research and technology organizations (EARTO, 2007). They suggest that RTOs, which are often partially publicly funded (but are receiving increasing levels of revenues from firms), have an important position connecting basic research to more applied applications. The various sources of innovation thus have specialized areas of knowledge and work together to deliver new products, improved product improvements, or improved production processes.
readily available scientific study could have significant practical impact for a firm, this knowledge may only be found if a firm has internal researches to find, understand and make use of this knowledge. Finally, Freitas et al. (2010) further study the complexity of the R&D process. A recent dataset from Eurostat’s Community
Innovation Statistics (CIS) contains dichotomous variables for whether a firm uses various sources of knowledge (universities, suppliers, buyers, competitors, government research institutes and other enterprises) as well as whether the firm successfully developed a new product, new process, or both. Their model shows that information sources are complements rather than substitutes. Furthermore, firms with different innovation strategies successfully use different combinations of knowledge sources and the innovative capabilities of each knowledge source depend on its effectiveness in interacting with the various market actors.
Finally, there are several more recent studies that have studied various facets of innovation complementarity. For example, multiple authors have found complementarities between product, process, and organizational
innovation (Mairesse & Mohnen, 2010). Furthermore, others have used different models to study the
complementarity of internal and external R&D. Lokshin et al (2007) test the complementarities for 4 different knowledge sources and find positive effects. Cassiman and Veugelers (2006) find complementarities, but with the contingency that basic research from universities promotes this relationship. Hagedoorn and Wang (2012) find that internal and external R&D are complementary for higher levels of internal innovation, but substitutes for lower levels of innovation. In conclusion, considering the growth of broader innovation sourcing strategies, and the supporting findings of recent studies, the following hypothesis is suggested.
IV. D
ATA AND
M
ETHODOLOGY
The following sections build on from the literature review and introduce a statistical regression model to empirically test the hypotheses. The model uses the intensity of internal and external R&D investments along with industry-level control variables to predict sales growth. This type of model requires a dataset that extends beyond the traditionally available data on firm-level innovation activities. The Community Innovation Statistics (CIS) collects an impressive amount of information on firms’ innovation strategies across 16 EU countries, including a disaggregation of the expenditure on internal and external R&D activities. CIS is collected in ‘waves’, four years apart which cannot be linked to each other, thus resulting in cross-sectional data rather than panel data. To complete the regression model, other firm-level data, such as sales, and industry-level control variables, such as industry size and industry growth are also included. The regression model will be introduced first, followed by the econometric method, a discussion of the data and the proxies for each of the necessary variables.
IV.1
M
ETHODOLOGY
IV.1.1 PRIMARY REGRESSION EQUATION
In order to demonstrate why the specific model used here is chosen, the various previous attempts will be introduced and discussed. These other methods are all viable models that failed to be applicable for this study. In several instances this is due to the lack of available variables in the CIS dataset, whereas other problems arose due to the micro-aggregation methods. Finally, panel data also allows for more possible empirical methods than the cross-sectional data available here. Given the scale and scope of this study, the micro-aggregated CIS dataset and the method that is used seems to be the best and most appropriate method available. With other datasets, the methods introduced below may prove to be more suitable.
The first option that was considered is based on a Cobb-Douglas production function and adds a third input for the technical knowledge stock alongside capital and labor (Griliches, 1979; Kafouros, 2008). This model was abandoned for two reasons. First, a technical knowledge stock needs to be calculated for internal and external R&D, which is rather difficult based on just observation for R&D investments in CIS. Second, because CIS contains so much firm-level data, anonymization methods (which will be discussed in more detail later) are used. A Cobb-Douglas model requires a proxy for labor such as the number of employees. However, the micro-aggregation methods replaced the actual numbers by three categories: small (1-49 FTEs), medium (50-249 FTEs) and large (>250 FTEs).
improved product/total sales) using the firm’s innovation intensity (R&D expenditure/sales), combined with other factors such as industry competition and proximity to basic research. The residual in explaining innovation output in stage 2 is called “innovativity” of the firm and is comparable to TFP in a Cobb-Douglas production function. Sales growth would then become the dependent variable in stage 2 here; however there is still a selection bias for the dependent variables for R&D inputs in stage 2. There is no way to find a proxy for the missing R&D investments values for firms that answered ‘no’ to the selection questions.
Because of the various problems with the methods introduced above, a more straightforward empirical model is used here. The key difference is that the studies using the methods above have not yet been able to compare the effects of internal R&D and external R&D on firm performance. The method used here will allow an examination of the differences as well as the complementarities. The base model for this study is based on the method used by several others (Del & Papagni, 2003; Coad & Rao, 2008). The main explanatory variables are two independent variables for internal and external R&D investments (RDint and RDext). Although R&D is not a
perfect measure of innovation, as discussed in detail above, it is a commonly used measure to quantify innovation. A further independent variable is introduced in the form of an interaction variable to test for the complementarity of both knowledge sources. The use of an interaction variable is discussed in more detail in the later section on variables. The dependent variable, sales growth (ΔQ), is introduced as a measure of firm performance.
∆𝑄 = 𝛽!+ 𝛽! 𝑅𝐷!"# !+ 𝛽! 𝑅𝐷!"# !+ 𝛽! 𝑅𝐷!"# !× 𝑅𝐷!"# ! + 𝜀! (1)
Extending from this equation, several transformations must be made. First of all, the input for R&D is a growth in the R&D stock rather than the current total technical knowledge stock. The technical knowledge stock is difficult to measure as the depreciation rate for past investments is often arbitrary. Several authors (Pakes & Schankerman, 1984; Goto & Suzuki, 1989), study the issues of depreciation of R&D and find the depreciation rate generally ranges between 15 and 25 percent. However, technical knowledge cannot realistically be
depreciated at a constant rate like physical capital. Finally, because a cross-sectional dataset is used here, it is not possible to calculate a total stock, so investment in R&D is considered to be the growth in the technical
knowledge stock, assuming the depreciation rate is 0.
Furthermore, R&D investments must be controlled for firm size, as sales growth is also independent of firm size. The variables for R&D are thus transformed into a measure of R&D intensity, in line with Nolan et al. (Nolan et al., 1980), among others:
𝑅 =!"
! (2)
Sales growth is measured as the average annual growth rate between 2002 and 2004 and is calculated as follows:
∆𝑄 = !!""#
!!""!− 1 (3)
The regression equation thus becomes:
∆𝑄 = 𝛽!+ 𝛽! 𝑅!"# !+ 𝛽! 𝑅!"# !+ 𝛽! 𝑅!"# !× 𝑅!"# ! + 𝜀! (4)
Essentially, this equation tests the effect of innovation on firm performance, with three independent variables for innovation to test hypotheses 1, 2 and 3, respectively. The size, direction and significance of β1 and β2 will
provide an answer for hypothesis 1 and hypothesis 2 respectively, estimating the effectiveness of internal and external technical knowledge capital stocks on promoting labor productivity. For this study, hypothesis 1 and hypothesis 2 will be accepted if β2 and β3 are positive and significant at the 5% level (α=0.05).
Hypothesis 3 is addressed by the estimation of β3, the coefficient of an interaction variable of the two technical
the separate effects of internal and external R&D. In an OLS model, the interaction variable is the most common way to test for situations in which the effect of the independent variables on the dependent variable is not additive (complementarity). Internal and external R&D will be considered complementary if β3 is positive and
significant at the 5% level (α=0.05).
IV.1.2 CONTROL VARIABLES
Although the focus of this study is on the effects of R&D and innovation on firm performance, these factors are clearly not the only factors that affect performance. As with any regression equation, in order to achieve a correct estimation of the beta-coefficients, the aim is to isolate the effects of the main independent variables. Controlling for other variables in the dataset that affect performance does this; however, micro-level control variables are unavailable in CIS, so the dataset is matched to another dataset with industry-level control
variables. The following sections will introduce the different ways in which control variables can be added at the micro-level and industry-level to isolate the effects of R&D.
The industry-level control variables are collected from two sources. The first source discusses the various industry characteristics that affect firm performance (Capon, Farley, & Hoenig, 1990). This article contains an analysis of 320 past publications that discuss the various factors that affect firms at the industry, firm, and strategy levels. Capon et al. find several industry factors that significantly affect firm performance, including industry concentration, industry growth, industry capital investment, industry size, and industry advertising. Industry concentration is a measure of the competitiveness in the industry, thus controlling for sales growth with low innovation simply because there are low level of competition. Industry growth is similarly used to control for the effects of the entire industry growth. Industry size also affects industry competition, as firms can grow more easily in a large industry than in small industry. The second source for control variables, past studies on the effect of R&D on sales growth (Geroski & Toker, 1996; Del & Papagni, 2003), confirmed the use of these control variables. Industry growth is used in both of these studies and Geroski and Toker also include industry concentration and industry advertising intensity. This study will also include controls for industry size and industry growth; however, data on industry concentration, advertising and capital investments are not complete enough to use in this study. More details on the control variables and their respective data sources will be given in a later section on data and variables. Extending equation (3) with the control variables, the final regression equation becomes:
∆𝑄 = 𝛽!+ 𝛽! 𝑅!"! !+ 𝛽! 𝑅!"# !+ 𝛽! 𝑅!"# !× 𝑅!"# !+ 𝛽!𝐺𝑅𝑂𝑊!"#+ 𝛽!𝑆𝐼𝑍𝐸!"#+ 𝜀! (5)
This final equation will be tested separately from equation (4) to be able to differentiate the effects of the control variables from the focus variables in terms of the changes in the beta-coefficients, significance levels, and the R2.
IV.1.3 REGRESSION ANALYSES USING SUBSAMPLES
A further issue that should be controlled for is the difference in the innovation process between more developed countries and less developed countries. More developed countries are at the knowledge frontier and have businesses that develop truly new products and processes. On the other hand, it is generally argued that less developed countries gain from the technical knowledge spillovers from more developed countries rather than innovating themselves (Gerschenkron, 1962). Because of this, investments in innovation in both types of countries will be of a different nature and may have different effects on sales growth. On the one hand, it could be argued that investments in less developed countries will have a larger impact because copying or emulating has significantly less risk. However, in an open EU market, less advanced firms may also have difficulty competing with more developed firms from other countries.
1986, whereas newer members did not join until 2004 or 2007. This results in an even distribution of seven “more advanced” early members and nine “less advanced” newer members. Thus, the regression will be estimated separately for both groups of countries in order to examine the differences in the relationship between R&D and sales growth in more and less developed countries.
Table 1 – Advanced and Less Advanced Economies
Advanced Economies Less Advanced Economies Belgium Germany Spain Greece Italy Norway Portugal Bulgaria Czech Republic Estonia Hungary Latvia Lithuania Romania Slovakia Slovenia
Furthermore, it is commonly known that average rates of innovation, or innovation intensity, varies greatly by industry. For example, the textile industry is far less complex and has lower rates of innovation than the semiconductor industry. Because of the difference in the innovation intensity between industries, additional spending in innovation may also have drastically different effect on actual firm performance. A lack of investments in innovation may not result in an enormous deficit in the textile industry; however, in
semiconductors a firm can become obsolete when it does not constantly invest. As mentioned in the literature review, several studies (Adams & Jaffe, 1996; Kafouros, 2008) find differences in the results between industries, with results ranging from a coefficient of 0.04 for the plastics industry, to 0.13 for pharmaceutical companies. Eurostat divides the NACE-coded industries into four categories: high-technology, medium-high-technology, medium-low-technology, and low- technology (Eurostat Glossary, 2013). Because of this, the regression will be done separately for all four categories that are used.
Table 2 – Industry Categories
High Technology Medium-Low Technology
30 to 33 (DL) Manufacture of office machinery, apparatus, radio, television, communication equipment and apparatus, medical, precision and optical
instruments, watches and clocks
25 to 28 (DH, DI, 27, 28) Manufacture of rubber and plastic products; basic metals and fabricated metal products; other non-metallic mineral products
Medium-High Technology Low-Technology
23 and 24(DF_DG) Manufacture of chemicals and chemical product, coke, refined petroleum products and nuclear fuel
29 (DK) Manufacture of machinery and equipment n.e.c.
34 to 35 (DM) Manufacture of motor vehicles, trailers and semi-trailers and other transport equipment
15 to 22 (DA, DB, DC, 20_21, 22) Manufacture of food products, beverages and tobacco; textiles and textile products; leather and leather products; wood and wood products; pulp, paper and paper products, publishing and printing
36 to 37 (DN) Manufacturing n.e.c.
2003; Geroski & Toker, 1996). Although all three studies show that firm size does not affect sales growth, in line with Gibrat’s Law, an additional test with initial firm size is added here. In order to account for changes due to firm size in both the y-intercept, and the regression coefficients, the regression is done separately for all three groups.
IV.1.4 QUANTILE REGRESSION
A final consideration regarding the regression equation is made before introducing the data and the variables. Coad and Rao (2008) argue that the distribution of the sales growth rate is highly skewed, or tailed. They show that there are only a few firms in each industry that simply outperform all others in terms of sales growth performance. Because of this, the returns to innovation are also highly skewed. Coad and Rao thus find it misleading to focus on the average effects alone as this ignores the effects for top performing firms. The following sections will thus introduce a quantile regression model for examining the effects of innovation on sales growth. Mosteller and Tukey (Mosteller & Tukey, 1977) present the following passage on quantile regression:
“What the regression curve does is give a grand summary of the distributions to the set of x’s. We could
go further and compute several regression curves corresponding to the various percentage points of the distributions and thus get a more complete picture of the set. Ordinarily this is not done, and so the regression often gives a rather incomplete picture. Just as the mean gives an incomplete picture for a set of distributions”
A quantile regression model is a variation of OLS regression in which additional regressions are done for the bottom and/or top of the data set. The model estimates the coefficients for the various conditional quantiles of the dependent variable rather than the conditional mean of the dependent variable. A simple regression situation with one independent and one dependent variable can be used to explain the essentials of quantile regression. The figure below presents a figure displaying the use of quantile regression for the relationship between household income and food expenditure from Koenker and Hallock (2001, page 147):
Figure 1 – Quantile Regression Example 1
When performing the regression at various quantiles, the intercept or the beta-coefficients may change. The figures below show how the intercept and various coefficients change when examining the relationship between sales growth and Coad and Rao’s variable ‘innovativness’, which is a combined variable for R&D expenditures and patent data. Figure 2 presents the coefficients for innovativeness for four different industries against the quantiles (Coad and Rao, 2008, page 643):
Figure 2 – Quantile Regression Example 2
The dashed lines are the original OLS coefficients, accompanied by the dotted lines which represent the 90% confidence intervals. For this particular study, it can be seen that the coefficients fall well outside of the
confidence interval for the upper and lower quantiles, thus supporting the merit of additional examinations using quantile regression. A further interesting point that Coad and Rao make based on their empirical results is that for the average firm they do no find a strong positive relationship between innovativeness and sales growth. The coefficients using OLS are often close to 0, and at the lower quantiles, innovativeness has a negative effect on sales growth.
IV.1.5 TESTING ROBUSTNESS
The various methods above examine the data extensively by repeating the regression analysis for various subsets. Furthermore, quantile regression examines the data at various quantiles, and is thus already much more robust to outliers and skewed data than OLS; however, there are reasons to test for robustness. First of all, later sections on data analysis will show that the variables are fairly skewed to the right. For example, most firms in the sample will grow modestly, but some firms will experience more extreme growth, resulting in a skewed distribution of sales growth. Similar distributions exist for internal and external R&D intensities as only a few firms invest heavily in R&D. The skewed distributions can result in a non-normal distribution of residuals. These various factors will be discussed and demonstrated in more detail in the section on statistical tests later; however, the various tests for robustness will be introduced here. Ultimately tests for robustness are used as a final test to examine the robustness of the original regression results. Three methods that are different from OLS and are less affected by outliers will be introduced here: median regression, robust regression and the removal of all outliers. The section on descriptive statistics and statistical tests will examine the data using graphical representations of the distribution of the dependent and independent variables. Furthermore, residual plots are used to test whether the assumptions for OLS are not broken. Finally, the data analysis section will implement the three methods to test the robustness of the regression results, and compare the results with the results from OLS.
IV.1.5.1 MEDIAN REGRESSION
Median regression has indirectly been introduced in the section on quantile regression, as median regression is simply quantile regression at the 0.5 quantile. For any sample, there are three common ways to calculate the average: the mean, the median, and the mode. The mean calculates the arithmetic average, whereas the median ranks the data from small to large and uses the “middle” observation. Outliers can heavily affect the mean, whereas the median is less affected. For example, a sample with 0, 1, 1, 2, 4, 6, 7, 17, and 25, the mean is 6.8, whereas the median is 4. A similar difference exists when comparing OLS methods to median regression. Because it may be that some firms with either extreme growth or extreme levels of R&D investments (which may or may not be caused by errors in the survey) can affect the results greatly, the results from median regression may be of greater interest.
IV.1.5.2 ROBUST
REGRESSION
Robust regression is another way to reduce the effects of outliers on the regression results. In contrast to OLS, robust regression is still suitable when the underlying assumptions of OLS do not hold. Robust regression can be done in many different ways, a discussion of which is beyond the scope of this paper. The methods can be allocated to two main types: the first uses deviations of OLS methods, and the second uses a different
distribution such as a heavy-tailed distribution. Stata, the statistical software package used in this study, has its own method using the command for robust regression (rreg). Stata’s version of robust regression first runs the OLS regression and then calculates the Cook’s distance for each observation and removes any observations with a Cook’s distance greater than 1. Cook’s distance is a combination of the residual and the leverage of an individual data point and is thus essentially a measure of the influence of a particular observation on the regression results. It then runs iterations in which observations with large residuals are down-weighted.
Essentially, this method either removes outliers or gives outliers a lower weight in the regression results in order to reduce the effects of outliers.
IV.1.5.3 REMOVING
OUTLIERS
IV.2
D
ATA
The following sections will introduce the dataset, and further study the variables used for this study and the limitations of the proxy for each variable in the CIS dataset. Eurostat collects extensive amounts of statistics from the EU area. The portion of Eurostat’s data for Innovation is known as the Community Innovation Statistics (CIS). CIS collects micro-level data from both manufacturing and service firms throughout the EU area every four years (every other year since 2006) that extend far beyond the traditionally available data for R&D based studies. This has resulted in a wave of recent innovation studies using this new data set (Srholec & Verspagen, 2008; Stefan Lachenmaier & Rottmann, 2011). The fourth CIS data set (CIS-4) compiled data on innovation in 2004 and covers 15 Western European and Eastern European countries. These new types of datasets are very appealing to studies on innovation because of the vastly increased amount of information on firms’ innovation processes as well as a disaggregation of the actual R&D investments. For this study in particular, the
disaggregation of internal R&D and external R&D allows us to study the relative effectiveness of each knowledge source. Other benefits include the size of the dataset, as several thousand data points will provide fairly robust results. Furthermore, because of the Europe-wide nature of the data, all industries are covered, removing industry biases from the data.
There is another important facet of the dataset and its effect on the regression equation. The base year of the CIS-4 dataset is 200CIS-4; however, the sales data is only available for 2002 and 200CIS-4, whereas R&D investments are recorded for 2004. Thus, when estimating sales growth with R&D investments, the equation actually estimates the effect of current (2004) R&D investments on past sales growth (2002-2004). Because the data is collected anonymously, it is impossible to match R&D investments in 2004 with sales data from CIS 2006, which could allow for a better method. On the other hand; it should be noted that R&D investments change very little over time (Kafouros, 2008). Thus, it is not unreasonable to assume that the level of R&D investments in 2002 will not differ greatly from 2004. Some authors have also studied the time-lag of R&D; however, because of the
continuity of R&D investments, Kafouros and others find very little difference in the statistical results when changing the time-lag from 0 years to 5 years. Finally, as high-performance firms may also have an incentive to spend large amounts on R&D (reverse causality), it should be noted that arguments regarding causality are difficult to make with the discrepancy in time of the data. Because the industry control variables are available at more regular time intervals, the base year for the control variables is 2002. Industry growth will be matched to sales growth and is averaged over the period between 2002 and 2004.
Furthermore, to control for several industry effects, the CIS data is merged with a dataset containing industry-level data for manufacturing in the EU. INDSTAT 4 was used for this study as it provides annual information on sales by industry for all 16 countries in the CIS dataset. INDSTAT is compiled by the United Nations and includes industry-level data, including output, number of establishments and number of employees from 1990 to 2009. INDSTAT uses ISIC (Rev. 3) industry codes, but this is easily matched to the NACE industry codes used in CIS. The following paragraphs will provide more information on which variables are constructed using the compiled dataset.
Finally, it should be noted that the focus of this study is on the manufacturing sector, thus excluding services, agriculture and mining. The first reason for this is that other sectors do not record investments in R&D correctly or at all. For example, the services sector is centered on human capital, so improvements in the use, efficiency or education of human capital is generally not seen as an investment in R&D. Furthermore, although CIS does include industries beyond manufacturing, INDSTAT does not, thus not allowing for a comparison between manufacturing firms and non-manufacturing firms.
IV.2.1 VARIABLES
Firms generally do not want to disclose much information regarding their innovation strategies or expenditures, making it difficult to obtain true micro-level data. A full step to industry-level aggregated data would remove much of the variability in the data and limit the entire data set to no more than 70 or 80 data points. However, for CIS data, there is luckily a third, intermediate option; micro-aggregated data. Micro-aggregation is a method that allows researchers to use data almost identical to the original firm-level data, without disclosing any details about the individual firms or their sales and R&D expenditures. In order to satisfy the demands of both firms and researchers, Defays (1997) experimented with new methods of micro-aggregation with CIS data at Eurostat. A method was developed that “consists of replacing each observation of a given variable by an average of itself and the two adjacent observations in a ranking order of the observations for that variable”(Mairesse & Mohnen, 2001, page 4). This process is done separately for each variable, sales, internal R&D expenditure, external R&D expenditure, etc.). First, the data points are ranked individually for each variable (2002 and 2004 sales, internal R&D expenditure, external R&D expenditure for this study). Then, each data point is replaced by the average of the selected data point and the data points above and below it. This process thus ‘camouflages’ the original data, protecting its confidential nature, without significantly impacting the data itself.
In comparison to true micro-level data, the main caveat is slightly reduced variability; however, the method does not affect the mean. The reduced variability entails that the minimum and maximum move closer to the mean and the standard deviation shrinks slightly, which could negatively affect the correlations and regression estimates; however several studies show that correlations are generally close to the originals (Defays, 1997; Mairesse & Mohnen, 2001). Mairesse and Mohnen examine innovation performance in France using both a firm-level data set and a micro-aggregated CIS dataset. They randomly select 1000 data points from both sets and control for industries with high and low levels of R&D. Although they compare two completely different datasets, one of which is micro-aggregated, they find virtually no difference in the regression estimates. Studies using micro-aggregated data in a quantile regression model are not yet available; however, considering the information available on aggregated data, several suggestions can be made. As stated above, micro-aggregation affects the maximum and minimum of the data and the overall variability, which indicates that the upper and lower ends of the data are affected most. Because of this, results using quantile regression at the extremes of the data may be affected. Considering the previous findings for standard regression methods using micro-aggregated data, it can be said that the data provides statistical results that are very close to the original data and are robust compared to the original data for (the majority of) the types of regression used in this study (Hu & Debresson, 1998; J Mairesse & Mohnen, 2001). However, in anonymizing the data, several other concessions were made that affect the availability of data for particular variables, which will be described in the following sections.
IV.2.2 OUTPUT
Using sales data for 2002 and 2004, it is possibly to calculate the average annual growth in that period. Besides profits, sales growth is generally one of the most salient targets for firms, and thus one of the main reasons why firms will invest in innovation. Although various studies use other measures of firm performance such as productivity or employees, both measures are probably not the main motivation for investments in innovation. Furthermore, a time-series data set would allow for a more extensive study of the effects of R&D on sales growth over time. However, because CIS is a cross-sectional dataset, which cannot be matched to other CIS datasets (CIS-3, CIS 2006), it is only possible to study sales growth over a short time period. Overall, sales growth should provide a strong proxy for firm performance when studying the effects of R&D investments.
IV.2.3 R&D – INTERNAL AND EXTERNAL
The proxy of R&D investments for technical knowledge was discussed extensively in the literature review and will not be further addressed here. Internal and external (outsourced) R&D investments are taken directly from the data set. The disaggregation of R&D investments and publication of (micro-aggregated) data is quite a novelty and was not available before innovation surveys like CIS. As stated above, continuous numerical variables are micro-aggregated, which protects the individual firms from confidentiality issues and has minimal effects on the regression outcomes.
IV.2.4 INTERACTION TERM
To test for the complementarity of both knowledge sources, an interaction term of internal and external R&D intensities is introduced. Complementarity is a difficult effect to examine as it is impossible to establish a perfect laboratory setting to examine various effects and test for complementarities in social sciences. Furthermore, Lokshin et al. (2007) argue that interaction terms are an imperfect method for testing complementarities. They argue that if a particular input or independent variable has a small impact on firm performance, there is no method to test the complementarity of two such sources because the effects are too small to draw a conclusion. Furthermore, when there are various inputs that are being tested for complementarity (i.e. organizational innovation, internal innovation, external innovation, university innovation, etc.) interaction variables are insufficient. However, because only two variables are being tested for complementarity here, an interaction term should be an appropriate method.
Brambor et. al (2005) study the use of interaction terms and summarize the key mistakes and suggestions in the following four points:
1. Include interaction terms when the hypothesis is conditional/complementary in nature 2. Include all terms used in the interaction term separately in regression equation
3. Do not interpret the separate terms as marginal effects when interaction terms are present 4. Calculate new marginal effects and standard errors
The first two points have been implemented in the regression equation mentioned above, and the second two points will be adhered to when discussing the regression results in later sections.
Statistical programs present the regression results but not the marginal effects of the variables, which comprise the interaction term, on the dependent variable when an interaction term is present (Brambor et al., 2005). For example, here the interaction terms makes the effect of the internal R&D intensity on sales growth conditional on the external R&D intensity. Because of this, it is incorrect to interpret the regular regression coefficient for internal R&D intensity as the marginal effect on sales growth when an interaction term is present in the regression. Instead, the new coefficient for β1 must be calculated using the following equation in order for the
coefficient to show the marginal effects of the independent variables on sales growth:
𝛽!= 𝑏!+ 𝑏!𝑅!"# (6)
Where b1 and b3 are the original coefficients from the regression and β1 is the corrected coefficient. A similar
equation is used for β2. This equation shows that the coefficients for internal R&D intensity from the statistical
statistical software. The variance of the coefficients and the covariance between the coefficients need to be obtained from a variance-covariance matrix to calculate the correct standard errors using equation (7). Similar to equation (6), a similar equation is used to calculate the standard error for β2.
𝜎!"# !!!"#
= 𝑣𝑎𝑟 𝑏! + 𝑅!"#!𝑣𝑎𝑟 𝑏
! + 2𝑅!"#𝑐𝑜𝑣 𝑏!𝑏! (7)
In both equations, the values of Rint and Rext are continuous, which affects the coefficients and the standard
errors. Brambor et al. (2005) suggest the use of a graph with one independent variable on the x-axis and the coefficient of the other independent variable in the y-axis. (i.e. Rext and β1) to show the changes in the coefficient
at different values of the other variable. However, a different method is used here for two reasons. First, due to the number of subsamples here, this would result in over 20 graphs of coefficients. Second, the values of internal and external R&D intensities are generally very small with a mean of 0.023 and 0.004, respectively. As
distribution plots in the data analysis section will show, the majority of firms have R&D intensities that lie close to the mean. Because of this, the regression coefficients and standard errors reported in the data analysis section are calculated using the mean value of internal and external R&D intensities for each individual subsample (the median is used for median regression). Furthermore, a discussion of what happens to the regression coefficients for firms with very high levels of R&D intensity will be included in the text.
IV.2.5 CONTROL VARIABLES
The two control variables introduced in the methodology section above are: Industry Growth (2002-2004)
Industry Size (2002)
These variables will aim to isolate the effects of R&D on sales growth. In combination, they control for general market movements and the competitive environment. Because the industry data is collected from another source as the R&D and sales data, they are matched using the NACE industry codes. However, CIS replaced the more specific NACE codes by slightly broader groups. Thus, there are now 30 industry groups rather than the original 72 so the industry data is averaged to match the CIS data. Furthermore, because this study examines
manufacturing industries only, 14 different industries remain. Each variable will be introduced briefly in the following paragraphs.
Industry growth is measured by the overall growth in output or sales in that industry. As such, industry growth controls for general trend (growth or decline) in that industry. Although particular firms may grow in terms of sales growth, if the output of the entire industry is growing at the same rate, sales growth no longer accurately measures firm performance. Adding a control variable for sales growth thus controls for these changes, making sales growth a more accurate measure of performance. Annual industry growth is calculated using equation (3).
∆𝑄 = !!""#
!!""!− 1 (3)
However, for several countries in the dataset, industry output data is unavailable for 2002 so industry growth is replaced by the annual growth rate between 2003 and 2004: ΔQ=Q2004/Q2003.
V. D
ATA
A
NALYSIS
The following paragraphs will empirically analyze the data using the methodology presented above to test each of the three hypotheses. The data will hopefully provide some answers to the previous literature and the hypotheses. Furthermore, beyond the original regression estimates, the various other techniques will show whether the relationships between R&D and sales growth change because of firm size, industries, economic development, or firm performance. First, the data will be introduced and discussed using several descriptive statistics. This will be followed by an analysis of the initial regression estimates and the related adaptations:
V.1
D
ESCRIPTIVE
S
TATISTICS
Various descriptive statistics will provide a summary of the data and will give some preliminary indications regarding the hypotheses. The table below presents the number of innovating firms that invest in internal and external R&D:
Table 3 – Investments in R&D
Investments in R&D
No Investment 2432
Both Internal and External Investments 1983
Internal Investments 2964
External Investments 338
Total 7717
Next, the following table summarizes several descriptive statistics for all the original inputs from the data, as well as the final variables used in the regression:
Table 4 – Mean, Minimum and Maximum of Variables
Variable Mean Min Max
Sales 2002 106M € 100 57.2B € Sales 2004 112M € 187 60.1B €
Sales Growth 10% -‐50% 250%
Internal R&D 3.1M € 0 2.0B € External R&D 0.9M € 0 653M € Internal R&D Intensity 0.023 0 0.48
External R&D Intensity 0.004 0 0.47
Industry Output 2002 47B € 0 284B € Industry Output 2004 69B € 0 408B € Industry Growth 22.2% -‐7.2% 146.2%
The summary statistics above provide several interesting insights. It should be noted that the italicized statistics are limited by the selections made to control for outliers. First of all, the size of the firms varies greatly from 0 sales to 57B €. Second, the manufacturing industry as a whole grew rapidly in the EU between 2002 and 2004, with an annual growth rate of approximately 22%; however, it is peculiar that sales growth of the selection of innovating firms is only 10%. Because this study only includes innovating firms, average sales growth was also calculated for the entire CIS dataset (including non-innovators). For the entire dataset, sales growth was very comparable to the average industry growth rates. This discrepancy may be caused by the fact that selecting only innovating firms has resulted in a larger portion of firms in more advanced economies, as is confirmed in table 5. Table 5 also shows that manufacturing in Europe has been growing more rapidly in less advanced countries than in more advanced countries.
Furthermore, on average, firms invested over 3M € in internal R&D, and over 0.9m € in external R&D. However, as shown above, only a portion of all firms carried out these investments. The following sections will delve into more in-depth preliminary analysis.
The following table shows differences in the mean R&D internal and external R&D intensities, and sales growth for small, medium and large firms, high-tech and low-tech industries, and more advanced and less advanced economies:
Table 5 –Average Sales Growth and R&D Intensities for Subsamples
Firms Sales Growth Internal R&D Intensity External R&D Intensity
Small Firms 2356 12.9% 0.31 0.005 Medium Firms 2820 10.6% 0.20 0.003 Large Firms 2211 6.0% 0.19 0.004 Low-‐tech firms 2519 10.6% 0.012 0.003 Medium-‐low-‐tech firms 1588 11.0% 0.015 0.003 Medium-‐high-‐tech firms 2325 8.5% 0.027 0.005 High-‐tech firms 1285 10.1% 0.049 0.008
More Advanced economies 6982 9.0% 0.025 0.005
Less advanced economies 735 19.1% 0.006 0.002
high-tech industries than in low-tech industries, both internally and externally. On average, high-tech firms invest four times as much in internal R&D and more than twice as much in external R&D as low-tech firms. Notably, firms in more advanced economies invest much more heavily in R&D than firms in less advanced economies. This is most likely because firms in more advanced economies are the technical knowledge frontier, and new products are generally better protected by the institutions protecting intellectual property rights.
Furthermore, in line with expectations, it can be seen that sales growth is largest in small firms, which also invest most in R&D. The dispersion of sales growth is fairly balanced among various high- and low-tech industries. However, sales growth is lowest in medium-high-tech firms. Finally, although firms in more advanced
economies invest more in R&D, sales growth is more than double for firms in less advanced economies. Overall, the relationships found between R&D intensity and sales growth support the expectations for firm size, are ambiguous for industry technology levels, and are opposed by more advanced and less advanced economies. The following table presents additional information on the number of firms from each country and the mean sales growth, and R&D intensities:
Table 6 - Average Sales Growth and R&D Intensities by Country
Country
# of Firms
Sales Growth Int. R&D Intensity Ext. R&D Intensity
Belgium
343
5.0%
0.0247
0.0054
Bulgaria
692
22.2%
0.0033
0.0022
Czech Republic
819
10.1%
0.0166
0.0038
Germany
829
5.1%
0.0429
0.0048
Estonia
199
16.8%
0.0126
0.0024
Spain
1790
8.9%
0.0273
0.0051
Greece
91
9.0%
0.0233
0.0083
Hungary
234
13.4%
0.0147
0.0040
Italy
1160
6.1%
0.0240
0.0044
Lithuania
127
15.6%
0.0093
0.0041
Latvia
100
25.6%
0.0043
0.0003
Norway
337
3.5%
0.0513
0.0094
Portugal
382
5.5%
0.0147
0.0019
Romania
373
16.8%
0.0161
0.0031
Slovenia
38
8.6%
0.0077
0.0344
Slovakia
203
13.1%
0.0057
0.0023
Total
7717
10.0%
0.0234
0.0044
In this table, the boldfaced countries are the early EU members and are thus considered to be the more advanced economies, as introduced earlier. It is evident that there are many more firms from more advanced economies (6.982), compared to firms from the less advanced economies (735). In the original dataset with manufacturing firms the separate samples were more balanced; however, after removing firms with missing data and controlling for outliers, the sample of less developed countries shrank more relative to the sample with more developed countries. Both samples are still very large and will be analyzed separately in the regression model to examine any differences in the relationship between R&D and sales growth. It is also apparent that not all countries are equally represented. For example, the ratio of Spanish firms (1,790) to German firms (829) in the dataset is not realistic. Thus, some countries are overrepresented, whereas others may be underrepresented.
Furthermore, the R&D intensities are lower in less developed countries than the total average. Bulgarian firms invest the least in internal R&D, whereas Latvian firms invest the least in external R&D. Furthermore, firms in more advanced economies appear to invest more in R&D. Norwegian firms have the highest internal R&D intensity, whereas Slovenian firms outsource more on average.
V.2
S
TATISTICAL
T
ESTS
Before running regression analyses, several tests must be done to see whether the assumptions for an OLS regression described above hold. As with any OLS regression, there are several assumptions that need to hold to ensure that the least-squared estimators are the best linear unbiased estimators (BLUE) of the parameters according to the Gauss-Markov theorem (Carter Hill, Griffiths, & Lim, 2008).
First of all, it is expected that the distribution of both the dependent variable and the independent variables are not perfectly normal. A large portion of the firms will most likely invest relatively small amounts, while certain industries or certain firms will invest particularly large amounts. These skewed distributions are displayed in the following figure:
Figure 3 – Distribution of Sales Growth and R&D Intensities
The histograms show that internal and external R&D intensities are particularly skewed. Sales growth has an almost normal distribution, but is slightly skewed towards the right. The Jarque-Bera test suggests that the distribution of sales growth does not deviate too much from a normal distribution. One attempt to linearize the R&D inputs is to take the logarithm of all independent variables. However, this resulted in a very strange distribution of the residuals, with a large number of outliers. Furthermore, the residuals are almost split into two groups, suggesting that the error terms would be heteroscedastic using this method.
0 1 2 3 4
D
ens
ity
-1 0 1 2 3SalesGrowth
0 10 20 30 40 50D
ens
ity
0 .1 .2 .3 .4 .5Internal R&D Intensity
0 20 40 60 80
