1
Spillover effects of intangible capital:
A comparative perspective of Europe and the United States
Master’s thesis University of Groningen Faculty of Economics and Business
Abstract
This paper analyses spillover effects of intangible investment and the interaction effect of intangible investment with ICT investment, it also analyses these effects for both the United States and Europe. Using data from the EU-KLEMS and INTAN-invest database, the paper finds an elasticity for non-R&D intangible of 0.11 for the market sector for both Europe and the United States. The estimated elasticity is 0.16 for the manufacturing sector in Europe and -0.063 for the US. Additionally, there is an interaction effect of ICT investment and non-R&D investment for the manufacturing sector. Finally, the United States has fewer spillover effects in most cases compared to Europe.
Student: Leonie Stuve Student number: s3409112
Supervisor: prof. dr. H.H. van Ark Co-assessor: dr. T. Kohl
2
Table of contents
List of figures ... 3
List of tables ... 3
1. Introduction ... 4
2. Relation between production factors and TFP ... 6
2.1 Total factor productivity ... 6
2.2 Human capital ... 7
2.3 ICT capital ... 8
2.4 Intangible capital ... 9
2.5 Spillover effects in Europe and the United States ... 12
3. Methodology ... 13 3.1 Model ... 14 3.2 Estimation method ... 16 3.3 Calculation of variables ... 17 4. Data ... 18 4.1 TFP growth ... 18 4.2 Tangible capital ... 20 4.3 Labour ... 20 4.4 Intangible capital ... 21 4.5 Preliminaries ... 23 5. Results ... 25 5.1 Full sample ... 25 5.2 Time periods ... 29 6. Discussion ... 32
6.1 Comparison of results to literature ... 33
6.2 Comparison of sectors ... 34
6.3 Comparison of Europe and the United States ... 35
7. Conclusion ... 37
References ... 39
Appendix A. Description of variables and industries... 42
Appendix B. Descriptive statistics ... 44
Appendix C. Tests econometric specification ... 45
3
List of figures
Figure 1: Intangible and tangible investment in 2013 4 Figure 2a: Market sector TFP growth US and European countries 19 Figure 2b: Manufacturing sector TFP growth US and European countries 19 Figure 3a: Division of intangible capital manufacturing sector 22 Figure 3b: Division of intangible capital market sector 22 Figure 4: Preliminary results market sector and manufacturing sector 24
List of tables
Table 1: Mean of input factor shares 17
Table 2: Mean of labour compensation and hours worked shares 21
Table 3: ∆ln(TFP)c, t 1995-2013 26
Table 4: ∆ln(TFP)c, t 1995-2013 28
4
1. Introduction
“The knowledge economy is all around us but can we see it in the official statistics?” (van Ark, Corrado, Hao, & Hulten, 2009, p. 63). This quote is highly relevant as most statistics only include tangible investment. However, intangible capital is becoming increasingly important. Intangible investment is at the same level as tangible investment and even exceeds it in some countries (van Ark et al., 2009). As such, this paper takes a closer look at intangible capital and its effects on productivity. More specifically, the paper compares spillover effects of intangible capital and ICT capital in the US and European countries.
But why should we care about productivity and spillover effects to begin with? As stated by van Ark (2014), total factor productivity (TFP) growth is the driver of long-run economic growth. Therefore, it is important to study which effects cause productivity growth and which variables might not be beneficial for growth. When these factors are known, it is easier to focus on increasing productivity growth. In the past couple of years, the focus has shifted to include intangible capital as a source of growth.
Figure 1 shows that intangible investment exceeds tangible investment for some of the countries in the sample. Furthermore, intangible investment represents an even higher share of value added in the manufacturing sector compared to the market sector. As intangible investment is large, it should be included in sources of growth analyses.
To avoid any confusion, intangible capital consists of multiple components. The three main components are computer software and databases, innovative property and economic competencies (Corrado, Haskel, Jona-Lasinio, & Iommi, 2016). Innovative property is largely driven by R&D capital, but it also includes design, new product development costs in the
Figure 1. Intangible and tangible investment in 2013
5 financial industry, and entertainment, literary and artistic originals (Corrado et al., 2016). Economic competencies consist of brand, organisational capital and training (Corrado et al., 2016). Generally, innovative property are knowledge assets associated with intellectual property rights, whereas economic competencies consists of knowledge assets which are not associated with intellectual property rights (Corrado et al., 2016).
More recently, intangible capital has been included as an important source of growth in the literature (Corrado, Haskel, & Jona-Lasinio, 2017). Moreover, intangibles are expected to increase spillover effects from ICT, for example workers need to be trained to use technologies before firms start to benefit from them (Corrado et al, 2017; Corrado, Haskel, Jona-Lasinio, & Iommi, 2013). Due to the expected interaction effect between ICT and intangible capital, this paper also includes ICT capital as a secondary focus of the paper. Furthermore, intangible capital is split into research and development (R&D) and other intangible capital (non-R&D capital). Non-R&D capital consists of factors such as design, branding, and organisational capital (Corrado, Haskel, & Jona-Lasinio, 2017).
Investments such as ICT capital and intangible capital are known to have spillover effects to productivity as knowledge is a non-rival good (Corrado et al., 2017). As discussed by Corrado et al. (2017), when there are spillover effects, it implies there is a connection between the source of growth (e.g. ICT capital, R&D capital, labour, intangible capital) and TFP.
The recent availability of data on intangible capital with data from 1995 until 2013 and the expectations about spillover effects of both intangible capital and ICT capital has sparked an interest in researching this topic. Furthermore, current literature usually investigates the market sector (Corrado et al., 2017) and there is little comparative evidence between the United States (US) and Europe available. Therefore, this paper attempts to answer the following question: “How large are the spillover effects of investment in ICT capital and
intangible capital in the United States and Europe for the market sector and manufacturing sector from 1995 until 2013?”
When most of the literature on this topic was written, there was not much (if any) data available on intangibles. However, nowadays, the INTAN-invest database1 is publicly available. Therefore, the contribution of this paper to the existing range of literature is threefold. First, the paper evaluates the importance of intangible capital and the interaction effects between ICT capital and intangibles. Second, there have been few papers written which include intangible capital and its interaction with ICT capital (Corrado, Haskel, & Jona-Lasinio, 2017; Corrado, Carol, Hulten, & Sichel, 2013). However, these papers do not use their conclusions to evaluate differences between the United States and Europe. Third,
6 this paper attempts to take a step in the direction of looking at intangibles and ICT capital at an industry level by distinguishing the market sector and the manufacturing sector.
This paper combines data from the EU-KLEMS and the INTAN-Invest2 database to arrive at a dataset containing both tangible capital and intangible capital. By estimating TFP and regressing inputs on this estimation, the paper finds evidence for spillover effects for non-R&D intangibles. The estimated elasticity for non-non-R&D intangibles is 0.11 for the market sector for both European countries and the US, whereas the estimated elasticity is 0.16 for the manufacturing sector in Europe and -0.063 for the US. Furthermore, this paper finds an interaction effect between ICT investment and non-R&D investment for the manufacturing sector. The results show that the US has lower spillover effects than European countries for ICT investment and non-R&D investment in the market sector. In the manufacturing sector, the US has higher ICT spillovers and lower R&D investment spillover effects. Lastly, the paper shows the existence of labour spillover effects in the manufacturing sector with elasticities ranging in between 0.41 and 1.01.
The remainder of the paper is outlined as follows. First, it describes existing literature on how sources of growth (human capital, ICT capital and intangible capital) strengthen TFP growth. Hence, these are spillover effects. Second, the paper explains the method used to determine spillover effects. Third, it describes the data and shows preliminary results based on correlations. Fourth, the paper describes the results. Fifth, the results are compared to the current literature and the hypotheses are discussed. Lastly, it concludes by discussing the research question.
2. Relation between production factors and TFP
This section starts off by providing some background information on TFP growth. It then discusses existing literature about the core variables for the analysis: human capital, ICT-capital and the contribution of intangible ICT-capital to TFP growth. It is important to note that effects on TFP growth are interpreted as spillover effects to productivity (Corrado et al., 2017).
2.1 Total factor productivity
A definition of TFP is necessary to avoid any confusion; the remainder of the paper uses the following definition of TFP: TFP measures how efficiently inputs are used in production. Hence, it is the part of output that is not explained by the inputs of the production process, e.g. capital and labour (Comin, 2010, p. 260). As TFP is calculated as a residual in regression
7 analyses it should not be set equal to technical change (Hulten, 2001). It consists of a lot more factors, such as institutional factors, omitted variables and measurement errors (Hulten, 2001; Stiroh, 2002). Using TFP as the dependent variable in a regression allows one to assess spillover effects from inputs into the production process. The coefficients show the contribution to TFP growth in addition to their input factor shares; hence, these are spillover effects (Stiroh, 2002).
2.2 Human capital
One of the main variables to analyse spillover effects in this paper is labour, which is represented by human capital in the literature. Some of the literature connects human capital to economic growth and technology rather than TFP growth. However, both of these relations are still highly relevant to the current discussion. Nelson and Phelps (1966) link human capital to technology. They argue that education is essentially an investment in people and that education has a higher return when an economy is more technologically progressive. This implies that, given technology is improving, education has a positive payoff. Hence, the relation between education and technology goes both ways; education improves technology and technology improves education. Nelson and Phelps (1966) also point to spillover effects from education. It is stated that if innovations produce spillovers by allowing innovations to be used by imitators, education also has a spillover effect because it stimulates innovation. While TFP growth is not solely technological progress, part of TFP growth represents technology. Hence, increases in technology also boost TFP.
Similar to Nelson and Phelps (1966), Benhabib and Spiegel (1994) also link education to technological progress. Their hypothesis states that education generates growth by increasing the capacity to create innovations and adapt to new technologies. They state that education can generate productivity because human capital determines the capacity of a country to innovate, and human capital determines the extent to which a country is able to imitate and use new technologies from technological leaders. They conclude that the level of human capital positively affects TFP growth.
However, Benhabib and Spiegel (1994) used a dataset including both developed and developing countries. Therefore, results for European countries or the United States separate might be different compared to the sample containing developing countries as well. The result could differ because both the level of human capital and technology are higher in these developed countries compared to developing countries.
8 but is invested in socially inefficient activities (e.g. a corrupt public sector). Second, there could be a declining demand for high-skilled labour. In the face of an increasing supply of skilled labour, this leads to lower private returns to education. Third, if the quality of education is low, an additional year of education barely provides skills to an individual. Importantly, this study also uses data of both developing and developed countries. Therefore, as the US and Europe are both highly developed economies, not all explanations are as relevant as they might be for developing countries.
2.3 ICT capital
In addition to human capital, ICT investment is a core variable in sources of growth analyses. Solow (1987) observed that “You can see the computer age everywhere but in the productivity statistics”, which caused a lot of discussion in the literature. For instance, Brynjolfsson (1993) focuses on the measurement problem of ICT and recommends academic researchers to look beyond the productivity measures of ICT. There are benefits of ICT that are not in the statistics and this is where intangible capital becomes important (Brynjolfsson, 1993). Examples of such unmeasured benefits of ICT are found in the interaction between intangible capital and ICT capital, as this combination can reorganise production processes more efficiently (O'Mahony & Peng, 2011).
Furthermore, there are two main opposing views on the influence of ICT capital on productivity growth. First, according to Stiroh (2002), TFP growth is exogenous in the neoclassical view because it is defined as a residual. Therefore, it is by definition exogenous from inputs. Any contribution from ICT to output stems from capital deepening rather than TFP growth. Hence, ICT capital has no influence on TFP growth in this view. Second, proponents of the “New Economy” view argue that ICT investments cause higher productivity growth (Jorgenson & Stiroh, 2000).
The rapid economic growth of the U.S. in the 1990s has not gone unnoticed by scholars. Jorgenson and Stiroh (2000) study the impact of ICT on the accelerated TFP growth. They estimate the contribution of ICT to be 0.64 percentage point from 1995 until 1998. Additionally, Jorgenson and Stiroh (2000) find that TFP growth in the 1990s is also partly due to productivity gains in non-ICT industries. The authors note that this could either indicate technological progress in non-ICT industries, which is entirely independent of ICT industries, or it could indicate spillover effects from productivity gains in ICT industries to non-ICT industries. However, it is difficult to disentangle these effects empirically.
9 capital and TFP growth, which confirms the neoclassical view as this implies there are no spillovers from ICT capital. More specifically, the effect of ICT capital on TFP growth is estimated to be negative which could be explained by adjustment costs of new technology. A note that could alter the results is that the study is performed for the American manufacturing industry. As Stiroh (2002) observes, this is not the most ICT intensive industry. For this reason, manufacturing is a good industry to study for spillover effects from ICT industries to non-ICT industries.
In contrast to the results above, Inklaar, Timmer, and van Ark (2008) find no evidence for positive spillover effects of ICT capital in the market services sector. Using TFP growth as the dependent variable, they conclude that ICT capital earns its marginal product.
Whereas most studies use industry aggregates to evaluate the impact of ICT on TFP growth, Corrado, Lengermann, Bartelsman, and Beaulieu (2008) distinguish six sectors for the US. These sectors are high-tech, construction, industrial, distribution, finance and business, and other services. Corrado et al. (2008) find that each of these sectors show different trends in TFP growth. Regarding the impact of ICT on TFP growth, the authors find that previous investments in ICT capital explained American economic growth in between 2000 and 2004. The relation between prior investments in ICT capital is used because it takes some time to adjust to technology, meaning the investment will pay off in the future rather than having immediate effects.
While the previously mentioned studies were performed for the US, Inklaar, O'Mahony, and Timmer (2005) look at the different sources of TFP growth for four European countries (Germany, the Netherlands, France, UK) and for the US. They note that TFP growth in the US accelerated a lot more than in Europe in the 1990s. Moreover, TFP growth in Europe was similar to that of the US from 1995 until 2000 only because Europe had higher contributions from non-ICT industries than the US. TFP growth of the US was driven by ICT industries. Furthermore, Inklaar et al. (2005) point out that there is substantial variation among the European countries. For example, TFP accelerated over the period from 1979 until 2000 for Germany and the UK, while it decelerated for the Netherlands and France. Lastly, the authors address important issues not yet taken into account. These involve spillover effects of ICT capital to TFP growth and the inclusion of intangible capital.
2.4 Intangible capital
10 Following Corrado et al. (2017), this paper distinguishes R&D capital and non-R&D capital. This section first discusses R&D capital in detail, then elaborates on non-R&D capital and concludes by discussing the interaction of ICT capital with intangible capital.
2.4.1 R&D capital
Many studies have confirmed spillover effects of R&D capital, which is a large part of total intangible capital, but the literature on R&D goes back further in time compared to literature on total intangibles. One of such studies is performed by Sterlacchini (1989) in which he assesses the impact of innovative activities on TFP growth in the British manufacturing industry from 1954 until 1984. He finds that the productivity slowdown of 1973 is related to R&D activities. However, in the early 1980s, R&D variables are insignificant in his regression and he states that this result indicates that institutional factors also have to be taken into account while explaining productivity growth. Naturally, it is important to consider that this study is performed for the British manufacturing industry. Therefore, it is debatable whether these findings would also hold for other industries and other countries but it definitely provides some insights in the impact of R&D on TFP growth.
Griliches (1980) assesses whether the productivity slowdowns in the United States of 1965-1973 and 1973-1978 were due to slow growth in R&D investments. He concludes that there is the possibility that causality runs from the productivity slowdown to R&D investments rather than vice versa. It is also important to note that at the time the paper was written, there was less data on R&D expenditures available so results are more likely to be subject to measurement error than results in more recent papers.
11 Lastly, Edquist and Henrekson (2017) assess the impact of ICT and R&D investments on TFP growth. Using a sample of Swedish industries, they find that R&D investment has short-run effects on TFP growth whereas the benefits of ICT capital take about seven years to materialise. They further disaggregate ICT capital and find that hardware ICT capital has complementary effects, while software ICT capital takes even longer to pay off in terms of TFP growth. The authors suggest that hardware ICT investments need additional R&D investments in order to reap the benefits in terms of TFP growth. Complementary effects such as these will be discussed in more detail later.
2.4.2 Non-R&D capital
Intangible capital is a relatively new topic in the literature and the body of literature on the subject is growing. As such, Corrado et al. (2017) perform an analysis for ten European countries and confirm spillover effects of intangible capital, for both R&D and non-R&D capital. Additionally, they provide evidence in favour of the existence of spillover effects of workforce skills.
A study of British industries, performed by Goodridge, Haskel, and Wallis (2017), uses the existence of R&D spillovers as a starting point and analyses whether other intangible assets have spillover effects as well. They find that non-R&D capital has significant correlations with TFP growth. However, these correlations are not as strong when the authors distinguish between the separate components of intangible capital: software, economic competencies and innovative property. They suggest that this result is either due to multicollinearity, as the asset types are collinear, or that it might the case that total non-R&D investment shows spillover effects but that these simply do not exist for the separate asset types.
Furthermore, Niebel, O'Mahony, and Saam (2017) study spillover effects of intangible capital; they use a sample of ten European countries and focus on the effect of intangibles on value added growth. They find that the output elasticity of intangible capital is larger than its factor share. Hence, there is evidence of positive spillover effects of intangible capital. Additionally, they run regressions using TFP as dependent variable. They find significant spillover effects of intangible capital, while the ICT and labour coefficients are negative. Lastly, they used a sectoral breakdown according to NACE 1 and find that the manufacturing industry has a larger contribution of intangible capital to value added growth compared to other industries.
2.4.3 Interaction effect
12 efficient organisation of production in the presence of both ICT capital and intangible capital (Edquist & Henrekson, 2017)
O'Mahony and Peng (2011) assess the interaction of ICT capital and training of workers, which is part of intangible capital for European countries. They find positive interaction effects of ICT capital and firm specific intangible capital, however, O'Mahony and Peng (2011) perform their analysis using labour productivity as the dependent variable. Furthermore, they note that firm-specific intangible investment might be more important in service industries and they find that levels of ICT capital and intangible capital are much higher in the US than in most European countries.
With regards to the comparison between the US and Europe, Bloom, Sadun, and Reenen (2012) find that American multinationals have management practices more suitable for information technologies in comparison to Europe. Management practices are part of intangible capital and this difference in management style could explain why labour productivity of the United States accelerated in the mid-1990s relative to Europe. However, their paper focuses on organisational capital and labour productivity whereas the present paper focuses on intangible capital as a whole and on TFP growth.
Furthermore, Chen, Niebel, and Saam (2016) also show that intangible capital contributes more to labour productivity in ICT-intensive industries. Hence, this is evidence in favour of complementary effects of the two variables. However, they only confirm this effect for organisational structures and R&D investments.
Lastly, similar to the previously mentioned authors, Corrado et al. (2017) find a positive interaction effect of total intangible investments and ICT capital. However, they note that intangible capital is largely R&D capital, organisational capital and training of workers. Therefore, there is a possibility that the result is still largely driven by these three factors.
2.5 Spillover effects in Europe and the United States
This subsection briefly describes some possible reasons for differences in spillover effects between the United States and Europe.
van Ark, O'Mahony, and Timmer (2008) discuss why Europe is lagging behind relative to the US in terms of TFP. A way to improve TFP is to make better use of innovative capabilities by restructuring the model of innovation. They also state that market services are more specific to a country than industries in the manufacturing sector, an example given is the superior management styles in the US. Capabilities such as these are difficult to transfer internationally. Moreover, Europe should increase flexibility of labour, capital and product markets in order to improve the allocation of resources (van Ark et al., 2008).
13 such as firing and hiring regulations and limitations on shopping hours. Additionally, they argue that there are barriers to entry in Europe which limit the extent of competition. However, Aghion, Bloom, Blundell, Griffith, and Howitt (2005) show the existence of an inverted U-shape of competition and innovation. They argue that there should be sufficient competition to provide incentives for innovation, but the competition should not be too tough because this could hinder innovation. Firms still need to have enough earnings from innovations in order to investment in them. Hence, restrictive regulations for Europe cannot provide the full story because some regulations can improve productivity as well (van Ark et al., 2003).
Overall, the United States is shown to have higher productivity growth than Europe. Furthermore, Europe has many restrictive regulations compared to the US. Hence, it is expected that the US has higher spillover effects from intangible capital and ICT capital. The literature discussed above provides evidence in favour of spillover effects of intangible capital, and in favour of the existence of an interaction effect between the two. However, most of these studies are performed for one region only; either the US or Europe and they lack an explicit comparison between the two regions. Additionally, this paper combines the analysis of spillover effects of intangible and ICT capital with the complementary effect between the two. Lastly, this paper contributes to the literature by comparing the manufacturing sector and market sector, as a step towards an industry-level analysis.
Building on the literature, the present paper tests the following hypotheses:
1. Investment in intangibles has spillover effects to the rest of the economy; resulting in higher total factor productivity growth.
2. ICT capital and intangibles are complements in production.
3. There are significantly stronger spillover effects of both intangible capital and ICT capital in the United States compared to European countries.
3. Methodology
14 This is done to provide a step towards an industry level analysis. However, a full industry analysis is left to future research.
3.1 Model
The method applied in this paper closely follows the approach developed by Corrado et al. (2017). The authors run regressions of capital, labour, and intangibles on value added. However, they find that this method is subject to endogeneity. They solve this issue in two ways. First, they use data for the US as instrumental variable. Second, they estimate TFP and use this result as the dependent variable. The latter is also used to test for spillover effects. In contrast to the original article, the present paper uses data for the US in the sample which means that the data for the US cannot be used as an instrumental variable. Therefore, the TFP-approach is used to control for endogeneity issues.
Corrado et al. (2017) estimate TFP growth using equation 1 below. They essentially calculate TFP by weighing the contribution of each factor according to its factor shares which implies that TFP growth is calculated assuming constant returns to scale. Note that Corrado et al. (2017) use an adjusted measure of value added which includes intangible capital (Qc,t)
rather than the standard value added. Importantly, equation 2 does not include spillover effects; it is the calculation of TFP growth assuming constant returns to scale.
∆ ln 𝑇𝐹𝑃𝑐,𝑡𝑄 = ∆𝑙𝑛𝑄𝑐,𝑡− 𝑠𝑐,𝑡𝐿 ∆𝑙𝑛𝐿
𝑐,𝑡− 𝑠𝑐,𝑖,𝑡𝐾 ∆𝑙𝑛𝐾𝑐,𝑡− 𝑠𝑐,𝑡𝑅 ∆𝑙𝑛𝑅𝑐,𝑡 (1)
, where Qc,t is the value added adjusted for intangibles, 𝑠𝑐,𝑡𝐿 , 𝑠𝑐,𝑡𝐾, 𝑠𝑐,𝑡𝑅 are the factor shares of labour, tangible capital and intangible capital respectively. Lc,t represents labour, Kc,t represents capital, and Rc,t represents intangible capital.
Using the factor shares and assuming constant returns to scale, TFP growth is calculated as follows.
∆ ln 𝑇𝐹𝑃𝑐,𝑡𝑄 = ∆𝑙𝑛𝑄𝑐,𝑡− 𝑠𝑐,𝑡𝐿 (∆𝑙𝑛𝐿
𝑐,𝑖,𝑡) − 𝑠𝑐,𝑡𝐾𝑖𝑐𝑡∆𝑙𝑛𝐾𝑖𝑐𝑡𝑐,𝑖,𝑡− 𝑠𝑐,𝑡𝐾𝑛𝑜𝑛−𝑖𝑐𝑡∆𝑙𝑛𝐾𝑛𝑜𝑛−𝑖𝑐𝑡𝑐,𝑡 − 𝑠𝑐,𝑡𝑅𝑅&𝐷∆𝑙𝑛𝑅
𝑅&𝐷𝑐,𝑡− 𝑠𝑐,𝑡𝑅𝑛𝑜𝑛−𝑅&𝐷∆𝑙𝑛𝑅𝑛𝑜𝑛−𝑅&𝐷𝑐,𝑡 (2)
where Qc,t is the value added adjusted for intangibles, 𝑠𝑐,𝑡𝐿 , 𝑠𝑐,𝑡𝐾, 𝑠𝑐,𝑡𝑅 are the factor shares of labour, tangible capital and intangible capital respectively. 𝐿𝑐,𝑡, 𝐾𝑛𝑜𝑛−𝑅&𝐷𝑐,𝑡, 𝐾𝐼𝐶𝑇𝑐,𝑡 and 𝐾𝑛𝑜𝑛−𝐼𝐶𝑇𝑐,𝑡 represent labour services R&D capital, non-R&D capital, ICT capital and non-ICT capital respectively.
15 There are two sets of regressions estimated using robust standard errors. The first set of regressions is based on equation 3, these include the interaction effect and the main variables. Equation 3 is estimated using mean centred variables in order to interpret the interaction between ICT investment and non-R&D investment. Equation 3 is split in four models, where model 1 only contains the core variables, model 2 adds the interaction and model 3 adds lagged variables. Moreover, investments in ICT and intangible capital might take time to materialise in terms of spillover effects which would be captured by the lagged variables. Following Corrado et al. (2017), the first and second lag will be added to each of the regressions. To arrive at the preferred specification, the insignificant lagged variables will be deleted. This is done to arrive at model 4, which is the preferred specification.
The most relevant coefficient of equation 3 is β7. If it is positive and significant,
complementary effects between ICT capital and intangible capital are confirmed.
∆ ln 𝑇𝐹𝑃𝑐,𝑡𝑄 = 𝛽0+ 𝛽1∆ ln 𝐾𝑐,𝑡𝐼𝐶𝑇+ 𝛽2∆ ln 𝐾𝑐,𝑡𝑛𝑜𝑛−𝐼𝐶𝑇+ 𝛽3∆ ln 𝑅𝑐,𝑡𝑅&𝐷+ 𝛽4∆ ln 𝑅𝑐,𝑡𝑛𝑜𝑛−𝑅&𝐷+ 𝛽5∆ ln 𝐿𝑐,𝑡+ 𝛽7(∆ ln 𝐾𝑐,𝑡𝐼𝐶𝑇⋅ ∆ ln 𝑅𝑐,𝑡𝑛𝑜𝑛−𝑅&𝐷) + 𝜆𝑐,𝑡+ 𝜀𝑐,𝑡 (3) ,where 𝐿𝑐,𝑡, 𝐾𝑅&𝐷𝑐,𝑡, 𝐾𝑛𝑜𝑛−𝑅&𝐷𝑐,𝑡, 𝐾𝐼𝐶𝑇𝑐,𝑡 and 𝐾𝑛𝑜𝑛−𝐼𝐶𝑇𝑐,𝑡 represent labour services, R&D capital, non-R&D capital, ICT capital and non-ICT capital respectively, 𝜆𝑡 represents lagged values of the independent variables and 𝜀𝑐,𝑡 represents the error term.
The second set of variables stems from equation 4, which include the main variables and the interaction with the US. This is done in order to interpret the coefficients, otherwise there would be too many interaction variables included in one regression. Equation 4 is split in three models; model 5 includes the core variables and dummy (interaction) variables, model 6 adds lagged variables and model 7 deletes the insignificant lags. Similar as before, model 7 is the preferred specification because it also takes delayed effect of investment into consideration.
The main coefficients are β3 and β4 as these show the estimates for European countries of
intangible capital. Moreover, β8, β9and β10 are relevant because these show the differences in
terms of spillover effects between the US and Europe.
∆ ln 𝑇𝐹𝑃𝑐,𝑡𝑄 = 𝛽0+ 𝛽1∆ ln 𝐾𝑐,𝑡𝐼𝐶𝑇+ 𝛽
2∆ ln 𝐾𝑐,𝑡𝑛𝑜𝑛−𝐼𝐶𝑇+ 𝛽3∆ ln 𝑅𝑐,𝑡𝑅&𝐷+ 𝛽4∆ ln 𝑅𝑐,𝑡𝑛𝑜𝑛−𝑅&𝐷+ 𝛽5∆ ln 𝐿𝑐,𝑡+ 𝛽7𝐷 + 𝛽8(𝐷 ∗ ∆ ln 𝐾𝑐,𝑡𝐼𝐶𝑇) + 𝛽
9(𝐷 ∗ ∆ ln 𝑅𝑐,𝑡𝑅&𝐷) + 𝛽10(𝐷 ∗ ∆ ln 𝑅𝑐,𝑡𝑛𝑜𝑛−𝑅&𝐷) + 𝜆𝑐,𝑡+
𝜀𝑐,𝑡 (4)
16
3.2 Estimation method
Corrado et al. (2017) double-difference their regression while using TFP growth as dependent variable. Similarly, this uses TFP growth as dependent variable and all inputs as independent variables. This paper deviates from Corrado et al. (2017) by applying random effects where possible, as indicated by the Hausman test (shown in appendix C.3). For models based on equation 3, a fixed effects model is estimated if a random effects model is not applicable because there are no time-invariant variables that require estimates. For models based on equation 4, a Mundlak approach is used because estimates of the US dummy variable are needed. The Mundlak model is essentially a fixed effects model with the advantage of providing estimates of time-invariant variables. The time-invariant estimates are crucial to show differences in TFP growth between the US and the European countries. For models that use the Mundlak approach, the Mundlak test is performed (appendix C.4). The validity of the Mundlak approach is confirmed in all the cases in which the Mundlak method would be preferable.
Common problems in regression analysis are the presence of multicollinearity, heteroskedasticity, serial correlation and spurious regression. Each of them is discussed in more detail below.
In case of multicollinearity, individual effects of independent variables are difficult to estimate but no assumptions of regression analysis are violated (Woolridge, 2013, p. 314). Variance inflation factors are used to determine whether there is multicollinearity. As shown in appendix C.1, all the variance inflation factors among the independent variables are all below 10 (Woolridge, 2013, p. 94). Hence, multicollinearity is not an issue in this analysis.
As the sample is a panel dataset, serial correlation is expected because the same countries are tracked over time and the growth rates are most likely dependent on the previous year. In case of serial correlation, the standard errors are biased (Woolridge, 2013, p. 341) which invalidates inference. There is a test conducted which confirms the presence of first order serial correlation for both the manufacturing and market sector. Serial correlation is accounted for by using autocorrelation and heteroskedasticity robust standard errors.
Heteroskedasticity is also a common problem in regression analysis. The presence of heteroscedasticity invalidates inference because the standard errors are biased (Woolridge, p. 96). The current model suffers from heteroskedasticity but this is also controlled for by using heteroskedasticity and autocorrelation robust standard errors.
log-17 differenced variables, which reduces the likelihood of spurious regression. Even if level values of the variables would be integrated of order 1, first differences eliminate the problem of spurious regression (Woolridge, 2013, p. 620). To confirm the absence of spurious regression, cointegration tests are performed. If the variables are not cointegrated, the regression is spurious (Woolridge, 2013, p. 622). As shown in appendix C.2, the paper uses the Kao cointegration test which confirms that there is cointegration between the variables. Hence, the regression is not spurious.
3.3 Calculation of variables
The calculation of TFP growth, as shown in equation 2, requires factor shares and a labour variable that still need to be calculated. The remaining variables are extracted from the databases directly.
Factor shares are calculated for the following inputs in production; R&D capital, non-R&D capital, ICT capital and non-ICT capital. Each of the factor shares is calculated as follows.
𝑠𝑐,𝑡𝐾𝑅&𝐷 =𝐾𝑅&𝐷𝑐,𝑡 𝑄𝑐,𝑡 , 𝑠𝑐,𝑡 𝐾𝑛𝑜𝑛−𝑅&𝐷 =𝐾𝑛𝑜𝑛−𝑅&𝐷𝑐,𝑡 𝑄𝑐,𝑡 , 𝑠𝑐,𝑡 𝐾𝐼𝐶𝑇 =𝐾𝐼𝐶𝑇𝑐,𝑡 𝑄𝑐,𝑡 , 𝑠𝑐,𝑡 𝐾𝑛𝑜𝑛−𝐼𝐶𝑇=𝐾𝑛𝑜𝑛−𝐼𝐶𝑇𝑐,𝑡 𝑄𝑐,𝑡 (5)
, where Qc,t is the value added adjusted for intangibles, 𝐾𝑅&𝐷𝑐,𝑡, 𝐾𝑛𝑜𝑛−𝑅&𝐷𝑐,𝑡, 𝐾𝐼𝐶𝑇𝑐,𝑡 and 𝐾𝑛𝑜𝑛−𝐼𝐶𝑇𝑐,𝑡 are represent R&D capital, non-R&D capital, ICT capital and non-ICT capital respectively.
As the calculation of the labour variable leads to a log difference, rather than a level value, the factor share for labour is calculated by subtracting all the other factor shares from 1. This is justified because all the factor shares have to add up to one, as this model assumes these are the only inputs in production. Table 1 shows the mean values of these factor share calculations for each input variable.
Market sector Manufacturing
Labour 75.33% 73.67%
R&D capital 2.34% 6.30% Non-R&D capital 6.95% 6.42%
ICT capital 3.28% 2.52%
Non-ICT capital 12.10% 11.10%
To calculate the labour variable, this paper applies the approach of Corrado et al. (2017). It considers the different marginal products of workers and the changes in total hours worked. The marginal product of workers is accounted for by weighing total hours worked by the share of low skilled, medium skilled, and high skilled workers. The same calculation is
18 performed for labour compensation. The following equation is used to calculate the growth rate of labour: ∆ ln(𝐿) = [(𝑠𝐿𝑆𝑤∗𝑊) (𝑠𝐿𝑆𝐻∗𝐻) 𝑊𝐻 ∆ ln ( 𝐻𝐿𝑆 𝐻 ) + (𝑠𝑀𝑆𝑤 ∗𝑊) (𝑠𝑀𝑆𝐻 ∗𝐻) 𝑊𝐻 ∆ ln ( 𝐻𝑀𝑆 𝐻 ) + (𝑠𝐻𝑆𝑤 ∗𝑊) (𝑠𝐻𝑆𝐻 ∗𝐻) 𝑊𝐻 ∆ ln ( 𝐻𝐻𝑆 𝐻 )] + ∆ln (𝐻) (5)
, where W is the labour compensation, H is total hours worked by persons engaged, 𝑠𝐿𝑆𝑤, 𝑠𝑀𝑆𝑤 , 𝑠𝐻𝑆𝑤 represent the share of labour compensation of low skilled workers, medium skilled workers and high skilled workers respectively. Similarly, 𝑠𝐿𝑆𝐻, 𝑠𝑀𝑆𝐻 , 𝑠𝐻𝑆𝐻 represent the share of total hours worked by persons engaged of low skilled workers, medium skilled workers and high skilled workers respectively.
The data section elaborates on the variables that were extracted from the databases directly, and which did not require further calculations.
4. Data
This paper uses three databases: EU-KLEMS database, the INTAN-invest database, and WIOD3. The first is used for capital and labour variables, the second database is used for data on intangible capital and the last database is used to extract shares of hours worked and compensation based on the level of education. Table A.1 shows variables used from the databases. The remainder of this section elaborates on the data and the calculations involved.
Due to missing data, TFP growth could not be calculated for Belgium, Greece, Ireland and Portugal. Hence, these countries are excluded from the analysis and there are eleven European countries4 and the United States left in the sample. Furthermore, data on intangible capital is available for the time period 1995-2013. Therefore, this is the period of interest for the analysis. Lastly, the data is extracted for both the manufacturing sector and the market sector. A more detailed description of the sectors is included in appendix A.2.
4.1 TFP growth
As TFP growth is calculated according to equation 2, an output variable is necessary in order to calculate TFP growth. However, value added in the EU-KLEMS does not take intangible capital into account. To avoid biased results, value added adjusted for intangible capital is used in the calculation of TFP growth (Corrado et al., 2017). The remainder of the variables necessary to calculate TFP growth are capital, labour and intangibles. The data underlying these variables are discussed more extensively in the next subsections.
3 Data publicly available at www.wiod.org.
19 Figure 2a and 2b show TFP growth for both the manufacturing sector and market sector from 1995 until 2013. The first panel shows that, in the market sector, TFP growth in the US had more extreme fluctuations compared to Europe. Overall, TFP growth is 2.95% on average for the US and 2.11% for the European countries. With respect to the manufacturing sector, TFP growth in the US is more volatile than in Europe. On average, TPF growth is 3.37% for the US and 1.89% for Europe in the manufacturing sector. These results show that TFP is growing faster in the US than in Europe in the manufacturing sector. TFP growth in the market sector is comparable for the US and Europe, as a difference of 0.19 percentage point is quite small. Furthermore, TFP growth is more volatile in the manufacturing sector compared to the market sector in the United States.
The volatility of the US relative to European countries can be explained by a higher number of observations for Europe. There are eleven European countries aggregated into an average, while the US is a single country. Therefore, low TFP growth in one European country might be compensated for by another country in which TFP grows rapidly. The same effect applies to the comparison between the market sector and manufacturing sector. The market sector consists of more smaller industries than the manufacturing sector, therefore, slow TFP growth might be cancelled out by other industries which experience more rapid TFP growth.
It is striking that TFP growth hit a low in the manufacturing sector is 2001, while the market sector was still steadily improving its TFP growth. This observation is the case for both Europe and the US. Finally, the crisis of 2008 is also visible in the data as TFP growth collapses in both regions and both sectors in 2008, and steadily improves again until 2011.
Figure 2b. Manufacturing sector TFP growth US and European countries
Autohor’s calculations. Data for ICT investment, non-ICT investment, total hours worked, labour compensation from Jäger (2017. Data for value added, R&D investment and non-R&D investment retrieved from Corrado et al. (2016). Data for labour compensation shares and shares of hours worked retrieved from Timmer et al. (2005).
Figure 2a. Market sector TFP growth US and European countries
Author’s calculations. Data for ICT investment, non-ICT investment, total hours worked, labour compensation from Jäger (2017. Data for value added, R&D investment and non-R&D investment retrieved from Corrado et al. (2016). Data for labour compensation shares and shares of hours worked retrieved from Timmer, Dietzenbacher, Los, Stehrer, and de Vries (2015)
20
4.2 Tangible capital
Data on capital is extracted from the EU-KLEMS database, which has data available from 1970 until 2013. The database uses the NACE 2 industry classification (Jäger, 2017). Following Corrado et al. (2017), capital is split into ICT capital and non-ICT capital.
First, ICT capital consists of computing equipment5, communications equipment6 and computer software and databases. Following Corrado et al. (2017), computer software and databases are included as part of ICT capital rather than as part of intangible capital.
Second, following Corrado et al. (2017), non-ICT capital is constructed by adding transport equipment, other machinery7 and equipment, total non-residential investment, residential structures, and cultivated assets. The effect of cultivated assets is relatively small because these only occur in agriculture. Furthermore, residential structures are only relevant for real estate activities which are excluded from both the market sector and manufacturing sector. Hence, the data that essentially makes up this variable are both non-ICT equipment and buildings and structures which are not used for residency.
4.3 Labour
As explained in the previous section, labour input consists of both the change in total hours worked and change in labour services per worker hour. Data on total hours worked by persons engaged are extracted from the EU-KLEMS database, whereas data on labour compensation and the marginal product of workers are extracted from the WIOD. Marginal product of workers is calculated by weighing both labour compensation and total hours worked by the shares based on the skill-level of workers.
Table 2 below shows the share of labour compensation and total hours worked based on skill level. The shares are very similar when comparing the manufacturing sector to the market sector. In both industries, high-skilled workers represent the only category which has a lower share in total hours worked than its share in labour compensation.
21 Market sector Manufacturing
Labour compensation: Low skilled 24.61% 24.87% Labour compensation: Medium skilled 49.15% 48.26% Labour compensation: High skilled 26.24% 26.86% Hours worked: Low skilled 29.38% 30.27% Hours worked: Medium skilled 51.54% 50.55% Hours worked: High skilled 19.08% 19.18% 4.4 Intangible capital
As data on intangibles have recently become available, it is easier to use it in estimations because proxies for intangibles are no longer necessary. Data on intangibles is extracted from the INTAN-invest database, which also uses a NACE 2 industry classification and has data available from 1995 until 2013 for EU-15 and the United States (Corrado et al., 2016). Intangible capital consists of three major components; computer software and databases, innovative property, and economic competencies (Corrado et al., 2016). To increase understanding of the structure of intangible investment, a diagram containing the different components of intangible capital is added in appendix A.3.
First, economic competencies include brand, organisational capital and training. These variables represent factors that are knowledge assets of a business but do not have intellectual property rights (Corrado et al., 2016).
Second, innovative property includes ‘entertainment, artistic and literary originals + mineral explorations’, design, new product development costs in the financial industry, and research and development (R&D). It includes the variables that could have intellectual property rights (Corrado et al., 2016). Note that R&D is included in innovative property. However this paper distinguishes R&D and non-R&D intangible capital in the analysis. Therefore, non-R&D investment includes economic competencies and innovative property excluding R&D.
Figures 3a and 3b show that the shares of the major components of intangible capital have been relatively constant over time and similar for the European countries and for the US. Furthermore, the graphs show differences between the market and manufacturing sector. The
22 manufacturing sector has a larger contribution of innovative property than the market sector, while the market sector has a larger contribution of economic competencies. Lastly, the share of computer software and databases is larger for the market sector than for the manufacturing sector.
Figure 3a. Division of intangible capital manufacturing sector
Note: European values are calculated as averages of the eleven countries. Data source: Corrado et al. (2016).
Figure 3b. Division of intangible capital market sector
23
4.5 Preliminaries
As all of the variables are transformed into log growth rates, this section describes these values rather than the levels. The section discusses both descriptive statistics and the correlations between TFP growth and the independent variables.
4.5.1 Descriptive statistics
The descriptive statistics are included in appendix B and show the log growth rates over time (1995-2013). The descriptive statistics can show trends and irregularities in the data, but they cannot provide any conclusion on spillover effects as the causal relation between input variables and TFP growth cannot be determined from these statistics.
There are a few developments in the descriptive statistics over time. First, TFP growth is higher in the period 1995-2007 than in the period 2009-2013. Second, all inputs have lower growth of investment after 2008 than before 2008. This is also the case for both sectors and some annual growth rates even turn negative after 2008. These observations could be due to the financial crisis, which confirms the validity of the break year 2008.
Despite the fact that the descriptive statistics may look quite similar for the manufacturing and market sector for a few variables, there are a couple of striking differences. First, the growth of ICT investment and labour input is lower in the manufacturing sector in the period 1995-2013. The labour variable is even negative for this time period in the manufacturing sector, which could be a first indication of negative labour spillover effects. Second, the standard deviations are higher for the manufacturing sector than for the market sector. This means that observations are more spread out in the manufacturing sector, which implies that the data for the market sector seems more reliable. Third, after 2008, growth of ICT investment turned negative for the manufacturing sector. Moreover, non-ICT investment and the labour variable show negative values for both sectors.
4.5.2 Correlations
In order to investigate the data in more detail, figure 4 below plots TFP growth against ICT investment, non-ICT investment, R&D investment, non-R&D investment and labour services. These figures show the correlations between the mean values of the variables for each country over from 1995 until 2013.
The first panel shows negative correlations between TFP growth and ICT investment growth for both sectors, but the correlation is a lot stronger for the manufacturing sector than for the market sector. The correlation for the market sector is quite weak, so there may be constant returns to ICT investment.
24 investment are in contrast to expectation, the negative correlations might be explained by adjustment costs to new capital (Stiroh, 2002). This implies that capital pays off over time, but there are also direct costs associated with the investment.
Figure 4. Preliminary results market sector and manufacturing sector
25 The third panel shows the correlations between TFP growth and R&D investment growth, which is negative for the manufacturing sector and positive for the market sector. However, the correlation coefficient of the manufacturing sector is very small so there may not be any spillovers of R&D.
The fourth panel shows correlations between TFP growth and non-R&D intangible investment growth. It indicates that there is no relation for the market sector while the manufacturing sector shows positive spillover effects.
Lastly, the fifth panel shows a negative correlation between labour input growth and TFP growth in the market sector, while the manufacturing sector shows a positive correlation. A negative labour coefficient indicates decreasing returns to labour. There might already be enough labour, which means that additional worker hours do not produce more output relative to the additional hours.
As the focus of this paper is on the difference between the US and European countries, the label of the US is bold in figure 2. The figure shows that the US is consistently in the upper-end of the sample in terms of TFP growth, while the US is in the middle section of investment growth rates in almost all variables. An exception is labour, where the US has one of the lowest growth rates of the sample, but still the highest TFP growth rate in the manufacturing sector. Taken these observations together, one could argue that this is a first indication of higher spillover effects of inputs in the US than in European countries, because a similar investment growth in inputs (or lower in the case of labour) leads to more TFP growth in the US.
5. Results
Having discussed the method by which spillover effects are estimated, it is time to turn to estimation results. First, this section presents the results in detail for the period 1995-2013. Second, the results are distinguished into two separate time periods (1995-2007 and 2009-2013) to further investigate spillover effects.
5.1 Full sample
Dependent variable: ∆ln(TFP)c, t
Model 1 Model 2 Model 3 Model 4
Market RE Manufacturing RE Market RE Manufacturing FE Market RE Manufacturing RE Market RE Manufacturing FE Estimation method ∆ ln ICT 0.00617 -0.0718*** 0.00627 -0.0647*** -0.0165 -0.00201 -0.00701 -0.0494*** (0.0199) (0.0189) (0.0212) (0.0141) (0.0293) (0.0219) (0.0213) (0.0144) ∆ ln non-ICT -0.0967*** -0.0941*** -0.0969*** -0.113*** -0.0810*** -0.104*** -0.0915*** -0.0949*** (0.0244) (0.0112) (0.0204) (0.0299) (0.0215) (0.0143) (0.0187) (0.0151) ∆ ln R&D 0.0347 -0.0844* 0.0347 -0.0555 0.0302 0.0247 0.0342* 0.00259 (0.0220) (0.0509) (0.0214) (0.0481) (0.0266) (0.0575) (0.0192) (0.0679) ∆ ln non-R&D 0.0852** 0.177** 0.0852** 0.215** 0.128** 0.0991 0.115*** 0.168** (0.0394) (0.0710) (0.0397) (0.0756) (0.0559) (0.0820) (0.0405) (0.0595) ∆ ln ICT*non-R&D -0.0134 0.679* 0.0834 0.393 0.109 0.879** (0.380) (0.348) (0.350) (0.370) (0.436) (0.400) ∆ ln labour 0.0214 0.302 0.0205 0.386* 0.0603 0.354* 0.0529 0.453** (0.0717) (0.184) (0.0858) (0.197) (0.0745) (0.185) (0.0839) (0.172) ∆ ln ICTt-1 -0.0480** 0.00820 -0.0452** (0.0213) (0.0136) (0.0204) ∆ ln ICTt-2 -0.0208 -0.0103 (0.0241) (0.0109) ∆ ln R&Dt-1 -0.0291 0.0356 (0.0201) (0.0308) ∆ ln R&Dt-2 0.0186 -0.0296 (0.0408) (0.0284) ∆ ln non-R&Dt-1 -0.00537 -0.219*** -0.217*** (0.0400) (0.0530) (0.0500) ∆ ln non-R&Dt-2 0.0481 -0.0166 (0.0520) (0.0416) Constant 0.0184*** 0.0256*** 0.0219*** 0.0191*** 0.0220*** 0.0230*** 0.0222*** 0.0195*** (0.00259) (0.00356) (0.00152) (0.000444) (0.00159) (0.00232) (0.00170) (0.000481) Obs. 210 208 210 208 186 182 199 198 Within R2 0.1132 0.2586 0.1133 0.2934 0.1627 0.2976 0.1451 0.3336 Table 3: ∆ln(TFP)c, t 1995-2013
27
5.1.1 ICT complementary effects
Table 3 below shows the results for models 1 until 4. The first model shows the core variables only, which can be interpreted as elasticities with respect to TFP. The coefficients for non-ICT investment and non-R&D investment are significant in the market sector, where non-non-ICT has a negative effect and non-R&D investment shows a positive effect. These effects persist after adding lagged variables and an interaction term for complementary effects of ICT investment with non-R&D intangible investment, which is insignificant in the market sector.
In order to interpret non-R&D investment, one needs to take the interaction effect into account. As the variables are mean centred, the coefficient on ICT investment can be interpreted as the effect of non-R&D investment at the mean of the ICT investment growth rate. Hence, given that ICT investment growth is at the average value of the sample, a 1% increase of non-R&D investment leads to a 0.11% increase in TFP. Further, the coefficients of nonICT and R&D investment can be interpreted directly. The former has an elasticity of -0.09, whereas the latter has a coefficient of 0.03.
Compared to the market sector, more independent variables show significance in the manufacturing sector. First, non-ICT investment shows a similar effect as in the market sector. Second, R&D investment turns insignificant after adding the interaction effect between ICT investment and non-R&D investment. Third, there are significant labour spillovers with an elasticity of 0.45. Fourth, both ICT investment and non-R&D investment are significant in the preferred specification, where the former has a negative effect while the latter has a positive effect. Given that ICT investment is at the average value, a 1% increase in non-R&D investment is associated with 0.17% increase in TFP. Lastly, the manufacturing sector confirms the existence of an interaction effect of non-R&D investment and ICT investment as the coefficient is positive and significant.
It is interesting that the fit of the models, as measured by the within R2, is consistently higher for the manufacturing sector than for the market sector. This could be caused by the aggregation of many different industries in the market sector, whereas the aggregate of manufacturing consists of more similar industries.
5.1.2 Comparison of Europe and US
Models 5, 6 and 7 contain a dummy variable and dummy interaction variables for the United States in order to study the differences between European countries and the US. The results are shown in table 4 below.
28
Dependent variable:
∆ln(TFP)c, t Market Model 5 Model 6 Model 7
RE Manufacturing RE Market RE Manufacturing RE Market RE Manufacturing RE Estimation method ∆ ln ICT 0.00974 (0.0206) -0.0686*** (0.0205) -0.0150 (0.0295) 0.00563 (0.0208) -0.00385 (0.0215) -0.0637*** (0.0173) ∆ ln non-ICT -0.0964*** (0.0257) -0.0958*** (0.0120) -0.0815*** (0.0260) -0.106*** (0.0136) -0.0924*** (0.0254) -0.0829*** (0.00384) ∆ ln R&D 0.0399* (0.0220) -0.0795 (0.0520) 0.0366 (0.0269) 0.0274 (0.0497) 0.0393* (0.0202) -0.0486 (0.0559) ∆ ln non-R&D 0.0765** (0.0382) 0.172** (0.0743) 0.121** (0.0578) 0.0730 (0.0930) 0.107*** (0.0400) 0.162** (0.0674) ∆ ln labour 0.0434 (0.0651) 0.359* (0.188) 0.0847* (0.0469) 0.389** (0.166) 0.0687 (0.0609) 0.412** (0.178) Dummy US 0.0173*** (0.00282) 0.0228*** (0.00788) 0.0186*** (0.00424) 0.0199*** (0.00574) 0.0174*** (0.00306) 0.0220*** (0.00738) US * ∆ ln ICT 0.0297 (0.0193) -0.0822*** (0.0265) 0.0733*** (0.0222) -0.150*** (0.0335) 0.0515** (0.0200) -0.0831*** (0.0248) US * ∆ ln non-R&D 0.0893*** (0.0332) -0.187** (0.0915) 0.00729 (0.0555) -0.109 (0.0894) 0.0277 (0.0393) -0.225*** (0.0775) US * ∆ ln R&D -0.300*** (0.0271) 0.0132 (0.0554) -0.293*** (0.0333) -0.0150 (0.0559) -0.265*** (0.0332) 0.0393 (0.0616) ∆ ln ICTt-1 -0.0432* (0.0220) 0.0110 (0.0130) -0.0420** (0.0196) ∆ ln ICTt-2 -0.0204 (0.0251) -0.00849 (0.00912) ∆ ln R&Dt-1 -0.0261 (0.0202) 0.0334 (0.0361) ∆ ln R&Dt-2 0.0190 (0.0438) -0.0288 (0.0243) ∆ ln non-R&Dt-1 -0.0142 (0.0402) -0.206*** (0.0550) -0.186*** (0.0460) ∆ ln non-R&Dt-2 0.0465 (0.0554) -0.0231 (0.0389) Constant 0.0176*** (0.00270) 0.0249*** (0.00382) 0.0177*** (0.00286) 0.0310*** (0.00353) 0.0186*** (0.00273) 0.0299*** (0.00412) Obs. 210 208 186 182 199 198 Within R2 0.1241 0.2695 0.1760 0.3140 0.1551 0.2863
Estimation method: RE stands for random effects, FE stands for fixed effects and ML stands for Mundlak method, estimates of the mean of time-variant variables are excluded from the table for space considerations. Robust standard errors are shown in parentheses. The specification is estimated using a random effects model. The countries included are the EU15 and the US, with the exception of Belgium, Portugal, Greece and Ireland. All variables are measured in log differences. Stars denote the significance level of the results: * p < 0.1, ** p <0.05, *** p<0.01.
29 R&D investment leads to a 0.04% increase in TFP in the preferred specification (model 7). However, the dummy interaction effect shows that a 1% increase in R&D investment is associated with a 0.23% decrease in TFP in the US. The results for ICT investment indicates that ICT investment has constant returns in Europe, while it has positive spillover effects in the US.
Furthermore, non-R&D investment has positive spillover effects in European countries and there is no significant difference between the US and Europe. Moreover, the US has 0.02% higher TFP investment growth than European countries. Lastly, when performing an F-test on the US dummy (interaction) variables, it shows that all the variables are jointly significant at the 1% level. Hence, there is a significant difference between the two regions.
With respect to the manufacturing sector, the preferred specification shows negative spillover effects of both ICT investment and non-ICT investment for European countries. The elasticity of ICT investment in the US is -0.15. Furthermore, R&D investment is insignificant in European countries and there is no difference between the US and Europe. Non-R&D investment has an elasticity of 0.16 in Europe, while it shows a negative elasticity in the US. Moreover, there are positive and significant labour spillovers for the full sample which show an elasticity of 0.41. Lastly, the dummy variable for the US indicates that the US has a TFP growth of 0.02% higher compared to the European countries in the sample.
5.2 Time periods
In addition to running a regression for the period 1995-2013, the data has been split in two periods. The break year is 2008, as this was the start of the financial crisis which could mean that the period before 2008 had different results compared to the period after 2008. The full consequences of the financial crisis cannot be studied using these data, as the data run up to 2013 and the crisis continued for a couple of years after 2008. This section has the same structure as the previous one and the focus is on comparing the pre-2008 sample to the post-2008 sample.
A remark about the regressions is the low number of observations available, especially for the time period 2009-2013. The standard error of the period 1995-2007 is much higher than those of 2009-2013 and 1995-2013. Therefore, the results from 1995 until 2007 may be less reliable than for either the full sample or for the time period 2009-2013. However, the within R2 is much higher in both of the separate time periods compared to the full sample, which indicates that 2008 is a reasonable break year.
5.2.1 ICT complementary effects
30
Dependent variable: ∆ln(TFP)c, t
Model 4: 1995-2007 Model 4: 2009-2013 Model 7: 1995-2007 Model 7: 2009-2013
Estimation method Market RE Manufacturing RE Market RE Manufacturing RE Market ML Manufacturing RE Market RE Manufacturing RE ∆ ln ICT -0.0421*** -0.0482** -0.0747 -0.0210 -0.0447** -0.0346** -0.0572 -0.00139 (0.0130) (0.0166) (0.0646) (0.0384) (0.0203) (0.0170) (0.0795) (0.0277) ∆ ln non-ICT -0.0887*** -0.0999*** -0.0707*** -0.153*** -0.0693* -0.107*** -0.0780*** -0.183*** (0.0301) (0.0155) (0.0269) (0.0593) (0.0419) (0.0228) (0.0254) (0.0483) ∆ ln R&D 0.0192 -0.0837*** 0.0670* 0.0625 0.0103 -0.0751*** 0.0767** 0.0823* (0.0194) (0.0252) (0.0348) (0.0638) (0.0201) (0.0281) (0.0346) (0.0469) ∆ ln non-R&D 0.0752** 0.288*** -0.0165 -0.280** 0.0810 0.263*** 0.0246 -0.309** (0.0313) (0.0809) (0.143) (0.140) (0.0516) (0.0699) (0.143) (0.141) ∆ ln ICT*non-R&D 0.620** -0.0861 0.245 0.411 (0.289) (0.223) (0.156) (0.703) ∆ ln labour -0.166 -0.553 -0.413 0.916*** -0.181 0.0483 0.283** 1.014*** (0.111) (0.543) (1.152) (0.250) (0.200) (0.160) (0.138) (0.242) Dummy US -0.00716 0.0185** 0.0268*** 0.0217** (0.00802) (0.00843) (0.00919) (0.0100) US * ∆ ln ICT 0.0636* -0.0215 1.141 -0.146** (0.0354) (0.0163) (1.296) (0.0602) US * ∆ ln non-R&D 0.238* -0.293*** -0.412 -0.132 (0.134) (0.0603) (0.427) (0.313) US * ∆ ln R&D -0.271*** 0.0569 -0.806 -0.463 (0.0882) (0.0635) (0.667) (0.365) ∆ ln ICTt-1 -0.0629** -0.0474** -0.0629** -0.0460* -0.0229 (0.0245) (0.0192) (0.0245) (0.0235) (0.0225) ∆ ln ICTt-2 -0.0407* -0.000131 -0.0407* -0.0308 -0.0293*** (0.0246) (0.0246) (0.0246) (0.0209) (0.0103) ∆ ln R&Dt-1 -0.0507 -0.0346 (0.0323) (0.0444) ∆ ln R&Dt-2 0.0532 -.0399 -0.0640*** (0.0398) (.0245) (0.0226) ∆ ln non-R&Dt-1 0.123** -0.165** -0.399*** 0.0805* -0.205** -0.363*** (0.0517) (0.0717) (0.0664) (0.0453) (0.0802) (0.0556) ∆ ln non-R&Dt-2 0.0497 (0.0404) Constant 0.0276*** 0.0290*** 0.0113*** 0.00575 0.00815 0.0269*** 0.0188*** 0.0430*** (0.00179) (0.00139) (0.00129) (0.00490) (0.00621) (0.00301) (0.00304) (0.00558) Obs. 114 113 63 62 113 123 63 60 Within R2 0.3109 0.5349 0.3015 0.5884 0.2760 0.4734 0.3134 0.6683 Table 5: ∆ln(TFP)c, t 1995-2007 & 2009-2013
31 and non-R&D investment. Table 5 shows the preferred specification (model 4 and 7) for both sectors. The complete tables containing all models can be found in appendix D.
The market sector shows a negative spillover effect of non-ICT investment both before and after the crisis in the preferred specification (model 4), which is similar to the results of the full sample. The preferred specification shows a positive coefficient on non-R&D investment before 2008. This means that, given that the ICT growth rate is at the average growth rate, a 1% increase in non-R&D investment leads to a 0.075% increase in TFP. ICT investment shows a negative and significant effect at the mean value of non-R&D investment growth before 2008. Moreover, there is a positive interaction effect between ICT investment and non-R&D investment prior to 2008 as well. However, all of these effects vanish after 2008 and the coefficient on R&D investment turns positive which indicates spillover effects.
In the period before 2008, the manufacturing sector indicates negative spillover effects of non-ICT and R&D investment. Furthermore, it indicates negative spillover effects of ICT given that non-R&D investment is at its average growth rate. The only positive spillover effects are those of non-R&D investment, also given that ICT investment is at its average growth rate. The most striking differences are that R&D investment shows constant returns rather than negative spillover effects after the crisis and that non-R&D investment shows negative spillover effects given that ICT investment is at its average growth rate after 2008 as well. Lastly, the manufacturing sector after 2008 also provides evidence in favour of positive labour spillovers, which was not yet the case in the period before 2008.
5.2.2 Comparison of Europe and US
In the preferred specification (model 7) for the market sector, European countries do not show spillover effects of R&D investment whereas the US shows a negative and significant difference to the European countries. After 2008, European countries show positive spillover effects of R&D investments and there is no significant difference between the US and Europe anymore. In contrast to the results of the full sample, non-R&D investment is insignificant in both the period before and after 2008. The dummy interaction variables show a significant difference in terms of non-R&D investment before 2008, while this variable turns insignificant after 2008.
Interestingly, the first lag of non-R&D investment is positive and significant before 2008 but it turns negative and is still significant after 2008. Furthermore, after 2008, there are positive labour spillover effects while labour shows constant returns before 2008. Finally, an F-test confirmed the joint significance of the dummy (interaction) variables in the preferred specification both before and after 2008.
