IE&B Master Thesis: The impacts of international technology transfers on total factor productivity: Empirical analysis of developing countries in Asia

IE&B Master Thesis:

The impacts of international technology transfers on total

factor productivity: Empirical analysis of developing

countries in Asia

Yuting ZHOU S1941526

Email: s1941526@student.rug.nl

University of Groningen Faculty of Economics and Business

Supervisor Drs. Gaaitzen J. de Vries

g.j.de.vries@rug.nl

Faculty of Economics & Business University of Groningen

Methodology Supervisor Prof. dr. H.W.A. (Erik) Dietzenbacher

h.w.a.dietzenbacher@rug.nl

Abstract

This paper focuses the impacts of international technology transfers on developing country‘s total factor productivity growth. We examine the extent to which OECD developed countries‘ technology transfers contributes to Asian economies‘ total factor productivity (TFP) growth, via channels of imports, Foreign Direct Investments (FDI), international migration both inward and outward as well as licensing agreements channel, using bilateral cross-country information for a sample of 13 Asian countries over the period 1990 to 2008.

Key words: TFP growth, foreign R&D stocks, international technology transfers

Content

4.1 Theoretical and empirical framework ...11

4.1.1 A simple production framework ...12

4.1.2 An illustrative model of TFP...13

4.2 Methodology applied ...15

5. Data and Measurement ...21

5.1 Sample ...21

5.2 Variables ...22

5.3 The regression models...27

5.3.1 The simplest possible model ...27

5.3.2 The extension model ...28

5.3.3 Why I use fixed effects model...29

5.3.4 Pros and cons of fixed effects model. ...33

6. Results ...33

6.1 Summary statistics ...33

6.2 Estimation results ...35

6.3 Further discussion...41

6.3.1 Groupwise heteroskedasticity and autocorrelation ...41

6.3.2 Endogeneity problem...45

6.4 Economics implications...46

7. Conclusion and discussion ...50

Acknowledgement ...52

Appendix ...53

1. Introduction

This paper aims to study the extent to which OECD developed countries‘ technology contributes to Asian economies‘ TFP growth at the macroeconomic level across countries via imports-embodied ITT, ITT from Foreign Direct Investments (FDI), migration-embodied ITT and process-embodied ITT such as licensing agreements as channels for international technology transfers. The main focus is on the role of foreign R&D stocks toward Asian country‘s productivity growth. In this paper, productivity growth has been related to the openness of economies with the associated technology spillovers. Our analysis address international technology transfers as sources of technological progress in a framework in addition to enhancement captured by the time series as formulated by neoclassical theory. Technology progress possesses special inherent features such as positive externalities which enhance its position as a driver of TFP growth. Actually TFP is often seen as the real driver of growth within economies whilst capital inputs and labor inputs are important contributors.

According to the classification in Conference Board Total Economy Database, in our paper technological developing countries in this study are 13 Asian countries, excluding Japan and South Korean. And technological developed countries contains all the OECD countries including Japan and South Korean, considering most OECD countries are technological advanced. Technology transfer refers to the arrival or the transfer of a certain technology to a country, where it has not been used before. Together with subsequent national diffusion and wider use of this technology, technology transfer works in increasing a country‘s total factor productivity (TFP) (Hoppe, 2005) International technology transfer in our research question mainly denotes the four ITT channels through trade with capital goods; through FDI; through international migration of skilled workers; through trade in technology such as licensing agreements. In selecting an appropriate measure for national productivity growth, the commonly used measure of labor productivity is inappropriate because it is influenced by capital contribution and does not accurately re flect the true technological environment of an economy (Park, Shin and Sanders, 2007). High labor productivity in a country may result from technological superiority, but it may also be simply due to greater levels of capital equipment per labor with no technological advantage. As a result, we choose Total Factor Productivity as our measure, preserving the portion due purely to technological efficiency. Based on above, I obtain the following main research question:

‘Do international technology transfers from OECD’s technological developed countries influence Asian economies’ Total Factor Productivity growth? ’

people living overseas can be a channel of transferring technology back to their home country. Licensing agreements are taken into consideration as well. Information transparency might pose a problem for data gathering. Nevertheless we try to capture international technology transfers as fully as possible. Moreover, spillover is one of our concerns in this paper, which cannot be directly measured because of some reasons such as imitation goes unreported (Hoekman, Maskus and Saggi, 2004). Given that technological progress is the result of cumulative investments in R&D and that innovative activities are concentrated in few advanced economies, developing countries through theses channels, may not only access to foreign technology, but also appropriate it. The presence of FDI brings us technology transfer with spillover effects. Do these effects influence the economic development? Previous findings are mixed. Some find that firms in sectors with relatively high MNEs tend to be more productive (Kokko, Tansini and Zejan, 1997), while others show the opposite way (Aitken and Harrison, 1999). In addition, trade necessarily encourages growth only if knowledge spillovers are international in scope. Empirical evidence on the scope of technology spillovers is ambiguous (Saggi, 2002). Through different channels, each developing country may benefit of spillovers, passive or active. Passive spillovers defined as the import of specialized capital goods from source countries which are always developed countries, so that TFP increases simply because the production process with specialized inputs. Active technology spillovers occur when learning these technologies with some purposes. That is host countries not only adopt the technology but also possess relational capability to improve domestic production and inventive activities (Crispolti and Marconi, 2005). Whereas most of the studies on this field focus on technology spillovers among advanced co untries, we would like to address our research question about international technology transfers from advanced countries towards developing countries, containing both kinds of spillovers.

2. Literature review

Solow (1956), Romer (1990), and Grossman and Helpman (1991) have emphasized the role of technology in their neoclassical or endogenous growth models as one of the source for countries‘ growth rate. However, the role of technological factors might be overlooked, because some researchers, such as Barrro (1991), assumed that technology grow at a constant rate across countries. On the contrary, Klenow and Rodriguez-Clare (1997), and Hall and Jones (1999) show, important differences exist in technology across countries.

To enhance the competitiveness, nations have to catch up the world level by reducing the technology gap (Glass and Saggi, 1998). Countries that lag significantly behind the technology frontier lack of capability to provide much of threat to real competition. Most developing countries must rely largely on imported technologies as sources of new productive knowledge (Hoekman, Maskus and Saggi, 2004) for fulfilling the technology requirements in the process of their industrialization (Kumar, 1997). Crispolti and Marconi (2005) reminded that the developing countries always, as technological followers, prefer to adopt appropriate innovations produced by advanced countries rather than engage in R&D. Eaton, Kortum, and Connolly, (2001) finds empirical evidence that domestic innovation in less developed countries consistently depends on technology imports from developed countries. In the study of Borensztein, De Gregorio and Lee (1998) the growth rate of a backward country explained by a ‗catching up‘ process in the level of technology, which depends on the extent of adoption and implementation of new technologies that are already in use in leading countries.

Frame work for improving TFP in developing countries

Total factor productivity (TFP) measures the productivity progress from measured inputs and therefore separates the component of contribution due purely to technological efficiency (Park, Shin and Sanders, 2007). Technological progress possesses special inherent features such as positive externalities which enhance its position as a driver of TFP growth, so TFP is often seen as the real driver of growth within an economy. Solow (1956) puts forward that technological progress is the main source and driver of economic growth, revealing the importance of technological progress. Until "U.S. productivity growth trend, (1961)" was published, Kendrick removed the growth of factor input and concluded that the economic growth is partly due to the growth in factor productivity which mainly consists of technical progress, diffusion of technological innovation, improvement of resource allocation, economies of scale, etc. Then Denison, a well-known American economist, classified a broad content of technological progress into the following six categories: changes in the quality of production factors; progress in knowledge; reallocation of resources; the economy of scale; policy impacts; irregular factors. These approaches held by Solow or Denison, etc. still play an important role in recent studies.

growth literature has highlighted the dependence of growth rates on the state of domestic technology relative to that of the rest of the world. International technology transfer becomes increasingly important for nations‘ technology progress; most international technology transfers are embodied in international trade in goods, services and factors and not that easy to distinguish the prices or flows reflecting technology concept (Hoekman, Maskus and Saggi, 2004). Both the acquisition of technology and its diffusion foster productivity growth (Hoekman, Maskus and Saggi, 2004).. The process of technology transfer is composed of the transfer of a systematically developed set of organized information, skills, rights, and services from a supplier organization located in a developed country (typically in the West), as invention and creation processes remain overwhelmingly the province of the OECD countries, to a recipient organization located in a developing country (typically, but not always, in one of the third-world nations) (Kedia and Bhagat, 1988).

fifty percent of all engineers and sixty three percent of skilled workers that quit multinationals, left to join local firms (Glass and Saggi,2002). As an intangible assets, technology is traded internationally either in embodied or in disembodied form (Kumar, 1997). Process-embodied technology transfers, as a vehicle for international technology transfer, it is inevitable that the choice between licensing and direct investment or which kind of technology is more difficult to transfer (Kedia and Bhagat, 1988). Licensing of an innovation has two problems in technology exchange: a licensor has private information on the value of the innovation, and the transfer of this information facilitates imitation (Gallini and Wright, 1990). In a paper of the impact of foreign direct investment on the transfer of technology to the Republic of China (Taiwan), using a macroeconomic approach based on aggregate statistical data which finds a positive and significant association between foreign direct investment and industrial productivity (Hsu and Kern,1989). National and international policy options were analyzed to encourage the international transfer of technology (Hoekman, Maskus and Saggi, 2004) Borensztein, De Gregorio and Lee reveal that FDI contribute to economic growth only when a sufficient absorptive capability of the advanced technologies is available in the host economy ( Borensztein, De Gregorio and Lee, 1998).

Helpman‘s (1995), provided empirical evidence of the positive relationship between foreign R&D and TFP in a sample of twenty-two advanced countries over the period of 1970-1990. Dedicated investments in research and development (R&D) for innovation could affect total factor productivity (TFP). Innovator will have the monopoly power over its products. Xu and Chiang (2005) show that the rate of foreign patenting is determined not only by the growth of world R&D investment, improvement of capital goods imports from innovating countries, but also by the capability of domestic countries to adopt and implement new technologies that are already in use in leading countries. Engelbrecht (1997) introduces a general measure human capital which has a connection to a country‘s education level, as well as a catch-up variable to obtain the effect of domestic and foreign R&D on TFP. Xu (2000) concludes that it is necessary for the host country to have a minimum level of human capital to benefit from foreign technology. Openness is believed to benefit from international technology transfer by import of international goods and services and enable the recipient country to keep pace with the international technology level, plus reaping the benefits of global technology spillovers from technologically ad vanced countries (Glass and Saggi 1998, Rivera-Batiz and Romer, 1991).

3. Hypotheses

For the purpose of addressing our research question, the literature on the subject of country‘s total factor productivity growth determinants was reviewed. It is these studies‘ empirical analyses that contribute to our research methodology. A brief overview of their main findings is given in this section and based on them we put forward our hypotheses.

3. 1 Hypothesis 1:

pooled time series cross section data which consists of 21 OECD countries plus Israel during the period of 1971 to 1990, find that the more open an economy is to imports, the stronger the foreign R&D capital stocks have beneficial effects on domestic productivity. Coe, Helpman and Hoffmaister (1997) focus on 77 developing countries and the 22 industrial countries and conclude that inventive activity in developed countries has a large influence on developing countries through trade, increasing with ratio of imports to GDP and the foreign R&D capital stock. On the basis of Coe and Helpman (1995), Lichtenberg and van Pottelsberghe (1998) conclude a positive relationship between international technology transfer through trade and country‘s productivity growth, indicating that the more open to trade a country is, the more likely it is to benefit from foreign R&D. Lichtenberg and Pottelsberghe use developed country data while Wang and Xu (2000) find strong empirical support for capital goods trade as a channel for international technology diffusion among industrialized countries.

Other papers investigate the developing country‘s total factor productivity growth through another perspective, due to foreign direct investment from developed countries. Crispolti and Marconi (2005) find FDI as a potential channel of international technology transfer towards developing countries affect total factor productivity levels in a panel of 45 developing countries‘ TFP react to R&D performed by United States, Japan and the European Union (TRIAD) over the period 1980-2000.

As suggest by Pessoa (2005) TFP is also affected by the authorized use of intangible, non- financial, non-produced assets and productivity rights, such as patents, copyrights, trademarks, franchise and industrial process. Following the studies of Bernstein and Mohnen (1994) and Guellec and Van Pottelsberghe (2001), a country can absorb technology from other countries in different ways, in Pessoa study, the use of intangible assets and proprietary rights and the use of produced originals of prototypes through licensing agreements are controlled by royalties and licensing fees (R&L). Pessoa (2005) try to estimate the effect of Royalties and License fees (R&L) on TFP in a panel data of 16 OECD economies in the 1985-2002 periods, in which R&L is annual payments to the exterior of Royalties and License Fees. Pessoa find negative effect of R&L on TFP.

The above findings in general support different channel of international technology transfers have the effects on TFP, positively or negatively. Given the available empirical findings and theoretical explanations, I expect to find that international technology transfers from developed countries to developing countries through imports, FDI, migration and licensing agreements has beneficial effects on productivity, not only separately but also simultaneously. As we assumed, four different channels of international technology transfers are associated to TFP. Not take them into account simultaneously may lead us to overestimate1 their impacts on TFP. This does not mean four channels do not work separately. Indeed, in the upcoming section, we will investigate the impact of each channel on TFP separately first and then all of them simultaneously. As the results, we have our fir st hypothesis as follows: Following the existing study of OECD countries, I also expect the same relationship between country‘s total factor productivity growth and international technology transfers, when taking Asian developing countries into consideration.

H1: TFP versus ITTs

International technology transfers from OECD technological developed countries positively affect Asian economies’ Total Factor Productivity growth when considering imports, FDI, migration and licensing agreements channels simultaneously.

1

3. 2 Hypothesis 2:

Four different channels, in general, enable receiving countries access to international technologies directly, but also indirect one could make the overall effect on TFP differ among countries, according to the level of human capital. Cross-country study such as Borensztein, Gregorio and Lee (1998) utilizing data on FDI flows from industrial countries to 69 developing countries during the periods of 1970s and 1980s, find that the higher productivity of FDI holds only when the host country has a minimum threshold stock of human capital. Crispolti and Marconi (2005) find developing country, for a given amount of foreign R&D, enjoys bigger spillovers the higher its educational level. Furthermore, Le (2010) find human capital has significant impact in R&D diffusion process as it enhances a country‘s capacity to learn from a foreign technology base. Therefore, given the discussed empirical findings, I expect to find the similar results in Asian developing countries. To be specified, international technology transfers from OECD countries will have positive effects on Asian developing countries‘ TFP, under the circumstances the receiving country within limits of stock of human capital.

H2: Taking human capital into consideration

The effect of technology transfers from OECD countries on productivity growth positively depends upon the level of human capital in Asian developing countries.

4. Conceptual model:

4.1 Theoretical and empirical framework

4.1.1 A simple production framework

To start with, our Total Factor Productivity (TFP) measure is based on the following theoretical models. The production function for a final consumption good Y using labor L and capital K as production inputs is assumed to take the following Cobb-Douglas form:

Yit =Ait KitαLit1-α , 0<α <1 (1)

where i is a country index and t is a time index for different years, respectively. A is regarded as Total Factor Productivity. The specified production function exhibits constant returns to scale to both production factors but diminishing returns to each production factor employed. This implies that an index of Total Factor Productivity which is defined in the following logarithmic form:

log Ait = log Yit –α log Kit− (1-α) log Lit (2)

In addition, growth accounting is a procedure to measure the contribution of different factors to economic growth and to indirectly compute the rate of technological progress, measured as a residual in an economy, which was introduced by Robert Merton Solow in 1956. Following the growth accounting method which decomposes the growth rate of economy‘s total output in to increase in the amount of capital and labor as well as increases in productivity, we obtain that:

γA = γY − αγK − (1−α) γL (3) where γA ,γY ,γK and γL are rate of growth of Total Factor Productivity, final output,

capital stock, and labor force respectively. This implies relationship between Total Factor Productivity growth and output growth.

equation (2).

4.1.2 An illustrative model of TFP

In this part, we are going to explain how Total Factor Productivity is related to intermediate goods. In Le (2010) study, TFP is positively related to a number of differentiated intermediate goods used:

log Ait = βi + η log Zit (4)

where βi is country i ‘s specific efficiency factor, and Zit is the varieties of

intermediate goods employed in country i ‘s production. While varieties of intermediate goods can be increased through international trade which is physical inputs, it also can be interpreted by ideas and technologies that captured or absorbed from international technologies. Because the international mobility of goods and labor force, foreign direct investments as well as trade with technology itself, and in Coe and Helpman (1995), a country‘s productivity level depends on domestic and foreign R&D capital stocks, following Coe and Helpman‘s idea, intermediate goods fall into two big parts, domestic intermediate goods which represented by domestic R&D capital stock SDit and foreign intermediate goods which is captured by foreign

R&D capital stock SFit . As in Coe and Helpman (1995), the domestic R&D capital

stock SDt is calculated based on the domestic R&D expenditure Rt, using following

perpetual inventory model:

SDt = (1-δ) SDt-1 + R t-1 (5)

where t is a time index representing different years. δ is the depreciation or obsolescence rate, which was assumed to be 15 percent. Rt is the domestic R&D

expenditure.

For each year data of domestic R&D capital stock, SDt, is based on the previous year

data SDt-1. However, we cannot obtain the previous year data for our beginning year.

So we use SD0 as a proxy of the previous year data of domestic R&D capital stock at

stock SD for the beginning year. SD0 measure is derived from the method of Grilichs

(1980):

SD0=R0/ (g+δ) (6)

where SD0 is the benchmark for domestic R&D capital stock SD of the beginning

year, R0 is the first available data of domestic R&D expenditure. g is calculated by

the average annual logarithmic growth of domestic R&D expenditures over the period of our data. δ is the depreciation rate.

Coe and Helpman (1995) estimate the following econometric equation which demonstrate domestic and foreign intermediate goods parts as well as import ratio,

log Ait = βi + βd log SDit +βfm ( ) log SFitm+ εit (7)

where βi is country i ‘s specific efficiency factor, the elasticity of TFP with respect to

the domestic R&D capital stock is βd

while the elasticity of TFP respect to foreign

R&D capital stock from international trade channel equals to βfm ( ) . y it is GDP of

country i at time t and ε is the error term. This is an equation that reflects the level of

imports. Specifically, is defined as the ratio of imports over GDP. The

advantage of this equation is that the elasticity varies across countries in proportion to their import shares which captures the fact that given the same composition of trading partners, a country with higher openness to imports will benefit more.

It seems that taking into account of import shares is reasonable. On the contrary, Lichtenberg and van Pottelsberghe (1998) show that the equation (7) is misspecified. To be specified, as Lichtenberg and van Pottelsberghe (1998) indicated, basing on the basic properties of logarithm, Coe and Helpman‘s estimation equation (7) can be rewritten into the following specification (8), with SFitm measure in index number

log Ait = βi + βd log SDit +βfm ( )*log (SFitm / SFi,1985 m )+ εit

= βi + βd log SDit +βfm ( )*[log SFitm – log SFi,1985m ]+ εit

= βi + βd log SDit +[βfm ( )]1* log SFitm – [βfm ( )]2* log SFi,1985m + εit (8) Theoretically, in equation (8) the elasticities [βfm ( )]1should equalto[βfm ( )]2,

because they are based on polynomial transform and they are common factor have the

same origin βfm ( ). Nevertheless, when we do regression, the estimates of [βfm

( )]1 and [βfm ( )]2 would not be the same. The reason is as follows. As we

mentioned above, import shares (mit/yit) varies across countries at the same time they

are time-varying. Although SFi,1985m is constant, the term of {[βfm ( )]2* log

SFi,1985m} cannot be integrated into country‘s fixed effects, making [βfm ( )]1and

[βfm

( )]2 turn out to be unequal to each other. In sum, import shares (mit/yit) in

equation (7) causes a misspecification. Finally we decide to leave out the term (mit/yit) in our following analysis.

The explanations of the equation of Coe and Helpman (1995) are shown above, upon which we do further investigation. However, we find the term (mit/yit) in Coe and Helpman (1995) equation causes a misspecification and then we drop this term in our upcoming study. In addition, we choose to use the bilateral data between developed countries and developing countries in our analysis. And it is more explicit to capture the effects of foreign R&D capital stocks from OECD developed countries instead of the influences from the whole world.

4.2 Methodology applied

technology transfers.

Imports Channel

Firstly, employing Lichtenberg and van Pottelsberghe (1998) the imports-embodied foreign R&D capital stock SFitm , is constructed as

SFitm = (9)

where i is a country index and t is a time index ranging from year to year, respectively . m ijt is the value of imports of capital goods from country j to country i

at time t (Crispolti and Marconi, 2005), yjt is country j ‘s GDP at time t. SDjt is

country j‘s domestic R&D capital stock at time t. In this formulation, country j‘s R&D capital stock times the ratio of country j‘s capital goods exported to country i, depicting the stock of R&D capital stock received by country i from country j.

Comparing with the import-share-weighted of importing country‘s domestic R&D

capital stocks in Coe and Helpman (1995), SFitm = , in which only

reflects the direction of import of R&D capital stock but not their imports intensity. Direction is represented by mij which denotes the flow of imports from country j to

country i. Intensity is added to Coe and Helpman (1995)‘s modified version in which the foreign R&D capital stock interacts with the proportion of imports on an sending country‘s GDP, mijt/yjt. This is better than Coe and Helpman‘s calculation because in

the equation SFitm = , the component mijt/yjt shows the openness of

exporting country. This is also the reason why we choose SFitm =

Another problem exists in the equation SFitm = in Coe and Helpman

(1995), because of which we will not practice it to calculate SFitm. In this equation, mijt represent country i‘s imports from country j at time t while mit denotes total

import of country i from the whole world. This equation assumes that country i will benefit more from international R&D spillovers if large propotion of its imports come from country j with relative high domestic R&D capital stock. In extreme circumstances, suppose all country i‘s imports come from the same source, mijt=mi, which cause SFitm=SDjt, meaning country i‘s foreign R&D capital stock benefits from country j‘s whole domestic R&D capital stock. Obviously, this case is unrealistic and might be one of the drawbacks of this equation. This kind of

drawbacks also contribute to why I choose the equation SFitm = instead

of the equation SFitm = from Coe and Helpman (1995).

FDI Channel

Secondly, based on the idea of Crispolti and Marconi (2005), the foreign R&D capital stock through Foreign Direct Investments channel, SFitfdi, is shown as

SFitfdi = (10) where i is a country index and t is a year index. fijt is the country j ‘s outward FDI

stock in the developing country i at time t. Plus the R&D intensity of country j at time t is given as SDjt/yjt. , the ratio of OECD developed country‘s R&D capital stocks on

measure of bilateral FDI through which international technology transfer has effects on developing country‘s productivity growth is appropriate to our purpose.

Migration Channel: Inward migration

Thirdly, deriving from Le (2010) research that investigates international labor movement served as a channel for international technology transfer. Two new measures of foreign technology transfer from international migration are proposed by Le (2010). One of the new foreign R&D capital stocks is embodied in inward labor mobility, called as inward migration weighted foreign R&D capital stocks, calculating as follows:

SFitg = (11)

where i is a country index and t is a year index. gijt is the stock of country j ‘s citizens

living in country i and njt is country j ‘s population at time t . The reason why the

stocks of people are used rather than flows is that stocks are less volatile than flows. To be explained, foreign workers from country j with embodied technology can convey their technology to the country i not only for one year but also for many following years, considering they can continue communicating with their colleagues in country j and absorb new technology.

Migration Channel: Outward migration

The other new foreign R&D capital stock is embodied in outward labor movement which denotes people living abroad in country j also be a channel of transferring technology back to their home country i. Outward migration weighted foreign R&D capital stocks is calculated as follows:

SFitk = (12)

where i is a country index and t is a year index. kijt is the stock of country i ‘s citizens

contact with people in home country i. Technology obtained by these people abroad, to exact extent, contribute to home country i‘s productivity growth.

Licensing agreements Channel

Fourthly, following Pessoa (2005) research, one part of foreign R&D capital stock is international technology transfer through trade in technology itself, such as licensing agreements. R&L is denoted as annual payment to the exterior of Royalties and Licenses Fees. And Pessoa find a negative effect of R&L on Total Factor Productivity. Ideally, foreign R&D capital stock by trade in technology itself is calculated as follows:

SFitr = (13)

where i is a country index and t is a year index. rijt is the payments of Royalties and

Licenses Fees from receiving country i to country j at time t.

Because of the transparency of information, we only obtain data of Royalty and license fees, payments (BoP, current US$) of country i at time t. Source: International Monetary Fund, Balance of Payments Statistics Yearbook and data files. In this case, equation (13) is not a suitable method to calculate the foreign R&D capital stock from licensing agreements.

Alternatively, we use perpetual inventory method to compute foreign R&D capital stock by trade in technology. Specifically, we consider Royalty and license fees, payments (BoP, current US$) of country i at time t as the R&D expenditure of country i. Considering the basic characteristic of licensing agreements is the payments for buying technology. Similar to the perpetual inventory model for capital stock calculation, foreign R&D capital stock by trade in technology, SFitr, is calculated based on Royalty and license fees, payments RLt.:

SFitr = (1-δ) SFi(t-1)r+ RL t-1 (14)

obsolescence rate, which was assumed to be 15 percent. RLt is the Royalty and

license fees, payments.

For each year data of foreign R&D capital stock by trade in technology, SFitr, is based on the previous year data SFi(t-1)r. However, we do not obtain the previous year data for the beginning year. So we use SFi0r as a proxy of the previous year data of foreign R&D capital stock by trade in technology at the beginning year. In other word,

SFi0r is the benchmark for foreign R&D capital stock by trade in technology, SFitr, for the beginning year. SFi0r measure is followed by equation (6):

SFi0r=RL0/ (gi+δ) (15)

where SFi0r is the benchmark for foreign R&D capital stock by trade in technology,

SFitr of the beginning year, RL0 is the first available data of the Royalty and license

fees, payments.. gi is calculated by the average annual logarithmic growth of the Royalty and license fees, payments over the period of our data. δ is the depreciation rate.

To notice, average years of schooling of the working population in each country should be taken into account, as Crispolti and Marconi (2005) suggest that the level of education is particularly important for the FDI channel. We use average years of schooling over the population of age 15 and plus as a proxy of the level of human capital.

Economic equations

TFP= (SD, SFm , SFfdi , SFg , SFk , SFr ) (16)

When we take the level of human capital (H) into consideration, the regression model will be extended. The equation (17) is based on equation (16), aim to exam whether human capital is a relative factor to influence the total impacts on TFP growth.

TFP= (SD, SFm , SFfdi , SFg , SFk , SFr, H) (17)

where TFP is Total factor productivity, SDit is our measure for the domestic R&D

capital stocks. SFitm , SFitfdi ,SFitg , SFitl , SFitr are our measures for the foreign R&D

capital stocks of each Asian developing country through the channels of imports, foreign direct investments, inward migration and outward migration as well as licensing agreements. Hit is average years of schooling over the population of age 15

and plus used as a proxy of the level of human capital. The reason of adding human capital to this specification is to investigate the influence of the foreign R&D capital stocks on productivity when the domestic labor force becomes more educated and the effect of the level of educational attainment itself on productivity. As a result, the foreign R&D capital stocks will interact with marginal propensity to the level of human capital.

5. Data and Measurement

5.1 Sample

All the results presents in this paper are based on country- level data on total factor productivity, R&D stocks and technology transfers. We study technological progress at the country- level data and the sample period for this analysis covers the years 1990 to 2008 which we could obtain from The Conference Board Total Economy Database of Total Factor Productivity Country Details. All OECD countries (including Japan and South Korea from Asia)2 as technology senders and 13 Asian countries

2

(excluding Japan and South Korea)3 as recipients constitute the sample in our study. Most of these 13 Asian countries are technological developing countries excluding Hong Kong (China), Singapore which are seen as advanced economies. 4 Finally, we obtain the panel data covers 13 Asian economies for the period from 1990 to 2008. In principle, we should have 247 observations. However, missing data exist in our dependent variable TFP so that we only obtain 238 observations.

Based on the framework description, a large amount of information needs to be collected. Part of them could be used directly in our regression model because they are numerical figures; other kind of data should be transformed before we practice them in our model. The following paragraphs are going to describe which dependent and independent variables we constructed and how to practice them in the regression model. These variables have been classified according to framework model: Firstly, the dependent variable is total factor productivity in developing countries. Secondly, independent variables are measures of country- level characteristics.

5.2 Variables

Dependent variable: TFP

Our empirical model uses Total Factor Productivity as dependent variable Ait, to

measure Asian developing country‘s productivity growth. Total factor productivity A is defined as A=Y/ (K a L1−a) where Y is the Gross Domestic Product at market price in economies by adjusting the output levels for differences in relative price levels by using purchasing power parities (PPPs). Purchasing power parity (PPP) is defined as a theory which states that exchange rates between currencies are in equilibrium when their purchasing power is the same in each of the two countries. This means given a fixed basket of goods and services, the exchange rate between two countries should equal to the ratio of the two countries‘ price level. For example, when a country‘s domestic price level is increasing, in other word inflation exists; the country‘s exchange rate must

3

13 Asian countries : Bangladesh, Cambodia , China , Hong Kong (China), India , Indonesia, Malaysia , Pakis tan, Philippines, Singapore, Sri Lanka , Thailand, Vietnam.

4

decrease in order to return to purchasing power parity. K is the stock of capital, L is the number of workers employed, a and (1− a) are the average income shares of capital and labor.

We study technological change at the country-level data and the sample period for this analysis covers the years 1990 to 2008. We directly obtain data of EKS GDP to represent Gross domestic product. EKS GDP is Total GDP, in millions of 2009 US$ (converted to 2009 price level with updated 2005 EKS PPPs) from The Conference Board Total Economy Database, September 2010. Employment (in thousands of persons) from the same The Co nference Board Total Economy Database is used to represent labor input. The domestic concept includes all workers employed domestically, but excludes any nationals working abroad. The domestic concept is in line with the production boundary for GDP, thus is the consistent measure of employment as an input.

In order to calculate the stock of capital, we have to indirectly calculate capital stock by the perpetual inventory method with a depreciation rate, using investment data. In the equation Kit = (1-ζi)* Ki(t-1) + Iit. , the stock of capital at time t is Kit, i is the gross

domestic fixed investment at time t. ζi is the depreciation rate. Gross fixed capital

depreciation rates 0.12 to represent the depreciation rate of gross fixed capital investment. Plus, initial value of the capital stock is obtained using Ki0= Ii0/(gi +ζi),

where gi is average of GDP growth rates. Ii0 is the first available value of gross

domestic fixed investment in our data. For the emerging and developing economies, we simply use 0.5 as the labor share, following the methodology notes of The Conference Board Total Economy Database.

In sum, our dependent variable natural logarithm of total factor productivity deriving from the equation log Ait = log Yit –α log Kit− (1-α) log Lit. In this dataset, Total

factor productivity (TFP) growth accounts for the changes in output not caused by changes in inputs. Furthermore, the data of Taiwan (China)‘s Gross fixed capital formation (% of GDP) is missing. So we cannot calculate out the gross domestic fixed investment and then we cannot work out the log of TFP of Taiwan (China) during the period from 1990 to 2008.5

Independent variable: Imports channel

Domestic R&D capital stocks, SDit, are calculated from domestic R&D expenditures

Rt based on the perpetual inventory method assuming a depreciation rate of 15

percent following Griliches (1990) Our figures of domestic R&D expenditures Rt

come from OECD Main Science and Technology Indicators (MSTI) database 2009-2, using Gross Domestic Expenditure on R&D—GERD (in millions of current PPP $) as a proxy. This database contains 30 OECD countries excluding Chile, but including some Asia counties: Singapore, China, Hong Kong (China), and Taiwan (China). For years in which data is missing for a particular country, we use a linear interpolation method to interpolate missing data. In addition, data of Asian countries‘ Gross Domestic Expenditure on R&D (% of GDP) is base on UNESCO Institute for

5

Statistics, Data Centre. Because this data is in the form of ratio, we have to transform them in to in millions of current PPP$ before we practice them in our analysis. In this dataset, Gross domestic expenditure on R&D (GERD) is defined as the total intramural expenditure on R&D performed on the national territory during a given period. It includes research and development funds allocated by nearly all firms, organizations and kinds of institutions6.

Independent variable: FDI channel

Foreign R&D capital stock through FDI channel is based on the OECD country‘s FDI stock to Asian developing countries which comes from United Nations Conference on Trade and Development (UNCTAD FDI Stat). We cannot find a direct measure of FDI inward stock of Asian developing countries. So we practice a indirect way by using FDI inward stock from all of the world multiplying the ratio of each OECD country in proportion of the whole world to calculate the bilateral FDI inward stock from OECD country j to Asian developing country i.

The indicator of direct investment in reporting economy (FDI Inward) is measured by Percentage of Gross Domestic Product from the Major FDI indicators (WIR 2010) in United Nations Conference on Trade and Development database. In order to obtain country i‘s FDI stock from the entire world, we have to multiple it with Total GDP, in millions of 2009 US$ (converted to 2009 price level with updated 2005 EKS PPPs) from The Conference Board Total Economy Database, September 2010. Then we multiple country i‘s FDI stock from the entire world with each OECD country‘s Direct investment abroad (FDI outward)‘s percentage of total world to obtain the bilateral FDI stock from OECD country to country i. OECD country‘s Direct investment abroad (FDI outward)‘s percentage of total world is also from UNCYAD FDI Stat and the data of Indonesia from year 1990 to 2002 are missing.

Independent variable: Inward migration channel

6

National stocks of foreign population by country of origin come from different sources: LABORSTA Labor Statistics Database‘s International Labor Migration Statistics 1990-2008; World Development Indicator database from World Bank, OECD.Stat International Migration Database. There are no complete time series of stocks of foreign population by country of origin for every country over the period 1990-2008 so we have to combine different sources to get the most obtainable data. In World Development Indicator database from The World Bank, we use International migrant stock, total by country minus Net migration by country to obtain the emigrant population to the world by country. These two indicators are 5-year data and we use linear interpolation method to supplement rest missing data. Base on this emigrant population of country; calculate out the ratios of country‘s emigration population to the world. When multiplying International migrant stock of country by the ratios above, we could obtain gijt, the data of the developing countries‘

stocks of foreign population by developed OECD countries.

Independent variable: Outward migration channel

In OECD.Stat International Migration Database 1990 to 2007, we use developed countries‘ stock of foreign population by nationality to capture kijt, the date of

developing countries‘ citizens living in developed OECD countries.

OECD developed countries‘ population is represented by Midyear population (in thousands of persons) from The Conference Board Tota l Economy Database, Output, Labor and Labor Productivity Country Details, 1990-2008.

Independent variable: Licensing agreements channel

i at time t as the R&D expenditure of country i from International Monetary Fund, Balance of Payments Statistics Yearbook and data files. Considering the basic characteristic of licensing agreements is the payments for buying technology. Similar to the perpetual inventory model for capital stock calculation, foreign R&D capital stock by Licensing agreements is calculated based on Royalty and license fees, payments RLt.. The specific calculation procedure of perpetual inventory model is

shown by equation (14) and equation (15).

Independent variable: Human capital

Human capital, Hit, is used as an input into production but also influence growth by

two mechanisms: influencing domestic technological creation and affecting absorbing foreign technology. The level of human capital is influence by different ed ucation levels: primary, secondary and tertiary education, vocational education, on-the-job-training and also working experiences. Social capabilities and health aspects of the workers could also contribute to the efficiency of workers (Hoppe, 2005). Coe, Helpman and Hoffmaister (1997) extend the Coe and Helpman (1995) regression equation by employed the secondary school enrolment ratios as a proxy for the level of human capital. Attempting to better quantify human capital level in different countries, Barro and Lee (2000) obtain education- levels and average schooling years as the main indicators of human capital. Human capital data refer to average educational attainment rates as measured by the average years of schooling of the total population aged 15 and over. They were obtained from The Work Bank Group, World Bank Education Statistics which are 5-yearly data and used as a proxy for the level of human capital.

5.3 The regression models

5.3.1 The simplest possible model

ln Ait =αi +β1lnSDit+β2 lnSFitm +β3 lnSFitfdi+β4 lnSFitg+β5 lnSFitk+β6 lnSFitr+εit (16)

In equation (16), independent variables of SFm , SFfdi , SFg , SFk , SFr are obtained by using equation (9), (10), (11), (12) and (14), respectively. Equation (16) assumes TFP is independent of the stock of human capital. Following Coe and Helpman‘s estimation equation, equation (16) above is a log- log model so β with different subscripts are the elasticities of TFP with respect to domestic R&D capital stocks, SD, and foreign R&D capital stocks, SF, through imports, FDI, inward and outward migration as well as licensing agreements channels, respectively. We apply a fixed effects model. The specific reasons why I apply a fixed effects model will be explained in section 3.3.3. αi represents the country specific dummies. ε is the error

term.

5.3.2 The extension model

In order to test whether our estimation equation (16) omitting some relevant variables, we take the impact of the level of human capital into our consideration, which is our most interested factor. Previous studies reminded human capital is an influential factor to international technology transfers, not only from the perspective of country‘s educational attainment but also from the absorptive ability of different transfers. In this context, we now introduce additional relevant variables in our framework, specifying in following equation (17):

ln Ait =αi + β1lnSDit+β2 lnSFitm +β3 lnSFitfdi+β4 lnSFitg+β5 lnSFitk+β6 lnSFitr

+β7 Hit +β8 Hit *lnSFitm +β9 Hit *lnSFitfdi+β10 Hit *lnSFitg

+β11 Hit * lnSFitk+β12 Hit * lnSFitr + εit (17) Following our analysis, we apply a fixed effects mod el. αi represents the country

specific dummies. The equation (17) extends the equation (16) by adding the level of human capital of country i in the simplest possible model. The level of human capital, H, is in level 7and interact with other foreign stocks of R&D. The interaction term Hit *lnSFitm means interaction between the level of human capital and imported

weighted foreign stocks of R&D. Similarly, other interactions terms Hit *lnSFitfdi , Hit

7

*lnSFitg , Hit * lnSFitk , Hit * lnSFitr, represent the level of human capital interacts with

FDI , inward and outward migration weighted, licensing agreements weighted foreign stocks of R&D, respectively. Equation (17) above is still a log- log model so that coefficients β with different subscripts are the elasticities of TFP with respect to relative capital stocks both domestic and foreign as well as their interactions with human capital level.

5.3.3 Why I use fixed effects model

From an economic perspective, a fixed effects model can represent individual effects result from several unobservable and time- invariant factors. In general, we pay more attention to entity fixed effects regression model, considering in this paper our observations are country- level data, which is only a small set. Furthermore, there is systematic pattern where various characteristics of host country influence international technology transfers, such as the influence of institutional variables, infrastructures and other missing country specific fac tors.

Based above, host country characteristics determine the impact of international technology transfers and that systematic differences between countries should therefore be expected. A reasonable explanation could be the ability and motivation of host country to engage in international technology transfers. It is an important factor to determine whether or not the potential spillovers will influence TFP growth.

From an econometric perspective, we compare estimation results of three different methods: OLS model, fixed effects model and random effects model by using Breusch and Pagan Lagrangian multiplier test for random effects and by using Hausman test to choose whether a fixed effects or random effects model is a suitable and valid regression model.

appropriate when the residuals are both time period and cross-section homoskedastic. But when heteroskedasticity or autocorrelation exits, OLS estimation is no longer valid. In column (b) we present the estimates of a fixed effects model which controls for unobserved heterogeneity.This heterogeneity is time invariant and correlated with independent variables. All individual specific effects were captured by differences in the intercept parameter. In column (c) estimations of a random effects model are displayed. All individual differences are also assumed to be captured by intercept parameters and uncorrelated with the independent variables, but we treat the heteroskedasticity across individual as a random component.

(a)OLS model (b)Fixed effects model (c)Random effects model

Constant ·-1.366* ·-1.366* lnSD `-0.025* `-0.011* `-0.025* lnSFm `-0.068* 0.078* `-0.068* lnSFfdi 0.025 0.003 0.025 lnSFg 0.112* 0.347* 0.112* lnSFk `-0.082* `-0.019 `-0.082* lnSFr 0.316* `-0.018 0.316* T 19 19 19 N 14 14 14 Obs 58 58 58 Adjusted R2 0.9874 0.9994

Table 1:TFP estimation results(panel data in 1990-2008 for 13 Asian countries)

Note: SD is domestic R&D capital stock; SFm is foreign R&D capital stocks embodied in imports, SFfdi is foreign capital stocks embodied in inward foreign direct investment, SFg is foreign R&D capital stocks embodied in inward migration, SFk is foreign R&D capital stock embodied in outward migration.

* Signif icant at the 5 percent level.

Magnitudes of the coefficients8 are shown in the above table 1. In principle, we should have 247 observations of 13 countries covering the year 1990 to 2008. However, missing data exist in our dependent variable TFP reduce our sample size to 238. In addition, our independent variables also have missing data. Especially, the data of domestic capital stock SD have only 69 observations during the period from 1990 to 2008. Furthermore, data of other independent variables cannot be guaranteed to show up whenever these 69 observations emerge. Finally, we only obtain 58

8

observations for regression. Because of data missing, STATA automatically neglect them when we run our regression model.

OLS estimation and fixed effects estimation use t va lue while random effects estimation shows Z value. In column (c), estimation of Random effects model does not show R-squared statistic, because when using generalize least squares (GLS)9, the total sum of squares cannot be break down into the sum squares o f the model and the sum squares of the residual. At this moment, eliminating or adding variables in a model does not always change the parameter of R-square, making it less useful. Generalized least squares (GLS) is an extension of the OLS method, that allows efficient estimation of coefficients when either heteroskedasticity or correlations, or both are present among the error terms of the model, as long as the form of heteroskedasticity and correlation is known independently of the data. As we know from table 1, the coefficients of OLS estimates equal to those of random effects estimates. It seems that heteroskedasticity of residuals or contemporaneously uncorrelated of residuals which is independent to the data could not be the factor to influence our estimation results.

In order to choose an explainable and reasonable model, we will do following relative test. The tests results are shown in table 2.

Tests Null hypotheses P-value

Breusch-Pagan

Lagrangian multiplier test H0: no random effects 0.028

Hausman test H0: difference in coefficients not

systematic 0.000

Table 2: Tests and Results

Firstly, we do a Breusch and Pagan Lagrangian multiplier test for random effects. We test for the presence of heterogeneity by testing the null hypothesis H0: variance of error term equal to zero against the alternative hypothesis H1: variance of error tem is larger than zero. If the null hypothesis does not hold, in other word null hypothesis

is rejected, and then we conclude that there are random individual differences among sample numbers. In this case, the random effects model is appropriate. On the other hand, if the null hypothesis holds, then we have no evidence to conclude that random effects are present. The p-value of Breusch and Pagan Lagrangian multiplier test turns out to be 0.028, which is less than 0.05. So we reject the null hypothesis and conclude that random individual differences are present. The random effects model seems to be appropriate.

In addition, we use a Hausman test to choose whether a fixed effects or random effects model is a suitable and valid regression model. Hausman test is to check any correlation between the error term and the independent variables in a random effects model by compare the coefficients from the random effects model to those from the fixed effects model. The null hypothesis H0: difference in coefficients is not systematic meaning there is no correlation between error term and the explanatory variables so that in larger samples the random effects and fixed effects estimate should be similar. In this case if H0 holds, both the random effects and fixed effects estimators are consistent. If H0 is rejected, the error term is correlated with any explanatory variable, the random effects estimator is inconsistent. Under the circumstances, the fixed effects estimator remains consistent, thus we should use fixed effects estimator. The chi-square statistic of Hausman test has a p-value equal to zero, leading us to reject the null hypothesis that the coefficient estimates are equal to one another. And we conclude that the fixed effects model is reasonable model for our analysis.

independent variables. Fixed effects model is more reasonable and generally explanatory.

5.3.4 Pros and cons of fixed effects model.

In this paper we choose fixed effects model to estimate. The advantage of a fixed effects model is taking the individual effects into consideration. However their drawbacks should not be ignored. Firstly, a reduction in the degrees of freedom could cause a problem. If it requires too many dummies to fulfill the specification, the model may not be analyzable and has less power to statistical tests. Secondly, too many dummies might cause multicollinearity which increases the standard errors. Furthermore, if the model contains factors that are invariant within the groups, parameter estimation may be influenced. Plus, these models might encounter groupwise heteroskedasticity and autocorrelation.

As we assumed our fixed effects model is under the assumption that no groupwise heteroskedasticity and autocorrelation. In the Results part, we will use Modified Wald test for groupwise heteroskedasticity in fixed effects regression and Wooldridge test for autocorrelation. Furthermore we will check that whether these problems influence our results. Relative responses will be shown in section 4.

6. Results

The panel data covers 13 Asian economies for the period from 1990 to 2008. Finally, we receive a time-series-cross-section data which is two-dimensional panel set of 58 observations. Although our dependent variable lnTFP has 238 observations, independent variables in our regression also have missing data.

6.1 Summary statistics

Variable Mean Std. Dev. Min Max Observations ln TFP overall 1.104993 0.4645416 0.4123012 2.256699 N = 238 between 0.4637993 0.6183855 2.01221 n = 13 within 0.1297865 0.7508483 1.450916 T-bar = 18.3077 ln SD overall 11.2737 3.031076 4.932415 16.2797 N = 69 between 2.509626 8.612667 14.49871 n = 6 within 1.959433 6.966119 13.17367 T = 11.5 ln SFm overall 8.648402 1.305322 5.822126 12.31342 N = 152 between 0.70582 7.854733 9.677652 n = 8 within 1.124753 5.650473 11.28417 T = 19 ln SFfdi overall 9.12224 1.847954 3.165566 12.9493 N = 234 between 1.653394 6.159287 11.5463 n = 13 within 0.8823354 6.128519 11.28805 T-bar = 18 ln SFg overall 6.698742 1.427467 3.066471 9.307039 N = 247 between 1.387028 3.989183 8.949159 n = 13 within 0.5045643 4.741903 7.932347 T = 19 ln SFk overall 4.255692 3.212128 -5.832056 8.701023 N = 229 between 3.380204 -4.77976 7.375842 n = 13 within 0.9798126 1.666476 8.214091 T = 17.6154 ln SFr overall 7.120559 2.159773 3.130673 11.10458 N = 156 between 2.182864 3.761886 10.09285 n = 11 within 0.7281167 5.476205 8.735661 T-bar = 14.1818 H overall 6.412777 1.927244 2.92 10.49 N = 247 between 1.898099 3.966421 9.424737 n = 13 within 0.6124589 4.827619 8.012356 T = 19

Table 3: Summary Statistics

Note: ln represents natural logarithm. TFP is total factor productivity, SD is domestic R&D

Variable Skewness Kurtosis Pr(Skewness) Pr(Kurtosis) adj chi2(2) Prob>chi2

lnTFP 0.8849147 2.798933 0 0.615 20.69 0

Table 4: Skewness/Kurtosis tests for Normality

`-joint

---Sktest: Skewness and kurtosis test for null hypothesis of variable is normality. For each variable, sktest is an overall test statistic combines skewness test and kurtosis test. Skewness/Kurtosis test requires a minimum of 8 observations. H0: variable is normality. Especially, when it is test for small sample, the significant level is 10 percent.

Before our regression, we will test the normality of our dependent variable. Normality distribution is the theoretical basis for some statistic analyses. Although most of statistic analyses do not have the mandatory requirement of normality, normality distribution is regarded as the theoretical basis especially for large sample statistic analysis. Based on Coe and Helpman (1995), the dependent variable in our analysis is the log of total factor productivity. Histogram of the log of TFP seems not normally distributed. Histogram is the visual analysis while statistic test is needed to give a reasonable conclusion. Skewness and kurtosis test is used to exam our dependent variable with the null hypothesis H0: variable is normality distributed. We choose Sktest test because it requires a minimum of 8 observations instead of a large sample. The p- value of Skewness/Kurtosis test equal to zero and we conclude that our dependent variable is not normality. Although log of TFP is not normally distributed, it still has asymptotically normal distributions if the sample size is larger enough. In simple regression, 50 observations might be enough while in multiple regression model the number might be much higher. In our dataset, lnTFP have relative large sample which are 238 observations. So we consider lnTFP as approximate normality distributions. The details of skewness test and kurtosis are shown in Table 4.

6.2 Estimation results

equation from 0.9874 to 0.9994. Plus, when we use fixed effects estimation, the results of P value for parameters jointly significance is zero and the P value for the test of equality of all the fixed effects is zero. This result is consistent with our analysis that fixed effects model should be preferred. Estimates of SFfdi, SFk and SFr are insignificant. Although insignificances of coefficients exist in regress all channels simultaneously, when we estimate each of them individually, all the coefficients are found to be positive and significant10. The results show lnSFfdi, lnSFk and lnSFr is statistically insignificant, we should not conclude that lnSFfdi, lnSFk and lnSFr are economic insignificance. So we decide not to drop these three variables in our equation.

Another possible explanation is that macroeconomic variables always have collinearity. Collinearity makes us not easy to reveal the parameter estimates and various other statistics with much precision. The correlation matrix is shown in table 5. Among the potential explanatory variables, the correlations in table 5 are less than 0.8, except the correlation between lnSFfdi and lnSFm is 0.8787. The inclusion of both imports channel and FDI channel makes the coefficient of FDI weighted R&D capital stock statistically insignificant meanwhile the p-value for parameters of fixed effects estimation jointly significance is equal to zero, rejecting the null hypothesis of joint insignificance of all the variables.. This problem can be explained by the strong correlation between lnSFfdi and lnSFm.

ln TFP ln SD ln SFm ln SFfdi ln SFg ln SFk ln SFr H ln TFP 1 ln SD -0.5522 1 ln SFm 0.3532 0.2549 1 ln SFfdi 0.3173 0.2552 0.8787 1 ln SFg 0.1565 0.1911 -0.211 -0.1893 1 ln SFk -0.8686 0.6459 -0.1849 -0.2694 -0.2444 1 ln SFr 0.5691 0.0009 0.6621 0.4665 -0.3053 -0.1408 1 H 0.8032 -0.3916 0.685 0.7206 -0.2629 -0.7315 0.6168 1 Tabel 5: Correlation (obs 58)

In addition, some of the bivariate correlations in table 5 are larger than 0.6 and less than 0.8. For example, the correlation between lnSFk and lnSD is 0.6459 and the

10

correlation between lnSFr and lnSFm is 0.6621. This might indicate that multicollinearity could be more complex between our explanatory variables.

Multicollinearity always attribute to the variables selection or dataset itself. In last paragraph, we have mentioned these independent variables we selected show joint significance. This could indicate that multicollinearity in our regression cannot be attributed to the introduction of inadequate variables into model. On the other hand, the common trend of economic data.

Multicollinearity cannot lead to invalid estimation 11 . However, when multicollinearity is serious, the variances of estimation results might not small, even increasing sharply with the improvement of multicollinearity. When an explanatory variable has strong correlations between other explanatory variables, this indicates the regression model has strong collinearity. At this moment, R-squares access to 1 and the variance of this explanatory variable turns out to be very large. The increase of variance leads the instability becomes high and then possibly causes the abnormal situations of the signals of coefficients or estimates.

Variable VIF lnSD 7.5 lnSFm 7.11 lnSFfdi 8.06 lnSFg 3.45 lnSFk 7.92 lnSFr 2.2 Mean VIF 6.04

Table 6: Variance Inflation Factor (VIF) to test multicollinearity

In table 6, we use Variance Inflation Factor (VIF) to test multicollinearity. Actually, multicollinearity always produces unfavorable effects on multiple linear regression through inflations of the variances of estimation results. Depending on the degree of inflation, we can find out whether there is multicollinearity and which explanatory variable causes. Variance Inflation Factor (VIF) is defined as VIF (bk) = 1/ (1-Rk

11

square). k denote the kth explanatory variable. Usually, whether VIF is larger than 10 is the Criteria for multicollinearity. In other words, if R square is larger than 0.9, indicating 90 percent of kth explanatory variable can be explained by other explanatory variables, multicollinearity should be addressed. In table 6, all the variance inflator factors are less than 10. In this case, multicollinearity does not cause adverse influences on our regression.

In extension equation (17), imports of capital goods, as a mechanism to access to international technology, are expected to have positive effects on TFP, when interacting with the human capital stock. Hence, it indicates a positive signal for coefficient on new interaction Hit *lnSFitm. Similarly, we expect the improvement of the level of human capital will increase the benefits from international technology transfers. So we expect a positive signals on the new interaction coefficients for Hit*lnSFitm, Hit*lnSFitfdi, Hit*lnSFitg, Hit*lnSFitk and Hit*lnSFitr, thinking of higher level of absorptive ability makes the technology spillover easier to occur.