Changing determinants of FDI location choice in China: a panel data analysis

(1)

Changing determinants of FDI location choice in China: a panel

data analysis

Word count: ±14.900

Student: Charlotte Blankenberg, 11393017 Supervisor: D. Veestraeten

Second reader: N.J. Leefmans Faculty of Economics and Business University of Amsterdam

(2)

2

Statement of originality

This document is written by Student Charlotte Blankenberg who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used creating it. The faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

Executive summary

Since opening up to foreign direct investment (FDI) in 1978, China has been a large recipient of FDI inflows. Because of the rapid expansion of China’s FDI inflows, this phenomenon received increasing attention in the academic world. Previous research mainly focuses on identification of determinants of FDI, rather than looking into identifying trends in FDI patterns. Therefore, the goal of this research is to show how important FDI determinants in China changed over time. Using a panel data random effects model over the period 1997-2015, including 31 Chinese provinces, the main determinants are identified for three different samples: entire China, East China and West China. A significant determinants in all samples are market size and trade openness. In addition, it appears that the global crisis starting at the beginning of 2008 affected FDI in China negatively. Using rolling regression analysis to observe trends, it can be concluded that coefficients of determinants are highly unstable over time, which is an interesting complement to the existing literature. Instead of focusing on which determinants are important within a specific time frame, future research in this field should focus on how FDI inflows in transition countries develop over time in order to make accurate predictions. Extended knowledge about this could be used to design more specific policies aimed to attract FDI.

(3)

3

1. Introduction

Since the adoption of the open-door policy in 1978 in China, there has been a large increase of foreign direct investment (FDI) inflows into China (Belkhodja, 2016). Opening up to FDI inflows was possible because of the implementation of two important laws: the Law of the People’s Republic of China (PRC) in 1978 permitted foreign investment in the form of joint ventures and in 1986 the PRC Law on Foreign Enterprises permitted foreign investors to wholly own firms in China. Due to with more advantageous tax regulations, growing GDP and expanded freedom for foreign companies, China became a favorable business environment with steadily increasing FDI inflows. It is visible in Figure 1 that since 1982 the FDI inflows increased from almost zero to 280 billion dollars in 2011 (Belkhodja, 2016).

Figure 1: FDI inflows into China, current USD. Scale in billion USD. Note: Adapted from https://data.worldbank.org/indicator/BX.KLT.DINV.CD.WD?locations=CN. Copyright 2017 by The World Bank Group. Reprinted with permission

Because of this rapid expansion, China’s FDI inflows received increasing attention in the academic world (Hu & Owen, 2005; Salike, 2016). Research in the field of FDI is not new: since the 1950s many theories have been developed to explain these international investment flows between countries (Denisia, 2010). Until now, most research into the Chinese FDI inflows focused on the identification of determinants in order to find motives for companies to invest in China (Hu & Owen, 2005; Salike, 2016). However, the academic world is lacking information about how determinants of FDI inflows in China changed over time. Markusen (2000) developed a theory discriminating between two types of FDI, vertical and horizontal FDI, which has now been broadly accepted. Vertical FDI is better in explaining investment in developing countries, because the foreign company uses the low factor prices of the host country. On the contrary, in more developed countries horizontal

(5)

5 FDI inflows are more common: starting local production to serve the local market might be a solution to for example high transportation costs (Popovici & Calin, 2014). Different stages of economic development create different opportunities for foreign direct investors leading motives for FDI to change (Dunning, 2003).

Also in China recent trends in FDI inflows indicate changing motives for investors (He & Zhu, 2010; KPMG, 2016). Initially, the main motivation of foreign investors in China was to benefit from low production factor prices, specifically from low labor costs. The largest share of FDI inflows was into the labor intensive manufacturing sector, with Hong Kong and Taiwan as largest investors in this industry (Ng &Tuan, 2006). At a later stage, the inflow of manufacturing FDI stagnated and in addition the traditional labor-intensive manufacturing was replaced by technology-intensive industries (Pingyao, 2002). Competitive advantage for foreign firms in the labor intensive manufacturing sector declined due to increasing wages: Morrison (2014) reported that the annual average wage increase over the period 2000-2013 was 11 percent. This trend is visible in Figure 2: until 2005 FDI in manufacturing accounted for 70 percent of total FDI inflows, but this declined to only 31 percent in 2015. FDI in manufacturing FDI has been replaced by FDI inflows into the service sector, especially in real estate and financial intermediation. Other upcoming sectors are healthcare, automation and e-commerce (KPMG, 2016).

Figure 2: Share of FDI inflows into the largest three sectors. Note: Adapted from http://data.stats.gov.cn/english/easyquery.htm?cn=C01. Copyright 2014 by National Bureau of Statistics of China. Reprinted with permission.

(6)

6 Next to the trends in the sectorial dimension of Chinese FDI inflows, the changing composition of source countries is important in explaining FDI inflow determinants (OECD, 2000). FDI inflows are mostly coming from Hong Kong, Macau and Taiwan (HMT). At the beginning of the 21st century there was an increase in FDI inflows from other, mainly OECD, countries (Pingyao, 2002). This increase stimulated the increase in horizontal FDI further because Western Countries are more involved in horizontal FDI in China (Whally & Xian, 2010). However, after 2007 the share of HMT rose again (Figure 3). Although China’s FDI inflows did not experience a drop in FDI inflows as large as in other countries, it experienced a decrease of FDI inflows from several countries other than HMT, for example the US (Kekik, 2011).

Figure 3: Share of FDI inflows from different source Countries. Note: Adapted from http://data.stats.gov.cn/english/easyquery.htm?cn=C01. Copyright 2014 by National Bureau of Statistics of China. Reprinted with permission.

FDI inflows in different sectors and from a changing source country group leads to the expectation that the motives of investors over the years have changed and therefore other determinants became relatively less or more important. For example: low wages decreased in importance due to the decreasing share of FDI into labor intensive manufacturing. When researching determinants of FDI in China, it has to be taken into account that FDI inflows into the Eastern (coastal) and Western (inland)

(7)

7 region in China are distributed very unequally (Madariaga & Poncet, 2007). The coastal region was growing at a rapid pace, leaving the Western region behind and causing an economic growth gap between the two regions (Lui, Luo, Qui and Zhang, 2014). However, Huang and Wei (2016) report that the Western and Central regions have had a faster increase of FDI inflows then the Eastern region since 1990: the share of FDI inflows in the Western region increased from 1.1 percent in 1990 to 8.6 percent in 2010. Because of the uneven development of FDI inflows in China, FDI determinants of the two regions are often analyzed separately (Boermans, Roelfsema, & Yi, 2009).

Aiming to fill the gaps in the existing literature, this research answers the following question: “What are the determinants of FDI inflows in entire China, East China and West China and how did their importance change over time?” Using a panel data method and rolling regressions, 31 provinces in China are researched over the period 1997-2015. Gaining more knowledge in this topic provides new understanding about how FDI flows into China develop over time, which is an extension of the current empirical literature about this topic. Furthermore this research could be important for practical purposes: to implement or expand policies to attract FDI in specific regions. Some policies might be more effective in a certain period than in others, because the preferences of investors change over time.

The structure of the thesis is as follows: theoretical models for explaining FDI inflows are discussed in the second chapter. The third chapter provides more information about empirical results from previous research, giving insights in which determinants were important. The empirical strategy, including the panel data analysis, rolling regression methods and data are explained in the fourth chapter. Lastly, the last two chapters include the estimation results and the main conclusions.

(8)

8

2. Theories of FDI

This chapter summarizes the main theories for explaining FDI. The first part of this chapter discusses the main general theories of FDI, developed since the Second World War. The second part of the chapter discusses more recent theories. Additionally, in the context of this research it is evaluated which theories are applicable to the Chinese FDI inflows.

2.1 General theories of FDI

After the Second World War, globalization became a growing economic phenomenon. The increasing globalization has led to an increase in research on FDI and the driving forces behind it. FDI takes place when a company invests in foreign business (in the host country) while having a controlling motive, thereby becoming a multinational enterprise (MNE). Actual involvement in the management of the foreign business distinguishes FDI from foreign portfolio investment (Blonigen & Piger, 2011). All theories about FDI should explain why a firm prefers actual investment into a foreign business, rather than just partly outsourcing its production process to a foreign firm. Economists formulated several theories describing these motives of MNEs. These FDI theories can be broadly divided into four categories: the product cycle theory, the exchange rate theory, internationalization theory and the eclectic paradigm (Denisia, 2010).

Production cycle theory

The first assumption of the product cycle theory, developed by Vernon in 1966, is that either a threat or a beneficial situation in the market is the basis of a firms’ motivation to innovate (Vernon, 1979). This stimulus for innovation originates from the home market, for example the company must overcome competition or wants to exploit opportunities here. The home market is therefore also the preferred location to employ this innovation (Vernon, 1966; Vernon, 1979). Innovation is the first of four stages: innovation, growth, maturity and decline of the production cycle. In the second stage (growth) the company expands internationally. Because of the competitive advantage gained due to the innovation in the first stage, the firm will start production in other countries. This international expansion is beneficial because of higher profits due to a larger market. In addition, the increased international market share prevents copying of the product by foreign firms. The international expansion goes on until the third stage (maturity) is reached, which is the case when foreign companies in other advanced countries will start to produce the product locally. The firm will stop growing because of declining competitive advantages. In the fourth stage the company loses market

(9)

9 share to international competitors (Vernon, 1966). The theory of Vernon (1996) was mainly used to explain US FDI into Western Europe in the period 1950-1970, but lost a large part of its predictive power in the years after (Vernon, 1979, Denisia, 2010).

Exchange rate approach

Among others, Froot and Stein (1991) formulated and empirically tested a theory in which FDI was explained by imperfect capital markets. For the US they show that FDI inflows of a country are related to a depreciation of its currency and that an appreciation of the currency is related to a decrease in FDI inflows (Froot, & Stein, 1991). However, when there are FDI flows between countries in both directions this theory only explains the FDI flow into the depreciating currency. It does not explain the opposing FDI flows that are present: from the country with the depreciating currency into the country with the appreciating currency (Denisia, 2010).

Internalization theory

Internalization means that a company rather handles the process within the company than outsourcing it to another (international) company. This theory focusses on strategic decision making of the investing firm and demonstrates in which situations individual decisions might differ (Rugman, 2010). The first economist publishing a theory of internalization was Ronald Coase in 1937. An important assumption in internalization theory is that internalizing is often done because a company does not want to share specific knowledge. To grow as much as possible, a company chooses a location to produce at low costs and keeps expanding until the costs of internalization exceed the benefits (Buckley, 2016). Although there are some unanswered questions when applying internalization theory to international business, internalization theory predicted the FDI inflows in China and other emerging countries at a relatively early stage: production costs in these countries were low and property right legislation was not highly developed. For this reason, internalization of the process was a better alternative compared to outsourcing (Buckley, 2016).

Eclectic Paradigm

Although the eclectic paradigm has some similarities with internalization theory, the aim of this theory is to explain outward FDI using not one, but a set of three variables (Dunning, 2000; Rugman, 2010). According to Dunning (2000), the three key drivers behind FDI are ownership advantages, locational advantages and internalization advantages. This theory is referred to as the OLI

(10)

10 theory (Dunning, 2000; Jadhav, 2012). Ownership advantages describe why companies might engage in FDI, such as competitive advantages. When the company has large competitive advantages, owning the business might be preferred over outsourcing production to a foreign partner company. Moreover, there are locational advantages that might attract companies to choose for foreign production. For example, locational advantages are present when an investing firm can obtain specific endowments at lower costs when operating abroad, which could be the case for natural or produced endowments, or endowments that are difficult to transport. Internalization is advantageous when a company wants to avoid sharing knowledge with a foreign firm, as explained previously. Owning a foreign company should be more beneficial than providing a local company a license and starting for example a franchise or technical service agreement (Dunning, 2000).

FSA-CSA matrix

In the early 1990s, Rugman develops the firm specific advantages and country specific advantages (FSA-CSA) matrix as a dynamic approach to explain FDI, based on the product cycle theory. The matrix has two dimensions, one for the location specific advantages and one describing the firm specific advantages. This theory compares the relative advantages of one host country over another and of the firm over other firms and hence uses comparative advantages. In different stages of the production cycle relative importance of firm specific against location specific advantages might vary. For example, for companies in a later stage of the production cycle, location specific advantages are more important, because the firm specific advantages might be weak. However, when a firm has a strong comparative advantage in an earlier production cycle stage, location advantages might be less important (Popovici & Calin, 2014).

Review of general theories

The four theoretical views can explain some of the Chinese FDI trends, such as the initial interest in low cost production factors, but are unable to explain the recent trends. With the introduction of the product cycle theory in 1966, Vernon (1966) describes how FDI flows can develop over time. He states that eventually other developed countries will start producing locally. His theory was able to explain FDI between the US and West-Europe during a specific period, but lost its predictive power later (Denisia, 2010). The internationalization theory and the OLI theory, explains the fast expansion of the Chinese FDI inflows. FDI increased because of initially low production costs, lack of property right protection and locational advantages (Buckley, 2016; Dunning, 2000). In order to explain the recent trends in Chinese FDI inflows, a theory describing the development of FDI

(11)

11 over time, such as the FSA-CSA matrix, might be more accurate. However, the FSA-CSA matrix still only describes the preferences of an individual company over time and does not give any explanation for possible changes in FDI patterns in a FDI receiving country (Popovici & Calin, 2014). Thus, it seems that the general FDI theories are able to explain the initial grow in FDI inflows, but are not able to explain development of Chinese FDI patterns.

2.2 Newer theories of FDI

The more recent developed FDI theories are generally based on the assumptions of increasing returns to scale and imperfect competition due to firm-specific advantages. Following the internalization and OLI theory, newer FDI theories describe FDI choices according to a general equilibrium model in which benefits must outweigh the costs (Jadhav, 2012). An important, more recent development in the theory building of FDI is the identification and acknowledgement of the so called ‘soft’ determinants, such as political environment, cultural factors and policies. Additionally, Markusen (2000) introduced the distinction between horizontal and vertical FDI (Popovici & Calin, 2014). The most important recent contributions to FDI theories are discussed below.

New Economic Geography

The theory of New Economic Geography (NEG) exists since 1991 and has been introduced by Krugman (Krugman, 1998). Although clustering of economic activity can be observed, at some point geographical concentration stops. This phenomenon is what can be explained using the NEG theory. There are centripetal and centrifugal forces explaining why economic activity tends to concentrate in certain areas and why economic activity spreads in others respectively. Forces that cause companies to locate their foreign business near other (foreign) business are called centripetal forces. The agglomeration effect depends on the demand side, because the assumption is that consumers favor a large variety of goods, which leads to product differentiation and monopolistic competition (Rosser, 2011). Because of the large local market, the production of intermediate goods becomes less costly, there is for example a large labor market and firms might be able to benefit from positive externalities such as knowledge spillovers. However, there are also centrifugal forces that cause companies to locate their firm further away from clustered industries. This might be advantageous when there are, for example, immobile production factors, or when a company wants to avoid diseconomies present in clustered areas, such as high land rents. NEG theory describes the regional distribution of FDI and explains agglomeration and the historical development of cities fairly well (Krugman, 1998; Rosser, 2011).

(12)

12 Institutional theory

As an addition to the importance of economic factors of the host country, which are described in the OLI theory, Dunning (2003) already acknowledged the importance of a new dimension: political and social infrastructure. In order to explain FDI patterns in transition countries, it appeared that these ‘soft’ determinants, such as institutional quality, were also important. The market of those countries had to develop strong institutions in order to adapt to a market economy (Popovici & Calin, 2014). Du, Yi, and Tao (2012) show that FDI in China is mainly located in regions having stronger institutions. FDI is also related to cultural factors: it is more difficult for the foreign investor to operate in a country with a culture that deviates much from the culture in its home country. This is why Du, Lu and Tao (2012) found a moderating effect of culture affecting the relation between institutional quality and FDI. When the cultural difference between the source and the host country was large, institutional quality was seen as more important by the foreign investor (Du, Lu, & Tao, 2012). When researching US companies in China, Du, Lu and Tao (2008) find four significant factors of institutional quality explaining location choice: contract enforcement, protection of property rights, low government intervention in business activities and low corruption rates.

Horizontal and vertical FDI

The theory of Markusen (2000) states that countries are likely to be involved in FDI in two situations: when their economies are similar with respect to size and factor endowments, or when one country is smaller but abundant in skilled labor. FDI between two similar economies (often developed economies) is called horizontal FDI. International firms undertaking this form of FDI are willing to service the local market and produce the same products in different countries (Markusen, 2000; Popovici & Calin, 2014). Vertical FDI is taking place when the international firms split up the production process geographically, in order to benefit from low production costs (Markusen, 2000). To benefit from vertical FDI, a minimum share of skilled labor must be present in the host country and transportation costs cannot be too large (Popovici & Calin, 2014). When an FDI recipient country develops over time, its FDI composition changes with the expansion of its market: first vertical FDI determinants are more important and later the share of horizontal FDI grows (Popovici & Calin, 2014).

(13)

13 Review of newer theories

More recent additions to FDI theory are relevant extensions to explain certain patterns in China. NEG theory helps to explain clustering in the Chinese Coastal area. The phenomenon of clustering is an interaction between centripetal and centrifugal forces described by Krugman (1998). There are several diseconomies from clustered industries and therefore companies may choose an FDI location outside the clustered area (Krugman, 1998). This could explain why FDI in West China is increasing. Additionally, institutional development could attract FDI in West China. Du, Yi, and Toa (2012) showed that FDI flows towards regions with stronger institutions, especially in an early stage of economic openness in transition countries. Markusen (2000) adds the distinction between horizontal and vertical FDI and explains how FDI can develop over time: when a country is less developed, vertical FDI is more important, but in more developed economies, horizontal FDI is more prominent (Popovici & Calin, 2014). This explains the Chines trend on the sectorial dimension: from manufacturing to service (KPMG, 2016).

(14)

14

3. Literature review

This chapter summarizes empirical findings with respect to FDI in China. First, empirical findings with respect to the main determinants of FDI flows in China are discussed. The second part of this chapter focusses specifically on the differences between East and West China.

3.1 Determinants of Chinese FDI inflows

Market size

Market size is identified as an important FDI determinant in various researches with different methodological approaches (Ali & Guo, 2005; Boermans, Roelfema, & Yi, 2009; Ho, 2004). A common approach for researching FDI determinants is panel data analysis. Ho (2004) uses this methodology to find out whether market size was an important determinant for attracting FDI in entire China and the province Guangdong, thereby distinguishing between different sectors. Gross domestic product (GDP) and gross regional product (GRP) respectively were used as proxy for market size. Ho (2004) found that the level of GDP has a statistically significant effect on inward FDI, for China as a whole and for the province in the period 1997-2002. Ali and Guo (2005), who analyzed the responses to questionnaires of 22 US MNE’s undertaking FDI in China, found that market size and growth potential of the Chinese market was the main important reason to invest in the country.

Boermans, Roelfema, and Yi (2009) used a research methodology that is not widely used among economists: instead of using a pre-selected set of variables he used a factor analysis, in which determinants can be identified without prejudice. Their dataset included the period 1995-2006 and the 31 Chinese provinces and they find that market size is one of the two most important factors in FDI attractiveness. As a critical note to the previous findings Belkhodja (2016) states that the effect of market size on FDI attractiveness is not always positive. When performing a more detailed analysis for FDI in the manufacturing sector, using a database of 1218 manufacturing firms involved in FDI in China from all over the world, a negative relation between market size and FDI is found for firms originating from the US and Europe. Belkhodja (2016) explains that investors form the US and Europe associate higher gross regional product with higher standards of living and higher costs, therefore these manufacturing firms are reluctant to invest in a region with a higher GRP. Interestingly, this relation was not significant for investors from Japan and South Asia (Belkhodja, 2016).

(15)

15 Labor costs

Wages are mostly used as proxy for labor costs. The effect of wages on FDI appears to be ambiguous when evaluating previous research. With a panel data analysis including 31 regions in China, Salike (2016) finds that average wage is a significant determinant, which is negatively related to FDI, but that the coefficients are very small. The small coefficient could be explained, because higher wages are at the same time attractive and unattractive to the foreign investor. Local labor costs are lower if wages are low, but when a company is willing to serve the local market, higher average wages could be attractive because they are an indication for higher local purchasing power (Salike, 2016). Boermans, Roelfema, and Yi (2009) using the factor analysis approach over 1995-2006, mentioned previously, identified wage as second most important determinant for FDI attractiveness of Chinese provinces. Low wages are attractive to foreign firms that want to benefit from low regional production costs, especially when supported by a good local infrastructure.

Hu and Owen (2005) and Liu, Daly and Varua (2012) perform more detailed analyses on the effect of wages on FDI. Hu and Owen (2005) research the effect of wages on FDI using provincial data for the 1990s, thereby discriminating between the coastal and inland region and the source country of the foreign investor. They show that indeed wage levels have various effects on FDI attractiveness: in the coastal area, they found that wage levels are negatively related to FDI for investors from HMT, but investors from other countries are indifferent. For this last group, wages are not significantly predicting FDI stock (Hu & Owen, 2005). Liu, Daly and Varua (2012) highlight a sectorial difference in their research discriminating between four Chinese regions over the period 2001-2009: labor costs are negatively related to FDI in low-technology manufacturing in the coastal area, but they find no significant relationship between wages and FDI in high-technology manufacturing in all regions.

Labor quality

Better labor quality in the host country can be attractive for foreign investor for several reasons. Firstly, better human capital boosts productivity and therefore the profit of the company. Secondly, better human capital is an indicator for higher living standards and higher local demand (Salike, 2016). In China, a rapid increase of the availability of high skilled labor is observed. Li, Whalley, Zhang, and Zhao (2011) describe an increase in the number of undergraduates and graduates of 30 percent each year since 1999, and conclude that this increase in high skilled labor is especially advantageous for high-technology manufacturing (Li, Whalley, Zhang, & Zhao, 2011).

(16)

16 Salike (2016) studies the effect of human capital using a very detailed method. Six factors are used to measure human capital: human capital productivity, human capital endowment, geography and human capital utilization, human capital support and human capital health. Four factors (endowment, utilization, demography and health) were positively and significantly related to FDI inflows. Hence, Salike (2016) confirms that there might be a positive relationship between human capital and FDI attractiveness.

However, when performing a more aggregate research into the effect of labor quality on FDI inflow, this positive relationship is not found (Gao, 2005). Gao (2005) used percentages of employees’ education levels (primary school, junior secondary school and senior secondary school) to define labor quality. In the panel data research for the period 1996-1999, Gao (2005) finds a negative and insignificant relation between labor quality and FDI, which could be explained by the fact that this analysis has been done on a very aggregate level (Blonigen, 2005; Gao, 2005). In addition, the fact that a negative relationship was found could be due to early period of investment; for more recent years, the relationship between education and FDI inflows might be more important.

Trade openness

Trade openness is frequently defined as the ratio of trade to GDP (Zhang & Zhang, 2003). Zhang and Zhang (2003) summarize economic developments in China and show that trade openness in whole China has increased from 0.05 to 0.30 over the period 1978-98. During this period, trade openness in East China increased by 26 percent more than in West China (Zhang & Zhang, 2003). Kinoshita and Campos (2003) study determinants of FDI in transition countries, using both panel data analyses and GMM and find that trade openness has a significant and positive effect on FDI stocks. More FDI activity is related to relatively more international trade, because companies must transport (intermediate) goods. Liargovas and Skandalis (2011) who also find a positive relationship between trade openness and FDI in developing countries, explain that trade openness is especially important for firms that are export orientated, while high tariffs are more likely to attract FDI with a market seeking nature (tariff-jumping). This is because if there are low barriers to international trade, it is easy to benefit from other regional advantages and later transport the (intermediate) goods. However, when trade barriers are high, transportation of goods to serve the local market is expensive. When this is the case, a firm might be more motivated to start horizontal FDI (Liargovas & Skandalis, 2012) Besides this, Kinoshita and Campos (2003) state that the international trade attracts FDI for another reason: firms could be reluctant to invest in a certain region because they do not have much information about the area and therefore perceive a higher risk. This is especially the case when an economy has only recently opened up to foreign investment, such as a transition economy, and there

(17)

17 is not much information about the local conditions. More international trade from a specific area will inform the foreign investor and therefore reduce the perceived risk: the foreign investor will be less reluctant to invest.

Clustering

Clustering is commonly formed via a bottom up process, which occurs spontaneously. However, clustering can also be induced using to governmental policies, via a top down process. The Chinese government stimulated clustering by introducing the so called Special Economic Zones (SEZ) in order to attract FDI. The government facilitated the increase in FDI inflows by favorable tax agreements, flexible labor contract options and favorable ownership agreements (Zeng & Zeng, 2011). Empirical research into the five SEZs over the period 1978-2008 shows that the attracted FDI inflow in these regions is positively related to increased GRP, employment and labor productivity. Delis and Kyrkilis (2016) use a cluster analysis in order to investigate the regional concentration of Chinese FDI inflows and find that agglomeration is especially present in East China. This is not surprising given the fact that most FDI inflows are in this region and most SEZ regions are located here. For the international firms, the agglomerated areas have the following advantages: high local demand, better institutional quality, well developed infrastructure, a larger share of high skilled labor and a proper supply network (Delis & Kyrkilis, 2016).

Another way to investigate clustering is with the Moran’s I measure, measuring spatial autocorrelation (Huang & Wei, 2016). Huang and Wei (2016) show that for each year from 1989 to 2010, this statistic was greater than 0.1 and significant, implying that during this period, Chinese FDI inflows show clustering patterns. They also show that over the years, agglomeration advantages become more important than institutional factors. Initially, after a country opens up to FDI, institutional factors are more important. To improve institutional quality, the Chinese government equalized tax regulation across the country and this decreased the relative importance of institutional factors in the investors’ decision making. The investors are now aiming to benefit from clustering advantages. There is an ongoing trend in the location of these Chinese clustering areas: they move from the coastal regions to central and western cities (Huang & Wei, 2016).

Infrastructure

There are roughly two ways to define infrastructure: physically and digitally (Salike, 2016). Efficient infrastructure in China is already realized in the coastal regions, but the quality in inland areas remains lacking. Infrastructure is therefore mentioned as an important reason for inland

(18)

18 provinces to attract less FDI than the coastal provinces (Zeng, 2011). In the panel data analysis of Salike (2016), including 31 provinces over the period 2002-2013, the infrastructure index is a combination of fixed telephone and internet subscriptions and the length of paved roads in the region. The coefficient of infrastructure in this research is high and significant. Performing a panel data analysis over the period 1990-2002 when discriminating for different source countries, Fung, Lizaka and Siu (2005) show that total kilometers of highroads per square kilometer of the province is a significant determinant for every group of source countries (USA, Japan, Hong Kong, Taiwan and Korea).

However, Donaubauer and Dreger (2016) find no significant relation between infrastructure and FDI when using the same measurement as Fung, Lizaka and Siu (2005) over the period 2005-2013. An explanation for the ambiguous results is that good infrastructure mainly plays a moderating role in the relation between low wages and FDI, as concluded by Boermans, Roelfema, and Yi (2009). Besides that, infrastructure development is a consequence of clustering. Thus, controlling for clustering might abolish the effect of infrastructure on FDI (Delis & Kyrkilis, 2016).

Inflation

In most studies, inflation is added because it is a suggested confounding variable. During high inflation, it becomes riskier to invest in the economy and this might FDI or other variables in the regression. In their panel data analysis, Kang and Jiang (2012) also considered inflation an important control variable when studying location choice of Chinese multinationals in East and Southeast Asia in the period 1995-2007. For developed economies in Asia, inflation shows to be significantly and positively related to FDI inflow. For developing economies however, there was no significant relation and the coefficient was very small (Kang & Jiang, 2012). The authors did not give any explanation for this finding. However, the fact that inflation rate might affect the outcomes indicate that it might be wise to include this control variable. It appears that China has very stable inflation rates over the recent years. Since 1997, according to the official statistics, the Chinese inflation rate was less than 2 percent on average and rates did not exceed 6 percent (Nakamura, Steinsson, & Liu, 2016).

Overview of determinants

All determinants discussed have shown to be related to FDI and will therefore be included in the empirical analysis. From the literature it can be concluded that the estimation results of previous research are ambiguous and that the coefficients largely depend on the empirical approach (Gao, 2005). In most research it is found that market size, trade openness and clustering are positively

(19)

19 related to FDI in China (Boermans, Roelfema, & Yi, 2009; Delis & Kyrkilis, 2016 Kinoshita & Campos, 2003). For wages, labor quality and infrastructure the relationship with FDI is less clear (Fung, Lizaka, & Siu, 2005; Hu & Owen; Salike, 2016). Wages could affect FDI via two channels (via labor costs and purchasing power), having opposite effects, and although the second channel is expected to become more important, wages are still often found to be negatively related to FDI (Salike, 2016). Labor quality is theoretically expected to be positively related to FDI, but from the literature it seems that very detailed analysis is necessary to find this positive relation (Salike, 2016). Infrastructure has been identified as determinant in some research, but this determinant seems more, if not only, important for FDI into West China (Zeng, 2011). Lastly, inflation is included because it is suggested as important confounder (Kang & Jiang, 2012).

3.2 Empirical differences between East and West China

Many previous studies separately analyzed the different regions within China (Boermans, Roelfsema, & Yi, 2009). It appears that for different regions, different factors are important for explaining FDI inflows, especially if the distinction between East and West China is made (Boermans, Roelfsema, & Yi, 2009; Luo, Brennan, Liu, & Luo, 2008; Liu, Daly, & Varua, 2012). As mentioned in the introduction, the largest share of FDI is in the Eastern region, although the share of FDI into the Western region is growing (Lui, Luo, Qui, & Zhang, 2014).

Motives for companies that choose to invest in inland China are different from investors in East China. A first reason for this might be that investors investing in West-China are mainly HMT investors. Compared to for example OECD investors, the HMT investors attach less value to the better education, intellectual property rights protection and infrastructure in East China (Belkodja, 2016). In line with Belkodja (2016), Hu and Owen (2005) show that for those investors agglomeration economies seem to be less important and those companies are responding more to changes in tax rates. Huang, Jin, and Qian (2013) state that HMT firms are lacking innovation and technology advantages compared to firms from OECD countries. The fact that the main investors in West-China are HMT investors, with relatively less technology advantages, might lead to an increase the importance of low production costs in West-China compared to the Eastern-region.

A second reason for differences between East-and West-China is that investors in East-China are investing in this area because of the high local demand. Boermans, Roelfsema, and Yi (2009) perform separate panel data analysis for each region over the period 1995-2005. The Eastern provinces in their research are Beijing, Fujian, Guangdong, Hainan, Hebei, Heilongjiang, Jiangsu, Liaoning, Shandong, Shanghai, Tiajin and Zhejiang. The other 18 provinces are labeled as Western provinces. They confirm that while the Eastern region mainly attracts FDI due to market potential and

(20)

20 institutional quality, the Western region is more likely to attract FDI due to geographical advantages, such as natural resources. However, the findings of Boermans, Roelfema, and Yi (2009) are not supported by the research of Luo, Brennan, Liu and Luo (2008), who examine the location choice determinants for 98 Chinese inland cities over almost the same period (1995-2006). They find that FDI inflows in inland cities are not significantly related to the abundance of natural resources and GDP per capita. Labor quality, GRP, policy quality and agglomeration are however found to be positively related to inland FDI. From their results, Luo, Brennan, Liu and Luo (2008) conclude that the inland Chinese FDI inflows have a market seeking nature, such as firms in the product industry and business services. Luo, Brennan, Liu, and Luo (2008) conclude that even though international firms located in West China might have a market seeking motive, this motive might be relatively more important for investors in East China when comparing the two regions (Luo, Brennan, Liu, & Luo, 2008).

A third important reason for the differing results between East and West China, is that infrastructure is less developed in West China (Zeng, 2011). As mentioned previously, infrastructure might play a moderating role for the relation between low wages and FDI attractiveness, especially in the manufacturing industry. Liu, Daly and Varua (2012) perform separate analyses for Eastern and Western China when researching the manufacturing sector. When discriminating between low technology and high technology manufacturing, they find that market size, labor costs and agglomeration are important for FDI into low-technology manufacturing, while telecommunication was important in explaining high technology manufacturing in the coastal area. Within the inland regions infrastructure was the main important determinant to explain all manufacturing FDI. This raises the expectation that, compared to East China, infrastructure plays an important role in explaining the choices for a specific province in West China.

(21)

21

4. Empirical methods

This chapter presents the methodological choices made in order to answer the research question. Firstly, the panel data model is discussed together with several statistical checks. The second part of Chapter 4 describes the methodology for the rolling regressions. The third part of this chapter explains the data and variables. Lastly, some descriptive statistics of the data are provided.

4.1 Panel data analysis

The dataset includes annual data over the period 1997-2005, of 31 Chinese provinces. Panel data analysis is used for datasets in which the observations differ on both a time dimension and a cross-sectional dimension. Therefore, panel data analysis is a suitable method for this research. As opposed to other methods of analysis, in which the observations vary in only one dimension (time-series or cross-sectional), the use of panel data allows to draw conclusions about causal relationships (Jadhav, 2012). Two important methods to conduct panel data research are the fixed effects estimation model and random effects estimation model (Ranjan & Agrawal, 2011). The difference between the fixed and random effects estimation model is that both methods use different approaches to model the group specific effects, which will now be further explained.

Fixed effects model

On the cross-sectional dimension, there may be so called fixed effects. Fixed effects are characteristics that vary across groups (in this research the groups are the different provinces), but are stable over time. This means that the fixed effects model allows for different constants per group (Ranjan & Agrawal, 2011). The model for this method can be written as:

𝑦_𝑖𝑡 = 𝑎 + 𝛽𝑋_𝑖,𝑡+ 𝜇_𝑖+𝑣_𝑖,𝑡

Where 𝑦_𝑖,𝑡 represents the dependent variable and 𝑋_𝑖,𝑡 represents a vector of the independent variables. The vector of the coefficients 𝛽 provides information about how a change in the independent variable affects the dependent variable. The province specific constant term is 𝜇_𝑖. In this analysis, these fixed effects are estimated separately and are not included in the error term. The error term 𝑣_𝑖,𝑡 only consist of individual random errors varying across time (Jadhav, 2012).

(22)

22 Random effects model

In the random effects model those fixed effects are not included. The model for the random effects model can be written as:

𝑦_𝑖,𝑡= 𝑎 + 𝛽𝑋_𝑖,𝑡+ 𝜔_𝑖,𝑡 Where 𝜔_𝑖,𝑡 = 𝜀_𝑖+ 𝑣_𝑖,𝑡

As opposed to the fixed effects model above, all group specific fixed effects are included in the error term 𝜔𝑖,𝑡 and not in the separate, group specific, intercept term 𝜇𝑖 (Jadhav, 2012). This model

considers the group specific effects as random parameters rather than stable characteristics over time (Ranjan & Agrawal, 2011). Which estimation method is more appropriate must be investigated with specific tests, such as the Hausman test.

Hausman test

The Hausman test is designed to identify which of the two models (fixed or random effects estimation model) provides the most consistent and efficient estimates (Baltagi, Bresson, & Pirotte, 2003). In order to test which method has to be used, the Hausman test investigates whether the group specific effects 𝜇_𝑖 are correlated with the explanatory variables 𝑋_𝑖,𝑡. The null hypothesis is that there is no correlation between the group specific effects and the independent variables and the opposite is true for the alternative hypothesis (Sheytanova, 2015):

𝐻₀: 𝑐𝑜𝑣(𝜇𝑖, 𝑋𝑖,𝑡) = 0

𝐻1: 𝑐𝑜𝑣(𝜇𝑖, 𝑋𝑖,𝑡) ≠ 0

When the null hypothesis is rejected and hence the explanatory variables are correlated to the group specific effects, the fixed effects model should be used. With the fixed effects model, the explanatory variables are robust to group specific effects because these effects are separately included in the regression as 𝜇_𝑖. If null hypothesis cannot be rejected, the random effects model can be used because the group specific component is not correlated to the explanatory variables and therefore not causing a bias in the parameter 𝛽, which estimates the effects of the explanatory variables on the dependent variable (Sheytanova, 2015).

(23)

23 Time fixed effects

Besides fixed effects on the cross-sectional dimension, it is possible that there are fixed effects on the time dimension: variables are likely to be affected by time specific events, but these effects do not depend on the group specific differences. It is therefore recommended to test whether including time fixed effects is necessary in the panel data. Without controlling for time related effects, it is likely that aggregate trends affect the outcome of the analysis, causing for example spurious relations (Torres-Reyna, 2007). Several events during the period 1997-2015 might have had an influence on the economic conditions in China. An example is the global crisis starting in 2008, which affected the world economy and international trade (Kekic, 2011).

The model for time fixed effects can be derived in a same way as the fixed effects panel data model, only the group specific effects 𝜇𝑖 have been replaced with the time fixed effects 𝜆𝑡 (Torres-Reyna,

2007):

𝑦_𝑖𝑡 = 𝑎 + 𝛽𝑋_𝑖,𝑡+ 𝜆_𝑡+ 𝜔_𝑖,𝑡

Note that this is the same as adding dummy variables for each year starting in the second year of observation:

𝑌_𝑖𝑡 = 𝑎 + 𝛽𝑋_𝑖,𝑡+ 𝜆₂𝐷2 + 𝜆₃𝐷3 + ⋯ + 𝜆_𝑇𝐷𝑇 + 𝜔_𝑖,𝑡

Where D(X) represents a dummy variable for each year and T is the number of time periods present in the sample. For each year a parameter is estimated which describes its relation with the independent variable.

The following step is to do the Wald test, to observe whether the coefficients of the time specific dummy variables are equal (𝜆2= 𝜆3 = ⋯ = 𝜆𝑇). If this test provides a significant result, the

time coefficients (𝜆𝑡) are significantly different from each other and therefore it is likely that the

parameters of the explanatory variables show a bias due to time fixed effects when those are not included (Torres-Reyna, 2007).

Assumptions in panel data

To calculate all parameter estimates, several assumptions about the data are made. These assumptions have to be satisfied in order to obtain valid estimates (Schmidheiny & Basel, 2011).

(24)

24 Important assumptions in panel data analysis are: no perfect multicollinearity, exogeneity, homogeneity and no autocorrelation (Schmidheiny & Basel, 2011).

Multicollinearity causes the parameter for the effect of an individual predictor to be less reliable, because in this case multiple individual predictors are highly correlated with each other (Alin, 2010). The parameter estimate of an explanatory variable is now largely affected by a small change in another related predictor. Multicollinearity is tested calculating the variance inflation factor (VIF), which is computed using the formula 1/(1-R2). R2is the quotient of the variance in the variable that is caused by other variables. When the VIF value for each variable does not exceed 10, multicollinearity issues should not be present (O’Brien, 2007). Please note that a multicollinearity test is different from a simple correlation test, because it estimates a correlation between multiple (at least three) variables.

Endogeneity issues could exist when an explanatory variable is related to the error term. This could be due to omitted confounding variables1, or this could be due to the fact that the dependent variable and the independent variables affect each other both ways. If the latter is true, causal interpretation becomes difficult: it is not certain that the independent variables 𝑋𝑖,𝑡 affect the

dependent variable 𝑦𝑖,𝑡 and not vice versa. To overcome a situation of endogeneity where 𝑦𝑖,𝑡 affects

𝑋_𝑖,𝑡, all predicting variables are lagged with one period 𝑋_{𝑖,𝑡−1}. The lagged independent variables cannot be affected by the values of the dependent variable, which is the measurement one period later. Therefore it can be concluded that when the explanatory variable and the independent variable are related, the variability of the explanatory variable is causing the variability in the independent variable and not the other way around (Baltagi, 1995).

The assumption of homoscedasticity of error terms is true when the variability of the values of a variable is equal across the values of the explanatory variable. If the standard errors are different across values of a variable in terms of the explanatory variable, this is called heteroscedasticity. To determine whether heteroscedasticity is present, a likelihood-ratio (LR) test is performed. The LR-test examines which of two models provides a better estimation, thereby comparing a model in which homoscedasticity is assumed with a model assuming heteroskedastic error terms2 (StataCorp, 2007). In this test the null hypothesis is that the assumption of homoscedastic error terms is true, so if rejected, this assumption is violated (StataCorp, 2007).

Lastly, serial correlation is present when the error term of a certain period 𝜀𝑡 is related to the

error term of a previous period (𝜀𝑡−1, 𝜀𝑡−2, etc.). The Wooldridge test has been designed to test for

1_{Omitted variables causing the independent variable to be related to the dependent variable could for example}

be group- or time specific fixed effects.

2

(25)

25 serial correlation in the model. This is a method in which the first difference of 𝑦𝑖,𝑡, the difference

between 𝑦_𝑖,𝑡 and 𝑦_{𝑖,𝑡−1}, is regressed on 𝑋_𝑖,𝑡 according to the following model (Drukker, 2003): 𝑦_𝑖,𝑡− 𝑦_{𝑖,𝑡−1}= (𝑋𝑖,𝑡− 𝑋𝑖,𝑡−1)𝛽1+ 𝜀𝑖,𝑡− 𝜀𝑖,𝑡−1

∆𝑦_𝑖,𝑡 = ∆𝑋_𝑖,𝑡𝛽₁+ ∆𝜀_𝑖,𝑡

If the error terms are not serially correlated, then the coefficient (𝛽1) of the first-difference equation

must be equal to -.5, which is a rule of thumb for this test as designed by Wooldridge (2010). If the this null hypothesis, that cov (𝜀𝑖,𝑡, 𝜀𝑖,𝑡−1) is equal to -.5, is rejected, there is serial correlation present

in the error term and in order to prevent biased results the model has to be adjusted for this (Drukker, 2003).

Cluster-robust standard errors

To overcome biased results due to correlated residuals, various adjusting measures have been designed. Choosing an appropriate adjustment measure is very important for the accuracy of the predictions (Hoechle, 2007). Statistical adjustment by means of cluster-robust standard errors is often recommended if both heteroscedasticity and serial correlation are present (Drukker, 2003; Hoechle, 2007). Furthermore, this method has been used before in research into FDI (Büthe & Milner, 2008;

Helpman, Melitz, & Yeaple, 2004). Using cluster-robust standard errors therefore seems an appropriate adjustment method in this research. However, it has to be taken into account this is only one out of many ways to control for serial correlation and heteroscedasticity and an important downside to keep in mind using this method is that the estimates are less efficient and that less reliable if the number of clusters is below 40 (Esarey & Menger, 2017; Imbens & Kolesar, 2016). It will now be explained what the use of cluster-robust standard errors means for this research. Firstly, the using this method, the standard errors are calculated differently than for a normal regression model in order to be robust to heteroskedasticity3. Secondly, the standard errors are also clustered. The clustering variable here is the province, because the variable for clustering should be the grouping variable in panel data (Hoechle, 2007). Clustering causes the model to be robust to serial correlation. For panel data the presence of serial correlation means that individual observations are not independent of each other: for the same province there are multiple measurements on the time dimension. Thus, the error terms over time are correlated within the province. The assumption that the observations are independent of each other is relaxed with clustering and hence, this approach allows

3

For the exact calculation method please consult Stata user’s guide section 20.21 about Obtaining Robust Variance Estimates.

(26)

26 for the presence of serial correlation (Croux, Dhaene, & Hoorelbeke, 2003; StataCorp, 2007). Recall that clustering the error terms on a province level differs from a fixed effects model, because this model only allows differences between provinces and not within (over time).

4.2 Rolling regression

In order to observe changing coefficients over time, and thereby answering the research question, a rolling regression method is used. Rolling regression techniques allow an illustration of the stability of the parameters over time, because the model is repeatedly estimated over different consecutive periods (Baum, 2004). In this research, the estimation window, or the time frame of estimation, is held fixed and set equal to 8. This implies that the regression over period 1997-2004 will be followed by the regression of 1998-2005 and this will continue until 2008-2015 (in total 12 replications). The aim of this approach is to plot line graphs in which the coefficients (β) over time can be observed.

An important downside in using rolling regressions is that the chosen window will affect the outcome. A larger window shows a smoother pattern in the coefficients, while a shorter window will show a more fluctuating pattern in the coefficients. When interpreting the rolling regression results it has to be taken into account that the coefficient plots only show the magnitude of the coefficients over time, but not the statistical significance (Zanin & Marra, 2012). However, when taking into account that the rolling regression results are very sensitive to the researcher’s interpretation, rolling regressions could provide a clear image of how coefficients are changing over time.

4.3 Data and variables

The dataset covers 31 provinces4 of China over the period 1997-2015. The main source for the data was China Statistical Yearbook. Two variables where not retrieved from China’s Statistical Yearbook: the yearly average exchange rate, which is used to convert USD measured variables into yuan, is retrieved from OANDA Corporation (2017) and the size of the provinces (km2), used to calculate the infrastructure measure is retrieved from Benewick and Donald (2009). Three separate samples are analyzed: entire China, East China and West China. The sample of entire China includes all provinces and the distinction between Eastern and Western provinces is made according to Boermans,

Roelfsema, and Yi (2009): the Eastern provinces are Beijing, Fujian, Guangdong, Hainan, Hebei, Heilongjiang, Jiangsu, Liaoning, Shandong, Shanghai, Tiajin and Zhejiang and the other provinces are defined as Western. The model including all variables can be written as follows:

4

The provinces included are: Anhui, Beijing, Chongqing, Fujian, Gansu, Guangdong, Guangxi, Guizhou, Hainan, Hebei, Heilongjiang, Henan, Hubei, Hunan, Inner Mongolia, Jiangsu, Jiangxi, Jilin, Liaoning, Ningxia, Qinghai, Shaanxi, Shandong, Shanghai, Shanxi, Sichuan, Tiajin, Tibet, Xinjiang, Yunnan, Zhejiang.

(27)

27 𝐹𝐷𝐼_𝑖,𝑡 = 𝑎 + 𝛽₁𝐺𝑅𝑃_{𝑖,𝑡−1}

+ 𝛽2𝑊𝐴𝐺𝐸𝑖,𝑡−1+ 𝛽3𝐸𝐷𝑈𝐶𝑖,𝑡−1

+ 𝛽₄𝑇𝑅𝐴𝐷𝐸_{𝑖,𝑡−1}+𝛽₆𝐶𝐿𝑈𝑆𝑇_{𝑖,𝑡−1}+𝛽₅𝐼𝑁𝐹𝑅𝐴_{𝑖,𝑡−1}+𝛽₇𝐼𝑁𝐹𝐿_{𝑖,𝑡−1}+ 𝜔_𝑖,𝑡

The dependent variable 𝐹𝐷𝐼 is the level of Total Investment of Foreign funded enterprises registered per year by the State Administration for Industry and Commerce. FDI is initially expressed in million USD, but is converted into million yuan using the average exchange rate of that specific year. The independent variable for market size is Gross Regional Product (𝐺𝑅𝑃), just like in Ho (2004) and is expressed in 100 million yuan. 𝑊𝐴𝐺𝐸 is the variable for labor costs and this is the average yearly wage of staff and workers in yuan. Furthermore, the variable for labor quality is represented by 𝐸𝐷𝑈𝐶 in the model. For the education variable, the number of graduates with degrees or diploma’s in institutions of higher education (in 10.000 people) is divided by the population of the province. For the ease of interpretation of this coefficient, the number of graduates is expressed per 1000 persons.

In accordance with the measures of other research, trade openness (𝑇𝑅𝐴𝐷𝐸) is the calculated dividing the value of imports of a province by the market size (Zhang & Zhang 2003). The value of imports and exports of firms (USD) for each province, registered by Custom Statistics, is converted into yuan and divided by GRP. Clustering (𝐶𝐿𝑈𝑆𝑇), is measured slightly different in this model than in other research: it is measured in which degree international firms tend to be attracted by the number of other international firms. Salike (2016) measures clustering as the number of industrial enterprises located in a province in relation to GRP. Based on this research, the variable clustering is constructed using the number of international enterprises registered by the State Administration for Industry and Commerce divided by 100 million GRP.

The variable for infrastructure (𝐼𝑁𝐹𝑅𝐴) is calculated based on the measurement of Liu, Daly and Varua (2012): the length of highways and railways (km) divided by the total area of the province (km2). Last but not least, the variable inflation (𝐼𝑁𝐹𝐿) is the percentage of change in the regional Consumer Price Index (CPI). Table 1 summarizes the expected relation for each explanatory variable with FDI, following from the literature discussed in Chapter 3.

(28)

28 Table 1

Expected relationships between the explanatory variables and FDI Variable Expected relation Literature

GRP Positive Ali and Guo (2005); Boermans, Roelfema, and Yi

(2009); Ho (2004)

WAGE Negative or insignificant Liu, Daly, and Varua (2012); Salike (2016) EDUC Positive or insignificant Gao (2005); Salike (2016)

TRADE Positive Kinoshita and Campos (2003); Liargovas and Skandalis (2012)

CLUST Positive Delis and Kykilis (2016); Huang and Wei (2016)

INFRA Positive or insignificant Fung, Lizaka and Siu (2005); Donaubauer and Dreger (2016)

INFL Positive or insignificant Kang and Jiang (2012)

4.4 Descriptive statistics

The dataset is strongly balanced, which means that there are not many missing values. All variables have a value for each province and each year, except for the variable inflation which has one missing observation for 1997 in the province Tibet. In Table 2 the differences between East and West China can be observed: on average the FDI inflows into East China are 9 times as large as the FDI inflows into West China, which is in accordance with the literature stating that around 90 percent of FDI flows into East China (Huang & Wei, 2016; Madariaga & Poncet, 2007). The average amount of GRP into East and West China also shows a huge difference: GRP is more than twice as high in East China.

The average wages over the years in Western China are almost 20 percent lower compared to the average wages in Eastern China, which could be explained by the higher share of graduates having a degree in higher education. Generally, better educated employees are more productive and receive higher wages. Eastern provinces are much more open to trade: their average trade volume relative to GRP is six times as large as for the Western provinces. Zhang and Zhang (2003) already stated that trade openness increased more for East China than for West China from the moment that the country opened up to FDI. Given the more developed FDI climate in East China, it is not surprising to find that the number of international companies relative to GRP is larger for this region and that the length of infrastructure per km2 present in these Eastern provinces is greater. With respect to the numbers for inflation there are no large differences for East and West China: both regions have an average inflation rate of around 2 percent.

(29)

29 Table 2

Summary statistics per region

Region FDI GRP WAGE EDUC TRADE CLUST INFRA1 INFL

East China Min 61737.28 411.16 4889 0.47 0.06 0.23 1.17 -3.2 Max 4862323 72812.55 113073 8.88 1.72 16.92 21.61 6.9 Mean 960957.7 15287.65 30557.6 3.57 0.62 2.27 8.38 1.86 SD 981010.7 15328.06 22311.13 2.27 0.46 2.33 4.59 2.10 West China Min 1981.27 77.24 4900 0.28 0.03 0.13 0.18 -3.6 Max 661978.7 37002.16 110980 7.92 0.41 2.99 17.31 10.1 Mean 117644 6605.76 24552.34 2.42 0.10 0.61 4.68 2.15 SD 114450.8 7000.96 17253.54 1.77 0.05 0.44 4.01 2.35 Entire China Min 1981.27 77.24 4889 0.28 0.03 0.13 0.18 -3.6 Max 4862323 72812.55 113073 8.88 1.72 16.92 21.61 10.1 Mean 444088 9966.50 2687.96 2.86 0.30 1.25 6.11 -3.6 SD 740653.5 11773.9 19570.37 2.05 0.38 1.70 4.61 2.26 1

(30)

30

5. Estimation

This chapter discusses the results of the empirical analyses. Firstly, a correlation matrix is presented in order to show how the variables used in the estimation models are related. The rest of the chapter shows the estimation results of random effects models with clustered standard errors. In the first model year fixed effects are added and in a second model a dummy variable is included, which divides the period 1997-2008 in a pre-crisis and post-crisis period. Lastly, the results of the rolling regressions are presented.

5.1 Correlations and assumption checks

The relationships between the variables are plotted in the correlation matrix in Table 3. Except for the variable inflation, all independent variables are significantly and positively related (p<0.05) to the dependent variable FDI. In addition all explanatory variables significantly related to FDI are significantly correlated to the other independent variables, except for clustering and infrastructure. None of the correlations are higher than 0.76 (education and wage).

Table 3

Correlation matrix

FDI GRP WAGE EDUC TRADE CLUST INFRA

FDI 1 GRP .74* 1 WAGE .44* .55* 1 EDUC .37* .52* .76* 1 TRADE .68* .31* .30* .38* 1 CLUST .21* -.10* -.20* -.15* .43* 1 INFRA .60* .66* .58* -.72* .52* .08 1 INFL 0.06 .15* .29* .34* .03 -.22 .17 *p<0.05

In order to choose the right approach for statistical analysis, several assumption checks have been done. For these tests the entire database, including all years and provinces has been used and the results can be found in Appendix A. To test whether there are multicollinearity problems, the VIF value has been calculated as explained in Chapter 4. No value exceeds 3.53, so there is no concern of multicollinearity (Alin, 2010). The Hausman test was proposed to clarify whether a fixed or random effects model provides better estimations. This test provided a non-significant result, which indicates

(31)

31 that using a random effects model should give more accurate predictions. With the random effect model, the group specific fixed effects are uncorrelated with the independent variables and do not have to be included. Although, the use of fixed effects models is more common in research into FDI determinants, the random effects model has been used previously in this research field, for example by Ranjan and Agrawal (2011).

The significant results of the likelihood ratio test show that the assumption of homogeneous standard errors has to be rejected. This means that the standard errors are not equal across the values of the dependent variable in terms of the values of an explanatory variable. Additionally, the significant result for Wooldridge test shows that there is serial correlation present and the null hypothesis of no first-order autocorrelation can be rejected. Given that the assumptions of homoscedasticity and no-serial correlation for the standard errors are not met. The model is adjusted using clustered standard errors, as suggested in the methods chapter. Although this method is suitable to make the necessary adjustments, it has to be taken into account that a disadvantage of the use of this method is that it could cause biases in the estimated parameters, because the number of panels in this research is smaller than 40 (Esarey & Menger, 2017).

5.2 Random effects model including year effects

Table 4 reports the random effects model estimation results for three samples: entire China, East China and West China in which time fixed effects (time dummies for individual years) are included.

Changing determinants of FDI location choice in China: a panel data analysis