China's transport infrastructure : the effect on regional income growth patterns

(1)

China's transport infrastructure: The

effect on regional income growth

patterns

Name: Peter John Robinson Student Number: 10604863 Supervisor: Lennart Ziegler Email address: pete6660@hotmail.com Date: 07/08/2014 Number of words: 11,440

(2)

1

1. Introduction

Over the past three decades China has been one of the fastest growing economies, achieving real gross domestic product (GDP) per capita growth rates in excess of 9% per annum between 1982 and 2012 (World Bank, 2014). International trade and foreign direct investment (FDI) to complement China's open-door policy have been crucial in China's transition from a command economy to a decentralized market based economy (Demurger, 2001). However, immediately following the reform and opening of China in 1982, the growth rate of regional real GDP per capita between coastal and inland provinces was markedly unequal.

Geographical constraints cause disparities between inland and coastal regions of China (Yu et al. 2012). Inland regions need adequate transport infrastructure to integrate with the outside world and to

compensate for their geographical constraints. This would allow such inland province areas to take advantage of growth enhancing international trade reforms. As Fan et al. (2011) point out,

infrastructure investment is not only a cause of regional inequality, but it could also be used to contain further rising regional inequality.

It was only in 1999 (after two decades of rigorous coastal development) that central government shifted their attention towards promoting interior region development in the “Western Development Program” (Lai, 2002). This was followed by the “Northeast China Revitalization” in 2003, and the “Central China Plan” in 2004. The advancement of transport infrastructure in the interior was one of the three programs several initiatives to promote the lagging inland provinces. In despite of these efforts, transport infrastructure endowments are still heavily concentrated on the coast of China (Jin et al. 2010).

It is exactly the aim of the thesis to research just how transport infrastructure endowments can be attributed to growth inequality between regions located on the coast and in the interior following the several policies to develop inland regions. Therefore, the central research question is:

 What is the impact of transport infrastructure endowments on the differences in economic growth rates between inland and coastal provinces of China?

The thesis contributes to existing literature in the following ways: First, the introduction of a coastal dummy variable and three separate transport infrastructure density interaction terms in a cross region economic growth regression will be used to investigate the coastal-inland economic growth disparity attributed to transport infrastructure. Furthermore, the nonlinear relationship between economic growth and transport infrastructure endowments is investigated in a separate cross region economic growth regression with the introduction of various squared endowment terms as in the Demurger (2001) and Hong et al. (2011) papers.

Second, given the recent enormous infrastructure investments, and its contribution to Chinese real GDP per capita growth rates, a large dataset of 31 provinces, on a recent time period (2004 to 2012) is em-ployed.

Third, many studies do not solve the problem of reverse causality between transport infrastructure and economic growth. The thesis addresses this issue by instrumenting for transport infrastructure in two-stage least-squares (2SLS) estimation equations.

(4)

3 Finally, some studies include the impact of one type of infrastructure on economic growth (Mody and Wang, 1997 focuses solely on roadway transport infrastructure). Similarly, other papers compile a sin-gle transport infrastructure density term from several types of transport infrastructure (Demurger, 2001). Proxies for three types, namely: railway density, waterway density and highway density transport infrastructure are included in the thesis. This is done to separately investigate the effect of each type of transport infrastructure on economic growth as in Banerjee (2012), and Hong et al. (2011). A combination of the four contributions outlined above has not been investigated in any other academic paper on China.

The thesis will progress as follows: First, a literature review will give the theoretical and empirical evidence on the link between transport infrastructure and economic growth. Second, a brief history of China’s transport infrastructure investment strategy is outlined. Third, several figures are presented to explain the regional distribution of transport infrastructure and real per capita GDP growth, as well as the causal interaction between transport infrastructure and economic growth between regions. Fourth, the research design and methodology are presented. The research design consists of data collection methods, sample design and the presentation of the model. Fifth, the main results from four regression models are shown to allow for a more in depth discussion of the link between transport infrastructure and growth between coastal and inland provinces. Sixth, various checks are utilized to check the robustness of the results obtained. Seventh, a conclusion will proceed on the said causal interaction, and the relative importance of each type of transport infrastructure towards economic growth. In doing so the central research question is addressed. Policy recommendations will also be stated in the

conclusion section based on the main results obtained. Finally, the thesis will state limitations of the research carried out, as well as areas to be given greater attention in future research.

2. Literature review

2.1 Theory

There are several ways in which the development of transport infrastructure is conductive to economic growth: First, better transport infrastructure allows companies of a similar industry to cluster closer together, therefore positive spillover effects such as information transfer occur (Banister and Berechman, 2000). Furthermore, firms located within close proximity attract more suppliers and consumers than a single firm could alone. This raises the demand for goods and services (Hong et al. 2011).

Second, when transport infrastructure is of higher quality, or distances between transportation routes are shorter, this reduces travel times. Therefore, firms who transport freight benefit through vehicle operating cost and time savings (Gunasekera et al. 2008). Moreover, a decrease in transport costs will enhance market integration to allow domestic markets to flourish, and economies of scale to be

achieved (Ming and Chen, 2006). Passenger transport time savings allow individuals to engage in more economic activities to benefit society as a whole (Ming et al. 2013).

Third, highway infrastructure investment allows companies to raise the turnover of goods occupied by inventory, which lowers inventory costs and raises firm productivity (Shirley and Winston, 2004).

(5)

4 Fourth, regions which are endowed with better transport infrastructure attract substantial FDI (Hong, 2007). Such investment can be crucial in the acceleration of growth rates.

Fifth, transport infrastructure increases the value of urban land. This attracts firms in search of easy access to roads and railways, so as to enable them to market their products to a wider base (Rives and Heaney, 1995). As well as this, residential house building projects may follow the building of adequate infrastructure to facilitate the effects of urban agglomeration. This is the process by which affluent regions who are endowed with better infrastructure systems attract the migration of productive capital and labor.

One might also consider the potential economic losses transport infrastructure development could bring. First, Elite capture in fiscal transfers from central to local government in road building projects is a very serious problem in Indonesia (Olken, 2007). Mauro (1995) cites corruption such as elite capture as a major cause of low growth rates.

Also, the addition of roads and railways could raise the incidence of accidents, which may place an added burden on healthcare services (Ming et al. 2013). The health of individuals in a country may suffer due to this added burden. Because health positively contributes to the human capital of an individual, according to economic theory, a substantial healthcare burden may cause growth rates to fall.

2.2 Empirical

Previous literature gives mixed evidence on the strength of the link between transport infrastructure and economic growth. Mody and Wang (1997) utilize data on the output of twenty three industrial sectors in seven Chinese coastal provinces between 1985 and 1989. They find transport infrastructure

significantly positively influences economic growth rates. Demurger (2001) include a more

comprehensive dataset of 24 coastal and inland provinces between 1985 and 1998. The author also finds positive growth effects of transport infrastructure. On the contrary, Banerjee et al. (2012) finds that there is no effect of transport infrastructure on economic growth. They explain this zero effect as being caused by the Chinese government placing stringent rules which restrict the mobility of labor. As a consequence the benefits of urban agglomeration are severely limited.

Fleisher and Chen (1997) realized that economic growth may in fact vary between coastal and inland provinces. They take a total factor productivity (TFP) approach to explain coastal-inland differences in economic growth in China. In coastal regions, TFP is roughly twice as high as the interior, but to stimulate growth, transport infrastructure investment is needed to attract FDI and to prevent the migration of university graduates from inland provinces.

Yu et al. (2012) explain coastal-inland differences in economic growth by using a different methodology. They investigate the causal relationship for China using a panel unit root, panel cointegration and Granger causality approach. The authors find that, at the regional level, there is bidirectional causality between transport infrastructure and growth in affluent coastal provinces, while poorer inland provinces exhibit unidirectional causality from growth to transport infrastructure. Based on this result it is important for Chinese policy makers to realize that transport infrastructure on its own

(6)

5 does not generate growth in lagging inland regions. To better realize the economic benefits of transport infrastructure in landlocked provinces, government agents should pay more attention to complementary factors. Nevertheless, the bidirectional causality is clear evidence of a degree of endogenous reverse causality between transport infrastructure and economic growth.

To address the endogeneity problem between transport and growth, Banerjee et al. (2012) exploit the fact that well-endowed regions are located closer to railways connecting historical cities. The authors compare those regions with low proximity to historical lines, with those situated further away as the total average treatment effect of transport infrastructure on economic growth. Note: other types of transport infrastructure were built more recently near these historical lines, so proximity to historical railway lines is also correlated with other nodes of transport. The method used by the authors therefore provides a source of exogenous variation in the endowment of transport infrastructure.

Demurger (2001) and Hong et al. (2011) take a different approach to solve the endogeneity problem. They instrument for the transport infrastructure variables in their IV fixed effects model using exogenous variables, such as lagged investment and FDI variables, annual real wage, agricultural to industrial relative prices, exports plus imports over GDP per capita, transport related lagged variables, and lagged values of GDP per capita. The thesis will use some of these exogenous instruments, as outlined in section 4.3.

Many previous studies explain the transport infrastructure, economic growth relation as linear (Fleisher and Chen, 1997). Ding et al. (2006) however recognized that the infrastructure to growth relation is nonlinear, and found positive, but diminishing economic returns to telecommunications infrastructure. Chinese provinces therefore benefit relatively more from their endowment earlier in the development process.

Hong et al. (2011) include squared land and water transport infrastructure terms. They found that there are diminishing returns to land transport, implying that provinces with initially poor land transport infrastructure benefit relatively more in terms of economic growth from the addition of such land transport infrastructure. Demurger (2001) found a similar result. Furthermore, Hong et al. (2011) found a U-shaped relation between water transport infrastructure and economic growth. One possible

explanation for this result posed by the authors is steep startup costs of water transport infrastructure, therefore investment in infrastructure is only beneficial once the investment exceeds a certain threshold level.

Authors often differ in the types of transport infrastructure they use as their explanatory variables, as well as how each variable is proxied for. In despite of the positive growth effects of transport

infrastructure found by Mody and Wang (1997), they solely focus on the impact of roadway transport infrastructure. This is a limitation given that both railway and waterway infrastructure have the potential to influence growth rates. Demurger (2001) took in to account the potential influence of railway and waterway infrastructure, and proxies for transport infrastructure endowments using transport density measures (i.e. the ratio of railway, navigable inland waterway and highway length to the surface area of the province). However, in contrast with Demurger (2001), who compiles a single transport infrastructure variable from several types, Hong et al. (2011) explores each transport infrastructure proxy variable separately. They also realize that the transport infrastructure density measure is a quantitative measure, and therefore does not account for the quality of such infrastructure. So instead of the infrastructure density measure utilized by Demurger (2001) they developed a more comprehensive index which takes in to account the quality of infrastructure.

(7)

6

3. Transport infrastructure provision

3.1 History

Since the People’s Republic of China was established in 1949, the country has experienced a huge expansion in its transportation systems. It is now endowed with the largest highway and navigable inland waterway system, as well as the third largest railway network in the world. However, in 1949 much of the transportation systems in China were destroyed or had been severely deteriorated as a result of decades of war. It became a national priority to restore such transport networks to boost the economy back into working order.

Prior to the economic reform and opening of China in 1982, investment decisions based on

infrastructure were under the control of the central government. To guide the decision making process they followed several priorities established in a strategy of general development. Among the goals set forth in the strategy were for provinces to achieve self-sufficiency, and to develop heavy industries which were located in the northern regions of China. Railway construction took preference because of the ability of freight carriages to carry raw materials at a low cost from resource rich provinces

(Demurger, 2001). However, instead of upgrading existing routes, the central government expanded rail network lines and paid little attention to highways and waterways.

At the start of the economic reform period, transport infrastructure investment lagged behind economic growth. This was due to poor incentives, and the inability of local government to mobilize sufficient infrastructure expenditure funding (Lin, 2001). Jin (1994) describes how investing in industrial capacities was seen as more important to raise productivity levels. This however left transport infrastructure largely neglected.

At the beginning of the early 1990’s there was a renewed interest in transport infrastructure spending due to large deficiencies which resulted in rising urban congestion (Demurger, 2001). Consequently, to ease the congestion related issues there were substantial increases in transport spending as a share of state fixed asset investments (Hong et al. 2011). Furthermore, despite railway investments being given priority up until 1982, highway construction played a dominant role from 1990 onwards to make rural areas more accessible (Demurger, 2001).

Throughout the entire reform period, central government have recognized the need to address transport quality issues by improving existing routes as opposed to expanding networks. Moreover, there was a gradual decentralization process which took place in China’s transition towards a market based economy. Decentralization enabled more specialization, and enhanced the ability of local government agents to collect funds for infrastructure investment.

Figure 1 gives an illustration of transport infrastructure development in recent years. The indicators of highways, waterways and railways are density measures, i.e. the transport length in kilometers divided by China’s entire surface area in kilometers squared. The variables are standardized, and the values lie between -2 and 2 with a mean of 0. A positive value signifies that the transport density variable is in excess of the mean value throughout the entire period, whereas the opposite is true for a negative value.

(8)

7 Figure 1 – The development of transport infrastructure between 2004 and 2012

Figure 1 shows that there is an acute spike in highway infrastructure development between 2005 and 2006, but then it grows at a fairly constant rate thereafter. Waterway infrastructure remains fairly stagnant between 2004 and 2007, then there is a sharp decline between 2007 and 2008, however there is a sharp increase in the rate of development from 2008. Upon investigation it is unclear what exactly causes this apparent fall in waterway infrastructure density. A possible explanation is that some of the waterway systems were declared no longer accessible by boat (since the indicator is a measure of navigable inland waterways). However, this is pure speculation. With regards to railway infrastructure, the rate of increase in development is slightly increasing throughout the entire period.

3.2 The regional distribution of transport infrastructure and real GDP per capita

Figure 2 illustrates the regional distribution of transport infrastructure density. There are several points worth highlighting. First, transport infrastructure density seems to be concentrated in the eastern coastal regions. This is due to the preferential treatment given to the coast following the opening of China, as well as the decentralization process which gave greater autonomy to local government. The propensity of local governments to collect funds largely depends upon their ability to bargain with central

government (Demurger, 2001). It appears that provinces located on the coast have higher bargaining power, and therefore have more to gain from decentralization. Because of low demand, mountainous terrain, and the vast areas covered by western provinces, transport density appears to be low in Xinjiang, Tibet, Gansu, Qinghai, Inner Mongolia, Ningxia, and Heilongjiang.

Second, the non-coast provinces which also have high levels of transport infrastructure density tend to be situated next to coastal provinces. This is the case for Anhui and Henan. The capital of the Hubei

2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 Year

Development of Transport Infrastructure over time

Highways Waterways Railways

(9)

8 province, Wuhan, has received a great deal of publicity over the past four years due to the incredible amount of infrastructure spending which has taken place, including the construction of a 140 mile long brand new subway system (New York Times, 2011). However, this is not an isolated case in China. Many cities in the interior are undergoing similar infrastructure investments because of the several policies launched by central government to develop interior provinces.

Third, the Yangtze River Delta (YRD) Economic zone is an important manufacturing base in China. The area covers sixteen cities in Shanghai, south Jiangxi, as well as the east and north of Zhejiang. Despite the area merely covering 1.1% of China’s entire land mass, it is estimated to have an

approximate GDP of 8,214 billion Chinese Yuan, or in equivalent terms, 17.4% of China’s economy in 2011 (Hong Kong Trade Development Council, 2013). To raise the economic unification of cities within the YRD, policy focusses on several initiatives, one of which is to construct adequate

transportation networks to bring about a more integrated marketplace. Shanxi is also endowed with a high concentration of transport infrastructure because of its role as a coal producer (Demurger, 2001). Fourth, it seems that provinces with a large surface area have below average transport infrastructure density. This is true for the autonomous regions of China (i.e. Inner Mongolia, Tibet and Xinjiang). It is possible that these atypical regions (in terms of surface area) have the potential to affect the results which will be presented in section 5. Hong et al. (2011) recognized that these regions had the potential to skew results, and left them out of their analysis. Nevertheless, no region will be excluded from the analysis of the main results, because it is in the interest of the thesis to explore all regions. However, atypical regions will be excluded in the robustness checks section.

Figure 2 – The sum of average highway, railway and navigable inland waterway transport density,

2004 – 2012

(10)

9 is somewhat surprising given the vast literature which finds a positive correlation between transport infrastructure and economic growth. Coastal provinces such as Beijing, Shanghai, Zhejiang and Guangdong, which are heavily endowed with transport infrastructure, seem to be experiencing below average growth rates between 2004 and 2012. On the contrary, Inner Mongolia and Shaanxi, who have a relatively low transport infrastructure concentration have had tremendous average growth rates in excess of 20% per annum throughout the same period.

There could be several reasons for this result. First, as briefly mentioned in the introduction, a range of policies were launched to enhance the growth potential of China’s inland regions. The Chinese

government engaged in a five year Western Development Plan in 1999. The plan was designed to attract foreign investment, prevent the migration of skilled workers to richer coastal provinces and to build infrastructure for poor interior regions (Schrei, 2003). The Northeast China Revitalization was adopted in 2003 to rejuvenate industrial bases in Northeast China and coordinate their development strategies (Climate Connect, 2011). The Central China Plan was announced in 2004 to accelerate development in China’s central provinces (China Business Review, 2010).

Second, a highly developed transport infrastructure system could be one characteristic of the region already being highly developed, and therefore such infrastructure may have a smaller effect on

economic growth rates than less developed areas. In other words, there may be a catch up phenomena whereby provinces with a relatively low level of real GDP per capita in the previous period, grow at a faster rate in the current period. This would support the idea that China has experienced some form of convergence. Comparing figure 2 and 4, one can clearly see that high levels of real GDP per capita are concentrated in coastal provinces, where such regions are also endowed with dense infrastructure systems. Furthermore, figure 5 shows a moderate negative correlation between the average annual growth rate of real GDP per capita between 2004 and 2012, and the natural log of the level of real GDP per capita in 2004. Therefore, providing reasonable evidence of a degree of convergence between Chinese provinces.

Third, one explanation put forth by Banerjee et al. (2012) (who found zero effect of transport infrastructure access on economic growth), is poor factor mobility. They proposed that considerable transport infrastructure could bring about sizeable growth effects to the whole Chinese economy. However, the localization of the gains to relatively well transport connected provinces is limited due to a lack of factor mobility.

Low labor mobility could stem from the Chinese “hukou system”, which inhibits the ability of China to reap the full benefits of urban agglomeration (Bosker et al. 2012). A hukou is a record which registers the household and individual to a particular area. To have full access to schools and hospitals at the lower urban costs the individual needs to hold an urban hukou. If the individual was born in and resides in a rural area, then his/her hukou is also registered there. Changing hukou status is a very difficult and bureaucratic endeavor. The Communist Party of China instigated a command economy in 1949, using the hukou system as one of its tools to control the migration of labor, keeping the cost of urban labor low, and allowing the urban middle class to retain their privileges (The Economist, 2014). It is now described as a caste system and in need of desperate reform.

(11)

10 Figure 3 – The average annual growth rate of real GDP per capita, 2004 – 2012

(12)

11 Figure 5 – Growth performance and initial level of real per capita GDP

4. Research Design and Methodology

4.1 Data collection and sample design

To investigate the impact of transport infrastructure on economic growth, data from several sources are utilized. Data from the National Bureau of Statistics (NBS), CEIC data, and the populstat site are the main data sources drawn for the independent, dependent and control variables.

Annual data from 2004 to 2012 is used in the thesis on a total of 31 Chinese provinces. The year 2013 is not included in the analysis due to limited data availability. The thesis will only explore the years following the several policies launched by the central government to enhance the growth potential of inland provinces, therefore the years preceding 2004 are left out of the analysis.

Table 1 gives a comprehensive view of how each variable is constructed in the thesis, the units by which they are measured, as well as the mean and standard deviation of each variable.

8 8.5 9 9.5 10 10.5 11 7 9 11 13 15 17 19 21 23 25

Natural log of real GDP per capita 2004

A ve rag e an n u al g ro wth rate o f r e al G D P per c ap ita (% ) b e twe e n 20 04 an d 20 12

Provincial real GDP per capita in 2004 and growth

performance between 2004 and 2012

(13)

12 Table 1 – Descriptive statistics and construction of variables

Variable Construction Unit Mean Standard

Deviation

Annual real GDP per capita growth

The percentage change in the level of real GDP per capita based on the previous year

Percentage .1659 .0645

The natural log of one period lagged level of real GDP per capita

The natural logarithm of one period lagged level of per capita real GDP in Chinese Yuan

Logarithm of per capita Chinese Yuan

9.8077 .6391

Highway density Length of highways divided by

province surface area Kilometers divided by kilometers squared

.6974 .4526

Railway density Length of railways in operation divided by province surface area.

Kilometers divided by kilometers squared

.0190 .0164

Waterway density Length of navigable inland waterways divided by province surface area. Kilometers divided by kilometers squared .0368 .0729

GFCF Gross fixed capital formation over GDP level

Population density Total provincial population divided

by province area 10,000 persons divided by kilometers squared

.0413 .0603

Level of education The ratio of the population aged 6 and over with at least secondary education to the total population aged 6 and over

FDI Foreign direct investment as a share of real GDP

The main explanatory variables are highway, railway and waterway density. The thesis investigates each of these variables separately to determine the relative importance of each variable, and their impact on annual real GDP per capita. By far, the highest density of transport infrastructure is that of highways, with a density of 0.6974 km/km2_{. This is followed by waterways, and then railways, with a}

density of 0.0368 km/km2_{, and 0.0190 km/km}2_{respectively. There is also the largest variation from the}

mean in the density of highways, with a standard deviation of 0.4526. This is followed by waterways, and then railways, with a standard deviation of 0.0729 and 0.0164 respectively.

The measure of transport density is the best proxy for transport infrastructure given the data available. Moreover, given the huge differences in the surface area of each province in China, it is important to incorporate the size of each province in to the proxy. This is precisely why the length of each node of transport is divided by the provincial surface area.

(14)

13 However, as briefly discussed in section 2.2, transport infrastructure density as a proxy is not without its limitations. To be more precise, it is a quantitative measure, and therefore does not include the quality of such infrastructure. The use of the proxy highway density in the thesis only measures the highest quality of roadway infrastructure in terms of road conditions. There are two other lower forms of roadway infrastructure not taken into account by this proxy, i.e. first class roadways and second class roadways. Despite these two other types exhibiting a lower roadway quality, they do also have the potential to also influence growth rates.

Alternative measures may include all forms of roadway infrastructure in despite of its possible lack of quality. To incorporate such quality measures, proxies may be developed using a weighting system, by which higher levels of infrastructure are given a greater relative significance.

4.2 Methodology

The thesis uses panel data because the variables utilized are measured across a range of entities (provinces), over a period of 9 years. To ensure valid research outcomes it is important to use the correct method when measuring the value and significance of the variables of interest, as well as the controls. Pooled ordinary least squares (POLS), random effects, fixed effects, and IV with fixed effects regression methods are used frequently on panel data, and will also be used in the thesis, following a similar methodology used by Demurger (2001) and Hong et al. (2011).

POLS techniques are subject to endogeneity, such as reverse causality and omitted variable bias. They also make the dubious assumption that individual effects are identical and constant across entities. However, they do provide benchmark results.

Fixed and random effects models take in to account the panel structure of some forms of data (such as that used in the thesis). However both models are subject to several advantages and disadvantages. These have been concisely summarized by Nerlove (2007). More precisely, the advantage of the fixed effects model is that it allows for the individual- and/or time-specific effects to be correlated with the explanatory variables. The disadvantage of this model is that when the number of observations become large, the number of unknown parameters increase, and so the estimator may fail to yield consistent estimates, assuming a finite number of observations (this is commonly referred to as the classical incidental parameter problem, Neyman and Scott, 1948). Furthermore, the estimator does not allow the estimation of coefficients when the variable does not vary across time.

The advantage of the random effects model is that the number of unknown parameters stay constant when the number of observations becomes large. Moreover, efficient estimates can be derived using both between and within entity variation. It also allows for the impact of time-invariant variables. The disadvantage of this model is that if individual- and/or time-specific effects are correlated with the explanatory variables, it results in a misspecification of the model, and the estimates will be biased. In summation, the benefits of the fixed effects model are limitations of the random effects model, and vice versa.

(15)

14 effects and random effects regressions.

4.3 The model and measurement

Barro (1991) derives his growth equation in the investigation of the impact of various variables on economic growth. Namely, he finds a positive correlation between school enrollment and growth, as well as a negative relation between initial level of GDP and growth. Furthermore, the author finds high growth rates are positively related to higher ratios of physical investment to GDP.

The Barro-type conditional convergence framework is frequently used to estimate a growth equation on panel data (Demurger, 2001; Hong et al. 2011; Ding et al. 2006). In the thesis Barro’s growth equation is extended to account for the effects of transport infrastructure endowment. A coastal dummy variable and three separate transport infrastructure density interaction terms in the cross region economic growth regression is used to investigate coastal-inland economic growth differences attributed to transport infrastructure in equation (1). The coastal dummy variable is also included separately to capture coast-inland differences not explained through transport infrastructure. Furthermore, the possible nonlinear relationship between economic growth and transport infrastructure endowments is investigated with various squared endowment terms in equation (2). Hence, the models have the following form:

git = Ai + Nt + α1(Xit) + α2(Xit.Cit) + α3(Cit)+ α4(Tit)+ εit (1)

git = Bi + Mt + β1(Xit) + β2(Xit2) + β4(Tit) + υit (2)

The entity (province) is denoted by letter i (i = 1,...,N), and time (year) by t (t = 1,...,T).

Annual real GDP per capita growth is g (which is the outcome variable of interest). Province specific parameters are A and B for equations (1) and (2) respectively. Time specific parameters are N and M for equations (1) and (2) respectively. The vector of proxies for transport infrastructure endowments is X (i.e. railway, waterway, and highway density). The coastal dummy is C (which equals one if the province is located on the east coast – Hebei, Beijing, Tianjin, Shandong, Jiangsu, Shanghai, Zhejiang, Fujian, Guangdong, and Hainan).

The vector of control variables is T, which includes the natural log of one period lagged level of real GDP per capita. This variable is used to test the conditional convergence hypothesis – a negative significant coefficient on this variable supports the idea that the higher last period real GDP per capita, the lower present real GDP per capita growth. Gross fixed capital formation (GFCF) is another control variable, and is a proxy for physical capital (GFCF only measures spending on fixed assets minus disposal of fixed assets. It does not take in to account consumption of fixed capital, i.e. depreciation of fixed assets). Population density is a proxy for labor force availability, and is another included control variable. Finally, the last control variable is the level of the population with at least secondary

education, which is a proxy for labor quality. The error terms are ε and υ for equations (1) and (2) respectively.

To understand the 2SLS procedure it is important to grasp the concepts of endogeneity and exogeneity. When a variable is correlated with the error term it is endogenous. Conversely, when a variable is not

(16)

15 correlated with the error term it is exogenous. Moreover, endogenous variables are determined within the model, whereas exogenous variables are determined outside. Reverse causality is a type of

endogeneity which is likely present in equation (1) and (2).

To illustrate, rapid economic growth and urbanization can itself place a strain on transport

infrastructure systems (Yu et al. 2012). Section 2.1 highlighted how causality runs from transport infrastructure to economic growth. However, we may consider that economic growth can itself boost government funds to be spent on the purchase of better transport systems (Hong et al. 2011). In other words, reverse causality is present when movements in economic growth causes variation in transport infrastructure density. Transport infrastructure will then be determined within the model, and is

endogenous. Reverse causality between transport infrastructure and economic growth can be addressed using a 2SLS procedure.

Due to the 2SLS procedure deployed in the thesis (as well as the various other methods mentioned), equation (1) and (2) can be treat as second stage regression equations. The first stage will take the form:

Xit = π0 + π1(Zit) + νit (3)

Z denotes a vector of IVs. As explained in section 2.2, the thesis uses a range of IVs which were utilized by Demurger (2001) and Hong et al. (2011). Namely: the one year lagged values of: GFCF over GDP level; FDI as a share of real GDP; as well as highway, railway and waterway transport density variables.

There are two conditions which ensure that the included IVs are valid (i.e. the instrument measures what it is intended to measure). They are instrument relevance and instrument exogeneity. When the instrument is relevant, then it must be correlated with the endogenous explanatory variable. In our case, valid instruments must be correlated with the endogenous transport infrastructure density term.

Furthermore, instrument exogeneity is satisfied when the part of the variation in the endogenous variable captured by the instrumental variable is exogenous. So the instrument must be uncorrelated with the error term in equation (1). When the instrument is both relevant and exogenous then it captures the movements in the endogenous variable that are exogenous. This exogenous variation can then be used to estimate a consistent and unbiased coefficient of interest, which is done using an IV estimator commonly referred to as 2SLS.

To be more clear as to just how the 2SLS procedure works, in the first stage regression (equation (3)), the endogenous variable is regressed on all instrumental variables, as well as all exogenous variables using an OLS procedure. The same method is carried out for all endogenous regressors until we obtain our desired number of predicted values. In the second stage regression (equation (1) and (2)), the outcome variable (real per capita GDP growth), is regressed on the predicted values of the endogenous variables along with the included exogenous variables also using OLS. The 2SLS estimators are the second stage regression equation estimators.

Upon investigation of equations (1) and (2) it is clear that alongside the linear transport infrastructure endogenous terms, the equations contain quadratic terms of these endogenous variables, and interaction terms which contain the endogenous variables. The instrumentation of these quadratic variables, and interaction terms require a more complex method. More specifically, in the first stage regression equation, the endogenous linear transport infrastructure variable is regressed on all exogenous

(17)

16 instruments. The square of the predicted value, as well as the coastal interaction with the predicted value is then used to instrument for the quadratic transport infrastructure variable, and the transport-coast interaction term respectively. As explained in Wooldridge (pp. 236, 2010), some fall in to the trap of using the “forbidden regression”. This is a phrase coined to describe the mistake of replacing the nonlinear endogenous variables with the square of the same nonlinear predicted values from the first stage equation in the second stage equation. The use of the forbidden regression will lead to completely incorrect coefficient estimates.

In theory, given that the lagged endogenous variables are past values or predetermined, they can be treat as exogenous (the present dependent variable does not influence the predetermined lagged variable – so reverse causality is not an issue). The non-transport related density variables are treat as exogenous in the 2SLS system of equations.

The Hansen-Sargan J-statistic of over identifying restriction test is used to test for the validity of the IVs (Note: this test is only useful if the number of IVs exceeds the number of endogenous variables).

5. Results

5.1 Main results table

Table 2 and 3 report the main panel data estimation results for equation (1) and (2) respectively, using POLS, fixed effects, random effects, and fixed effects 2SLS. The cluster robust command was used in STATA to indicate that the observations are clustered in to provinces, and also that the observations could be correlated within provinces, but remain independent between provinces. Therefore, the calculated standard errors are robust to problems associated with the regression residuals exhibiting heteroskedasticity and/or serial correlation. In other words, robust to the residuals variance not being constant between provinces and/or being correlated over time.

In tables 2 and 3, columns (1), (3), (5) and (7) are the model estimates which do not include control variables: natural log of one period lagged level of real GDP per capita; gross fixed capital formation; population density and secondary education level. Columns (2), (4), (6) and (8) do include these control variables. The thesis utilizes multiple specifications for each method to explore which control variables are the most relevant to the analysis. Certain controls could be crucial to the analysis if they for

(18)

17 Table 2 – Estimation results (Equation (1))

POLS (1) POLS (2) Random effects (3) Random effects (4) Fixed effects (5) Fixed effects (6) Fixed effects, 2SLS (7) Fixed effects, 2SLS (8) Constant .1743*** (18.47) .3914*** (4.36) .1747*** (14.15) .3298*** (7.38) .0478 (0.39) .0347 (0.17) N/A N/A Highway densityit -.0031 (-0.17) .0141 (0.70) -.0031 (-0.21) .0513 (0.81) -.0029 (-0.11) .0014 (0.04) .0764 (0.91) .0351 (0.42) Highway densityit*Coastit -.0396 (-1.22) -.0287 (-0.88) -.0435** (-2.16) -.0672 (-1.42) -.0644** (-2.11) -.0555* (-1.72) -.2121 (-1.11) .1367 (0.54) Railway densityit .1182 (0.14) -.1417 (-0.13) .0793 (0.12) -.1113 (-0.14) -5.6290** (-2.23) -4.3755 (-1.36) -17.2687 (-1.61) 32.4853 (1.13) Railway densityit*Coastit -.4588 (-0.50) .3856 (0.36) -.4027 (-0.54) .3900 (0.43) 4.5713* (1.72) 4.2770 (1.22) 16.2778 (0.87) -48.0610 (-1.26) Waterway densityit .1890 (0.68) .2984 (0.98) .1935 (0.94) .2137 (1.49) 22.5927** (2.53) 27.3051*** (2.97) 8.5548 (0.32) 10.2297 (0.58) Waterway densityit*Coastit -.1852 (-0.65) -.1208 (-0.38) -.1822 (-0.85) -.1831 (-0.54) -24.3345** (-2.53) -28.7076*** (-2.84) 3.9083 (0.09) 22.1470 (0.55) Coastit .0223 (1.08) .0150 (0.69) .0247 (1.52) .0084 (0.79)

N/A N/A N/A N/A

Ln(GDPi-1) -.0333*** (-2.69) -.0376*** (-4.23) .0065 (0.27) -.2107* (-1.77) GFCFit .0814* (1.84) .0878** (2.52) .0168 (0.27) .1682 (1.48) Population densityit -.3097 (-1.51) -.2334*** (-3.02) -.4096 (-0.51) 3.6821 (0.93) Secondary Educationit .0917 (1.55) .0913 (1.39) -.2138 (-1.26) .6870 (1.10) Observations 279 279 279 279 279 279 248 248 R-Squared 0.1003 0.1286 Hausman test (p-value) 54.416*** (0.0000) 26.324** (0.0178) Hansen-Sargan J-statistic of over identifying restriction (p-value) 1.239 (0.2060) 2.561 (0.1723)

*, **, and *** are significance values at 10%, 5%, and 1% respectively. The t-statistics are in brackets for each row of explanatory variables.

The instruments used in the 2SLS model are one year lagged values of highway density, railway density, waterway density; gross fixed capital formation over GDP level; and FDI as a share of real GDP.

The STATA command xtserial was used to test for serial correlation in the panel data. The test statistics are 1.236 and 0.580, with probability values of 0.3586 and 0.4521 for the regressions without and with control variables respectively.

(19)

18 Table 3 – Estimation results (Equation (2))

POLS (1) POLS (2) Random effects (3) Random effects (4) Fixed effects (5) Fixed effects (6) Fixed effects, 2SLS (7) Fixed effects, 2SLS (8) Constant .1674*** (15.79) .4050*** (5.14) .1672*** (12.06) .4089*** (9.64) .0259 (0.09) .0462 (0.13) N/A N/A Highway densityit .0168 (0.36) .0470 (1.00) .0257 (0.63) .0426 (1.09) .0720* (1.70) .1400** (2.52) -.0657 (-0.69) -.0110 (-0.03) Highway densityit2 -.0209 (-0.84) -.0266 (-1.06) -.0267 (-1.25) -.0242 (-1.25) -.0552** (-2.02) -.0836*** (-2.72) .0362 (0.59) .0533 (0.21) Railway densityit .8187 (0.71) .0668 (0.05) .6466 (0.82) .0619 (0.06) -3.5157 (-1.46) -.1436 (-0.06) -5.3474 (-0.79) -1.4456 (-0.16) Railway densityit2 -19.5253 (-1.46) -1.9796 (-0.13) -17.4917* (-1.90) -1.9769 (-0.18) 18.0643 (0.87) -11.0965 (-0.33) -35.9725 (-0.40) 368.5845 (0.94) Waterway densityit -.0559 (-0.25) .1929 (0.78) -.0723 (-0.31) .1905 (1.10) 5.8498 (0.53) 10.1268 (0.81) 4.6138 (0.20) 5.0397 (0.08) Waterway densityit2 .0181 (0.03) -.2717 (-0.28) .1072 (0.15) -.2793 (-0.39) -4.7782 (-0.27) -12.0407 (-0.60) 46.2568 (0.61) -131.7847 (-1.17) Ln(GDPi-1) -.0364*** (-3.23) -.0368*** (-5.85) -.0156 (-0.65) -.1862* (-1.92) GFCFit .0978** (2.17) .0912*** (3.51) .0320 (0.58) .0376 (0.37) Population densityit -.1553 (-0.55) -.1581 (-0.99) .5780 (0.46) -13.9417 (-1.11) Secondary Educationit .0857 (1.42) .0866 (1.36) -.1489 (-0.83) .8595 (0.97) Observations 279 279 279 279 279 279 248 248 R-Squared 0.0871 0.1299 Hausman test (p-value) 15.251** (0.0184) 20.519** (0.0126) Hansen-Sargan J-statistic of over identifying restriction (p-value) 1.385 (0.2361) 3.953 (0.1386)

The STATA command xtserial was used to test for serial correlation in the panel data. The test statistics are 1.192 and 0.915 with a probability values of 0.3151 and 0.3463 for the regressions without and with control variables respectively.

(20)

19

5.2 POLS and random effects model estimates

Both the POLS estimates and the random effects estimates rely on rather similar methods, so it

therefore makes sense to discuss them both together. There is however, one crucial difference, which is that the POLS estimates are more efficient when assuming no heterogeneity among provinces.

Conversely, the random effects model produces better estimates when there is heterogeneity, but it is not related to transport infrastructure.

For the POLS models, there are no significant coefficients for any of the transport infrastructure density related variables both with and without control variables. This implies no relationship between

transport infrastructure endowment and economic growth. But in particular, the estimates imply no coastal-inland growth inequality attributed to transport infrastructure.

For the random effects model estimates in table 2 the coefficient on the highway density, coast interaction is negative and significant at the 5% level without control variables. However, neither the highway nor the highway squared variables enter the model significantly. The interaction term coefficient can therefore be interpreted as a negative economic growth effect of an additional unit of highway density, but only in the coastal provinces. This result should be treat with a degree of caution because when control variables are not omitted from the random effects regression equation, the coefficient become insignificant (and similar to the POLS estimate).

There could be several reasons for the insignificant transport infrastructure variables in the POLS and random effects regression equations. For example, despite the low standard errors which may be present for some variables, they will remain insignificant if the coefficient estimates are particularly low. Upon investigation this appears to be the case for the highway density variables. Another reason could be that the coefficient is in fact rather large, but due to a small sample size is measured

inaccurately. Given that the data utilized in the thesis is measured across quite a short time frame, this reason is also possibly true.

If we turn our attention to the control variables, when such variables are added to the POLS and ran-dom effects equations, there is a significant negative coefficient at the 1% level for the natural log of one period lagged level of real GDP per capita. Implicitly, those regions with low levels of real GDP per capita tend to grow at a faster rate. This is strong evidence in support of the conditional conver-gence hypothesis. Conditional converconver-gence is when each region converges to its own steady state. This differs from absolute convergence, which states that in the long run each region will attain not only an identical rate of economic growth, but also the same level of per capita income. The result does not im-ply absolute convergence between provinces.

Another significant coefficient at the 10%, 5%, 5% and 1% level for the POLS and random effects regressions in table 2, as well as the POLS and random effects regressions in table 3 respectively, is on the proxy for physical capital – GFCF. This time the coefficient is positive, suggesting that higher spending on fixed assets (minus disposals), as a fraction of GDP boosts economic growth.

Population density (the proxy for labor availability) enters the POLS regression equations

insignificantly. Hong et al. (2011) explains the insignificant coefficient on population density as the variable not capturing the effect of labor availability, but rather a crowding out effect. In other words, the crowding out effect of additional labor offsets the labor availability effect. This result is also true

(21)

20 for the random effects equation in table 3. However, interestingly there is a negative significant

coefficient at the 1% level on the population density variable in table 2. According to this result, a denser population may have negative growth effects. Malthus (1798) explains how population growth may exceed the growth in food supply over time, leaving a country in the grips of famine, war and negative economic growth. Nevertheless, this view is seen as outdated, and is widely disregarded. Secondary education level (our proxy for labor quality) enters the POLS and random effects equations insignificantly, suggesting no positive effect of labor quality.

5.3 Fixed effects model estimates

In table 2 there are negative significant coefficients on the highway, coast interaction terms at the 5% and 10% level without and with control variables respectively. Implicitly, additional highway

infrastructure will have negative growth effects in coastal provinces. However, table 3 reports that there are positive significant coefficients on the highway infrastructure density variable at the 10% and 5% significance levels, without and with control variables respectively. But also, negative significant coefficients on the squared highway density variables at the 5% and 1% significance levels, without and with control variables respectively. This can be interpreted as a positive but diminishing overall effect of additional highway infrastructure. Therefore, those provinces with relatively less developed highway infrastructure, benefit relatively more in terms of economic growth from the addition of such infrastructure.

It seems puzzling that table 2 implies a negative growth impact of additional highways in coastal provinces, but no effect in inland provinces, and table 3 reports a positive overall effect of additional highways. Upon inspection, in table 2 the significance of the interaction term actually falls with the addition of control variables. It is possible that if an extended set of control variables were used in the thesis, the significance of the interaction term would no longer be present, and additional highway infrastructure would not contribute to differences in economic growth rates between inland and coastal provinces. Nevertheless, this hypothesis is left untested.

Table 2 reports negative and positive significant coefficients at the 5% and 10% level for the railway density and railway density, coast interaction term respectively (without controls). This suggests that additional railway infrastructure will have negative growth effects for both inland and coastal

provinces, but the effect is greater for inland provinces. That is, the impact of an additional unit of railway infrastructure equals -5.6290 and -1.0577 (-5.6290 + 4.5713) for inland and coastal provinces respectively. This result should be treat with a degree of caution given that when control variables are included, the regressors become insignificant.

Table 2 also indicates positive significant coefficients at the 5% and 1% levels for the waterway density variables without and with control variables respectively. Furthermore, there are significant negative coefficients on the waterway density, coast interaction terms at the 5% and 1% levels without and with control variables respectively. This suggests that there are positive economic growth effects of

additional waterways in inland provinces, but on the coast the impact is negative. To illustrate, without control variables one unit of additional waterways has a growth effect of -1.7418 (22.5927 - 24.3345) on the coast, and with controls equals -1.4025 (27.3051 - 28.7076).

(22)

21 The coast dummy variable is omitted from the fixed effects regression equations in both tables. The reason behind this is that such regressions involve subtracting the average entity value from the regressors. In the case of a dummy variable like coast (which equals either one or zero), when

subtracting the entity mean from the variable, the resulting value will be zero. It is therefore apparent that fixed effects models can only include variables which vary over time.

The fixed effects models have insignificant coefficients on the natural log of one year lagged level of real GDP per capita, GFCF, population density and secondary education explanatory variables. Meaning that there is no evidence of conditional convergence; and no positive growth effects of physical capital, labor availability and labor quality according to this model.

A robust Hausman test is carried out to determine which model is preferred between the fixed effects and random effects regressions. The original form of the Hausman test in STATA is non-robust, and also assumes the random effects estimator to be efficient. This assumption is only fulfilled if the errors are independent and identically distributed (i.i.d.). This is certainly not the case if the cluster robust standard errors vary greatly from the default standard errors. The xtoverid command (which does allow for clustered robust standard errors) was used after the robust random effects regression. This is

because a test of whether the random effects model is preferred to the fixed effects model can also be seen as a test for overidentifying restrictions, and is numerically identical to a robust Hausman test. The probability values of 0.0000, 0.0178, 0.0184, and 0.0126 indicate that we can reject the null hypothesis that the random effects model is preferred at the 1%, 5%, 5% and 5% significance levels for the

random effects models in table 2 without and with controls, as well as table 3 without and with controls respectively. So in fact the fixed effects models yield more consistent estimates both with and without control variables in both model specifications.

5.4 Fixed effects 2SLS model estimates

For the fixed effects 2SLS estimates, the Hansen-Sargan J statistic tests the null hypothesis that the excluded instruments are valid. That is, uncorrelated with the error term (exogeneity) and rightly excluded from the estimated equation (1). The reported probability values in the fixed effects 2SLS are 0.2060, 0.1723, 0.2361 and 0.1386 for table 2 regressions without and with controls, as well as table 3 regressions without and with controls respectively. We can therefore conclude that the excluded instruments are valid in this model because the probability values are greater than the 10% critical value, so the null hypothesis is consequently accepted.

Note: The xtivreg2 command was used in STATA, as created by Schaffer (2007) to allow for clustered robust standard errors whilst using the fixed effects model and testing for the validity of excluded instruments. There is no constant term reported when using the fixed effects model and xtivreg2, because such a model relies on within variation to estimate the value of the coefficients. Therefore, it cannot estimate coefficient values which do not vary within panels, such as the constant term (and coastal dummy variable). Interestingly, and somewhat rather counterintuitively, the xtreg fixed effects STATA routine does report a constant by utilizing cross sectional variation. In theory, the fixed effects model should generate a separate constant for each entity (province). Upon investigation it appears that the constant term reported in this case is an average of the fixed effects.

(23)

22 once the endogeneity problem is corrected for using the instrumental variables procedure, the

insignificant coefficients on the transport variables imply that the significant results which were found previously in the POLS, random effects and fixed effects models are merely a product of reverse causality. So it can be concluded that there is no transport-growth relationship, and so no growth inequality between the coastal and inland regions attributed to transport infrastructure according to the fixed effects 2SLS model.

Again, it is important to explore the reasons behind the insignificance of the transport infrastructure variables. As mentioned in section 5.2, despite low standard errors, the coefficients will remain insignificant if the coefficients are too low. This appears to be the case for the highway infrastructure variables. Furthermore, the coefficients may in fact be sufficiently large, but measured imprecisely because of the small sample size. Moreover, we cannot rule out a weak instrument problem. This is present when the excluded instrument is only weakly correlated with the endogenous regressor. The Stock-Yogo weak ID test critical values are normally used to test the null hypothesis of weak IVs. However, the STATA routine xtivreg2 only shows critical values for up to 3 endogenous variables and 100 excluded IVs (Baum, Schaffer and Stillman, 2007). Given that equation (1) and (2) contain 6 endogenous variables each, the test could not be carried out.

With the exception of the natural log of one year lagged level of real GDP per capita variable, the other control variables enter the regression insignificantly, suggesting that the 2SLS procedure again

invalidates the significance of such variables which were found significant in the POLS and random effects models. Table 2 and table 3 only report the significance of the natural log of one period lagged level of real GDP per capita at the 10% level, therefore providing only moderate evidence in support of the conditional convergence hypothesis.

5.5 Robustness test

Demurger (2001) excluded municipalities (Beijing, Tianjin, and Shanghai) from their data analysis because they are highly atypical in terms of how they are governed, their land area and economic structure. Chongqing was given municipality status after 1997, and so was treat as part of Sichuan in the last two years of the author’s analysis. Hong et al. (2011) exclude Inner Mongolia, Tibet and Xinjiang autonomous regions because of their huge relative size, and their poor transport systems (which can be attributed to low demand and unsuitable terrain). Given the characteristics of the above regions they have the potential to skew results.

Robustness is confirmed by identifying variables whose importance remains constant across various specifications, and with reduced (or expanded) datasets. To check the robustness of the results obtained in table 2 and 3, the sample is adjusted by removing municipalities, as well as the Inner Mongolia, Tibet and Xinjiang autonomous regions in table 4 and 5.

(24)

23 Table 4 - Estimation results (Equation (1)) without atypical municipalities and Inner Mongolia, Tibet and Xinjiang autonomous regions.

POLS (1) POLS (2) Random effects (3) Random effects (4) Fixed effects (5) Fixed effects (6) Fixed effects, 2SLS (7) Fixed effects, 2SLS (8) Constant .1795*** (16.34) .3432*** (3.54) .1593*** (20.81) .3902*** (6.11) .0942 (0.58) .2508 (0.91) N/A N/A Highway densityit .0006 (0.03) -.0032 (-0.13) .0003 (0.04) -.0153 (-0.13) -.0050 (-0.18) -.0207 (-0.55) -.2104 (-1.07) .0117 (0.11) Highway densityit*Coastit -.0468 (-1.30) -.0269 (-0.73) -.0634** (-2.33) -.0705 (-1.32) -.0703** (-2.41) -.0385 (-1.16) .2539 (0.92) .1500 (0.47) Railway densityit -.3213 (-0.35) .0547 (0.04) -.3190 (-0.57) .0551 (0.06) -4.8329** (-1.97) -5.4273 (-1.44) 24.475 (0.85) 31.7537 (0.98) Railway densityit*Coastit 1.5840 (1.03) .4621 (0.28) 1.5163* (1.72) .4600 (0.54) 4.7456* (1.79) 5.9867** (2.00) -43.9265 (-1.13) -36.7448 (-0.81) Waterway densityit .1723 (0.56) .2897 (0.79) .1115 (0.91) .3165 (1.27) 24.0415* (1.80) 19.2551 (1.57) -27.9563 (-0.41) -4.1976 (-0.07) Waterway densityit*Coastit -.0338 (-0.10) -.1906 (-0.52) -.0274 (-0.17) -.1362 (-0.79) -31.5023** (-2.25) -26.2395** (-2.01) .2697 (0.00) -22.6630 (-0.19) Coastit -.0164 (-0.64) .0024 (0.08) -.0016 (-1.31) .0021 (0.16)

N/A N/A N/A N/A

Ln(GDPi-1) -.0231* (-1.74) -.0362*** (-2.92) .0111 -.2911 (-1.56) GFCFit .0897* (1.92) .0889*** (3.53) .0113 .2208 (1.01) Population densityit .3959 (0.61) .4670 (0.96) -6.8040 -4.2674 (-0.51) Secondary Educationit -.0129 (-0.16) -.0132 (-0.21) -.0310 1.4478 (1.38) Observations 216 216 216 216 216 216 192 192 R-Squared 0.0532 0.0855 Hausman test (p-value) 23.2649** (0.0130) 56.2137*** (0.0000) Hansen-Sargan J-statistic of over identifying restriction (p-value) 1.579 (0.1183) 2.049 (0.1524)

The STATA command xtserial was used to test for serial correlation in the panel data. The test statistics are 1.193 and 0.585, with probability values of 0.3071 and 0.4522 for the regressions without and with control variables respectively.

(25)

24 Table 5 - Estimation results (Equation (2)) without atypical municipalities and Inner Mongolia, Tibet and Xinjiang autonomous regions.

POLS (1) POLS (2) Random effects (3) Random effects (4) Fixed effects (5) Fixed effects (6) Fixed effects, 2SLS (7) Fixed effects, 2SLS (8) Constant .1555*** (9.68) .3910*** (4.23) .1551*** (10.60) .3540*** (9.70) .1588 (0.57) .3869 (1.01) N/A N/A Highway densityit .0364 (0.70) .0703 (1.29) .0381 (0.84) .0630 (1.60) .1190*** (3.05) .1214** (1.99) .1103 (1.37) .5367* (1.83) Highway densityit2 -.0290 (-0.95) -.0439 (-1.39) -.0302 (-1.25) -.0297* (-1.81) -.0902*** (-4.10) -.0842*** (-2.59) -.1172** (-2.03) -.2440 (-1.41) Railway densityit 3.4052 (1.31) 1.8138 (0.54) 3.4260 (1.43) 1.8846 (0.84) 4.5853 (1.20) 6.3421 (1.48) .9128 (0.09) 3.0980 (0.35) Railway densityit2 -123.1619 (-1.58) -57.6222 (-0.64) -124.0323* (-1.76) -57.5703 (-0.97) -202.0339* (-1.92) -207.4955* (-1.95) -61.1327 (-0.26) 189.9504 (0.47) Waterway densityit -.5726* (-1.82) -.1049 (-0.27) -.5762* (-1.95) -.1152 (-0.37) -.0863 (-0.01) .0740 (0.01) 10.7533 (0.38) 50.3632 (1.22) Waterway densityit2 2.3020* (1.76) .7490 (0.53) 2.3233** (1.99) .7950 (0.77) -7.6394 (-0.29) -9.4334 (-0.34) -172.032 (-0.61) -132.0113 (-0.54) Ln(GDPi-1) -.0321** (-2.44) -.0586*** (-4.52) -.0117 (-0.47) -.2013** (-2.52) GFCFit .1082** (2.34) .1256*** (5.64) .0299 (0.51) .1034 (1.12) Population densityit .2449 (0.45) .1952 (0.60) -5.9560* (-1.94) -14.1216*** (-3.77) Secondary Educationit -.0093 (-0.11) -.0171 (-0.15) .0218 (0.13) .6329 (1.56) Observations 216 216 216 216 216 216 192 192 R-Squared 0.0335 0.0898 Hausman test (p-value) 39.880*** (0.0000) 17.8329** (0.0327) Hansen-Sargan J-statistic of over identifying restriction (p-value) 1.211 (0.3492) 2.608 (0.2715)

The STATA command xtserial was used to test for serial correlation in the panel data. The test statistics are 1.758 and 1.129 with a probability values of 0.1502 and 0.2989 for the regressions without and with control variables respectively.

(26)

25 The results from the robustness check in tables 4 and 5 yield some results which are inconsistent with those main results reported in tables 2 and 3. In particular, the random effects model yields a positive significant coefficient on the railway density, coast interaction term in table 4 without controls at the 10% level. Furthermore, in table 5 the waterway density and waterway density squared variables yield negative and positive significant coefficients at the 10% and 5% level without controls in the random effects model. However, when control variables are added all variables become insignificant.

Interestingly, the population density variable, which was significant at the 1% level in table 2 is insignificant and the opposite sign in table 4. Therefore, we can conclude that the significance of the variable is entirely driven by atypical provinces.

With regards to the fixed effects model (preferred according to the Hausman test), table 4 reports a significant positive coefficient at the 5% level on the railway density, coast interaction term with controls. The result implies that once atypical regions are excluded then there is a positive growth effect of additional railway infrastructure on the coast. However, the waterway density variable and the highway density, coast interaction term which were significant at the 1% and 10% significance levels in table 2, become insignificant in table 4 with the exclusion of atypical regions. Therefore, the results are entirely driven by atypical regions.

Possibly the most surprising result from table 5 is that the highway density variable becomes positive and significant in the fixed effects 2SLS model at the 10% level including control variables. This implies that when atypical regions are excluded, additional highway infrastructure will have a positive impact on economic growth, albeit this is only moderate evidence given the 10% significance level. Moreover, the population density variable also becomes significant and negative at the 1% level in the fixed effects 2SLS model excluding atypical regions with controls in table 5. This suggests that the insignificance of population density in table 3 is driven by atypical regions.

There are some results which do remain consistent between tables 2, 3, 4 and 5. The insignificance of secondary education (labor quality) holds true in both tables. Therefore, labor quality does not

influence growth rates even with the exclusion of atypical regions.

The use of lagged variables to instrument for transport infrastructure endowment terms may not be valid under certain circumstances. According to Angrist and Krueger (2001), if the equation error or the omitted variables exhibit serial correlation, then the use of lagged endogenous variables as instruments becomes problematic.

To investigate whether or not serial correlation is in fact prevalent in the panel dataset the xtserial STATA command (as developed by Drukker, 2003) was carried out on the full panel of 31 provinces, as well as the reduced panel (excluding atypical regions). Note that, a significant test statistic provides evidence of the presence of serial correlation. The test statistics are 1.236 and 0.580, with probability values of 0.3586 and 0.4521 for table 2 without and with controls respectively. In table 3 the test statistics are 1.192 and 0.915 with probability values of 0.3151 and 0.3463. In table 4 the test statistics are 1.193 and 0.585, with probability values of 0.3071 and 0.4522. Finally, in table 5 the test statistics are 1.758 and 1.129 with a probability values of 0.1502 and 0.2989. The probability values exceed the 10% critical significance level. Therefore, we can accept the null hypothesis of no first order

autocorrelation in the regressions. So it appears that the use of lagged endogenous variables to

instrument for transport infrastructure endowment is quite a valid method in both the full and reduced panel datasets.

China's transport infrastructure : the effect on regional income growth patterns