1 Introduction

Transportation and economic growth are closely linked. On the one hand, economic growth stimulates transportation demand. On the other hand, transportation improvement can drive economic growth by unlocking new sites for development, enlarging the existing product and labor markets and opening up new and larger markets (Rietveld and Nijkamp 1992; Vickerman et al. 1999), although these benefits cannot guarantee increased economic development. China’s economy has gained astonishing growth in the last few decades, which catapulted the country to its current status as the world’s second largest economy. However, investment-led growth has long been criticized as a key development strategy in China (Qin et al. 2006). In 2008, approximately USD 220 billion was spent on transport-related projects to counter the global economic crisis. In 2014, the total investment amount for highway, railway and waterway transportation stood at USD 393.25 billion in China to promote employment and grow investment circles (Xinhuanet 2015).Footnote 1 However, many critiques have been raised as to this investment-led economy and whether these projects are economically reasonable with enough considerations of travel demand (Martinsons 2008). Thus, the necessity of understanding the mutual impact of transportation in terms of transport supply as well as travel demand and economic growth is of considerable significance.

Although previous studies have emphasized on the economic effect of transport infrastructure at macrolevels since the ground-breaking study by Aschauer (1989), mixed results are shown in Table 1 because of different methodologies, geographical scales, time coverage and study contexts (Blum 1982; Boopen 2006; Farhadi 2015; Fujita and Krugman 2004; Lall 2007; Lall et al. 2004; Pradhan et al. 2013). On the US aggregate economy, transport infrastructure being viewed as public capital (Garcia-Mila et al. 1996), can promote specialization and long-run economic growth with the elasticity of 0.24, although its impact on the latter is not monotonic over time (Bougheas et al. 2000; Aschauer 1989). Tatom (1993) used the corrected statistical data and showed no evidence that an increasing public capital could produce higher private productivity. Fernald (1999) and Jiwattanakulpaisarn et al. (2010) find that when the growth of road infrastructure changes, the productivity growth disproportionately changes in different sectors of the economy in US. Crescenzi and Rodrigues-Pose (2012) took the European as the example and confirmed that transport infrastructure endowment cannot predict regional growth efficiently. Besides the direct effect, transport infrastructure would bring spatial spillover effect (Esfahani and Ramı́rez 2003) and agglomeration economies (Chatman and Noland 2011), which is conducive for increasing agricultural output (Crihfield and Panggabean 1995; Lall et al. 2004; Tong et al. 2013). However, clear evidence indicates a negative employment spillover from interstate lane-mile additions (Jiwattanakulpaisarn et al. 2010).

Table 1 Review of selected literature on the nexus between transportation (T) and economic growth (D)
Full size table

Methodologically, the cost function approach is employed by most recent approach to examine whether and how transport infrastructure affects economic growth. Both positive and negative impacts of public capital on economic growth are described in related research (Seitz 1993; Vijverberg et al. 1997; Jiang et al. 2017) and the output elasticity is estimated from − 0.92 to 0.27 (Jiang et al. 2017). However, this method is inadequate in uncovering their reverse causality from transport infrastructure to economic growth, thus producing biased estimates of the structural parameters because of the endogeneity problem (Beyzatlar et al. 2014; Fujita and Mori 2005; Pradhan 2019). Unlike the cost function approach, the production function approach as well as the OLS estimation, the causality models such as vector auto-regression, are able to investigate their mutual interaction between transport infrastructure and economic growth with the estimated elasticity from − 0.31 to 0.52 (Berndt and Hansson 1991; Boarnet 1996; Randolph et al. 1996). The results about the sign, elasticity and the direction are inconsistent in related studies as seen in Table 1.

To date, these techniques have been utilized mostly to analyze macroeconomic time series. Panel data turns out to be more popular than time series data or cross-sectional data since it shows more accuracy such as increasing the degrees of freedom and reducing the collinearity among independent variables (Erdil and Yetkiner 2004). Some techniques examining panel data causality have been constructed and widely used, because they allow for nonstationary individual effects in the panel data sets (Holtz-Eakin et al. 1988). Zhang and Fan (2004) conducts a Granger causality test to examine their relation between road density and agricultural productivity in 290 districts of rural India and finds a unidirectional causality from the former to the latter. Using Vector Error Correction Model (VECM), Pradhan and Bagchi (2013) find different causality relations, respectively, for road and rail transport with taking India as a case study for 1970–2010. Mutual causality is more frequently existent for road infrastructure despite a unidirectional causality from rail infrastructure to economic growth. By running a panel vector autoregressive model, Arvin et al. (2015) prove that their causality relationship is also closely correlated with economic conditions for G-20 countries for 1961–2012, namely no causality for the developed group and only a unidirectional causality from transport infrastructure to economic growth for the developing group.

The panel causality model proves to be effective and practical to examine the direction of causality. However, the use of cross-sectional information in panel data implies to consider the heterogeneity across individuals in defining the causal relation (Venet and Hurlin 2001), which has been neglected in most previous studies. Following Erdil and Yetkiner (2009) and Beyzatlar et al. (2014), this paper proposes an extension of conventional Granger causality—a Granger non-causality test for heterogeneous panel data models. The extension of Granger causality allows considering both the heterogeneity of the causality and the heterogeneity of the data generating process (Hurlin 2004a, b). Erdil and Yetkiner (2009) used the Granger non-causality analysis to investigate the Granger-causality relation between health care expenditure and output. The results indicated the bidirectional causality is the dominant type and the pattern of unidirectional causality is not homogenous, that is, one-way causality generally is from health to income in high-income countries while the reverse holds for low- and middle-income countries. Beyzatlar et al. (2014) investigated the Granger-causality relation between GDP and transportation of EU-15 countries in 1970–2008. They verified the dominant type of bidirectional causality, but found instances of one-way or no Granger-causality in countries with the lowest income. A heterogeneous bias is easily introduced if the use of panel data assumes a homogeneous relationship across individuals in defining causal relationships (Pesaran and Smith 1995). Venet and Hurlin (2001) proposed an extension of the Granger causality definition to test homogeneous causality against heterogeneous causality. This method has been applied to investigate the causality between carbon emissions and economic growth (Maddison and Rehdanz 2008), between government expenditure and growth (Gregoriou and Ghosh 2009), between exports and economic growth (He and Zhang 2010) and between energy consumption and GDP (Akkemik and Göksal 2012). Therefore, the heterogeneity problem in examining the causal relationship in panel data should be paid enough attention.

Many studies have analyzed the influence of transport infrastructure on economic output in China at the national level or regional level or provincial level. Sahoo et al. (2010) justified China’s high investment on infrastructure since the early 1990s by confirming the unidirectional causality from infrastructure development to regional growth. Based on a panel data of 31 provincial units in China for 1978–2008, Yu et al. (2012) proved that the effect of transport infrastructure on productivity is spatially varied. Specifically, the estimated elasticity parameters of transport infrastructure are 0.09, 0.24 and 0.08, respectively, in the eastern, central and western regions. The study of Song and van Geenhuizen (2014) indicates the positive effect of port infrastructure on economic growth in China’s four regions by performing the production function although the elasticity varies considerably among four regions of China. Using the technique of VECM, Shi et al. (2017) found that the contribution of infrastructure investment to regional growth demonstrates significant regional differences and temporal effect. Their study further verified the inverse U-shaped relation between infrastructure investment and growth which supposes a “crowding-out effect” of private capital. Using a panel dataset across regions of China for 1986–2011, Jiang et al. (2017) concluded that the magnitude of transport infrastructure on economic growth is sensitive to spatial scales. These differences could be explained by phases of development, transport investment policies, local spatial spillovers from neighboring areas as well as reform institutions from the central Chinese government.

Overall, it is of vital significance to understand their interplay between transport infrastructure and economic output in the context of a growing economy like China for the necessity of formulating appropriate transport policies (Soja 2009). This paper aims to provide a comprehensive empirical analysis of the speculated nexus between transport infrastructure and economic output observed in city-level China during 1999–2014. This study tries to contribute in the following aspects. First, we employ the extended panel Granger non-causality model to examine the heterogeneity problem embedded in their causal relation between transport infrastructure and economic output. Moreover, four types of causality relations are also examined. Second, empirical analysis at the city-level China is conducted and transport supply as well as travel demand represents major components of transport infrastructure. Third, regional differences of their mutual causality are discussed and further investigation of transport-led growth or growth-led transport infrastructure in four regions of China is made to formulate some policy implications.

The remaining content of this paper is organized as below: Sect. 2 presents the data sources and study area. The methodology and testing procedures are explained in Sect. 3. Section 4 describes the results and Sect. 5 clarifies the main findings. Section 6 concludes with the main findings and policy implications.

2 The study area and data description

2.1 The study area

China is often known for undergoing fast economic growth stimulated by the reform and opening-up policy since 1978. According to official statistics, from 1978 to 2013, the country achieved an annual growth rate of Gross Domestic Product (GDP) of 9.77% and an annual growth rate of per capita GDP (PCGDP) at 8.67% (CSB 2014). Transport infrastructure also expanded significantly during the same period (Chinahighway.com 2011), with the national total length of highway growing at an average rate of 4.64% per annum. Yet such economic and transportation progress has been made in a geographically highly uneven manner. Taking differences in geographical context and level of development into account, China can be divided into four regions as seen in Fig. 1: eastern, central, western and northeastern.Footnote 2 In terms of economic growth, the eastern and western regions are on the lead. The growth rate of GDP for the same period is 10.40% in eastern region, 9.72% in central region, 9.82% in western region and 8.43% in northeastern region.

Fig. 1
figure 1

The study area

Full size image

However, in terms of transport development, infrastructure is slightly improved more quickly in the central and western regions. From 1982 to 2013, while highway length has grown at an average annual rate of 5.55% in the central region, 5.21% in the western region, it was only 4.97% in the eastern region, and 4.17% in the northeastern region. There does not seem to be any one-to-one correspondence in the pace of transport infrastructure and economic output. This raises the question of whether transport improvement equally facilitates economic growth in different regions.

2.2 Data sources and data description

In this study, we collected data on GDP, population, transport infrastructure, capital input and labor input for the 1999–2014 period across cities. The data source is mainly from the China City Statistical Yearbook and the China Statistical Yearbook. Our database includes 216 cities from across China. We categorized these cities into four groups by geographical location to ascertain whether any structural differences exist at the subnational scale. These groups are the eastern region (78 cities), central region (58 cities), western region (47 cities) and northeastern region (33 cities) (Fig. 1).

In this paper, we prefer to infer from a theoretical framework instead of unrestricted choice of control variables. Specifically, we employ the widely used production function where economic output can be expressed as a function of capital, labor and transport infrastructure. The variables capital and labor are added to the panel causality regressions as the control variables. By adding these two variables, the impact of transportation as an independent variable to economic output, as well as vice versa, relative to other independent variables can be gauged. In so doing, the multivariate heterogeneous panel Granger causality analysis could more accurately measure their causal relationship (Leightner and Inoue 2012).

Through dynamic factor analysis, transport infrastructure is measured by the weighted summation of the total highway length, the total passengers by rail and highway and the total volume of rail and highway freight, ranging from 0 to 100. The eigenvalues for the above three indicators are all larger than 1. The first indicator namely the total passenger capacity by rail and highway, accounts for 70.18% of transportation development; the second indicator namely the total volume of rail and highway freight, accounts for 17.68% of transportation development; the third indicator namely highway length, accounts for 12.14% of transportation development. Economic output is measured by real GDP in constant 1990 RMB currency. Capital input refers to the fixed asset investment at the comparable price in 1990, while labor input refers to the number of employed persons.

In total, all panel data are available for the time period from 1999 to 2014. Table 2 presents descriptive statistics of all variables, namely T representing transport infrastructure, D representing economic output, K representing capital input and L representing labor input. From a geographic perspective, the eastern region has the highest mean value of all variables, indicating the strength of its economy, adequate provision of transport infrastructure, as well as the well-functioning labor and capital market in the coastal China. However, the northeastern region has the lowest level of transport infrastructure and the western region has the lowest levels of economic strength, fixed asset and labor supply. Furthermore, the eastern region has the highest standard deviation for transport infrastructure and GDP, indicating the large disparity of transportation and economic output in the coastal China.

Table 2 Descriptive statistics for variables about transport infrastructure, economic output, capital and labor
Full size table

3 Methodology

We employ the extended panel Granger non-causality model as developed by Venet and Hurlin (2001). The fixed effect estimator and lagged value of the dependent variable are also included in the panel model. We apply this technique to a balanced panel of 216 cities in China over the 1999–2014 period. The following model is utilized to test the heterogeneous causality from transport infrastructure to economic output:

$$\ln D_{i,t} = \mathop \sum \limits_{k = 1}^{p} \gamma^{k} \ln D_{i,t - k} + \mathop \sum \limits_{k = 1}^{p} \beta_{i}^{k} \ln T_{i,t - k} + v_{i,t} \quad v_{i,t} = \alpha_{i} + \varepsilon_{i,t}$$

Likewise, the reverse causality can be elaborated as shown below:

$$\ln T_{i,t} = \mathop \sum \limits_{k = 1}^{p} \gamma^{k} \ln T_{i,t - k} + \mathop \sum \limits_{k = 1}^{p} \beta_{i}^{k} \ln D_{i,t - k} + v_{i,t} \quad v_{i,t} = \alpha_{i} + \varepsilon_{i,t}$$

where i refers to the cross-sectional unit; t denotes time; k is the number of time lags; \(\varepsilon_{i,t}\) are i.i.d. \(\left( {0,\sigma_{\varepsilon }^{2} } \right)\); and \(\alpha ,\beta ,\gamma\) are the parameters to be estimated. In this model, we assume that the autoregressive coefficient \(\gamma^{k}\) and regression coefficient slope, or elasticity, \(\beta_{i}^{k}\) are constant for \(\forall k \in \left[ {1,p} \right]\). Therein, \(\gamma^{k}\) is identical for all cross sections, whereas \(\beta_{i}^{k}\) could be heterogeneous across all individuals. lnK and lnL are entered into the equations as exogenous variables.

For a heterogeneous panel data framework, four types of causality hypothesis can be tested: homogenous non-causality (HNC), homogeneous causality (HC), heterogeneous non-causality (HENC) and heterogeneous causality (HEC). In this paper, we test for these hypotheses for each direction of causality between transport infrastructure and economic output of cities in the following order. First comes the test for HNC, which means that transport infrastructure and economic output are causally unrelated for all cities. If the HNC hypothesis is rejected, a second test for HC is followed to determine if transport infrastructure and economic growth has a causal link with homogeneous elasticity among all individuals. If the HC hypothesis is also rejected, a third test is undertaken for HENC. Acceptance of the HENC hypothesis means that the existence of a sub-group of cities for which transport infrastructure does not cause economic growth. On the contrary, if the HENC hypothesis is rejected, it implies a causal relation between transport infrastructure and economic output for all cities of the panel, but the data generating process is still heterogeneous. Then the HEC hypothesis is accepted.

4 Empirical results

4.1 Stationarity test and cointegration test for panel data

In order to examine the stochastic properties of the panel data, panel unit root test such as Dickey–Fuller (ADF) tests and Levin et al. (2002) (LLC) tests, are employed to detect the stationarity of panel data. Table 3 reports the results of ADF and PP tests for four variables (lnD representing economic output, lnT representing transport infrastructure, lnL representing labor, lnK representing capital). Results reveal that the panel data is nonstationary in level. However, the first differences for these four variables lead to stationarity.

Table 3 Panel unit root test results for lnD, lnT, lnL and lnK
Full size table

Furthermore, panel cointegration tests such as Pedroni and Kao tests are employed to detect their long-run or co-movement relation between transportation and economic growth. The result as shown in Table 4 confirms the presence of their long-run co-movement between transportation and economic growth in eastern region, central region, western region and northeastern region. This indicates the strong long-run association between transport infrastructure and economic output in China although further studies are required to clarify the underlying causal mechanisms and causal directions.

Table 4 Panle cointegration tests
Full size table

4.2 Panel causality analysis from transport infrastructure to economic output

In order to determine the optimal lag length for the dependent variable and independent variables, we utilize the vector autoregressive (VAR) lag length criteria, such as Akaike (AIC), Schwarz (SC) and Hannan–Quinn (HQ) (Lütkepohl 2005). In the VAR model, the variables of transport infrastructure and economic output are entered as endogenous variables while the variables of capital and labor as exogenous variables. In most VAR models, we chose SC information criterion to choose the optimal lag order. Heteroscedasticity in panel data model is corrected using cross section weights estimation method. F tests are employed to examine whether each causality hypothesis is accepted or rejected.

Homogeneous (non)causality hypothesis is first tested and the results are shown in Table 5. Homogeneous non-causality hypothesis is rejected at 1% level for all cities and homogeneous causality hypothesis gets rejected as well at 1% level for all cities. According to geographic divisions, homogeneous non-causality hypothesis is rejected at 5% level in these four sub-regions according to FHNC statistics, whereas eastern region, western region and northeastern region accepts homogenous causality hypothesis at 5% level. The fourth column indicates that only central region finds panel heterogeneity.

Table 5 F statistics for HNC, HC and HENC hypotheses from transportation to economic growth
Full size table

In the next step, heterogeneous non-causality hypothesis is tested and this hypothesis cannot get rejected for central region at 1% level. Only 18 out of 58 cities exhibit heterogeneous causality from transportation to economic growth in central China. 42 cities from central region demonstrate heterogeneous non-causality. For this sub-group, transportation does not Granger cause economic growth.

Overall our results suggest some evidence of causality from transport infrastructure to economic growth and this causal relation is only heterogeneous in central China. Although homogeneous causality exists in eastern, western and northeastern regions, their elasticity coefficients are different, which is 0.007, 0.006 and 0.053.Footnote 3 The facilitation of transportation on economic growth is positive and is only statistically significant in northeastern region. As the major heavy-industry base in the past, economic development of northeastern region has been challenged by low marketization, insufficient vitality of state-owned enterprise, single industrial structure, resource depletion, etc., emphasized by President Xi (SinaLiaoning 2016). Therefore, the strategy of reviving northeastern region was proposed in 2004. Therein, improving transport network is recognized to be prerequisite for pulling economic growth. During the 11th 5-year plan (2006–2010) northeastern region totally invested 369 billion yuan in constructing transport infrastructure, accounting for 7.85% of the nationwide.

However, for the eastern and western China, in spite of homogeneous causality, the role of transportation on economic growth is relatively marginal. The eastern region as the mostly developed region in China has sufficient transport supply and transport demand, which plays an important role in strengthening regional economy in coastal cities. However, the continuing investment of transportation may not lead to a sustaining growth of economic development for the eastern region. More development strategies should be oriented at technological innovations, industrial structure upgrading, environmental protection and so on. The western region as the relatively backward area in China is constrained by the mountainous terrain and fragile ecological environment. The findings of this paper imply that the transport infrastructure is adequate for its economic development in western region and further transport investment cannot significantly stimulate its economy.

An important finding of this paper is that the causality from transportation to economic growth is highly heterogeneous in central China. This finding proves the significance of panel heterogeneity since individual predictors for the causal relation differs among panel members. Central region as the major supplier of grains in China, transportation is significant for shipping goods and labors, thus increasing total factor productivity and promoting economic growth in China. As a result of the strategies—“Rise of Central China” in 2004, a continuing large sum of fiscal investmentFootnote 4 is flown to transportation industry to conquer the bottlenecks of economic growth in central China. However, although these transport investments have been always believed to be effective, some cities in central China cannot find any causality from transportation to economic growth. This finding point to divergent results in causality studies listed in Table 1. In other words, panel heterogeneity may be an important determinant of the causal relation since the causality relationship only holds for a sub-group of cities in the panel.

4.3 Panel causality analysis from economic output to transport infrastructure

The results for homogenous non-causality test running from economic growth to transportation are shown in the third column of Table 6. This hypothesis gets rejected for all sub-groups at 1% level, which is slightly different from the results for the reverse causality explained in the previous subsection. As revealed in the fourth column of Table 6, homogenous causality hypothesis is rejected as well for all sub-samples at 1% level, indicating no homogeneous (non)causality running from economic growth to transportation.

Table 6 F statistics for HNC, HC and HENC hypotheses from economic growth to transportation
Full size table

Consequently, we further look at the causality at the individual level by testing heterogeneous non-causality hypothesis and heterogeneous causality hypothesis. Heterogeneous non-causality is presented in eastern, central, western regions, whereas heterogeneous causality exists in northeastern region. A total of 53 cities, including 11 out of 78 (14.10%) cities from eastern region, 15 out of 58 (25.86%) cities from central region, 27 out of 47 (57.45%) cities from western region, and 18 out of 34 (52.94%) cities from northeastern region exhibit heterogeneous non-causality, which means economic growth does not Granger cause transportation in these sub-groups.

Overall we find that the causal relationship from economic growth to transportation is highly heterogeneous across each subregion in China. In eastern region, 67 out of 78 cities demonstrate heterogeneous causality mostly with positive elasticity coefficients. The average elasticity coefficient is 1.28 with the highest value of 7.41 in Laiwu City in Shandong province. In central region, 43 out of 58 cities find heterogeneous causality mostly with positive elasticity coefficients. However, the average elasticity coefficient is only 0.29 with the highest value of 1.67 in Wuhan city of Hubei province. In western region, 20 out of 47 cities find heterogeneous causality mostly with negative elasticity coefficients. The average elasticity coefficient is − 0.16 with the highest value of 0.80 in Meishan city in Sichuan province. In northeastern region, 16 out of 33 cities find heterogeneous causality mostly with positive elasticity coefficients. The average elasticity coefficient is 1.31 with the highest value of 4.01 in Baishan city in Jilin province.

Therefore, the significance of economic growth on transport improvement is more evident in developed areas like the coastal/eastern region. Regions with better performance of economic growth have more tendencies to improve transportation, which is understandable that developed regions recognized the pulling effect of transportation and the government is financially capable of providing transportation. The effect from economic growth to transportation is roughly positive but only evident in a few cities of the western China. It may be attributable to the mode of economic growth as well as its insufficient marketization within the western region. The mode of economic growth in western region is still extensive and resource-relied, resulting in increased pollution and environmental destruction. Negative impacts of economic growth and insufficient marketization within the western region cannot promote the flow of labor elements. Overall, this finding provides another evidence for the heterogeneity of the causal relationship in the panel. Appropriate policy-making should stress on both geographical difference and the internal socio-economic differences within each subregion.

5 Discussion

We extend the causality research between transportation and economic growth by introducing panel heterogeneity and apply it to the China context across cities. The panel consists of 216 cities covering the period from 1999 to 2014. Most previous studies tend to assume the homogeneous causality in a panel data framework and the finer-scale research for the causality between transportation and economic growth in China is still lacking. The contribution of this paper is to enrich the existing empirical evidences for their causal relationship by examining how geographic heterogeneity and panel heterogeneity work and to shed some new lights on transportation-led or development-led policy-making for the Chinese government.

Previous studies implicitly assume panel homogeneity and geographic homogeneity in examining the causal relationship between transportation and economic growth. Previous studies in the China context as shown in Table 1 mainly conclude the unidirectional causality from transportation to economic growth mainly focusing on provinces. Only Yu et al. (2012) mention that their causality from transportation to economic growth is spatially heterogeneous. Our paper tries to enrich existing studies at the city level in four geographic regions by uncovering the sign, elasticity and direction of their causality in mutual directions. This paper verifies the extensive existence of panel heterogeneity in both causality directions between transportation and economic growth, especially in the reverse direction. More specifically, the causal relation between transportation and economic growth differs among four geographic regions. Unlike previous studies, we cannot find homogeneous causality from transportation to economic growth in central China.

Although Beyzatlar et al. (2014) take panel heterogeneity into account when investigating the Granger-causality between transportation and GDP in EU-15 countries, this study only investigate the heterogeneity in the direction of their causality between transportation and economic growth regardless of the sign and the elasticity. This paper tries to enrich existing studies by providing the sign, elasticity and direction of their causality in order to uncover their complex relationship in China context. An important finding of this paper is that the pulling effect of transportation on economic growth is positive in all geographic regions but relatively marginal in eastern and western regions. This partly proves the Chinese slang “pave the road to get rich”. However, to what extent transportation can stimulate economic growth is heterogeneous across geographic regions (He et al. 2017). In the reverse causality, only western region finds negative impact of economic growth on transport improvement. The mode of economic growth in western region is still extensive and resource-relied, resulting in increased pollution and environmental destruction. Negative impacts of economic growth and insufficient marketization within the western region cannot promote the flow of labor elements. Overall, the reverse causality differs among each city and each geographic region. It suggests that we take individual city perspective rather than the aggregated country when devising policy recommendations.

Panel causality tests show that bidirectional causality is observed in 53.24% of the cities (115 out of 216) in the panel. Unidirectional causality is observed in about 39.35% (85 out of 216) and no causality in 7.41% (16 out of 216). We conclude for the 115 cities exhibiting bidirectional causality that there is an interaction between transportation and economic growth. This is called “feed-back hypothesis”, which implies that transportation serves as an engine of economic growth and is also determined by the scale of economic activities. The share of cities with bidirectional causality is mainly from relatively developed eastern and central regions. The interaction between transportation and economic growth is important in this respect. On the one hand, conforming to previous cross-country studies as shown in Table 1, bidirectional causality is likely to occur in developed countries such as US, EU, Netherlands or South Africa. This study proves this viewpoint with taking city-level China as a case study. On the other hand, China’s resource-rich provinces in the under-developing western region exhibit bidirectional causality. Different from eastern region, the construction of transport network in western region is more dependent on fiscal transfers from the central government. This paper finds evidence for the significance of transportation for sustaining economic growth in these less-developed cities, which is consistent with the studies of Hart (1993) and Zhang et al. (2007) that the local government is motivated to drive economic growth within its jurisdiction in less-devleoped regions.

There is unidirectional causality from transportation to economic growth in 11 cities in eastern region, 4 cities in central region, 27 cities in western region and 17 cities in northeastern region. The share of cities with unidirectional causality from transportation to economic growth is mainly from western and northeastern regions. Given that these cities are mainly resource-exhausted or topography-constrained; transport infrastructure in this group is relatively lagged, which also reflect the transport inequality in four sub-regions of China. Thus, in order to stimulate their growth, some preferential policies from the central government should target at these cities when proposing transport-related projects. In the case of causality from economic growth to transportation, cities in this group include 26 cities in central China. In these economies, blind investment and repeated construction will result in a waste of resources, traffic jam, even destroying urban landscape and ecology, etc. Even these transport investment cannot see any real economic feedback. A possible solution is to put in place relevant policies to enhance the usage efficiency of transport infrastructure and promote smart transportation and this may rather promote economic development.

There is no Granger causality in either direction in 16 cities in central China. In these cities, transportation and economic growth are independent from each other and investment in transport infrastructure has a neutral effect on economic development. These cities do not seem to share a common characteristic in terms of development trajectories.

6 Conclusions and policy implications

In this paper, heterogeneous Granger non-causality model at the national level and region-specific heterogeneous Granger non-causality models are established with the panel data of 216 cities from 1999 to 2014. The mutual interaction between transportation and economic output are estimated. According to the results of this empirical study, we can conclude the following findings.

First, this paper verifies the extensive existence of panel heterogeneity and geographic heterogeneity in both causality directions between transportation and economic output, especially in the reverse direction. Second, in the direction from transport infrastructure to economic output, this paper finds homogeneous causality in all geographic regions except the central China and it proves that the pulling effect of transportation on economic growth is positive in China but relative marginal in highly developed eastern region and the relatively backward western region. Third, in the reverse direction, heterogeneous causality can be found in all geographic regions. Western region finds negative impact of economic growth on transportation. Four, over a half of cities in China exhibit bidirectional causality and they are mainly from relatively developed eastern and central regions.

This paper tries to enrich existing studies by providing the sign, elasticity and direction of their causality in order to uncover their complex relationship in China context. The presence of geographic heterogeneity and panel heterogeneity sheds some important insights into devising effective transportation-led or development-led policies. Geographic targeting should be emphasized when making region-specific policies. The following policy implications can be proposed according to the empirical findings. First, although the bidirectional causality is evident in the coastal region, the impact of transportation on economic output is relatively marginal. This implies that the development-led strategies in coastal China should be more oriented at technological innovation and industrial structure upgrading instead of transport investment. Second, the central region as the main supplier of China finds positive causality from economic output to transportation but Granger non-causality in most cities in the reverse direction. Thus, the local government in central China should reduce the continuing transport investment and more fiscal revenue spending should be directed at agricultural mechanization, improving rural environment and stimulating new-type urbanization. Third, the western region as the relatively backward region in China finds negative impacts of economic growth on transportation and marginal influence of transportation on economic growth. Thus, further transport investment cannot significantly stimulate the economy of the western China. The mountainous terrain and fragile ecological environment are still the major constrains for developing the western region. The development-led strategies in western China should pay attention to regional carrying capacity of resources and environments as well as the sustainability of regional economy. Four, the northeastern region as the heavy industrial base in China, finds a significant and positive impact of transportation on economic growth. This indicates that the transport-led development strategies can be effective to stimulate a new-round economic growth in northeastern region.

This paper also has some limitations which can be addressed in the future. First, transport infrastructure can be measured from multiple aspects including waterways and airports. As constrained by data availability, this paper only incorporates three transport variables including highway length, the total volume of rail and highway freight as well as the total passenger capacity by rail and highway. The geographic big data about airlines can be used to enrich the empirical analysis. Second, the high-speed railway network is under rapid construction in China. How the high-speed railways are associated with economic output requires further studies.