4. Empirical Analysis

4.1 Benchmark Regression Results

First of all, we examine the average treatment effect (ATE) of HSR on the economic growth of counties served by HSR based on Model (1). To address the potential heteroscedasticity and serial correlation problems, we cluster standard errors at the county level in all our regression analyses.

Table 2 is the benchmark regression results. Column (1) is full sample regression. Results suggest that the coefficient of core explanatory variable T is significantly negative at the 1% level, i.e. HSR launch has restrained economic growth in counties along the route. Based on Column (1), HSR launch would reduce countywide GDP per capita by 2.6 percentage points.

Furthermore, we divide counties in the experiment group into counties with stations and those without for regression based on Model (1), and the results are shown in Columns (2) and (3) of Table 2. The control group is consistent for both regressions, i.e. regression results in Column (2) are obtained by excluding county samples without stations, and regression results in Column (3) are obtained by excluding county samples with stations. Coefficients of core variables in both Columns (2) and (3) are significantly negative and share similar coefficients. Results in Columns (2) and (3) demonstrate the creation of experiment group to be reasonable from a regression perspective.

Faber’s (2014) study on China’s National Trunk Highway System found that China’s expressway network construction had restrained economic growth in “peripheral counties”. Since HSR routes are similar to those of the expressway network, the negative effect in our benchmark regression analysis could also stem from expressway. To exclude such effect, we perform two robustness tests: First, we limit the samples to counties already served by expressway in 2008, and if the regression coefficient of T remains significantly negative in the subsample regression, the negative effect of HSR launch on the countywide economy will be proven to exist. Second, we include a cross term between the dummy variable for counties served by expressway in 2008 and the dummy variable for year to control for the impact of expressway on economic development over time.

Our data include 737 counties that were connected to expressway in 2008. Column (4) is the result

of regression performed with data of the 737 counties from 2007 to 2016. Column (5) includes a cross term between the dummy variable for expressway and the dummy variable for year. Obviously, the coefficient of T remains significantly negative. This result has once again proven that HSR launch indeed has a negative effect on the countywide economy.

4.2 Re-identification of the Causal Relationship

HSR’s effects on countywide economic development investigated in this paper could be subject to the problem of sample self-selection. To address the endogeneity problem thus incurred, this paper creates a time-dependent instrumental variable (IV), i.e. the instrumental variable of minimum cost ( cost_ ivit) for an IV regression. In calculating the instrumental variable of minimum cost, we define two cities at both ends of an HSR line put into service during 2008-2016 as node cities, and based on the node cities, create the minimum cost pathways for newly launched HSR lines. Then, we generate cost_ivit

based on the minimum cost pathways following the same method for generating Tit.

Table 3 is the regression results of the instrumental variable method. Column (1) is the first-stage regression results. Obviously, the instrumental variable cost_ivit is significantly positively correlated with T. Column (2) discusses the exogeneity of the instrumental variable referencing Sun and Chen (2017), i.e. T and instrumental variable cost_ivit are simultaneously included into regression. At this moment, the regression coefficient of cost_ivit is insignificant, which indicates the exogeneity of the instrumental variable. Column (3) is the second-stage regression results, in which the coefficient of T is significantly negative. Regression results in Table 3 have once again proven the negative effects of HSR launch on countywide economic development.

4.3 Parallel Trend Test

Lastly, we perform a parallel trend test for the DID regression referencing Li et al., (2016) and examine the dynamic changes of countywide economy before and after HSR launch.

We draw the regression coefficient of the parallel trend test into a dynamic diagram, as shown in Figure 1. After the HSR launch in 2008, the economic development trends of counties with and without HSRstarted to diverge, and the negative effects of HSR launch on the countywide economy began to emerge. Before HSR launch, the economic development trends of counties with and without HSR were parallel, i.e. our DID model is consistent with the parallel trend hypothesis. In addition, Figure 1 shows that the negative effects of HSR launch on countywide economic development started to diminish in the

fourth year after the launch and vanished after the sixth year. The implication is that regional economic development may gradually reach a new equilibrium four or five years after HSR launch.

In this section, our analysis demonstrates that HSR launch has a negative effect on countywide economic development. After a series of robustness tests, this result still holds true, which indicates the correctness of Hypothesis 1.

5. Mechanism Analysis: Agglomeration Shadows

Section 4 demonstrates the negative effects of HSR launch on countywide economic development, thus verifying Hypothesis 1. In this section, we investigate whether such negative effects are a real reflection of the agglomeration shadows, i.e. whether Hypothesis 2 holds true.

5.1 Agglomeration Shadows: Empirical Analysis

From the perspective of distance, we demonstrate whether agglomeration shadows exist in the

6

context of HSR launch. Referencing Hodgson (2018), we modify our econometric model as follows:

Σ 5

Yit = α0+ βj dumj×Tit + θXit + regioni + yeart + εit (2)

j=1

Where, dumj is a group of dummy variables measuring the distance between counties in the experiment group and their respective central cities. It is created as follows: First, counties in the experiment group are ranked in the ascending order by the distance to a central city. If the distance between a county and a central city is in the left 20th percentile, the value of this county sample is 1 in dum1; otherwise, it is 0. Similarly, if the distance between a county and a central city is in the left 20th to 40th percentiles, the value of this county sample is 1 in dum2; otherwise, it is 0, and so on and so forth.

Second, dumj is respectively cross-multiplied with T and included into the regression. Referencing Faber (2014) and Baum-Snow et al., (2020), we identify central cities as provincial capital cities, regional capitals, municipalities directly under the central government, and cities specially designated in the state plan already served by HSR in 2016. Distance is the shortest straight-line distance between

7

a county’s government seat and its nearest central city’s government seat. Following the concept of agglomeration shadows, if the negative effects of HSR launch on countywide economic development are a real reflection of agglomeration shadows, the coefficients of variables on both ends of the five variables from dum1×T to dum5×T should be greater than those in the middle. It is also likely that the coefficients of variables at the center are significantly negative, and those of variables at both ends are insignificant.

Table 4 shows the results of regression following equation (2). Results of Column (1) are consistent with the expectations, i.e. the coefficients of dum1×T and dum2×T are insignificant, and the coefficients of dum3×T and dum4×T in the middle is significantly negative. Yet the coefficient of dum5×T is also insignificant. Since there was no HSR in Yunnan Province, Ningxia Hui Autonomous Region and Inner Mongolia Autonomous Region by 2016, we perform another sub-sample regression analysis by excluding these three provinces

in Column (2). In Column (3), we include a cross term between the dummy variable for expressway and the dummy variable for year to control for the impact of expressway. Similarly, agglomeration shadows still exist.

In addition, although straight-line distance serves as a more exogenous grouping criterion (Faber,

2014), it may introduce a measurement error. Based on China’s HSR network in 2016, we calculate the railway distances between counties with HSR and their respective central cities, and group our samples according to railway distance for a robustness test with regression results shown in Columns (4)-(6). Obviously, agglomeration shadows still exist.

The above results indicate that the negative effects of HSR launch on countywide economic development were primarily suffered by counties at a medium distance from central cities. Measured by the distance to central cities, the negative effects of HSR launch on the countywide economy exhibit an inverted U-shaped pattern. As shown in Column (2) of Table 4, agglomeration shadows are roughly in the range between 97 km ( dum3×T) and 195 km ( dum4×T) to a central city.

5.2 Agglomeration Shadows: Competing Hypothesis and Robustness Test

As noted in our theoretical analysis, agglomeration shadows originate from the centripetal and centrifugal forces of economic factors from the perspective of central cities. Hence, we introduce central cities for a heterogeneity analysis of the distance between counties with HSR and their nearest central cities to verify the existence of agglomeration shadows. However, there is one competing hypothesis that may threaten the credibility of the results, i.e. access to HSR may also reduce GDP per capita of central cities, i.e. HSR has

8 a negative impact on economic development of all regions, and no agglomeration shadow exists.

To verify the competing hypothesis, we replace the explained variable with the ratio between GDP per capita of each county and GDP per capita of central cities (“GDP per capita ratio”, the same below). If the GDP per capita ratio decreases after HSR launch, the implication is that the relative GDP of counties with HSR will decrease more, and the competing hypothesis does not hold true. Table 5 shows the regression results. Where, Column (1) shows that HSR launch has indeed led to a decrease in the GDP per capita ratio of counties. Results in Column (2) suggest that decrease in the GDP per capita

ratio of counties has also followed a pattern of agglomeration shadows, i.e. the range of 97 km ( dum3×T) to 195 km ( dum4×T) from a central city is an agglomeration shadow. In Column (3), we also include a cross term between the dummy variable of expressway and the dummy variable of year, and the results remain robust.

This section has verified the correctness of our Hypothesis 2, i. e. HSR’s negative effect on countywide economic development can be explained with the agglomeration shadow theory. That is, among counties with HSR, regions with a certain critical distance from the central cities (roughly from 97 km to 195 km) experienced the sharpest decreases in their regional economic indicators.

6. Equilibrium amid Agglomeration

After Hypothesis 2 is verified in Section 5, this section will discuss the negative effects of HSR

10 launch on the countywide economy from the perspective of change in permanent county population.

By introducing the variable of permanent county population, we will verify Hypothesis 3.

6.1 HSR Launch and Change in Permanent Population

Compared with traditional modes of transportation, HSR is rapid, punctual, safe and comfortable. For these strengths, HSR is conducive to the flow of population, especially highly qualified labor force (Lin, 2017). In this section, we put together data of permanent population11 in most counties across China from 2010 to 2016 to investigate whether HSR launch will affect the size of permanent population with the logarithm of permanent county population as the explained variable.

Due to limited data availability, we cannot access the county-specific permanent population data before 2010, and pre-phase data are lacking for some experiment group samples. Compared with the DID method, results estimated with the instrumental variable method could be more reliable. Hence, we

12

have employed the instrumental variable method for regression. Table 6 is the regression results with the logarithm of permanent county population as the explained variable. Explained variable in Column (1) is the logarithm of permanent population, and the regression coefficient of T is significantly negative, indicating that HSR launch would cause an outflow of permanent population from counties with HSR stations.

For a comparative analysis, we have replaced the explained variable with the logarithm of registered population. Column (2) is the regression results of the instrumental variable (IV) method from 2010 to 2016, and Columns (3) and (4) are IV and DID regression results for samples from 2007 to 2016. In those three columns, T’s coefficients are all insignificant and very close to 0, which indicates that HSR launch would not affect the registered population of counties with stations. Regression results in Table 6 suggest that HSR launch would influence the size of permanent county population, and that in the absence of material change to the hukou system, HSR launch would not have any impact on the size of registered population (Au and Henderson, 2006; Faber, 2014).

Lu and Chen (2008), Lu et al., (2012), Lu and Xiang (2014), Lu (2017) and many others have called for China’s overall economic development to pursue “equilibrium amid agglomeration”, a key aspect of which is to promote infrastructure connectivity and the free flow of factors across regions. Our findings suggest that HSR has played a very positive role in promoting the free flow of economic factors, not least population. Long et al., (2017) suggests that HSR development also has a positive effect on the free flow of capital.

To further verify the existence of agglomeration shadows, we employ the logarithm of permanent populations as the explained variable for an IV regression following equation (4). Referencing Jedwad et al., (2017), we perform a two-stage regression by generating instrumental variables from cost_ivit

corresponding to dumj×Tit. Results of Table 7 are similar to results of Table 4, i.e. the coefficient of dum3×T is significantly negative while the other four coefficients are insignificant. This also suggests that the negative effects of HSR launch on permanent county population were primarily suffered by counties with HSR in the middle between central cities. Among various economic factors, population is relatively vibrant, and the effects of HSR launch on permanent county population can be regarded as a manifestation of the agglomeration shadows.

6.2 Heterogeneity Analysis: HSR Launch for Regions with Differentiated Resource Endowments

Differences in resource endowments influence the development pathways of various regions in

profound ways (Ellison and Glaeser, 1999; Faber and Gaubert, 2019). The notion of “equilibrium amid agglomeration” also calls for fostering competitive industries based on local resources (Lu, 2016). Lin (2017) demonstrated that HSR launch would significantly increase employment in the tourism industry in prefecture-level cities. Hence, we create a dummy variable ( resourcei), which is defined as 1 if a county boasts superior tourism resources; otherwise, it is 0. This dummy variable is measured by whether a county had any national 5A tourist attraction from 2007 to 2018. This period is extended till 2018 because the recognition of 5A tourist attractions was a lengthy process. Similarly, we introduced the product of multiplication between T and resource into regression with results shown in Column (1) of Table 8. According to Column (1), T’s coefficient remains significantly negative, T×resource’s

coefficient remains significantly positive, and their absolute values are greater than T. That is to say, counties rich in tourist attractions may benefit from HSR launch by attracting tourists from adjacent cities. With an inflow of factors, those counties registered an increase in GDP per capita by 0.7%.

Regression results of the above sections have once again verified Hypothesis 2, as well as the first part of Hypothesis 3, i.e. HSR launch is conducive to the free flow of economic factors, and counties with advantageous resource endowments served by HSR may reap development dividends from access to HSR.

6.3 HSR Launch, Change in Permanent Population and Economic Development

In the above analysis, we use GDP per capita for registered population as the explained variable. Given that each region’s registered population is relatively constant, decrease in GDP per capita for registered population is more indicative of the economic size of counties with HSR.As shown in the results of the previous section, HSR launch may also cause a county’s permanent population to shrink. If change in permanent population is taken into account, HSR may have a smaller negative effect on the GDP per capita of counties served by HSR.

To verify this guess, we will perform a regression analysis after re-calculating countywide GDP per capita for permanent population and comparing with the regression results calculated with GDP per capita for registered population. Table 9 presents the regression results. In comparing Column (1) with Column (3), we may find that if change in permanent population is taken into account, HSR indeed will have a much smaller impact on GDP per capita, i.e. change in permanent population will partially offset the negative effect of HSR on countywide GDP per capita, which is an essential element of the “equilibrium amid agglomeration” concept. As shown in Columns (2) and (4), agglomeration shadows still exist after samples of 2010 to 2016 are employed for an instrumental variable (IV) regression; in Column (4), the absolute values of regression coefficients of the five cross terms are all somewhat smaller than the values in Column (2). With changing permanent population taken into account, the phenomenon of agglomeration shadows will diminish, which is consistent with the “equilibrium amid agglomeration” concept.

As mentioned before, this paper aims to unravel the negative effects of HSR launch on the countywide economy and the consequent agglomeration shadows. Faber ( 2014) considered transportation infrastructure improvement as a necessary condition for regional economic integration. After the reduction of trade cost, the relative decline of economic volume in the hinterland was

accompanied by increases in overall economic efficiency and economic output. From the perspective of “equilibrium amid agglomeration”, decreasing trade cost will facilitate the free flow of economic factors and lead to the convergence of per capita income between central and peripheral regions. Hence, the negative growth effect of HSR launch on the countywide economy could be a necessary adjustment process (Lu, 2016, 2017). According to Figure 1 based on the event study method, the negative effects of HSR start to diminish three to four years after HSR launch, and such a gradual decrease of negative effects may represent a necessary adjustment process. Our research indicates that with changing permanent population taken into account, HSR’s negative effect on countywide GDP per capita would be less severe, and counties with advantageous resource endowments could directly benefit from HSR launch. These findings can be seen as indirect empirical evidence of “equilibrium amid agglomeration”. Research in this section has demonstrated the correctness of Hypothesis 3.

7. Conclusions and Policy Implications

With high-speed railway (HSR) lines put into operation during 2008 and 2016 as a quasi-natural experiment, this paper investigates the effects of infrastructure development on the countywide economy. Results suggest that HSR launch since 2008 was accompanied by a decrease in countywide GDP per capita by 2.6 percentage points. Results of the parallel trend test indicate that the negative effects of HSR on the countywide economy became evident three to four years after HSR launch.

Further mechanism analysis uncovers that the negative effects of HSR launch on the countywide economy are consistent with the agglomeration shadows of the core-periphery model, the scope of which is primarily 97 km to 195 km from the central cities. Moreover, the freer flow of economic factors after HSR launch will reduce permanent population of counties with HSR,and change in permanent population also follows a pattern of agglomeration shadows.

With change in permanent population taken into account, HSR has a smaller negative impact on the GDP per capita of counties, which is consistent with the concept of “equilibrium amid agglomeration”.

This study further finds that although HSR will cause resources to concentrate in central regions, when counties also boast advantageous resource endowments, HSR may also draw resources to those counties and spur their economic growth. Specifically, counties with favorable tourism resources may benefit from HSR launch.

In the new era of China’s transition from a moderate to a high degree of urbanization, HSR will link central cities with the hinterland, reduce trade cost, facilitate the free flow of economic factors, and induce a readjustment of China’s overall economic structure following the core-periphery model. HSR is conducive to not only the attractiveness of central cities, but the flow of resources to counties with favorable resource endowments (such as tourist attractions) for their economic growth. In the long run, HSR will deepen regional economic integration and finally contributing to the “regional economic equilibrium amid agglomeration”.

This study is of great policy significance: While some peripheral regions suffer from overall economic restructuring, a key question for local governments is how to cope with such shocks based on their respective strengths. Local governments in the agglomeration shadows should not only increase their transportation connectivity, but more importantly, take HSR as an opportunity, give play to local resources, increase exchanges with central regions by means of HSR, proactively attract an inflow of factors, and adopt development strategies suitable to local conditions for economic growth.