Internet, transportation infrastructure and the spatial structure of urban employment in China

ABSTRACT This study revealed the non-linear impact of the internet on the spatial structure of intracity employment and how transportation infrastructure moderates this non-linear impact. Using data from 22.47 million enterprises from the China Economic Census of 2004, 2008 and 2013, we found that (1) on average, the internet promotes urban employment agglomeration, but this agglomeration effect diminishes marginally as internet penetration increases; (2) the internet promotes the secondary sector to agglomerate first and then disperse, while it only has an agglomeration effect on the tertiary sector; and (3) improvements in the transportation infrastructure diminish the internet’s agglomeration effect.


INTRODUCTION
As two key technological innovations of the 20th century, transportation and internet technologies have been important forces influencing the spatial structure of urban employment by reducing the constraints of distance (Glaeser, 2020).After 1950, the development of transportation technology promoted the migration of urban employment to the suburbs in some developed countries (Baum-Snow, 2007).Since 1980, the development of internet technology has enabled the instantaneous and long-distance transmission of information and further reduced the distance costs of communication (Shen, 1999).Therefore, some futurists proposed the idea of the 'death of cities' (Toffler, 1980) and the 'death of distance' (Cairncross, 2001).However, there is an opposite view in the urban economics literature that telecommunications improvements will complement face-to-face interactions and thereby promote urban agglomeration (Gaspar & Glaeser, 1998;Kolko, 1999).The theoretical contradiction is also reflected in reality.In practice, we can observe both the dispersal of some traditional manufacturing firms to urban peripheral nodes (Zhang et al., 2022) and the ubiquitous landscape of intracity employment agglomeration (Glaeser, 2020).The dual paradoxes at the theoretical and practical levels require us to re-examine the real impact of the internet on the spatial structure of intracity employment.
Theoretically, the internet has two opposite effects on the spatial structure of urban employment (Dadashpoor & Yousefi, 2018).First, online communication can be a partial substitute for face-to-face communication and thus generates a dispersion force.Second, due to the limited ability of the internet to transmit knowledge (Panahi et al., 2013), frequent online communication may trigger a subsequent demand for more face-to-face communication (Gaspar & Glaeser, 1998;Glaeser, 2020), thus generating an agglomeration force.The relative magnitude between dispersion and agglomeration forces determines the net effect of the internet on the spatial structure of urban employment.
Although several studies have examined the impact of the internet on the spatial structure of urban employment, the conclusions reached are mixed.Some have found dispersion impacts (Ioannides et al., 2008;Nilles, 1991;Qin et al., 2016;Tranos & Ioannides, 2020;Zhang et al., 2022), while others have found agglomeration impacts (Hong & Fu, 2011;Huang et al., 2020;Kolko, 1999;Sinai & Waldfogel, 2004;Sohn et al., 2002).The likely reason for this is that the impact of the internet is non-linear, with non-linearity reflected in two aspects.One aspect is that the internet's impact on the spatial structure of urban employment could change as its penetration increases.With the penetration of the internet, not only will its dispersion force be enhanced via the improved knowledge-encoding ability (Tranos, 2020), but the proportion of tacit knowledge in economic activities will also be increased by the ensuing development of knowledge-intensive cities, which increases the demand for face-to-face communication (Glaeser, 2020) and thereby enhances the internet's agglomeration force.Since the extents of enhancement of these two forces are different during the penetration process of the internet, the net effect of the internet is likely to be non-linear.Furthermore, given the different dependence of different sectors on close interaction, the non-linearity of the internet's net effect could vary across sectors.In the literature, Tranos and Mack's (2016) study of US knowledge-intensive firms at the county level and Wang et al.'s (2021) study of the national city size distribution have considered the non-linearity of the impact of the internet.However, our study differs from their studies in terms of research object and spatial scale.Analysing geographical phenomena must consider spatial scale, and the results of studies at different spatial scales cannot be directly compared.At the intracity level, the non-linear impact of the internet on the spatial structure of employment is what this study aims to reveal.
The non-linearity of the internet's impact on the spatial structure of intracity employment is additionally reflected in the fact that this impact changes with improvements in the urban transportation infrastructure.As long as the need for face-to-face communication between economic agents remains, firms and workers tend to be spatially close to each other to save the transportation cost of initiating face-to-face communication (Fujita & Thisse, 2013).Improvements in transportation infrastructure help reduce transportation costs (Gokan et al., 2019), thereby reducing the need for the 'spatial agglomeration' of economic agents and enhancing the dispersion force of the internet.A few studies have considered the impact of both transportation and telecommunications on urban space, but they mostly treat the internet and transportation as two independent influencing factors (Gokan et al., 2019;Shen, 1999;Tayyaran & Khan, 2003) or control for transportation when examining the impact of the internet (Ioannides et al., 2008), without considering the interaction between the two in reshaping the spatial structure of urban employment.In summary, to the best of our knowledge, existing studies have neither fully revealed the non-linear impact of the internet on the spatial structure of employment within cities nor considered the moderating effect of the transportation infrastructure on the impact of the internet, which leads to the fact that the identified dispersion or agglomeration effects are only the net effects of the internet at its specific stage of penetration.
To fill these gaps, and in the process respond to the longstanding controversy over the impact of the internet on the spatial structure of urban employment, we aim to reveal the non-linear impact of the internet on the spatial structure of intracity employment in its continuous penetration in addition to the impact's sectorial heterogeneity, as well as how the transportation infrastructure moderates such a non-linear impact.We conduct an empirical study using data from 22.47 million enterprises in 289 prefecture-level cities from the China Economic Census in 2004, 2008and 2013 and construct instrumental variables (IVs) to alleviate endogeneity problems.
In addition, the contribution of this paper to the previous literature is reflected in the fact that we used more granular and 'long-term' data to conduct an empirical analysis to improve the reliability of the findings.Due to the limited availability of data, previous studies mostly used the size of the urban population or employment to reflect a city's overall level of agglomeration (Ioannides et al., 2008;Tranos & Ioannides, 2020, 2021;Wang et al., 2021).Even if some studies considered the intracity spatial structure, they only used a case of a certain city (Qin et al., 2016), cross-section data for a specific year (Hong & Fu, 2011) or employment data from specific industries (Tranos & Mack, 2016;Zhang et al., 2022) for their analysis.The China Economic Census data used in this study, which is the most comprehensive micro database in China to date, allow for the use of postal districts (on average, there are over 100 postal districts per city) as the basic spatial unit.Compared with previous studies that used districts or counties as spatial units to calculate urban spatial structure, our study uses data that allow us to paint a more accurate picture of the spatial distribution of employment within cities.In terms of the time span of our sample, different from developed countries whose urban systems are already mature and evolving slowly, China experienced 'the largest migration in human history' (Tombe & Zhu, 2019) from 2004 to 2013, with a rapid concentration of workers in urban areas and a nearly 13 percentage point increase in the urbanisation rate, which led to dramatic changes in the spatial structure of intracity employment (Yu et al., 2022); during the same period, internet penetration rate in China also jumped from 6.0% in 2004 to 45.8% in 2013.Although our sample is not a long-term sample in the absolute sense, the rapid urbanisation and informatisation in China during the sample period still provides an important observation to capture the non-linear impact of the internet.

LITERATURE REVIEW AND HYPOTHESIS DEVELOPMENT
The urban spatial structure is the pattern in the spatial distribution of various urban activities, which depends on the trade-off between various agglomeration economies and urban costs (Fujita & Thisse, 2013).Firms and workers agglomerate in the city centre for agglomeration economies but are dispersed by the rising urban costs caused by agglomeration.Knowledge spillovers triggered by proximate exchanges are important sources of agglomeration economies in cities (Marshall, 1920).The internet has two opposite effects on the agglomeration effect of knowledge spillovers, thereby changing the urban spatial structure: (1) the 'substitute effect', which refers to some interactions that would otherwise take place face-to-face being gradually replaced by online communication, thereby promoting the dispersion of urban space (Shen, 1999); and (2) the 'complement effect', which refers to online communications leading to a subsequent demand for more face-to-face communication, thereby enhancing urban agglomeration (Gaspar & Glaeser, 1998;Kolko, 1999).In addition, the internet may also enhance the agglomeration of urban employment by lowering entry costs to businesses, increasing intracity trade and creating labour demand (for details, see Appendix A in the supplemental data online).
Early empirical studies provided preliminary evidence for the above two theoretical perspectives.Nilles (1991) found that telecommuting increased the proportion of people living in the suburbs and commuting to work in the city centre, and Tayyaran and Khan (2003) also supported this idea, confirming the existence of a 'substitute effect'.In contrast, Sohn et al. (2002) found that the impact of information technology on the urban spatial structure of the Chicago area is dominated by the agglomeration effect.Sinai and Waldfogel (2004) and Panayides and Kern (2005) also found that the internet is a complement to, rather than a substitute for, urban strengths and ultimately manifests itself through an agglomeration effect, supporting Gaspar and Glaeser's (1998) view.However, early studies were conducted when internet technology was immature, so the impact of the internet on urban structure may not have been fully realised (Tranos & Ioannides, 2021).
With the penetration of the internet, its impact on the intracity spatial structure has gradually become prominent; however, the direction of the impact is still unclear.On the one hand, more studies have captured the dispersion effect of the internet.For example, Rachmawati et al. (2015) found that the use of information and communication technology (ICT) promoted the migration of economic services and residents from urban centres to areas with urban sprawl.A study on Nanjing, China, also found that the use of the internet led to the dispersion of residential space within the city (Qin et al., 2016).On the other hand, some studies have found that even online communication is characterised by spatial proximity (Huang et al., 2020), suggesting the 'complement effect' between online and offline communications.Therefore, scholars currently tend to hold that the internet has both dispersion and agglomeration effects (Dadashpoor & Yousefi, 2018).
The dispersion and agglomeration forces of the internet are not static; the internet's net effect will change with relative changes between the two forces.The literature has not revealed the non-linear effect of the internet on the spatial structure of intracity employment.Although not conducted at the intracity spatial scale, Tranos and Mack's (2016) study at the county level in the US considered the internet's dynamic penetration effects, and Wang et al.'s (2021) study at the country scale found a non-linear effect of the internet on the size distribution of cities of countries, which provides some insights for our study to consider the non-linear impact of the internet at the intracity scale.
The non-linearity of the internet's impact on the spatial structure of intracity employment is first reflected in the fact that, with the penetration of the internet, its agglomeration and dispersion forces are both enhanced; however, the net effect could be non-linear due to the different degrees of enhancement of the two forces.The literature usually uses internet penetration rate to characterise the degree of application of the internet in people's lives, that is, the proportion of internet users in the total population, which to a certain extent reflects the broader level of digitisation (Wu et al., 2022).On the one hand, with the rapid penetration of the internet, the internet's ability to encode knowledge continues to improve, which allows more knowledge that would otherwise rely on face-to-face communication to be transmitted through the internet (Tranos, 2020), thereby enhancing the internet's dispersion force.Specifically, knowledge can be divided into explicit knowledge, which can be coded and shared through the internet, and tacit knowledge, which is entrenched in the local innovation environment and social culture and can only be acquired through personal experience (Panahi et al., 2013).Tacit knowledge is a continuum and can be characterised by different degrees of tacitness (Panahi et al., 2013;Tranos, 2020).With the increased ability of the internet to encode knowledge, it has become sufficient to enable the sharing of knowledge with medium-and low-degree of tacitness (Tranos, 2020).
On the other hand, the internet has also accelerated the development of knowledge-intensive and service-oriented cities (Glaeser, 2020;Kolko, 1999), which helps to increase the proportion of tacit knowledge in economic activities and thereby enhance the internet's agglomeration force.The internet can not only drive the development of innovative service-based and knowledge-intensive industries that rely on the input of high-degree tacit knowledge (Glaeser, 2020;Kolko, 1999) by promoting knowledge spillovers and technological innovation (Wu et al., 2022) but also makes the world more information intensive.Prolonged immersion in an information-intensive environment in turn generates a wealth of high-degree tacit knowledge that is difficult to replicate online (Glaeser, 2020).Since knowledge with a high degree of tacitness has been difficult to access and transfer via information media thus far (see Appendix B in the supplemental data online), an increase in the proportion of knowledge with a high degree of tacitness in economic activities not only increases the dependence of economic agents on face-toface communication but also drives the same online communication to trigger more face-to-face communication needs (i.e., a stronger 'complement effect'), thereby driving existing urban employment to concentrate in a few highdensity areas and exhibit a more agglomerated spatial distribution pattern.In summary, the trade-off between the internet's dispersion and agglomeration forces as its Internet, transportation infrastructure and the spatial structure of urban employment in China 1035 REGIONAL STUDIES penetration rate increases determines that it has a non-linear net effect on the spatial structure of intracity employment.Accordingly, we propose Hypothesis 1: Hypothesis 1: As internet penetration increases, the net effect of the internet on the spatial structure of intracity employment is non-linear.
Given that different sectors' needs for proximity interaction are different, the non-linearity of the net effect of the internet may vary across sectors.In general, knowledge-intensive industries require more face-to-face connections and benefit more from knowledge spillovers (Tranos & Mack, 2016); therefore, the internet could have a stronger agglomeration effect in the tertiary sector, which is based on knowledge-intensive activities, and a stronger dispersion effect in the secondary sector, which relies mainly on conventional information for its economic activities.The development of traditional industries relies mainly on large-scale inputs of land and labour, has less demand for agglomeration economies and is less able to afford the costs of agglomeration (Zhang et al., 2022).
The internet can relax the geospatial constraints on the location of industrial firms by facilitating the optimal allocation of resources over a larger space: (1) the internet expands the spatial scope of knowledge spillovers.As the knowledge base of the secondary sector is easier to codify, it is easier for online communication to replace face-toface communication in the secondary sector; (2) the internet enhances firms' efficiency in matching workers with specific skills in a larger space, as well as their ability to connect with potential suppliers and customers of intermediate products (Goldfarb & Tucker, 2019); (3) the internet promotes working from home; and (4) the internet helps to reduce the cost of communication between headquarters and factories, allowing them to organise and execute discrete activities from different locations.As a result, under the context of rising urban land rents and labour wages, the internet may have a stronger dispersion force on employment in the secondary sector than in the tertiary sector.Accordingly, we propose Hypothesis 2: Hypothesis 2: As internet penetration increases, the dispersion force of the internet on the secondary sector increases faster than that on the tertiary sector, leading to its net effect on the spatial structure of employment in the secondary sector shifting faster from agglomeration to dispersion.
The non-linear impact of the internet on the spatial structure of urban employment exists on the premise that online communication cannot completely substitute face-to-face communication.As long as there is a demand for face-to-face communication, the impact of the internet is moderated by the transportation cost of initiating face-to-face communication and will change as the urban transportation infrastructure improves.Specifically, face-to-face communication and product delivery activities incur transportation costs, and when transportation costs are too high, firms and workers choose to gather in space.In contrast, when transportation costs are reduced, economic entities can communicate face-to-face by paying lower transportation costs and without having to move to urban centres with higher urban costs.Many studies have confirmed that improvements in transportation infrastructure have reduced transportation costs and consequently led to the suburbanisation of the urban population and employment in many countries (Baum-Snow, 2007;Duranton & Turner, 2012).Thus, as the transportation infrastructure improves, even if the 'complement effect' of online communication generates more demand for face-to-face communication, it would ultimately increase the frequency of intracity commuting and would not (entirely) generate a spatial agglomeration effect.This suggests that in the dynamic trade-off between the internet's dispersion and agglomeration forces, improvements in transportation infrastructure would reduce the demand for spatial agglomeration triggered by the 'complement effect', leading to the net effect of the internet tending towards dispersion.Accordingly, we propose Hypothesis 3: Hypothesis 3: Improvements in urban transportation infrastructure will weaken the agglomeration force of the internet, causing the net effect of the internet to tend towards dispersion.

Construction of agglomeration indexes
We calculated the agglomeration indexes for the employment distribution in prefecture-level cities using data on the amount and location of employment for 22.47 million enterprises from the China Economic Census in 2004, 2008 and 2013.We used the postal district as the basic spatial unit to measure how the workers in cities are distributed across areas of different densities rather than simply the overall employment of the city.Two indexes are constructed to measure the degree of intracity employment agglomeration: DELTA index (Galster, 2001;Massey & Denton, 1988) where j is the postal district in the city; n is the number of postal districts in the city; e j is the employment of postal district j; E is the total employment in the city; a j is the area of postal district j; A is the total area of the city; e j /E is the proportion of the employment of postal district j in the total employment of the city; and a j /A is the proportion of the area of postal district j in the total area of the city.The larger is the Delta index, the greater the difference in the distribution of employment per unit area within the city and the 1036 Sixu Wu et al.

REGIONAL STUDIES
greater the concentration of employment in one or a few high-density areas.
Gini index (Gordon et al., 1986;Small & Song, 1994) where j is the postal district in the city and n is the number of postal districts in the city.We rank the postal districts in each city from smallest to largest according to employment density as postal district 1 to postal district n.Ce j is the cumulative proportion of employment from postal district 1 to postal district j (the cumulative employment from postal district 1 to postal district j/total employment in the city), and Ca j is the cumulative proportion of the land area from postal district 1 to postal district j.A larger Gini index indicates a more agglomerated distribution of employment, that is, employment is distributed in 1 or a few high-density areas.

Construction of the basic econometric model
First, according to Hypothesis 1, as penetration increases, the relative change in the internet's dispersion and agglomeration forces may lead to a marginal or directional change in its net effect; that is, the marginal impact of the internet on the spatial structure of intracity employment varies with the penetration rate.Based on statistical tests (see Appendix C in the supplemental data online), we use a quadratic term function to test the non-linear impact of the internet on the spatial structure of intracity employment, which is expressed as follows: where Y it is the employment agglomeration index (DELTA or GINI) for city i in year t.According to Hypothesis 2, in the analysis of different sectors, Y it is the DELTA index of the secondary and tertiary sectors (see Appendix D in the supplemental data online for China's industrial classification).internet it is internet penetration rate, which is the number of internet broadband subscribers per 100 residents.b is the coefficient vector of the city's other characteristics.X it is a set of control variables.To alleviate omitted variable bias, the following variables were controlled for.(1) Considering the moderate effect of transportation infrastructure, we controlled for road density.(2) Considering the impact of cities' existing monocentric versus polycentric structures, we controlled for the polycentricity of employment (see Appendix E in the supplemental data online).
(3) Considering the demographic factors closely related to the urban spatial structure, we controlled for population size and population density.(4) Considering economic development factors, we controlled for the real GDP per resident, the industrial structure, the average wage of employed staff and workers and the house price.( 5) Considering that the development of urban space in China is strongly influenced by the planning decision-making factors of city governments in terms of land use, environmental governance and development strategies, etc. (Sun & Lv, 2020), we controlled for the area of land used for urban construction, green coverage, environmental governance capacity and the degree of government intervention.g i and l t are city fixed effects and time fixed effects, respectively, and 1 it is the random error term.As the data for the above variables were all non-normally distributed, logarithmic treatment was applied to all variables.Second, according to Hypothesis 3, we examine the moderating effect of urban transportation infrastructure on the non-linear impact of the internet by introducing the interaction term between transportation infrastructure and internet penetration rate into equation ( 3), which is expressed as follows: where M it denotes six proxy variables of transportation infrastructure: (1) road density; (2) expressway density; (3) the opening of a high-speed railway (HSR), which is a dummy variable for the opening of an HSR in each city, taking 1 if the HSR was opened on or before 30 June of the year, and 0 otherwise; (4) the number of public buses; (5) the number of taxis; and (6) to comprehensively reflect the level of transportation infrastructure, we used principal component analysis to transform the above five transportation infrastructure proxies into one composite index, denoted by PCA_transport (for details, see Appendix F in the supplemental data online).X it is the same control variable as that in equation (3); g ′ i and l ′ t are city fixed effects and time fixed effects, respectively; and 1 ′ it is the random error term.

Construction of the instrumental variable (IV)
The robustness of the econometric model estimation in section 3.2 may be affected by the endogeneity between the internet and the urban spatial structure.The endogeneity mainly arises from two phenomena: (1) reverse causation, as the more densely populated a city is, the easier it will be for internet penetration rate to increase; and (2) omitted variable bias, as investment in internet infrastructure by city governments is often accompanied by investment in other infrastructure affecting the spatial structure, and these omitted variables are difficult to completely control.Therefore, we constructed exogenous penetration rates for the robustness test using the relatively exogenous initial telephone network and a logistic curve function.Specifically, we performed two separate regressions.The first regression is used to construct exogenous fitted penetration rates (hereafter, internet fitted ) using a logistic curve function, and the second regression is used to estimate the effect of the exogenous fitted penetration rates on urban employment agglomeration, that is, uses the fitted penetration rates to replace actual penetration rates for regressions.The results of the second regression are referred to in the text as IV results.In the fitting process, the maximum penetration rate for each city is determined by its initial telephone network.On this basis, we use the internet fitted as an IV for the actual internet penetration rate and perform standard twostage least squares (2SLS) regressions for robustness tests.
An effective IV must first satisfy the correlation assumption.In 2000, the former Ministry of Posts and Telecommunications of China built an '8-horizontal-8-longitudinal' fibreoptic cable backbone network.The subsequent branch networks of cities are connected via this trunk line, and thus, the broadband network was largely developed based on the '8-horizontal-8-longitudinal' network (for details, see Appendix G in the supplemental data online).
Furthermore, the diffusion of a new technology follows the form of a logistic curve: it diffuses slowly in the early stage; when the number of users reaches a certain scale, its diffusion accelerates; and when the mature stage is reached, the rate of diffusion slows again (Geroski, 2000).The process of internet technology diffusion is similar.Inspired by Wu et al. (2022), we use logistic curves to fit the exogenous penetration process of internet broadband, assuming that the maximum coverage of the internet is limited by the pre-existing voice infrastructure and that diffusion follows a logistic curve.Due to the lack of access data for television subscribers, we used the number of fixed telephone subscribers per 100 residents in the starting year (2000), telephone i0 , to fit the maximum value for internet penetration rate: where i is the city.Then, the logistic curve for predicting the exogenous penetration rates is shown below: where b ′′ 2 is the diffusion speed, b ′′ 3 is the inflection point and 1 ′′ it is an error term.By observing the data, we find that internet penetration rate of Chinese cities has not yet reached the inflection point of the logistic curve during the sample period; therefore, we set the last period of the sample, 2013, as the inflection point, that is, b ′′ 3 = 2013.In addition, the IV also satisfies the assumption of exogeneity.First, the function of the logistic curve is set exogenously, and it does not contain any parameters related to urban spatial structure during the sample period other than the city's initial telephone data.Second, the initial telephone data are also relatively exogenous.(1) The original intent of the '8-horizontal-8-longitudinal' network was to connect China's cities and regions; thus, it is not directly related to the spatial structure of intracity employment.We regressed the urban employment agglomeration index during the sample period on the initial telephone data in 2000 and found that there is no significant correlation between the two variables, ruling out a direct association between them in a statistical sense.(2) China's early telecommunication infrastructure was mainly dominated by the central government; by the end of 2000, there were 256 million telephones in China, of which 126 million were owned by the central government and 45 million by local governments, together accounting for approximately 67% of the total number of telephones.At this point in time, China's telecommunications infrastructure was relatively mature and had strong attributes of a public facility; it was mainly dominated by the central government and was not completely related to independent characteristics such as the stage of urban development and the structure of employment agglomeration.(3) As lagged historical data for the year 2000, it can be assumed that these initial data have no direct impact on the explained variables for 2004, 2008 and 2013 in the sample used in this study.(4) The IV satisfied the exclusion restriction, that is, the fitted penetration rates only affected the city's economic growth through the actual penetration rates.Specifically, we regressed agglomeration indexes to the fitted penetration rate (internet fitted ), and the results show that the coefficients of internet fitted and internet 2 fitted are significant; however, after controlling for internet and internet 2 , the coefficients of internet fitted and internet 2 fitted became insignificant, proving the good exogeneity of the fitted penetration rate.Furthermore, by adding relevant control variables, we tried our best to exclude the cases where internet fitted affects urban employment agglomeration through channels other than internet.

Data sources and descriptions
This study was conducted at the level of 289 prefecturelevel cities in China for 2004, 2008 and 2013.The employment data used to calculate the agglomeration indexes are sourced from a micro database in the China Economic Censuses in 2004, 2008 and 2013, which provides information on a firm's postal code and the number of people employed.In the 289 prefecture-level cities included in the sample, there were 30,628 postal districts in 2004, 30,666 postal districts in 2008 and 30,374 postal districts in 2013.On average, there are over 100 postal districts per city.The highest frequency of the area of postal districts in the sample falls into the 59-88 km 2 range, and the next highest frequency ranges from 28 to 59 km 2 .Among them, approximately 40% of the postal districts occupy less than 2% of the total area of the city in which they are located, and approximately 99% or more of the postal districts occupy less than 10% of the total area of their city.This result means that the postal district is a smaller spatial unit, thus providing high-precision basic data for this study to construct a spatial agglomeration index based on postal districts.Other data were obtained from the China City Statistical Yearbook and the China Statistical Yearbook.The statistical descriptions of the main variables are shown in Table 1.

Non-linear impact of the internet on urban employment agglomeration
To test Hypothesis 1, we estimate the non-linear impact of internet penetration rate on urban employment 1038 Sixu Wu et al.Internet, transportation infrastructure and the spatial structure of urban employment in China 1039 REGIONAL STUDIES agglomeration based on equation (3).The ordinary least squares (OLS) results are reported in columns ( 1)-( 4) of Table 2.Then, the IV is used to re-estimate to alleviate endogeneity problems, and the results are reported in columns ( 5)-( 10) of Table 2.

REGIONAL STUDIES
The OLS results in Table 2 show that, on average, there is a significant inverted 'U'-shaped impact of internet penetration rate on urban employment agglomeration, and the penetration rate at the symmetry axis of the inverted 'U'-curve is equal to 6.6% (exp.(0.049/(-2 × (-0.013)))), suggesting that as the penetration rate increases, the internet first promotes employment agglomeration and then promotes employment dispersion after the penetration rate exceeds 6.6%.Of the 289 cities examined, only 25 cities' penetration rates were lower than 6% in 2013, indicating that the net effect of the internet in most Chinese cities has changed from an agglomeration force to a dispersion force.However, possible endogeneity between the internet and the spatial structure of urban employment means that the estimates in the OLS results are often biased and inconsistent and thus need to be analysed in conjunction with the results of the IVs.
The results of the first regression predicting exogenous fitted penetration rates using a logistic curve function in Table 2 show that the fitting coefficients of the IV are significant, with an R 2 of 0.828, and the F-test for the two parameters in the first regression shows an F-value of 3586, indicating that the IV does not suffer from weak identification problems, satisfying the prerequisites of IV correlation.The results of the second regression, which estimates the effect of the fitted penetration rate on urban employment agglomeration also show that there is a significant inverted 'U'-shaped impact of the internet but that the penetration rates at the symmetry axes of the inverted 'U'-curve are 284% for the DELTA index and 392% for the Gini index.Since the penetration rate is saturated at 100%, a 284% penetration rate is impossible to achieve in reality.Thus, the impact of the internet on urban employment agglomeration is located only on the left side of the inverted 'U'-shaped curve, that is, the internet significantly promotes urban employment agglomeration, but as internet penetration increases, its promotion effect on employment agglomeration tends to marginally diminish.Our results support Hypothesis 1.
The coefficient of ln(internet) in the IV results is 0.115, which is larger than that in the OLS results.The economic implication of this coefficient is the elasticity of the agglomeration index to the penetration rate, that is, for a 1% increase in the penetration rate, the estimated agglomeration index DELTA increases by 0.115%.Assuming that a city's internet penetration rate increases from 9 (%) (the sample mean) to 18 (%), the agglomeration index for that city would increase by 11.5%.An 11.5% rate of increase is relatively large for the agglomeration index, as the mean of the agglomeration index in our sample (0.67) would reach its maximum value (0.97) when increased by 45%.Similarly, the smaller values of the two symmetry axes estimated by the OLS method suggest that if the endogeneity problem is not addressed, the dispersion force of the internet will exceed its agglomeration force early, leading to an early shift of its net effect from agglomeration to dispersion force as well.
Possible explanations for why the coefficients of the IV regression are larger than those of the OLS regression are as follows.First, the omission of unobservable variables in the OLS estimates that are positively correlated with the penetration rate and would promote the dispersion of urban employment has led to their negative effect on urban employment agglomeration being included in the effect of the internet, thereby underestimating the positive effect of the internet on urban employment agglomeration (smaller coefficient on ln(internet) and symmetry axes), while the IV corrects the resulting estimation bias better and identifies the agglomeration force of the internet more accurately.Second, IV estimates the local average treatment effect, while OLS estimates the average treatment effect.The local treatment effect on the margin for the IV internet users could exceed that of the internet's users' average treatment effect.
The results considering sectoral heterogeneity show that the internet has an inverted 'U'-shaped impact on the agglomeration of employment in both the secondary and tertiary sectors.The difference is that the impact of the internet on the secondary sector shifts from agglomeration to dispersion at a penetration rate of 37%, but it needs to achieve a higher penetration rate (116%) to generate a dispersing force on the tertiary sector, which implies that the net effect of the internet on the tertiary sector is only an agglomeration effect, confirming Hypothesis 2. This may be because economic activities in the tertiary sector are more knowledge intensive than those in the secondary sector and require more agglomeration to obtain knowledge spillovers.Our tests indirectly demonstrate this mechanism (see Appendix I in the supplemental data online).

Moderating effect of urban transportation infrastructure on the net effect of the internet
To test Hypothesis 3, we further examine whether improvements in urban transportation infrastructure weaken the internet's agglomeration force based on equation (4).Due to space limitations, we report only the IV results in Table 3.The OLS results are largely consistent with the IV results (see Appendix J in the supplemental data online for OLS results).
The results in Table 3 show that for all transportation infrastructure variables other than the dummy variable of the opening of the HSR, the coefficients of ln(internet) × M are significantly negative, while the coefficients of ln(internet) 2 × M are significantly positive.Taking road density as an example, the results indicate that the inverted 'U'-shaped curve of the internet gradually flattens out as road density increases and that the impact of the internet on employment agglomeration is positive 'U'shaped when road density exceeds a certain threshold. 1This evolutionary process can be drawn as shown in Figure 1.Since the maximum urban road area per resident in the sample is 60.29 m 2 , we took the maximum value for  Note: The diffusion speed does not vary across regions.The control variables are the same as those in equation (3).Robust standard errors are shown in parentheses.The regression models controlled for both city and time fixed effects.The first regression is used to construct exogenous penetration rates using a logistic curve function.The second regression is used to estimate the effect of the internet on urban employment agglomeration.See Appendix H in the supplemental data online for complete results.Significance levels: *p < 0.1, **p < 0.05 and ***p < 0.01.
Internet, transportation infrastructure and the spatial structure of urban employment in China 1041 REGIONAL STUDIES road density as only 60 m 2 .For most cities in China, since the road density has not yet exceeded the threshold to transform the inverted 'U'-shaped effect of the internet into a positive 'U'-shaped effect, the moderating effect of the transportation infrastructure on the net effect of the internet is mainly reflected in the flattening of the inverted 'U'shaped impact curve and the shifting of the symmetry axis to the right, that is, a gradual weakening of the agglomeration force of the internet, confirming Hypothesis 3.
In addition, we also conducted a subsector analysis, and the results show that the moderating effect of transportation infrastructure on the internet's impact is stronger in the secondary sector (see Appendix L in the supplemental data online).To check the robustness of the results, we re-estimated the non-linear impact of the internet on the spatial structure of intracity employment and the moderating effect of transportation infrastructure using standard 2SLS regression methods.The 2SLS results are generally consistent with the results in the paper (see Appendix M in the supplemental data online).

DISCUSSION
This study responds to the longstanding controversy over the impact of the internet on urban spatial structure and shows that the internet has promoted the agglomeration of urban employment, but as internet penetration increases, this net agglomeration effect tends to decrease marginally under the effect of a continuously increasing dispersion force.Figure 2 presents the dynamics of the internet's agglomeration and dispersion forces during its continuous penetration.Our results explain why there are two seemingly contradictory conclusions found in the literature, namely, the agglomeration effect and the dispersion effect of the internet.For example, we found that without addressing endogeneity, the net effect of the internet shifts from agglomeration to dispersion when the penetration rate reaches 6.6%, which explains why many early studies found agglomeration effects of the internet (Gaspar & Glaeser, 1998;Kolko, 1999;Sinai & Waldfogel, 2004;Sohn et al., 2002), while studies at a later stage after the widespread penetration of the internet concluded that the dispersion effects of the internet gradually became prominent (Qin et al., 2016;Rachmawati et al., 2015;Tranos & Ioannides, 2021;Zhang et al., 2022).
On a larger spatial scale, Wang et al. (2021) found that with the penetration of the internet, its impact on the distribution of a country's population across cities changed from decentralisation to concentration, which is contrary to our findings.Likely reasons for this are (1) distance costs and knowledge spillovers have different degrees of influence on migration decisions at different spatial scales and (2) the internet has different influence mechanisms on the spatial structure of population (Wang et al., 2021) and employment.However, even in terms of the impact of the internet on the spatial structure of employment, we found significant sectoral heterogeneity.The internet shows an overall agglomeration impact on the tertiary sector, while its impact on the secondary sector shifts from agglomeration to dispersion.This echoes the findings of Tranos and Mack (2016) on knowledge-intensive business services and of Zhang et al. (2022) on manufacturing.
Our findings also corroborate the actual evolution of China's urban spatial structure.With the rapid urbanisation of China in the last two decades, its traditional urban centre structure has changed and is shifting towards a polycentric structure (Sun & Lv, 2020).Sun and Lv's (2020) study of 287 prefecture-level cities in China found that the majority (187/287) of cities had three or more employment centres in their administrative areas.Since the formation of a polycentric structure is a concentrated expression of the decentralisation of monocentric cities (Fernandez-Maldonado et al., 2014), our findings that in the context of the dominant agglomeration effect, the dispersion force continues to increase with the penetration of the internetreflect the current trend in which the spatial structure of urban employment evolves from monocentric to polycentric.Moreover, as the improvement of urban transport infrastructure can further weaken the agglomeration force of the internet, the trend towards the polycentric evolution of urban spatial structure would be further reinforced in the internet era.This issue is also important for our future research.
Against the backdrop of a digital economy driving the increasing scale and share of the knowledge and service economy, it will also be a long-term trend for the internet to have an agglomeration effect in urban areas.However, we cannot exclude the possibility that ICT could generate a short-term dispersion force in some developed cities because the internet's reshaping of urban spatial structure is mainly based on its impact on the trade-off between the existing agglomeration economy and urban costs in cities (Dadashpoor & Yousefi, 2018).For cities in a state of excessive agglomeration, economic entities tend to use existing technologies to overcome temporal and spatial barriers and alleviate the rising urban costs brought by spatial agglomeration.In general, however, the increase in high-degree tacit knowledge is a general trend, which would allow the internet to generate a stronger agglomeration force to counteract its dispersion force, resulting in the net impact of the internet in China's cities being dominated by agglomeration effects.
From a public policy perspective, the current orientation of local governments in China to excessively pursue spatial expansion in urban planning runs counter to the trend of agglomeration of urban employment in the internet era revealed in this study, which may lead to a loss of urban efficiency and waste of resources.In the institutional context of local official promotion tournaments and fiscal decentralisation in China, local governments, as the main actors of urban development, generally pursue the construction of new cities and new districts in suburbs far from the main urban areas to improve short-term economic growth performance and maximise fiscal revenue through land concessions (Peng et al., 2017), which has led to the expansion of urban built-up areas much faster than population growth during the same period.During  Internet, transportation infrastructure and the spatial structure of urban employment in China 1043 REGIONAL STUDIES the sample period of this study, China's urban built-up areas expanded by 68.7%, which was much faster than the growth rate of the urban population (39.9%).As a result of the lack of population density in new cities to meet the agglomeration requirements for industrial development, a large number of these government-planned 'new cities' have become 'empty cities', which are out of step with market demand.The pursuit of urban spatial expansion by local governments aiming at short-term economic performance is obviously contrary to the law of urban spatial agglomeration dominated by market forces, and how to reconcile the two is a key issue that needs to be addressed by China's city managers and planners.One implication provided by our finding is that the dispersion force of transportation technology can be used to moderate or neutralise the agglomeration force of the internet.Nevertheless, the evolution of urban spatial structure is also affected by other factors, such as economic forces and geographical features, and the non-linear laws of those effects also need to be further examined.

CONCLUSIONS
Using data of 289 prefecture-level cities from the China Economic Census in 2004, 2008 and 2013, we examined the non-linear impact of the internet on the spatial structure of urban employment and the moderating effect of transportation infrastructure on this impact.By doing so, we filled the gap in the literature on the non-linear impact of the internet at the intracity spatial scale and provided a theoretical explanation for the mixed conclusions of the existing studies regarding the impacts of the internet.This paper not only analysed the dynamic impact mechanisms of the internet in depth but also used more granular and long-term data in empirical testing to improve the credibility of the findings.The results show that (1) on average, the internet promotes urban employment agglomeration but that the agglomeration effect diminishes marginally as internet penetration increases.
(2) The net effect of the internet on the agglomeration of both the secondary sector and the tertiary sector is inverted 'U'-shaped.The difference is that the net effect of the internet on the secondary sector shifts from agglomeration to dispersion once its penetration rate exceeds 37%, while in the tertiary sector, it shifts to dispersion only after its penetration rate exceeds 116% (which is impossible to achieve in reality).(3) Improvements in urban transportation infrastructure weaken the agglomeration effect of the internet.Overall, the internet reinforces the advantages of cities as the centre of production and exchange and thus becomes an enabler of urban employment agglomeration.However, the weakening of the internet's agglomeration force by improvements in transportation infrastructure may drive the spatial structure of urban employment to evolve from a monocentric to a polycentric structure.
Our findings provide useful insights for city managers and planners on how to use the internet to improve city governance.First, unlike the common view that the internet generates a dispersal force, we find that the internet promotes the agglomeration of urban employment and that the agglomeration effect will continue in the future as internet penetration continues to grow.Therefore, in urban planning or urban spatial strategies, the government's arrangement of land, functions and infrastructure should consider the possible trend towards agglomeration in the internet age.It is necessary to prepare in advance for the urban diseases that may be caused by the agglomeration of mega-cities and to improve the carrying capacity of cities by enhancing the capacity of urban public services and increasing the density of the urban transportation network on the supply side.
Second, urban spatial planning should not be one-sizefits-all but should be flexible according to the agglomeration status, industrial structure and level of internet penetration in different cities.For developed cities that are already in a high degree of agglomeration, the agglomeration advantage of internet technology should be used to create a more suitable environment for the city's functional positioning as a centre for knowledge creation and information exchange and to promote the agglomeration and development of intracity service industries.Information technology should be used to improve urban governance, and the layout of urban transportation networks should be planned or adjusted to alleviate the problems of  congestion, pollution and other rising urban costs, thus creating conditions for the migration of traditional industries to the suburbs to improve the quality of agglomeration and the carrying capacity of cities.For underdeveloped cities with insufficient agglomeration, increasing investment in internet facilities can provide an important impetus for the agglomeration of enterprises and economic development.Furthermore, city managers should devote themselves to improving public services and industrial supporting facilities to enable the virtuous cycle of enterprise agglomeration and development within their cities.
The limitations of this study are described as follows.First, due to data availability, our sample only covers the period from 2004 to 2013 and cannot reflect recent urban developments.Second, as this study aims to reveal the non-linear impact of the internet and the moderating effect of transportation infrastructure on the internet's impact in a general sense rather than its sectoral heterogeneity in a specific sense, we only test our arguments for the secondary sector and the tertiary sector.If more up-to-date Chinese economic census data becomes available in the future, using those latest data to examine the non-linear impact of the internet at a more disaggregated industry level will provide further insightful results.In particular, the current COVID-19 pandemic has exogenously promoted the widespread application of the internet and working from home, and examinations of whether and/ or how the non-linear impact of the internet has changed in the Chinese case during the COVID-19 pandemic would be insightful to predict the future spatial distribution of human economic activities.

Figure 1 .
Figure 1.Moderating effect of the urban transportation infrastructure on the impact of the internet.Note: The road in the legend is the road density.

Figure 2 .
Figure 2. Strength of the agglomeration and dispersion forces of the internet.

Table 1 .
Descriptive statistics of the main variables.

Table 2 .
Non-linear impact of the internet on urban employment agglomeration.

Table 3 .
Moderating effect of the urban transportation infrastructure on the impact of the internet: IV results.The diffusion speed does not vary across regions.The first regression results of the instrumental variable and control variables contained in the second regression model are the same as in Table2.Robust standard errors are shown in parentheses.The regression models controlled for both city and time fixed effects.The second regression is used to estimate the moderating effect of the urban transportation infrastructure on the impact of the internet.See Appendix K in the supplemental data online for complete results. Note: