Hukou and Highways WPS7350. Policy Research Working Paper 7350

Transcription

1 Polcy Research Workng Paper 7350 WPS7350 Hukou and Hghways The Impact of Chna s Spatal Development Polces on Urbanzaton and Regonal Inequalty Maarten Bosker Uwe Dechmann Mark Roberts Publc Dsclosure Authorzed Publc Dsclosure Authorzed Publc Dsclosure Authorzed Publc Dsclosure Authorzed Development Research Group Envronment and Energy Team June 2015

2 Polcy Research Workng Paper 7350 Abstract Chna has used two man spatal polces to shape ts geographc patterns of development: restrcted labor moblty through the Hukou resdental regstraton system and massve nfrastructure nvestment, notably a 96,000 klometer natonal expressway network. Ths paper develops a structural new economc geography model to examne the mpacts of these polces. Fttng the model to avalable data allows smulatng counterfactual scenaros comparng each polcy s respectve mpact on regonal economc development and urbanzaton patterns across Chna. The results suggest large overall economc benefts from constructng the natonal expressway network and abolshng the Hukou system. Yet, the spatal mpacts of the two polces are very dfferent. The constructon of the natonal expressway network renforced exstng urbanzaton patterns. The ntally laggng regons not connected to the network have not beneftted much from ts constructon. By contrast, removal of the Hukou restrctons, whch Chnese polcy makers are consderng, would result n much more wdespread welfare gans, allowng everyone to gan by movng to where he or she s most productve. Removal of the Hukou restrctons would also promote urbanzaton n currently laggng (nland) regons, mostly by stmulatng rural to urban mgraton. Ths paper s a product of the Envronment and Energy Team, Development Research Group. It s part of a larger effort by the World Bank to provde open access to ts research and make a contrbuton to development polcy dscussons around the world. Polcy Research Workng Papers are also posted on the Web at The authors may be contacted at bosker@ese.eur.nl, udechmann@worldbank.org, and mroberts1@worldbank.org. The Polcy Research Workng Paper Seres dssemnates the fndngs of work n progress to encourage the exchange of deas about development ssues. An objectve of the seres s to get the fndngs out quckly, even f the presentatons are less than fully polshed. The papers carry the names of the authors and should be cted accordngly. The fndngs, nterpretatons, and conclusons expressed n ths paper are entrely those of the authors. They do not necessarly represent the vews of the Internatonal Bank for Reconstructon and Development/World Bank and ts afflated organzatons, or those of the Executve Drectors of the World Bank or the governments they represent. Produced by the Research Support Team

3 Hukou and Hghways The Impact of Chna s Spatal Development Polces on Urbanzaton and Regonal Inequalty Maarten Bosker 1, Uwe Dechmann 2 and Mark Roberts 3 JEL Codes: R11, R23, R42 Key words: Chna, transport nvestments, hghways, mgraton, new economc geography Sectors: Transport, Urban Development 1 Department of Economcs, Erasmus Unversty Rotterdam, The Netherlands and CEPR <bosker@ese.eur.nl> 2 Development Research Group, World Bank, Washngton, D.C., USA <udechmann@worldbank.org>; 3 Socal, Urban, Rural and Reslence Global Practce, World Bank, Washngton, D.C., USA <mroberts1@worldbank.org>. The authors thank Harry Garretsen, Vernon Henderson, Bert Hofman, Laura Herng, Dego Puga, Karls Smts, Masy Wong, and partcpants at the 2 nd Urbanzaton and Poverty Reducton Research Conference and the 2014 North Amercan Regonal Scence Conference for helpful comments and suggestons. Fnancal support from the World Bank s Knowledge for Change Program s gratefully acknowledged.

4 1. Introducton In recent decades, Chna s economy has been characterzed by rapd economc development accompaned by equally rapd urbanzaton. The resultng benefts have not, however, been shared equally across the country. Many rural areas saw consderable mprovements n lvng standards and large reductons n poverty rates, but gans n urban areas and those n Chna s coastal regons n partcular have been consderably larger. The Chnese government has always had a keen nterest n the country s spatal economc development. Its man polces have been amed at restrctng the flow of mgrants to the bg ctes, whle at the same tme tryng to develop nteror regons by better connectng them to the boomng coastal regons. Two polces stand out n ths regard. 1 The frst s the Hukou, or household regstraton, system. Every Chnese ctzen s Hukou status s determned by that of ther parents at the tme they were born. A Hukou s ted to a partcular urban or rural locaton and represents an enttlement to welfare benefts and publc servces (such as educaton, health care, etc.) n that place. When mgratng to a dfferent cty, a person s Hukou status does not change, so that mgrants are unable to clam many welfare benefts or publc servces n ther destnaton cty. Despte the Hukou system, many rural mgrants have stll mgrated to the ctes. A 2013 survey by Chna s Natonal Bureau of Statstcs estmated a populaton that s separated from ther household regstraton of 289 mllon (NBS 2014, also Chan, 2013). There s growng consensus that f Chna s contnung urbanzaton process s to be economcally productve and socally nclusve, the Hukou system put n place to restrct mgrant flows nto the ctes wll have to be loosened or even abolshed (World Bank and DRC 2014). Chnese polcy makers seem to agree, and have recently ssued sgnfcant reform proposals (State Councel, 2014) The second polcy s the huge nvestment n large-scale nfrastructure. In recent decades Chna has constructed an extensve 96,000 km network of hghways connectng the largest ctes n the country, vast ntra-cty transport nfrastructure (rng roads, metros), and s currently buldng the longest hgh speed ral network n the world connectng ts man populaton centers (see Baum-Snow and Turner, 2012; Roberts et al., 2012, Baum-Snow et al. 2012; Zheng and Kahn, 2013; Faber, 2014; among others). The explct am of these projects, and that of the constructon of the hghway network n partcular, s to spread development from the more developed coast to the ctes n the nteror. They make t cheaper for these ctes to mport materals and ntermedate goods, and to shp ther own products to the rest of Chna and the world. Indrectly, ths polcy s also amed at allevatng mgraton pressures. By 1 Another s land reform. The rsk of losng one s plot of land keeps many rural resdents from movng to the cty, even f they were able to obtan an urban Hukou. Land reform, whereby people would be able to sell ther land would be a way to allevate ths problem. 2

5 contrbutng to the economc development of Chna s nteror, workers ncentves to mgrate to the coastal ctes are expected to dmnsh as they can now fnd (well-pad) jobs n ther own regon. In ths paper, we quantfy the effects of these two man spatal development polces on Chna s economc geography. Buldng on earler work by Roberts et al. (2012), Behrens et al. (2013), and Tabuch and Thsse (2002), we ncorporate labor moblty nto a structural new economc geography (NEG) model that allows us to assess how the rapd constructon of the ntercty natonal expressway network (NEN) and the Hukou system have jontly shaped Chna s spatal economy. Frst, we ft ths model to the data usng nformaton on the rural and urban part of each of Chna s 331 prefectures. Next, wth our estmates of the mportant model parameters n hand, we smulate varous counterfactual scenaros that allow us to pnpont the mpact of the NEN and the Hukou system on the dstrbuton of both people and economc actvty across Chna s prefectural areas, as well as on rural-urban ncome nequalty, and urbanzaton rates n dfferent parts of the country. We are not the frst to look at the aggregate and/or spatal mpacts of ether the Hukou system or the NEN. However, earler papers have looked at these ssues separately. Banerjee et al. (2012), Roberts et al. (2012), and Faber (2014) 2 focus on the effect of the NEN. Roberts et al. (2012), usng a smlar combnaton of estmaton and calbraton as n ths paper, fnd that the NEN dd ncrease aggregate Chnese welfare. However, ts constructon dd not decrease real ncome nequalty between prefectures, nor reduce urban-rural wage nequalty. Faber (2014) also fnds that the NEN has renforced the concentraton of economc actvty n the largest ctes. 3 But the perpheral regons, whle losng economc actvty, have also ganed better access to the products produced n the ndustral centers thanks to the NEN. Takng both nto account, he fnds postve welfare mpacts of the NEN n all Chnese regons. Whalley and Zhang (2007) and Bosker et al. (2012) nstead focus on the spatal economc consequences of the Hukou system. Gven that a counterfactual Chna wthout Hukou restrctons s fundamentally unobserved, these papers establsh the effect of the Hukou restrctons by comparng Chna s regonal economy today to varous smulated no Hukou-scenaros. 4 Whalley and Zhang s (2007) results 2 Baum-Snow, et al. (2012) s another promnent example. That paper focusses on the effect of the extent and confguraton of ntra-cty nfrastructure on urban form. The effect of ntra-cty nfrastructure les beyond the scope of our paper. Coşar and Fajgelbaum (2013) provde a model and some evdence that Chna s nternal economc geography s affected by the accessblty of the nteror ctes to the nternatonally well-connected coastal regons. An mprovement n domestc nfrastructure n ther setup would result n a further mgraton to the coast. They do not, however, explctly consder the effects of the NEN. 3 Faber (2014) combnes reduced form estmates of the causal effect of the NEN on economc actvty and populaton growth, wth the calbraton of a spatal economc model that allows hm to generalze hs reduced form estmates. Banerjee et al. (2012) use a very smlar emprcal strategy as employed by Faber (2014). Unlke Faber (2014), however, they do not calbrate the model they propose, only showng reduced form estmates. 4 Desmet and Ross-Hansberg (2013) can also be consdered to smulate spatal economc outcomes n Chna under a no-hukou scenaro. Ther model assumes free labor moblty across prefecture ctes throughout, but they argue that the mgraton restrctons can be captured by ther cty-specfc estmates of amentes. Ther equal 3

6 suggest that a removal of the Hukou-restrctons would ncrease overall Chnese welfare, whle at the same tme substantally ncreasng real ncome nequalty between provnces. The latter s the result of people movng from the currently underdeveloped nteror to the rcher provnces on the coast. Bosker et al. (2012) come to a smlar concluson when analyzng the effect of the Hukou system at a spatally more dsaggregated level (294 prefecture ctes). They focus exclusvely on the spatal dstrbuton of people across Chna s prefecture ctes, and fnd that a removal of the Hukou restrctons would strongly renforce the core-perphery pattern that already exsts today. Our paper s contrbuton to ths exstng body of lterature s twofold. Frst, by explctly consderng both of Chna s two man spatal development polces wthn the same framework, we are able to assess ther effects wthout abstractng from the other spatal polcy. 5 The earler papers consderng the mpact of the NEN (Faber, 2014 and Roberts et al., 2012) all assume that people are completely mmoble. Ths, n effect, represents an extreme Hukou scenaro 6 : n realty many people dd decde to mgrate even n the presence of the Hukou restrctons. Abstractng from any mgraton response to changes n real ncome as a result of the constructon of the NEN, may lead to serously over- or understatng the (spatal) economc mpact of the NEN. Some of the NEN s mmedate mpact on the agglomeraton of economc actvty may e.g. have been renforced as people, despte the Hukou restrctons, decded to move to places that, because of the NEN, now offer hgher expected real ncomes. The earler papers consderng the mpact of the Hukou system nstead abstract from Chna s enormous nvestments n transport nfrastructure. Whalley and Zhang (2007) base ther analyss on a regonal economc model that abstracts from the large dfferences n accessblty across Chna. Bosker et al. (2012) do take trade costs nto account, but approxmate them by usng the great-crcle dstance between prefectures. As such, they stll abstract from the large dfferences n actual trade costs across Chna that are the result of the unequal nvestment n both the qualty and quantty of nfrastructure (and that have been shown by Roberts et al. (2012) and Faber (2014) to sgnfcantly affect regonal economc outcomes). In our model nstead, we use detaled nformaton on the travel tmes between prefectures before and after the constructon of the NEN as a more drect measure of trade costs between prefectures 7. amentes across ctes could then be vewed as a tentatve no-hukou scenaro: they fnd an ncrease of overall welfare n that scenaro accompaned by a more unequal cty sze dstrbuton (large ctes become larger, and small ctes become smaller). Ther model however also abstracts from dfferences n ctes accessblty,.e. costs nvolved n shppng goods nto or out of the cty, focusng nstead on wthn-cty frctons (congeston). 5 Moreover, we can for example also assess whether the constructon of the NEN has changed the expected effect of abolshng the Hukou restrctons. 6 Both papers actually justfy ther assumpton of complete labor mmoblty by argung that the Hukou system effectvely keeps (almost) everyone from mgratng. 7 Note that usng these travel tmes can stll be argued not to be the deal measure of trade costs. It tells somethng about how long t takes to get somethng from one place to another, whch need not necessarly be reflected n the costs pad to shp somethng along that route. 4

7 Moreover, we explctly allow for mgraton both wthn (from rural to urban areas) and between prefectures. Here we follow Behrens et al. (2013) and Tabuch and Thsse (2002), and model people s mgraton decsons as dependng on real wage- and amenty-dfferences between locatons, as well as on ndvdual-specfc dosyncratc preferences for lvng n a partcular locaton. 8 Only relatvely few studes estmate the relatve mportance of these dfferent factors for people s mgraton decsons n the Chnese context. One mportant reason for ths s the absence of comprehensve data on blateral mgraton flows (ether urban-rural mgraton flows or nter-prefectural mgraton flows). 9 Relyng on the equlbrum condtons of our model, shows that we can estmate ther mportance usng readly avalable data on the stock of mgrants n the rural and urban part of each prefecture only. Our results confrm the noton that mgraton n Chna s predomnantly drven by people n search of hgher real wages, and better provson of publc amentes. Mgraton wthn the same provnce responds more strongly to these factors than mgraton to prefectures n other provnces (consstent wth fndngs n Zhang and Zhao, 2013). The second contrbuton of our paper s that our model and accompanyng data set allow us to provde a much more complete pcture of the mpact of the NEN and/or Hukou system on Chna s spatal development. Gven that earler papers consderng the mpact of the NEN all assume regonal labor mmoblty, they only focus on the NEN s effect on the spatal dstrbuton of real ncome - ether between prefectures n Roberts et al., (2012) and Faber (2014), or between the urban and rural areas wthn prefectures (Roberts et al., 2012). Bosker et al., (2012), nstead focuses exclusvely on the effect of the Hukou on the spatal dstrbuton of people and frms across Chna s prefectural ctes. They do not detal the effects on real ncome nequalty or urbanzaton rates. 10 In ths paper we consder the effect of the NEN and the Hukou system on ) real ncome nequalty between and wthn (urban vs. rural areas) prefectures, ) evaluate how they affect the spatal dstrbuton of people across Chna s prefectures, and ) assess ther mpact on the urbanzaton rates n dfferent parts of Chna Modellng and estmatng people s mgraton dynamcs n ths way mproves upon Whalley and Zhang (2007) and Bosker et al. (2012). Whalley and Zhang model people as movng n response to regonal (provncal) wage dfferences only, whlst Bosker et al. ncorporate mgraton dynamcs based on evdence relatng to nterprovncal mgraton flows and ad hoc assumptons on mgraton costs. 9 Poncet (2006) also reles on mgrant flows at the provncal level, focusng on rural to urban mgraton only. Ths hdes a lot of varaton at the nter-prefectural and ntra-prefectural level. The bulk of mgraton n Chna s wthn provnces. Other studes (notably Rozelle et al., 1999; Zhao, 1999a,b and more recently Zhang and Zhao, 2013 or Gulett et al., 2014) provde evdence on the determnants of people s mgraton by relyng on surveys n partcular areas of Chna. Although very nterestng, and despte the fact that these latter two studes use surveys coverng the 15 ctes that are the largest mgraton destnatons, t s not obvous that ther results generalze to other parts of Chna. 10 An excepton s Whalley and Zhang (2007). They do consder the effect of the Hukou restrctons on ncome nequalty, as well as on the spatal dstrbuton of people across Chna. However the most detaled spatal level at whch they conduct ther analyss s that of Chna s 31 provnces. These provnces are stll very large. As a consequence, ther analyss hdes a lot of varaton at a more detaled geographcal level. 11 In both Faber (2014) and Roberts et al. (2012), the NEN mplctly has no mpact on the rate of urbanzaton because of the assumed absence of mgraton. 5

8 Our man fndngs show substantal overall economc benefts of the constructon of the NEN and of abolshng the Hukou restrctons. The spatal mpact of these two polces s however very dfferent. We fnd that the constructon of the NEN has only renforced exstng urbanzaton patterns. The ntally largest and most urbanzed prefectures have ganed most from ts constructon. Although not necessarly the ntally rchest places, these ctes saw the largest mprovements n ther connectvty due to the constructon of the NEN. The ntally laggng regons not connected to the NEN have not beneftted much from ts constructon. Hukou reform would nstead beneft the ntally rchest prefectures along Chna s South-Eastern coastlne. These ctes are ntally not necessarly the most urbanzed nor those wth the largest urban populatons, but ther economc success wll attract mgrants from across the country when the Hukou restrctons are relaxed. However, many places n Chna s nteror also beneft more from Hukou reform than from the constructon of the NEN, be t mostly because of rural outmgraton. By allowng everyone to gan by movng to where he/she s most productve, the benefts of easng the Hukou restrcton are more wdespread. The structure of the remander of the paper s as follows. Secton 2 brefly revews Chna s two man spatal development polces transport nvestments and mgraton restrctons. We present our modelng strategy n Secton 3. Secton 4 presents the data, estmaton and modellng strategy, and Secton 5 dscusses the estmated mpacts of hghway constructon and mgraton restrctons on both overall and spatal economc outcomes. Secton 6 concludes. 2. A bref hstory of the NEN and the Hukou system 12 Chna s natonal expressway network, also known as the Natonal Trunk Hghway System, was conceved n 1988 wth the goal of establshng seven hghways radatng from Bejng, nne North- South connectons, and 18 East-West connectons, gvng t the unoffcal name 7918 network. The frst phase of the network connectng all ctes wth populaton above 200,000 people was completed by 2007 wth a length of about 40,000 km (see Roberts et al., 2012; World Bank, 2007). Snce then, the network has further expanded, reachng more than 96,000 km n 2012 (Chna Statstcal Yearbook, 2013) larger than the U.S. Interstate Hghway System. In ths paper, we restrct analyss to the mpact of the frst phase of constructon, whch was largely concentrated between 1997 and The man reason for ths s that n 2008 Chna ntroduced the frst hgh speed ralway lnes, and today Chna has the largest hgh speed ral network n the world. Although these two networks have dfferent user profles (more goods transported on hghways, more busness travel on hgh speed ral), separatng the mpact 12 We keep the dscusson n ths secton bref gven that both of these polces have been descrbed n much greater detal elsewhere n the lterature, see n partcular the references cted n ths secton. 6

9 of the NEN from those of ralway lnes would be dffcult (especally snce there s a substantal amount of overlap n the ctes they connect). Whle large scale transport nfrastructure nvestments have been a farly recent feature of Chna s spatal polces, Chna s populaton regstraton system, or Hukou, was ntroduced well over 50 years ago, prmarly as a way to control populaton movements and the allocaton of labor to state-controlled producton (Chan and Buckngham, 2008; Chan, 2009; Bosker et al., 2012). The man dstncton was between agrcultural (rural) versus non-agrcultural (urban) resdence status. It hstorcally consdered the rural populaton as self-suffcent, whle provdng food ratons, housng and educatonal and health servces to the urban populaton. Ths general dstncton perssted even as urban dwellers became far better off durng economc lberalzaton. After polcy changes n the 2000s, preventng populaton moblty s no longer the domnant motvaton for mantanng the Hukou system. The reason for ts persstence, despte frequent expectatons that t wll be abolshed snce at least the md-1990s, s a concern that gvng mgrants equal rghts wll exceed the fscal capacty of ctes to provde publc servces and welfare benefts to everyone. Ths problem s made more severe because Chna lacks a mechansm to transfer fscal resources from rural to urban areas n proporton to a changng populaton dstrbuton (World Bank and DRC 2014). Further reforms of the Hukou system are currently debated (State Councl 2014). These reforms would essentally abolsh the Hukou system for smaller ctes below 1 mllon populaton. Regstraton n larger ctes would be ncreasngly more dffcult, but the reforms would gve ctes flexblty, ncludng usng pont systems to determne elgblty based on job qualfcatons, length of resdence, and so on. The reforms would also strengthen the rghts ncludng land use rghts of rural resdents, ensure more equtable servce delvery across rural and urban areas, and mprove publc fnance transfer mechansms between regons of outmgraton and those whose populaton s growng. Despte ther restrctve nature, Chna s regstraton polces have not prevented large scale mgraton to urban areas snce the begnnng of the reform era. Between 1995 and 2000, for example, an estmated 50 mllon people moved from rural to urban areas (Chan, 2013). In 2013, Chna s Natonal Bureau of Statstcs estmated the total non-hukou populaton (people lvng n a place wthout havng a Hukou for that locaton) at 289 mllon (NBS, 2014). Nevertheless, by reducng or preventng access to benefts n the destnaton ctes, and by lmtng the portablty of accrued benefts and monetzaton of assets n the rural areas, mgraton under the Hukou system has lkely stll been consderably lower than t would otherwse have been. 3. The model Our model ntroduces labor moblty nto the NEG model developed n Roberts et al. (2012). We explctly allow for people to mgrate between prefectures as well as between the urban and rural part 7

10 of each prefecture. To do ths, we base ourselves on Tabuch and Thsse (2007) and Behrens et al. (2014), and assume that each person j s lkelhood to choose to lve n locaton (the urban or rural part of any of the 331 Chnese prefectures n our sample), s based on the utlty he or she derves from lvng n that place. Ths utlty depends lnearly on real wages, W, earned n locaton, locaton s amentes, A, and an ndvdual-specfc dosyncratc preference for lvng n locaton, ε j: 13 U j = W + A + ε j (1) People choose to lve n the locaton whch provdes them wth the hghest utlty 14, so that the probablty that ndvdual j chooses to lve n locaton s: P(U j > max k U kj) (2) Assumng that the ε j are drawn from a double exponental functon wth mean π 2 µ 2 /6, ths probablty can be wrtten n the followng logt form (see McFadden, 1974): P(U j > max k U kj) = exp((w + A )/µ) / k[exp((w k + A k)/µ)] (3) Here, µ determnes the mportance of a person s dosyncratc preferences n determnng hs/her locaton choce. If µ s very small, people bascally choose ther locaton only based on W + A. That s, everybody chooses that locaton offerng them the best combnaton of real wages and amentes. In ths scenaro, multple ctes can only exst f possble real wage dfferences between locatons are perfectly offset by dfferences n amentes. By contrast f µ s very large, people bascally choose each locaton wth equal probablty 1/K, where K denotes the number of possble locatons. In ths case people s dosyncratc preferences for each locaton are very heterogeneous and, as a result, real wages and amentes do not matter for each ndvdual s locaton choce. Spatal equlbrum n the model s reached when the probablty n (3) corresponds to the actual observed share of people lvng n locaton : L /( k L k) = exp((w + A )/µ) / k[exp((w k + A k)/µ)] (4) 13 The dosyncratc preferences could also reflect dfferences between people n the cost of mgraton cost, n the value they put on lvng close to ther frends or relatves, or n the nformaton they have about the potental benefts/cost of mgraton. 14 Many NEG-models (notably Krugman s orgnal 1991 contrbuton) assume that people only move n response to real wage dfferences. Ths would mean that n a spatal equlbrum real wages are equalzed, both across the urban and rural sectors, as well as across all prefectures n Chna. Introducng other motves for mgraton (here amentes and other unobserved locaton-specfc preferences) means that real wage dfferences between locatons can persst n equlbrum. Moreover, the locaton-specfc dosyncratc preferences ensure that, n equlbrum, at least some people lve n each and every locaton. Note that n Desmet and Ross-Hansberg (2013) people are also perfectly moble and decde where to lve based on more than just real wages. However, they do not consder ndvdual specfc heterogenous preferences for lvng n each cty. As a result, some ctes dsappear n ther counterfactual scenaros (the same holds for Bosker et al. (2012), where n the equlbrum wth labor moblty only a few ctes reman n exstence). 8

11 In our model we take each locaton s amentes A as exogenously gven. Real wages n each and every locaton are nstead endogenously determned n our model. 3.1 Determnng real wages the role of transportaton nfrastructure Real wages n each and every locaton are determned n the exact same way as n Roberts et al. (2012). Here we brefly set out the man features of ther model, referrng to ther orgnal paper for the full detals. We focus n partcular on the way that transport costs determne real wages n the urban and rural sectors of each prefecture. By nfluencng market access they are the man condut through whch the NEN affects spatal economc outcomes n our model. The model n Roberts et al. (2012) s an elaborated n-regon verson of the orgnal NEG model of Krugman (1991a, 1991b). Each locaton conssts of an urban and a rural part. The urban part of each prefecture houses the urban sector whch s characterzed by nternal economes of scale and monopolstc competton. Its rural part s nstead home to a constant returns rural sector n whch perfect competton prevals. Perfectly tyng each of the two sectors n Roberts et al. (2012) to the rural and urban part of each prefecture allows us to explctly model mgraton between the urban and rural part of each prefecture. In effect, n our model, the decson to move from the rural part of a prefecture to ts urban part necessarly nvolves a change of sector. 15 Both sectors face love of varety preferences and (ceberg) transport costs to the rural sector, whereas, n the orgnal Krugman model, these are confned to just the urban sector. It mples that the constructon of the NEN also has an mpact on the prces and wages n the rural sector. Fnally, followng Südekum (2005), the model allows for varatons n labor effcency both across regons and between the urban and rural sectors. As set out n Appendx A, equlbrum n the model s characterzed, for each prefecture, by a system of fve smultaneous non-lnear equatons. For each prefecture, these equatons determne wages and the prce ndces n both ts urban and rural part (sector). Importantly, these depend on two factors () the regon s (exogenous) level of labor effcency n ts urban/rural sector, and () the regon s level of real market access (RMA) n that sector. The RMA n each sector provdes the man channel through whch transport costs and, hence, the constructon of the NEN affects the overall spatal equlbrum. In partcular, a declne n transport costs assocated wth e.g. the constructon of a new hghway network lnk between two prefectures has two opposng effects on wages workng through RMA. To see ths, take the example of the urban sector. A reducton n the costs for urban frms n prefecture of 15 We thnk ths s a realstc assumpton because mgrants n urban areas of Chna predomnantly work n lowsklled manufacturng or servce sector jobs. See also World Bank (2013). 9

12 transportng ther output to another prefecture j, ncreases demand for prefecture s output. Ths puts upward pressure on urban wages. However, the reducton n transport costs also exposes urban frms n regon to greater competton from urban frms located n prefecture j. Ths resultng ncrease n competton results n countervalng downward pressure on real wages. The overall mpact on the urban wage depends on whch of these opposng forces domnates Spatal equlbrum wth and wthout the Hukou restrctons A spatal equlbrum n our model s defned by (4) and [A1]-[A5] n Appendx A 17. The spatal equlbrum that these sx equatons te down s however a spatal equlbrum that assumes perfect nterregonal labor moblty. Everyone freely chooses hs/her locaton. In other words, ths represents a no Hukou equlbrum : an equlbrum where every Chnese ctzen can choose where to work and lve wthout facng any restrctons. However, we know that the Hukou restrctons do stll pose sgnfcant barrers for many people, keepng them from mgratng to ther preferred place of lvng. Ths makes t unlkely that (4) holds n realty. Ths s especally mportant for the papers smulatng the mpact of the NEN n a Chna where the Hukou restrctons are n place (Roberts et al, 2012; and Faber, 2014). 18 These papers bascally deal wth ths ssue by assumng that the observed dstrbuton of people and economc actvty can be proxed by a spatal equlbrum wthout any nterregonal labor moblty,.e. an extreme Hukou scenaro. In the context of our model, ths means that equlbrum s defned by equatons [A1]-[A5] only (as n Roberts et al., 2012). Here we use a dfferent strategy to adapt the model to a stuaton wth restrcted, but not prohbtvely restrcted, labor moblty. Gven that many people n Chna dd choose to mgrate despte the Hukou restrctons, assumng that no one responds wth hs/her feet to changes n economc opportuntes across locatons brought about by the constructon of the NEN, mght result n over- or underestmatng the mpact of the NEN on regonal economc outcomes. To take restrcted labor moblty more serously, we use the nformaton that we have on each locaton s total mgrant and non-mgrant populaton, where. In partcular, we assume that, under the current Hukou system, only 16 Also, the mpact of a declne n transport costs n the urban sector tends to postvely spll over to the rural sector because these sectors are lnked through ncome. However, ths postve relatonshp can break down when the constructon of a network lnk reduces both urban and rural transport costs (as wll be the case n our NENscenaros). Ths, n partcular, can occur when there s a large degree of asymmetry between prefectures and j n the sense that one s heavly specalzed n the urban sector and the other n the rural sector. It generates the possblty that a prefecture s urban and rural wages may move n opposte drectons n response to the constructon of a new network lnk. 17 Equatons [A1]-[A5] correspond to equatons [3] [7] n Roberts et al. (2012). 18 For the papers consderng the mpact of the Hukou system (Bosker et al., 2012; Whalley and Zhang, 2007), ths s not so much an ssue. They smply compare today s spatal dstrbuton of people or economc actvty to the smulated dstrbuton under completely unrestrcted labor moblty. Ths means that they do not have to smulate any model outcomes under a scenaro of restrcted labor moblty. 10

13 those people that have actually been observed to mgrate wth the system n place, are potentally wllng to mgrate (.e. the revealed Hukou mgrants n our dataset, ). These revealed Hukou mgrants are assumed to choose ther locaton accordng to (1) (3). For the non-mgrants n our data, we nstead assume that they wll never mgrate as long as the current Hukou restrctons reman n place. Ths means that a spatal equlbrum wth the current Hukou restrctons n place s formally characterzed n our model by equatons [A1] [A5] n Appendx A n combnaton wth: L /( k L k) = + = P(U j > max k U kj) + (5) = [exp((w + A )/µ) / k [exp((w k + A k)/µ)]] + (5) clearly shows that when the constructon of the NEN results n changes n real wages, W, across prefectures, ths wll trgger a restrcted mgraton response by those people that have been shown to be wllng to mgrate despte the Hukou restrctons. The rest of the Chnese populaton stays were they are. Rewrtng (5) shows that t s bascally an equlbrum condton on the (relatve) sze of a locaton s mgrant populaton only: = [exp((w + A )/µ) / k [exp((w k + A k)/µ)]] (5b) 4. Estmatng the man model parameters Havng set out our model, detalng how we allow for dfferent types labor moblty, we now turn to how we estmate the man model parameters. These estmates are the crucal nputs nto the counterfactual exercses that we perform n the next secton to assess the spatal mpacts of the NEN and/or the Hukou restrctons on labor moblty. We brefly dscuss our data and revew the estmates that we take from Roberts et al. (2012). The man focus n ths secton, however, s on how we estmate the mportance of real wages and (natural) amentes n people s mgraton decson. 4.1 Data We buld on Roberts et al. (2012, Appendx A) for data on the NEN and on each prefecture s urban and rural wages, ncome, nvestment per worker, human captal (years of educaton), and land area. Ths data s avalable for 331 prefectural ctes (three prefectural ctes were excluded due to mssng data) 11

14 and smlar admnstratve unts. 19 Travel tmes as a proxy for trade costs between each par of provnces are derved from two detaled GIS road network data sets one wthout and one wth the NEN usng a standard shortest-path algorthm. Travel tmes n the before network (.e. wthout the NEN n place) wll always be no larger (usually substantally lower) than n the after (.e. wth the NEN n place) network, because hghways allow greater speeds and more drect connectons. One smplfcaton s that we assume that travel tmes wthn a prefecture are zero (t = 0). 20 The novel data that we use n ths paper concern nformaton on mgraton and some of ts alleged determnants. Chnese mgraton statstcs are notorously dffcult to nterpret (Chan, 2013). Ths s especally the case for blateral mgraton flows between locatons. A vrtue of our structural modelng approach s that we can nfer the mportance of dfferent mgraton determnants by lookng at mgraton stock only. Ths nformaton s avalable 21. Our data on the stock of mgrants comes from a comprehensve data set for 2000 and 2010 derved from offcal Chnese census publcatons and other statstcal data sets such as the provncal statstcal yearbooks (Chreod Ltd., 2013). The data set ncludes nformaton on total populaton of mgrants from the same county, total populaton of mgrants from other countes n the same provnce and total populaton of mgrants from other provnces, each for 2000 and It also provdes data on total populaton wth household regstraton,.e. wth a local Hukou, for both years. We complement t wth data on geographc characterstcs/amentes possbly nfluencng people s mgraton decsons. Informaton on terran ruggedness comes from Nunn and Puga (2012). Data on coolng and heatng degree days are from NASA (2009). Annual coolng degree days measure the total number of degrees by whch daly temperatures exceeded 18 C. It s often used as a measure of the need for ar condtonng. Smlarly, heatng degree days ndcate the accumulaton of degrees when the daly mean temperature s below 18. Informaton on amentes (the share of households wth access to tap water, tolets and natural gas supply) come from the same data set as our mgraton stocks. 22 Fnally, we have collected the domnant language for each prefectural regon from a GIS data set of lngustc regons called the World Language Mappng System (Ethnologue, 2004). 19 The 331 prefectures covers 280 of Chna s 283 prefectural ctes, 11 of ts 17 prefectures, all 30 of ts ethnc mnorty autonomous prefectures, and all three of ts leagues. Also ncluded are the muncpaltes of Bejng, Shangha and Tanjn. Three areas that together comprse Chongqng are also ncluded n the sample the One Hour Crcle, the Northeast wng and Southeast wng (see Roberts et al, p 590, for further detals). 20 Ths means that we abstract from any dfference n the qualty of nfrastructure wthn prefectures. See Baum- Snow and Turner (2013) and Baum-Snow et al. (2013) for detaled studes on nter-prefectural nfrastructure. 21 Its qualty can stll be debated. The offcal statstcs that we use are lkely to underreport actual mgraton stocks n some ctes gven the many unoffcal mgrants. 22 We use the 2010 fgures n most of our estmatons. Results are very smlar when usng estmates for 2007 nstead (the same year for whch we have our NEN-travel tme-nformaton), obtaned by lnearly nterpolatng the fgures from 2000 and

15 4.2 Estmaton Snce our only adaptaton of the Roberts et al. (2012) model s the ntroducton of labor moblty, the only parameters that we need to assgn values to here are those determnng the mportance of real wages and (natural) amentes n people s mgraton decsons. We dscuss the estmaton of these parameters below. Appendx B sets-out the estmates of all the other man model parameters that we adopt drectly from Roberts et al. (2012) The determnants of mgraton n Chna To dentfy the relatve mportance of real wages and amentes n people s mgraton decsons, we follow Behrens et al. (2013) and make use of the spatal equlbrum condton under the current Hukou system (5b). The strong assumpton underlyng our dentfcaton s that Chna s spatal economy s n equlbrum n Under ths assumpton, we can frst obtan real wages n each prefecture s urban and rural part n the exact same way as calbrated n Roberts et al. (2012). Ths step reles solely on parameters already dentfed n that paper (see also Appendx B). Next, we plug these calbrated real wages, W, nto equlbrum condton (5b). Ths allows us to back out the ndrect utlty levels that correspond to the 2007 Hukou spatal equlbrum: Frst normalze W 1 + A 1 0 (ths bascally means that we measure utlty relatve to that n a baselne locaton, whch we take to be the urban part of Shjazhuang prefecture 24 ). Next dvde each locaton s mgrant share n total Chnese populaton by that n the baselne locaton, use the normalzaton W 1 + A 1 0, and take logs on both sdes of the equaton. From ths t mmedately follows that each locaton s log mgrant populaton relatve to that n the baselne locaton s drectly related to the ndrect utlty derved from the real wages and amenty levels n that locaton: ln = (W + A )/µ (6) Now, assume each locaton s amentes consst partly of observed amentes (.e. those that we have data on), A obs, and partly of unobserved amentes, A unobs. We can then estmate the mportance of real wages and observed amentes n drvng people s mgraton decson by runnng the followng regresson of each locaton s log mgrant populaton relatve to that n the baselne locaton on ts calbrated real wage W and ts observed amentes, A obs : 23 In our defense, the exact same strong assumpton s made n all other studes usng calbraton exercses to gauge the mpacts of ether the NEN or Hukou restrctons (e.g. Bosker et al., 2012; Whalley and Zhang, 2007; Roberts et al., 2012; Faber, 2014; and Desmet and Ross-Hansberg, 2013). 24 The choce of reference locaton s arbtrary, and does not nfluence any of the results. 13

16 ln = β 0 + β 1 W + β 2 A obs + ε (7) where β 0 captures the reference locaton s share n Chna s mgrant populaton. Moreover, the resduals can be nterpreted as the unobserved part of each locaton s amentes, A unobs =. 25 We use the above outlned strategy to estmate the relatve mportance of real wages, varous geographcal amentes, and dfferent publc amentes n determnng ndvdual locaton choces. The geographcal characterstcs of each locaton that we are able to nclude n our analyss are ts ruggedness, temperature (number of heatng and coolng days), ranfall, a dummy varable for locaton on the Yangtze Rver, and avalablty of natural resources. The publc amentes we are able to nclude focus on the provson of tap water, sewage, and natural gas to households. Tables 1a and 1b below show the results of estmatng dfferent versons of equaton (7). Table A1 n Appendx C provdes some addtonal extensons to these results that show that people are more senstve to real wage and/or amenty dfferences n nearby locatons (.e. wthn the same provnce). We need to menton three detals before dscussng our fndngs. Frst, all regressons control for unobserved determnants of mgraton that do not vary between the rural parts of prefectures n the same provnce as well as unobserved factors that do not vary between the urban parts of prefectures n the same provnce (.e. we nclude provnce-rural and provnce-urban fxed effects). In addton, each regresson also ncludes longtude and lattude and a locaton s area as controls. Includng area as a control takes account of the fact that the same populaton on less land means a larger populaton densty, and thus more congeston, whereas ncludng longtude and lattude ams to control for unobservables that are related to absolute (and arguably also relatve) locaton. Fnally, we nclude dummes for four of the major languages spoken n Chna, to (roughly) capture possble language barrers to nternal mgraton. 26 Second, n almost all specfcatons we allow the geographcal amentes to dfferentally affect mgraton decsons n urban and rural areas. We do ths as most of our geographcal data reports the same value for each prefecture s urban and rural part. Thrd, and most mportantly, by regressng our real wage and amenty data on log mgrant populaton relatve to a baselne locaton, we effectvely dentfy people s wllngness to mgrate n response to real wage and/or amenty dfferences from the varaton n locaton choces of people who dd decde to mgrate. Ths mples that n our counterfactual no Hukou scenaro we make the mplct assumpton 25 Usng the resduals as a measure of each locaton s unobserved amentes does assume: () that our lnear utlty model s well-specfed (.e. wth real wages, observed and unobserved amentes enterng utlty n an addtvely separable way), and () that these unobserved amentes are uncorrelated wth both a locaton s real wages as well as ts observed amentes. 26 Full results are avalable upon request. 14

17 that current mgrants preferences are representatve of those of the entre Chnese populaton. 27 An alternatve could be to estmate (7) usng log total populaton as the dependent varable. We do show these results as a robustness check, but prefer our estmates usng log mgrant populaton. In fact, a complcaton that arses when estmatng (7) usng total populaton shares, s that the resdual does not only capture unobserved amenty dfferences, but also the Hukou restrctons capablty to keep people from movng out of each locaton. 28 Table 1a bulds up to the baselne results that we use to smulate our counterfactual scenaros n the next secton. Column 1 shows results when only ncludng real wages n the regresson. Subsequently, n columns 2 and 3 we consecutvely add our geographc and publc amenty varables. Column 4 complements the other columns by showng results when restrctng the mpact of the geographcal varables to be the same n urban and rural areas. Irrespectve of the amentes ncluded, we always fnd a large and sgnfcant postve effect of real wages on the sze of a locaton s mgrant populaton. 29 In our preferred specfcaton n column 3, a 1 percent hgher real wage corresponds to attractng a 0.75 percent larger mgrant populaton. Of the geographcal characterstcs we fnd that only ranfall, heatng days and ruggedness are sgnfcantly assocated wth a locaton s ablty to attract mgrants. However, wth the excepton of ruggedness, ths s only so n rural areas. Fnally, we fnd that the provson of publc servces, as captured by the three varables we were able to collect comprehensve data on, s another mportant determnant of people s mgraton decson. Especally the avalablty of sewage facltes, and a drect natural gas connecton for heatng and cookng are mportant. Some care s warranted n takng these results regardng publc amenty provson too lterally, however, as they may suffer from endogenety ssues (nduced by ether reverse causalty or omtted varables). For example, other amentes (that we do not observe) correlated wth both our ncluded publc amentes and mgrant stocks are lkely to exst. Overall, however, we take our fndngs as ndcatve of the mportance of both real wages and publc amenty provson n shapng mgraton patterns n Chna. 27 Ths s a strong assumpton. Whether or not t s a reasonable assumpton s hard to test gven the avalable data. One way to nfer ths would be to ask non-mgrants what factors keep them from mgratng, as well as what factors would determne ther mgraton decson n the absence of the current Hukou restrctons. Ths s beyond the scope of the current paper. 28 See also Desmet and Ross-Hansberg (2013) 29 Ths corroborates earler fndngs by Poncet (2006), Rozelle et al. (1999) and Zhao (1999). 15

18 Table 1a. Determnants of the stock of mgrants (1) (2) (3) (4) Dep. Varable: ln mgrants ln mgrants ln mgrants ln mgrants ln real wage [0.00]*** [0.00]*** [0.00]*** [0.008]*** urban rural urban rural no splt Geo ln ruggedness [0.023]** [0.001]*** [0.044]** [0.003]*** [0.003]*** ln coolng days [0.254] [0.537] [0.622] [0.933] [0.403] ln heatng days [0.056]* [0.001]*** [0.486] [0.045]** [0.394] ln ranfall [0.564] [0.003]*** [0.735] [0.004]*** [0.23] D yangtze [0.104] [0.273] [0.768] [0.12] [0.489] nat.res.ndex [0.37] [0.871] [0.55] [0.839] [0.71] % hh water [0.702] [0.098]* % hh tolet [0.009]*** [0.004]*** % pop gas [0.022]** [0.031]** nr.obs R Notes: all regressons nclude provnce-urban and provnce-rural fxed effects, as well as controls for a prefecture s ln(x-coordnate), ln(y-coordnate), ln(area) and four dummy varables denotng whether the domnant language spoken n each prefecture corresponds to one of four of Chna s man languages spoken (Mandarn, Yue, Wu and Jnyu (often consdered a dalect of Mandarn)). p-values, based on standard errors clustered at the provnce level, n brackets. ***, **, * denotes sgnfcance at the 1%, 5%, 10% respectvely. In all columns we allow the coeffcents for all geography-related varables to dffer between the urban and rural parts of prefectures respectvely (effectvely ths means that we nclude each geography varable nteracted wth our rural dummy as well as nteracted wth our dummy ndcatng the urban part of each prefecture). Table 1b shows results when usng dfferent dependent varables nstead of mgrant stocks when estmatng (7). They provde crucal perspectve on our choce of dentfyng the (relatve) mportance of dfferent mgraton determnants lookng at each locaton s mgrant populaton only. Column 1 and 2 show results when consderng the share of non-mgrants or the share n total Chnese populaton n each locaton as the dependent varable respectvely. 30 The most strkng dfference s the fact that we fnd a much weaker assocaton between real wages and these shares. 30 We also performed a regresson wth the share of Hukou-holders n each locaton as the dependent varable. The results from ths regresson were almost dentcal to those when usng non-mgrant shares. Ths s not that surprsng as both measures are hghly correlated (wth a correlaton coeffcent of 0.996). These results are avalable upon request. 16

19 Table 1b. Non-mgrants, Hukou holders, and total populaton (growth) Dep. Varable: ln non-mgrants ln tot. pop D tot. Pop ln real wage [0.100]* [0.046]** [0.022]** Urban rural urban rural urban rural ln ruggedness [0.005]*** [0.001]*** [0.006]*** [0.001]*** [0.278] [0.172] ln coolng days [0.814] [0.911] [0.968] [0.919] [0.628] [0.266] ln heatng days [0.043]** [0.001]*** [0.073]* [0.002]*** [0.626] [0.003]*** ln ranfall [0.847] [0.012]** [0.943] [0.012]** [0.178] [0.555] D yangtze [0.763] [0.331] [0.788] [0.241] [0.52] [0.842] nat.res.ndex [0.932] [0.062]* [0.87] [0.112] [0.465] [0.181] % hh water [0.012]** [0.066]* [0.008]*** % hh tolet [0.023]** [0.012]** [0.098]* % pop gas [0.029]** [0.081]* [0.004]*** nr.obs R Notes: all regressons nclude provnce-urban and provnce-rural fxed effects, as well as controls for a prefecture s ln(x-coordnate), ln(y-coordnate), ln(area) and four dummy varables denotng whether the domnant language spoken n each prefecture corresponds to one of four of Chna s man languages spoken (Mandarn, Yue, Wu and Jnyu). p-values, based on standard errors clustered at the provnce level, n brackets. ***, **, * denotes sgnfcance at the 1%, 5%, 10% respectvely. In all columns we allow the coeffcents for all geography-related varables to dffer between the urban and rural parts of prefectures respectvely (effectvely ths means that we nclude each geography varable nteracted wth our rural dummy as well as nteracted wth our dummy ndcatng the urban part of each prefecture). Both the sgnfcance, as well as the sze, of the estmated coeffcent on real wages falls substantally. Ths s very dfferent for the geographcal characterstcs, where we fnd smlar results as when consderng mgrant stocks. The same holds for our three publc amenty varables, except for a surprsng negatve effect of the percentage of households that have tap water n ther home. Ths most lkely reflects the fact that provdng tap water s easer when servng a smaller populaton. The results when consderng total populaton shares (mgrants + non-mgrants) n column 2 are more smlar to those usng non-mgrant shares, whch s not that surprsng gven that mgrants typcally make up only a small part of a prefecture s populaton. Fnally, when consderng populaton growth nstead (column 3), we agan fnd a postve effect of real wages. It s not as sgnfcant as n our mgrant-regressons, 17

20 but one has to keep n mnd that populaton growth s only partly drven by mgraton. 31 Combned wth our earler mgrants-results n Table 1a (as well as the extensons n Table A1), and the survey and provncal evdence n earler studes (Poncet, 2006; Rozelle et al., 1999; and Zhao, 1999a,b), all our evdence supports the dea that real wages are ndeed a very mportant drver of mgraton decsons n present day Chna. 32 In the counterfactual scenaros that allow for labor moblty, we always use the results shown n Table 1a, column 3, as shapng people s mgraton decsons. We do not only use the estmated parameters of (7), but also the correspondng estmates of the unobserved amentes n each prefecture s urban and rural parts (the sum of each locaton s resdual and ts urban-rural/provnce fxed effect). 5. The spatal mpacts of the constructon of the NEN and the Hukou system Wth all the structural model parameters n hand, we are now n a poston to assess the mpact of the NEN and/or Hukou restrctons on Chna s spatal economy. We do ths by smulatng our full structural NEG model under three dfferent scenaros. We dscuss n detal how we generate the counterfactual spatal equlbra that each of these scenaros rely on. In all three scenaros our analyss focuses on the dstrbuton of four man varables across Chna s prefectures: real ncome per worker, the urban-rural real wage gap, urbanzaton rates, and overall populaton. 33 For sake of comparson, we also sometmes report the Roberts et al. (2012) results that consder the mpact of the NEN under an extreme Hukou scenaro (.e. no labor moblty), and/or compare our no Hukou -fndngs to those found by Bosker et al. (2012). 5.1 Smulatng the mpact of the NEN and/or Hukou system We observe the current 2007 dstrbuton of people, urbanzaton and real ncomes across the urban 31 On average, only about 40% of Chnese cty growth s drven by mgraton. Another 40% s drven by urban expanson nto rural areas, leavng 20% as beng determned by natural populaton growth [World Bank and DRC, 2014]. 32 From the perspectve of our later smulaton exercses, the coeffcent on real wages n these regressons s most mportant. In our setup, the (constructon of the) NEN affects people s locaton decson solely through ts effect on real wages n the rural and urban part of each prefecture. We are aware that our man results regardng the mportance of real wages n people s mgraton decson mght stll be plagued by reverse causalty or omtted varable bas. We lack a convncng source of exogenous varaton to deal wth ths ssue. But, n lght of all fndngs n our robustness checks as well as n earler papers that have shown the mportance of wage/ncome consderatons when decdng to mgrate or not, we fnd t very unlkely that addressng these endogenety concerns wll entrely undo the postve effect of real wages that we dentfy n our regressons. It could result n a smaller postve effect however. Our results n secton 5 are qualtatvely robust to assgnng a smaller role for wages n people s mgraton decson. 33 Our smulatons also allow us to look at nomnal ncomes, prce levels, etc. n the urban and rural part of each prefecture. Results are avalable upon request. 18

21 and rural parts of each Chnese prefecture (n total we consder 662 locatons of whch 331 are urban and 331 rural). We take ths as our baselne spatal equlbrum wth the NEN n place and wth restrcted labor moblty because of the Hukou system,.e. a spatal equlbrum satsfyng equatons [A1]-[A5] and (5). Scenaro 1: The mpact of the NEN under the current Hukou system In ths scenaro, we compare the observed 2007 NEN + Hukou spatal equlbrum to the counterfactual no NEN + Hukou spatal equlbrum. We smulate ths counterfactual by changng the travel tmes between locatons back to the stuaton before the NEN was bult. To fnd the spatal equlbrum correspondng to ths new stuaton, we frst numercally solve for the (short run) equlbrum values for w U, w R, G U, and G R, n each prefecture, usng equatons (A1) (A5) n Appendx A1, the no-nen travel tme matrx, and the teratve procedure detaled n Roberts et al. (2012). In Roberts et al. (2012) ths would already consttute the counterfactual spatal equlbrum. Ths s not the case when allowng for (restrcted) labor moblty: now the new equlbrum wage and prce levels n each locaton may nduce people to change ther locaton accordng to the mgraton dynamcs we estmated n Secton 4. In partcular, we use the smulated (short run) equlbrum values for w U, w R, G U, and G R to calculate real wages n the urban and rural part of each prefecture: ω = w U /P f s the urban part of a prefecture and ω = w R /P f s the rural part of the prefecture respectvely, where P = (G U ) θ (G R ) 1- θ and G U and G R are the urban and rural prce ndces respectvely (see Appendx A). Subsequently we allow these counterfactual real wage changes to trgger a mgraton response. More specfcally, they change each locaton s mgrant populaton relatve to that n our baselne locaton (defned earler) accordng to equaton (7). 34 ln /,, (8) In a Hukou scenaro wth restrcted labor moblty, we can then calculate each locaton s counterfactual populaton as follows based on (5): / (9) 34 Note that ths assumes that the level of amentes n a locaton s ndependent of ts populaton. In partcular, we hold amentes fxed at ther 2010 levels. In dong so, we effectvely abstract from the fact that a populaton nflow may put a stran on local governments to keep up the same level of publc servce provson, or may gve rse to congeston. 19

22 Next, ths new spatal confguraton of populaton across the urban and rural parts of all 331 prefectures changes each locaton s real market access and thus ts real wages. To capture these changes, we numercally resolve for the equlbrum urban and rural wages and prce levels based on equatons [A1] [A5] and the new populaton levels n each locaton. These new real wages n turn may affect mgrants locaton choce, and so on. We repeat these steps teratvely untl convergence s reached (.e. real wages, and populaton shares no longer change between teratons). Upon convergence 35, we have obtaned our counterfactual no NEN + Hukou spatal equlbrum. Comparng t to the 2007 NEN + Hukou stuaton reveals the general equlbrum mpact of the NEN on Chna s spatal economy makng the more realstc assumpton of restrcted labor moblty nstead of the extreme Hukou assumpton of no labor moblty used n earler papers (Roberts et al., 2012, or Faber, 2014). By allowng for more realstc mgraton dynamcs, we can assess the possble mtgatng effect of labor moblty on the ncrease n regonal nequalty, assocated wth the constructon of the NEN found n Roberts et al. (2012). Scenaro 2: The mpact of abolshng the Hukou system Our second scenaro focusses on the effect of the complete abolshment of the Hukou system on Chna s spatal development. 36 It does so by comparng the observed 2007 NEN + Hukou spatal equlbrum to the counterfactual NEN + no Hukou spatal equlbrum. We smulate a counterfactual scenaro under a complete abolshment of the Hukou system as follows. In the 2007 NEN + Hukou spatal equlbrum condton (5) or equvalently (5b) holds. In a spatal equlbrum where the Hukou restrctons are abandoned, everyone s now a potental mgrant, and equlbrum s defned nstead by (4). Usng (8) and our calbrated real wages n 2007, we can easly verfy whether (4) holds. If t does not, we update each locaton s populaton accordng to the followng equaton that now takes all Chnese ctzens as potental mgrants: / (10) As n Scenaro 1, ths new spatal confguraton of populaton across the urban and rural parts of all 331 prefectures changes each locaton s real market access and thus ts real wages. To capture these changes, we numercally re-solve for the equlbrum urban and rural wages and prce levels based on equatons 35 To be precse, we defne convergence to be reached when the squared dfference between teratons n urban, rural and total populaton s smaller than for each of these three varables respectvely. 36 A complete abolshment of the Hukou system s unlkely to happen n the near future. In prncple, we could also look at scenaros nvolvng a partal abolshment of the Hukou system (e.g. allowng free mgraton to all but the already largest ctes, or n partcular provnces only). However, such results wll depend qute heavly on the specfc partal-hukou system consdered. Moreover, a complete abolshment allows us to most clearly contrast the effects of nfrastructure nvestment vs. mgraton restrctons on Chna s economc geography. 20

23 [A1] [A5] and the new populaton levels n each locaton. These new real wages n turn may affect mgrants locaton choce accordng to (8) and (10), and so on. We repeat these steps teratvely untl convergence s reached (.e. real wages, and populaton shares no longer change between teratons). Upon convergence, we have obtaned our counterfactual NEN + no Hukou spatal equlbrum. Scenaro 2 can be drectly compared to Scenaro 1 to assess whether, and f so how, Chna s two man spatal polces dfferently affect ts spatal economy. Also, t can be drectly compared to the results shown n Bosker et al. (2012) that also consder the spatal effects of the abolton of the Hukou restrctons at the prefecture level, but wthout explctly takng the large-scale nvestments n nfrastructure nto account. Note that the comparson to Bosker et al. (2012) s lmted nsofar as that paper focusses exclusvely on the changes n the spatal dstrbuton of people across prefectures, whereas we also consder the changes n, nter ala, real wages and urbanzaton levels. Scenaro 1b: The mpact of the NEN under free labor moblty (.e. no Hukou restrctons) Scenaros 1 and 2 are our man scenaros of nterest. Ths thrd scenaro complements Scenaro 1 (explanng why we call t Scenaro 1b). It provdes evdence on whether, and f so how, the constructon of the NEN would have had dfferent effects on Chna s spatal economy had there been no Hukou restrctons on labor moblty. It allows us to further assess the postulated mtgatng effect of free labor moblty on the ncrease n regonal nequalty assocated wth the constructon of the NEN that was found n Roberts et al. (2012). In ths scenaro, we compare the counterfactual NEN + no Hukou - spatal equlbrum that we already used n Scenaro 2 to a counterfactual no NEN + no Hukou - spatal equlbrum. We smulate ths last counterfactual usng the no NEN travel tme matrx to proxy transport costs between prefectures when solvng equatons [A1]-[A5], and combne t wth (8) and (10) to teratvely solve for ths no NEN + no Hukou spatal equlbrum. 5.2 Results the mpact of the NEN and/or the Hukou system on Chna s spatal economy In ths secton, we dscuss our man fndngs. Table 2 shows the changes n our man varables of nterest n each of the three scenaros defned n Secton 5.1. Table 3 complements these results by showng the correlaton between these changes, as well as wth the ntal levels of urbanzaton, populaton, and welfare n each prefecture. Fnally, Fgures 1 4 and Appendx D further llustrate some of the most nterestng correlatons and provde detaled maps of the spatal dstrbuton of these changes across Chna s prefectures. 21

24 5.2.1 The mpact of the NEN We start by consderng the mpact of the NEN under the current Hukou system (Scenaro 1). Table 2 shows that our Scenaro 1, that makes the more realstc Hukou assumpton of restrcted labor moblty, delvers results that are very close to those obtaned n earler papers that completely abstract from labor moblty (e.g. Roberts et al., 2012; but also Faber, 2014). We fnd a slghtly larger overall ncrease n real ncome per worker of about 6.6 percent, accompaned by an also slghtly larger ncrease n regonal nequalty (as measured by the standard devaton of real ncome per worker across prefectures) of about 9 percent. Also the urban-rural wage gap ncreases more n Scenaro 1 than n the Roberts et al. s (2012) extreme Hukou case. We also observe a slght ncrease n the agglomeraton of people when allowng for restrcted labor moblty. However, the populaton changes that we fnd are small. Chna s overall urbanzaton rate e.g. ncreases by a mere 0.4 percent pont. Ths small ncrease does however hde some sgnfcant spatal dfferences. Some of the fastest urbanzng places see an ncrease n ther urbanzaton rate of more than 4 percentage ponts. As shown n Fgure 1b and the maps n Fgure 2, these fastest urbanzng places are the ntally already most urbanzed places. They are also the ctes attractng most new mgrants from other ctes followng the constructon of the NEN (see the correlatons n Table 3). In sum, the constructon of the NEN has only strengthened the exstng agglomeraton pattern n Chna. It has renforced ts urban core at the expense of ts rural hnterland, both n economc as well as n populaton terms. Of course the fact that we only assume that only revealed Hukou mgrants, about a ffth of the total Chnese populaton, move n response to changes n real wages s an mportant part of the explanaton for these relatvely small dfferences between Scenaro 1 and the earler extreme Hukou results reported n Roberts et al. (2012). However, Scenaro 1b shows that even n the hypothetcal absence of any Hukou restrctons, we stll fnd almost the same overall welfare effect of the constructon of the NEN: an overall natonal ncrease of 6 percent n real ncome per worker. However, the spatal consequences of the constructon of the NEN do dffer substantally n ths no Hukou Scenaro 1b. Two thngs stand out n ths respect. Frst, we fnd a much smaller ncrease n regonal nter-prefectural ncome nequalty as a result of the constructon of the NEN n Scenaro 1b. The ncrease n the standard devaton of real ncome s only about half that n Scenaro 1 (see Table 2). Also, the relatonshp between ntal real ncome per worker and ts change due to the constructon of the NEN s much weaker (see Table 3). Fgure 1a shows that now many ntally wealthy places even observe a fall n real ncome per worker. And, on top of ths, we also see a drop n urban-rural ncome nequalty wthn the average prefecture, as well as a fall n the number of prefectures wth rsng urban-rural nequalty (only 36 percent compared to more than 50 percent n Scenaro 1). A hgher degree of labor moblty ndeed 22

25 appears to mtgate the rse n regonal ncome nequalty assocated wth the constructon of the NEN, whle delverng a smlar overall welfare ncrease for the average Chnese ctzen. Table 2 Counterfactual results Roberts et al. (2012) Scenaro 1 Scenaro 1b Scenaro 2 Counterfactual scenaro: NEN vs. no NEN NEN vs. no NEN NEN vs. no NEN Hukou vs. no Hukou (extreme Hukou) Hukou no Hukou NEN change n aggregate Chnese real ncome (pw) (%) sd real ncome pw (%) urbanzaton (ppt) sd urbanzaton (%) sd populaton (%) mean (std dev) change n real ncome pw (%) 4.0 (3.4) 3.9 (3.7) 2.7 (5.3) 23.0 (26.2) urban/rural wage gap (%) 0.4 (7.1) 1.3 (6.3) -1.1 (6.5) 11.4 (35.3) total populaton (%) (1.75) -1.7 (2.8) 3.7 (143.6) urbanzaton (ppt) (0.7) -0.3 (1.0) 21.8 (15.9) % prefectures wth ncreasng real ncome pw urban/rural wage gap populaton urbanzaton rate Second, also n Scenaro 1b, the places whose connectvty s most mproved by the constructon of the NEN gan most n terms of real ncome per worker and populaton (see the correlatons n Table 3). Wth the Hukou restrctons n place these now better connected places also wtnessed an ncrease n ntraprefecture urban-rural dspartes, both n terms of populaton (urbanzaton) and welfare (urban-rural wage gap). Ths s very dfferent n Scenaro 1b. In the absence of any restrctons on labor moblty ths relatonshp wth ntra-prefecture nequalty dsappears. Fgure 3 shows ths n more detal for the NEN s mpact on urbanzaton rates across Chna s prefectures. These dfferences between Scenaros 1 and 1b can be explaned by the fact that wthout any Hukou restrctons n place, many more people respond wth ther feet to the changes n real ncomes nduced by the constructon of the NEN. As a result, more people do move to the now better connected, ntally large prefectures than n the Hukou scenaro (aptly llustrated by the maps n Fgure 2c, and the hgher correlaton between the change n real ncome per worker and the change n populaton n Scenaro 1b n Table 3). It results n more spatal nequalty n terms of populaton (the standard devaton n prefecture populaton ncreases almost twce as much as n Scenaro 1b), renforcng the exstng agglomeraton pattern n Chna. Fgure 3. % reducton n travel tme because of the NEN and ppt change n urbanzaton 23

26 5 Restrcted labor moblty 2 Full labor moblty = no Hukou 4 0 ppt change urbanzaton Metro regons NE Coastal Central NW SW ppt change urbanzaton Metro regons NE Coastal Central NW SW % reducton travel tme NEN % reducton travel tme NEN Yet, t s also exactly these mgraton patterns that mtgate the rse n both ntra-prefecture and nterprefecture ncome nequalty assocated wth the constructon of the NEN that we saw n the Hukou Scenaro 1. On the one hand, the bgger nflow of people nto the ntally largest places mtgates the ncrease n ncome per worker n these places as a result of fercer competton for jobs. Ths explans the much lower assocaton between ntal real ncome per worker and both growth n real ncome per worker as well as n urban-rural nequalty n Scenaro 1b (see Table 3). On the other hand, regonal nequalty s also reduced because especally those people lvng n prefectures exhbtng the hghest levels of urban-rural nequalty are the ones movng out of ther prefecture n search of the hgher real ncomes offered elsewhere. Ths outflow of mostly rural nhabtants n prefectures losng populaton, results n a relatve ncrease of rural wages that mtgates the urban-rural wage gap n these places. 24

27 Table 3: Correlaton matrx outcomes dfferent scenaros Correlatons Scenaro 1 Scenaro 1b Scenaro 2 NEN vs. no NEN (Hukou) NEN vs. no NEN (no Hukou) no Hukou vs. Hukou (NEN) change n: [] [] [] [v] [] [] [] [v] [] [] [] [v] travel tme NEN (%) [] real ncome pw (%) [] urban/rural wage gap (%) [] total populaton (%) [v] urbanzaton (ppt) ntal: ln real ncome pw ln real ncome urban/rural wage gap ln populaton urbanzaton rate Notes: bold correlatons are NOT sgnfcant at the 5% level. 25

28 Fgure 1. Convergence or dvergence? Real ncome per worker and urbanzaton a. Real ncome per worker 30 Restrcted labor moblty 50 Full labor moblty = no Hukou 200 NEN - Hukou vs. no Hukou % change n real ncome per worker Metro regons NE Coastal Central NW SW % change n real ncome per worker Metro regons NE Coastal Central NW SW % change n real ncome per worker Metro regons NE Coastal Central NW SW Ln of real ncome per worker, wthout expressways Ln of real ncome per worker, wthout expressways Ln of real ncome per worker, Hukou b. Urbanzaton 5 Restrcted labor moblty 2 Full labor moblty = no Hukou 70 NEN - Hukou vs. no Hukou 60 ppt change urbanzaton, wth expressways Metro regons NE Coastal Central NW SW ppt change urbanzaton, wth expressways Metro regons NE Coastal Central NW SW ppt change urbanzaton Metro regons NE Coastal Central NW SW urbanzaton, wthout expressways urbanzaton, wthout expressways urbanzaton, no Hukou 26

29 Fgure 2. Spatal mpacts n each of the three scenaros a. Percent change n real ncome per worker b. Percent change n urbanzaton 27

30 Fgure 2, contnued c. Percent change n total populaton 28

31 DRAFT Not for ctaton Summng up, n both Scenaros 1 and 1b, the constructon of the NEN results n a smlar 6 percent ncrease n Chnese real ncome per worker. Ths average welfare gan s however assocated wth a substantal rse n regonal nequalty. The exstng urban and economc core regons of Chna gan most, whereas the ntally poorest, rural prefectures tend to lose populaton as well as economc actvty. A hgher degree of labor moblty does mtgate ths rse n spatal nequalty assocated wth the constructon of the NEN. But, even f every Chnese ctzen had been free to mgrate to the place of hs/her preference, the large-scale, spatally targeted, nvestments, n the NEN would stll have resulted n more, not less, spatal nequalty. 37 Ths s an mportant fndng, as the man reason to construct the NEN was the exact opposte: mtgatng the exstng regonal nequalty n Chnese economc development The mpact of removng the Hukou restrctons Now, we turn our attenton to the aggregate and spatal development mpacts of Chna s other man spatal development polcy: the Hukou system. In Scenaro 2, we take the constructon of the NEN as gven and smulate the (spatal) mpact of a complete abandonment of the Hukou system. 39 As can be seen n Table 2, allowng every Chnese ctzen to freely choose hs/her preferred locaton has a much bgger mpact than the constructon of the NEN. 40 The average Chnese worker experences an ncrease n real ncome of 170 percent. However, ths much larger ncrease n real ncome also comes wth a much more unequal spatal dstrbuton of real ncome and people across prefectures. Despte the fact that real ncome per worker goes up n bascally all prefectures, real ncome nequalty between prefectures rses as a few already rch prefectures experence much faster growth n real ncome per worker (Fgure 2a shows that these are manly located n Chna s coastal regons). Some ntally poor regons, n e.g. Chna s South West also beneft a lot (especally when compared to the mpact of the 37 The prefectures wth the largest gans n accessblty are typcally also the most urbanzed. Ths s not surprsng gven that the am of the NEN was to connect all ctes wth 200,000 or more nhabtants. A 1 percent hgher pre- NEN urbanzaton rate s assocated wth a 6.2 percent larger reducton n average travel tme as a result of the constructon of the NEN. Ths bas n NEN-nvestments towards the already exstng urban cores s an mportant explanaton for why we do not fnd an nequalty decreasng effect of ts constructon. 38 Of course our fndngs are based on a stylzed economc model that e.g. does not consder the constructon costs of the NEN. Nor does our model take nto account that dfferent ndustral sectors may respond dfferently to a change n transportaton costs. Moreover, we do not explctly take congeston costs nto account apart from the fact that compettve pressure on wages s hgher n larger places. 39 Results are very smlar when smulatng the mpact of the Hukou restrctons n a Chna wthout the NEN. We focus on the abandonment of the Hukou system takng the constructon of the NEN for granted as ths s the most polcy relevant scenaro gven that the NEN s already constructed whle the Hukou system s stll n place. 40 Do note the mportant caveat made at the end of ths secton that puts crucal perspectve on the often much bgger effects that we fnd n Scenaro 2. 29

32 DRAFT Not for ctaton constructon of the NEN, see Fgures 1a and 2a), but gven that they start from an ntally much lower base, the absolute change n ther ncomes s stll substantally lower than that n rch (coastal) Chna. Besdes ths rse n nter-prefectural real ncome nequalty, abandonng the Hukou restrctons also results n a much larger rse n the urban-rural wage gap of the average prefecture. Ths average does however hde substantal varaton between prefectures, the urban-wage gap actually falls n the same number of prefectures as n Scenaro 1. Interestngly (see Table 3), urban-rural nequalty ncreases the least n the fastest urbanzng prefectures. Ths s very dfferent from Scenaros 1 and 1b, where rsng urbanzaton and rsng urban-rural nequalty went hand n hand. It can be explaned by the fact that the fastest urbanzng places are also often places losng populaton (see below for more on ths). Ths outflow of mostly rural nhabtants results n a relatve ncrease n rural wages, mtgatng the rse n urban-rural nequalty n these places. Fgure 4. The spatal dstrbuton of people and urbanzaton n the absence of the Hukou system When abandonng the Hukou system, the spatal dstrbuton of people across Chnese prefectures also becomes much more unequal. As Fgure 2c shows, nland Chna sees a large outflow of people towards the coast (and some places n the North(-East)). As a result our smulatons predct that a Chna wthout Hukou restrctons would be a Chna where most people lve n the prefectures along ts (southern) coastlne (see Fgure 4). The man remanng large nland populaton centres, are the currently already very large nland ctes (notably Chengdu, Chongqng and Wuhan). 41 Interestngly, ths much more 41 Bosker et al. (2012) found an even stronger agglomeraton pattern emergng wth the abandonment of the Hukou system, wth only 52 ctes survvng n ther baselne smulatons. Moreover, ther agglomeraton pattern s qute dfferent from the one we fnd. Instead of our coastal + ntally large nland ctes core, ther core conssts of many more nland ctes as well as several place n Chna s North-East. These dfferences can be attrbuted to four mportant dfferences between our and ther counterfactual smulatons. Frst, Bosker et al. (2012) abstract from the qualty and quantty of nfastructure development n Chna, approxmatng travel tmes between ctes by smple geodesc dstances. Second, they consder only 264 prefectures, compared to the 332 that we consder (our sample also covers the prefectures n the west and north of Chna). Thrd, ther smulatons rely on ad-hoc mgraton dynamcs that, mportantly, do not consder ndvdual specfc dfferences n locaton preferences nto account. The fact that we do ths prevents prefectures from beng completely abandoned n a no-hukou scenaro (n ther baselne scenaro more than 200 prefectures empty out ). Fnally, we explctly consder each prefecture s 30