China Economic Review

Transcription

1 China Economic Review xxx (2017) xxx-xxx Contents lists available at ScienceDirect China Economic Review journal homepage: Fiscal spending and air pollution in Chinese cities: Identifying composition and technique e ects Yue Hua, Rui Xie, Yaqin Su a College of Economics and Trade, Hunan University, China b Center for Economics, Finance and Management Studies (CEFMS), Hunan University, China A R T I C L E I N F O JEL classi cations: Q53 Q55 R58 Keywords: Ed u ca tion spend ing R&D spend ing Air pol lu tion STIR PAT model 1. Introduction a a b, A B S T R A C T Fis cal spend ing has both di rect and in di rect im pact on the en vi ron ment. Us ing city-level data in China, this pa per in ves ti gates if ed u ca tion spend ing af fects air pol lu tion through hu man cap i tal ac cu mu la tion, known as the com po si tion ef fect, and if R&D spend ing af fects air pol lu tion through clean-tech nol ogy adop tion, known as the tech nique ef fect. Con trast ing the o ret i cal pre dic tions and pre vi ous em pir i cal ev i dence, we nd both ef fects of in ter est to be triv ial in ur ban China. Com po si tion ef fect ap pears to be slightly stronger rel a tive to tech nique ef fect, while subsam ple analy ses show some re gional het ero geneities. The re sults re main ro bust when we switch be tween pol lu tion mea sure ments, ex am ine only the re gional cen tral cities, in stru ment en doge nous co vari ates, and adopt the spa tial set tings. We fur ther dis cuss po ten tial chan nel-block ing mech a nisms that lead to weak es ti mates. Air pol lu tion plagues mod ern Chi nese cities. While Chi na's fast eco nomic growth has long been re ly ing on GDP-ori ented in dus tri al iza tion and ur ban iza tion, the de te ri o ra tion of eco log i cal and en vi ron men tal qual ity has be come in creas ingly no tice able in ur ban ar eas (Zheng & Kahn, 2017). Fac tory and ve hic u lar com bus tion of fos sil fu els gen er ate Sul fur Diox ide (SO 2 ), Ni tro gen Ox ides (NO X) and sus pended par tic u late mat ter (SPM) that lead to re cur rences of toxic smoggy weather and con t a m i na tion of ur ban aquatic sys tems. In 2014, the Chi nese cen tral gov ern ment re leased the Na tional New-Type Ur ban iza tion Plan ( ) in which the abate ment of ur ban air pol lu tion is of cially sched uled to be achieved in the con struc tion of pro-en vi ron men tal cities. The restora tion of en vi ron men tally amenable cities re quires an in-depth un der stand ing of the eco nomic de ter mi nants of ur ban air 1 pol lu tion. Among a broad set of can di date fac tors, this study ex am ines the ef fect of pub lic s cal spend ing on the level of lo cal air pol lu tants, mo ti vated by the fact that ex ces sive air pol lu tion is es sen tially the re sult of mar ket fail ures that un avoid ably ne ces si tate gov ern ment in ter ven tion. Ex ist ing the ory sug gests that the level and com po si tion of gov ern ment ex pen di ture may af fect en vi ron men tal pol lu tion through sev eral chan nels (Huang, Chen, Huang, & Yang, 2016; Lopez, Galinato, & Islam, 2011; Lopez & Islam, Corresponding author. addresses: huayue_ econ@ 126. com (Y. Hua); Xrxrui@ 126. com (R. Xie); yaqinsu@ hnu. edu. cn (Y. Su) 1 Studies on economy-pollution interactions have grown in recent decades since the seminal World Bank Development Report (1992) and Grossman and Krueger (1993). Sha k and Bandyopadhyay (1992), Seldon and Song (1994), and Grossman and Krueger (1995) are some of the earliest studies that show the inverted U- shaped relationship between environmental degradation and income, known as the Environmental Kuznets Curve (EKC). Kaufmann, Davidsdottir, Garnham, and Pauly (1998) report a quadratic relationship between the spatial intensity of economic activity and SO2 concentration. Technological improvements and high-tech rms are found to be helpful in relieving air pollution in developed countries (Suri & Chapman, 1998), while urbanization increase both CO 2 emissions and energy footprint (York et al., 2003) / j. chieco Received 28 December 2016; Received in revised form 2 June 2017; Accepted 19 September 2017 Available online xxx / 2017.

2 2008). First, eco nomic growth gen er ated by an in creased s cal spend ing cre ates more pol lu tion (a scale ef fect). Sec ond, re al lo ca tion of gov ern ment spend ing in fa vor of ed u ca tion tends to raise the share of cleaner hu man cap i tal-in ten sive ac tiv i ties rel a tive to the share of dirt ier phys i cal cap i tal-in ten sive ac tiv i ties (a com po si tion ef fect). Third, more s cal spend ing on R&D can re sult in a higher adop tion rate of cleaner tech nol ogy by rms that re duces the pol lu tion-out put ra tio (a tech nique ef fect). Fi nally, pri vate in come raised by pub lic-good ex pen di tures leads to higher de mand for a cleaner en vi ron ment and more strin gent reg u la tions (an in come ef fect). In this study we fo cus on the iden ti ca tion of com po si tion ef fect and tech nique ef fect, which re ceive much less dis cus sion in the lit er a ture rel a tive to the other two ef fects. Ex ist ing em pir i cal stud ies on the ef fect of gov ern ment size (or ag gre gate level of s cal spend ing) gen er ally sug gest that in creas ing to tal s cal ex pen di ture with out chang ing its com po si tion fails to im prove air qual ity (Bernauer & Koubi, 2006; Lopez et al., 2011). One in ter pre ta tion is that the re duced pol lu tion from rais ing pub lic-good ex pen di ture tends to be o set by the aug mented 2 pol lu tion from in creased pri vate-good ex pen di ture (Lopez et al., 2011). An al ter na tive ar gu ment is that gov ern ment ex pen di ture may grow for rea sons other than ex ter nal ity cor rec tion and pub lic-good pro vi sion, such as bu reau cratic in ef ciency and cor rup tion that can even ex ac er bate air pol lu tion (Bernauer & Koubi, 2006). Na tion-level ev i dence shows that re al lo cat ing gov ern ment spend ing in fa vor of pub lic goods tends to have de sir able im pacts on pol lu tion mit i ga tion. By em ploy ing a 15-year panel of 38 coun tries, Lopez et al. (2011) re ports a 10% in crease in the share of ex pen di tures on pub lic goods can re duce the SO 2 emis sion by 4% and the lead con cen tra tion by 7%. Us ing a 20-year panel of 77 coun tries, Halkos and Paizanos (2013) claims a neg a tive com po si tion and tech nique ef fect with out ex plic itly mea sur ing ed u ca tion and R&D spend ing. This study in tends to con tribute to the ex ist ing lit er a ture by rst in ves ti gat ing the ex is tence and mag ni tude of com po si tion and tech nique ef fect at city-level, with di rect mea sure ments of ed u ca tion spend ing and R&D spend ing. We use data from 284 Chi nese cities within the pe riod of for em pir i cal il lus tra tion, while the com bi na tion of s cal de cen tral iza tion, re gional het ero gene ity and en vi ron men tal di ver sity in China pro vides us with suf cient data vari a tion. Our nd ings show triv ial com po si tion ef fect and tech nique ef fect in ur ban China, with the for mer slightly stronger than the later. Re gional analy sis demon strates that the re sponse of air pol lu tion to wards ed u ca tion and R&D spend ing di min ishes from coastal area to in land area. The re sults re main ro bust when we switch be tween three types of air pol lu tion mea sure ments, con ne the sam ple to re gional cen tral cities, in stru ment en doge nous co vari ates, and in cor po rate spa tial lags to ac count for po ten tial lo ca tional spillovers. We fur ther pro vide both sup ply-side and de mand-side in ter pre ta tions for the un der ly ing mech a nisms that yield the weak ef fects of in ter est. The rest of the pa per is struc tured as fol lows. Section 2 pre sents an op ti mal con trol frame work that pro duces rel e vant com par a tive-sta tic hy pothe ses. Section 3 de scribes the data and em pir i cal mod el ing strate gies. Section 4 pro ceeds to the demon stra tion of em pir i cal re sults. Section 5 dis cusses po ten tial mech a nisms. Section 6 con cludes. 2. Conceptual framework The sem i nal Keeler, Spence, and Zeckhauser (1971) study rec og nizes that pol lu tion con trol has strong in tertem po ral as pects, and rec om mends the use of dy namic con trol ap proach in the o ret i cal mod el ing. We do not de vi ate from their pro posal and de scribe the 3 con trol of air pol lu tion as a dy namic op ti miza tion prob lem in this study. The model is con structed based upon the neo clas si cal Solow-Swan model of eco nomic dy nam ics and the ag gre gate econ omy-en vi ron ment in ter ac tion model de scribed by Hritonenko and Yatsenko (2013), with mod i ca tions that ac count for the ef fect of pub lic-good s cal spend ing. A so cial plan ner max i mizes a so cial util ity func tion by al lo cat ing nal out put across con sump tion, in vest ment and s cal spend ing 4 on pub lic goods, within a nite pe riod of time. A so cial util ity func tion U(C(t), P(t)) is set to be ad di tively sep a ra ble in the util ity of con sump tion C and the disu til ity of air pol lu tant emis sion P: lnc is a stan dard log a rith mic util ity func tion, and (with > 0) is a power util ity func tion, both sat is fy ing In ada con di tions. > 0 rep re sents the so ci etal tol er ance to wards pol lu tion. Fol low ing the con ven tional Solow-Swan set ting, P is mod eled to be in creas ing in out put Q and de creas ing in pub lic-good s cal spend ing F, with Q be ing pro duced by a neo clas si cal Cobb-Dou glas pro duc tion process us ing phys i cal cap i tal K as the sole vari able in put (as sum ing xed la bor-sup ply): 2 Lopez and Galinato (2007) proposes the taxonomy that divides scal expenditure into expenditures on public goods that correct for certain market failures (such as subsidies on education, research and development, environmental protection, transfer payments and other conventional public goods), and expenditures on private goods allocated to industries that are not privately under-invested. 3 Our model share some similarities with the Keeler et al. (1971) model in terms of basic settings, including the form of social welfare function and constraints associated with utility maximization, but signi cantly di ers from it by incorporating technological change and scal spending. 4 We model the problem as a nite-horizon mainly because of the xed tenure system for government o cials in China (10 years for two consecutive appointments, at most). The ending point of local policy planning usually coincides with the resignation of the person in charge. The Solow-Swan model considered on a nite horizon is also known as the Solow-Shell model. (1) (2) 2

3 where 1 < < 0. S stands for the adopted tech nol ogy. The as so ci ated cur rent-val ued ob jec tive func tion can thus be writ ten as sub ject to the fol low ing con strains: Eq. (4) de com poses nal prod uct Q into con sump tion C, in vest ment I, and pub lic-good s cal ex pen di ture F, with C and I be ing in de pen dent con trol vari ables and F be ing a state vari able de pend ing on C and I. Eq. (5) is the equa tion of mo tion that de scribes the move ment of cap i tal K as a func tion of the con trol vari able I and de pre ci ates at a con stant rate > 0. Con straint (6) con tains bound ary con di tions of the two state vari ables. The cur rent-val ued Hamil ton ian of this prob lem is given by: H is as sumed to be strictly con cave and dif fer en tiable in C and I to ex clude end point so lu tions. 1 and 2 are costate vari ables as so ci ated with con straint (4) and (5). Based on the Pon trya gin max i mum prin ci ple, the op ti mal ity con di tions are writ ten as: where Eqs. (8) and (9) are rst-or der ex tremum con di tions for the two con trol vari ables. Eqs. (10) and (11) are Euler con di tions as so ci ated with two state vari ables. For the pur pose of con duct ing com par a tive sta tic analy sis that yields hy pothe ses on the s cal spend ing-pol lu tion re la tion ship, we per form a steady-state study with out an a lyt i cally solv ing this model. By set ting K, C and F as pos i tive con stants, and, a unique sta tion ary state ex ists for the sys tem: From the above two steady-state equa tions we de rive two em pir i cally testable hy pothe ses re lated to this study : Hy poth e sis 1 (com po si tion ef fect): Eq. (12) dis plays that phys i cal cap i tal de creases with the pub lic-good s cal spend ing, a pre dic tion that the re al lo ca tion of pri vate-good s cal spend ing to wards ed u ca tion spend ing could fa cil i tate the re struc tur ing of the econ omy in fa vor of cleaner in dus tries. Hy poth e sis 2 (tech nique ef fect): Eq. (13) shows that to tal fac tor pro duc tiv ity S is pre dicted to be in creas ing in the level of s cal spend ing on pub lic goods, a sup port ive sign that R&D spend ing can lead to tech no log i cal im prove ments that might trans late into to a cleaner en vi ron ment. 5 The current conceptual framework proxies education spending and R&D spending with total public spending, and shows its relationship with physical capital and technology in steady-states. We recognize this modi cation of the traditional Solow-Swan model as the rst step to reach our ultimate goal of a more delicate model that explicitly incorporates certain spending components. Future works are anticipated to develop theoretical models that t in better with the research questions and data while keeping the system parsimonious and computationally simple. The other steady-state equation that speci es the relationship between C and K is not of primary interest and is omitted. 5 (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) 3

4 3. Empirical models and data 3.1. Baseline framework When study ing the im pact of an thro pogenic ac tiv i ties on en vi ron ment, a broadly em ployed an a lytic frame work is the IPAT iden tity (Stern, Young, & Druckman, 1992). The IPAT model spec i es that en vi ron men tal qual ity (I) are de ter mined by pop u la tion (P), af u ence (A) and tech nol ogy (T). Since the for mula I = PAT ap pears as an ac count ing equa tion that pre vents data-based hy poth e sis test ing, a sto chas tic ver sion of IPAT named STIR PAT (sto chas tic im pacts by re gres sion on pop u la tion, in u ence and tech nol ogy) is used to em pir i cally es ti mate the im pact of rel e vant so cioe co nomic fac tors (Dietz & Rosa, 1994). The base line STIR PAT model in log a rith mic form can hence be writ ten as: where co ef cients 1, 2 and 3 pos sess the com mon elas tic ity in ter pre ta tion. As noted by Dietz and Rosa (1997) and York, Rosa, and Dietz (2003), the STIR PAT model is quite ex i ble to al ter na tive func tional forms and can be re ned to suit dif fer ent needs. P, A and T can thus be spec i ed as vec tors en com pass ing a rich set of de mo graphic and eco nomic in for ma tion that might raise the ex plana tory power of the model. Built on the panel na ture of our data, we write the base line model in this study as fol lows: Emis sions of two lo cal air pol lu tants, SO 2 and in dus trial soot, are set as de pen dent vari ables. On the right hand side, E and T stand re spec tively for the s cal ed u ca tion spend ing and s cal R&D spend ing, which are two vari ables of in ter est. X rep re sents the set of con trol vari ables com pat i ble with the STIR PAT frame work. Over all gov ern ment size is ac counted for by con trol ling s cal spend 7 ing at ag gre gate level (in clud ing spend ing on both pub lic and pri vate goods). Ur ban pop u la tion and its den sity are added as de mo 8 graphic in for ma tion. Af u ence is rep re sented by per capita GDP, its qua dratic term, and to tal re tail sales of con sumer goods. Eco 9 nomic struc ture is con trolled for by in clud ing the per cent age of GDP gen er ated by the sec ondary sec tor. Due to the paucity of di rect data on en vi ron men tal s cal spend ing, we in tro duce the cov er age rate of ur ban green space to par tially proxy for in ten sity of en vi ron men tal reg u la tion. En tity xed-ef fects and time xed-ef fects are also in cluded Extended speci cations i t Past emis sion of air pol lu tants of one city can af fect its cur rent emis sion, ce teris paribus. For in stance, the his tor i cal tra jec tory of heavy pol lu tion may re lease the sig nal that the city is an ap pro pri ate choice for pol lut ing in dus tries, hence main tain ing and at tract ing pol lut ing rms for the next stage. Vary ing across time pe ri ods and cities, this non-ob serv able au tore gres sive process is un likely to be cap tured by city xed-ef fects or time xed-ef fects, but can be ac counted for by time-lagged terms of emis sion. We hence mod ify the base line model by adding a time lag of the de pen dent vari able 1 ln I it 1 on the right-hand side and mak ing the model dy namic. Air pol lu tion prop a gates and drifts across re gions. Given the like li hood that emis sion of air pol lu tants and the in de pen dent so cioe co nomic fac tors can be spa tially au to cor re lated, we fur ther con sider spec i ca tions that cap ture spa tial de pen dence. We choose the spa tial au tore gres sive (SAR) model and the spa tial Durbin model (SDM), with the for mer cap tur ing en doge nous in ter ac tion ef fects by in clud ing spa tial lags of the de pen dent vari able (W ln I jt) and the later han dling both en doge nous and ex oge nous in ter ac tion 10 ef fects by fur ther adding spa tial lags of in de pen dent vari ables (W ln E jt ; W ln T jt and W ln X jt). The spa tial arrange ment of a city and its neigh bors is char ac ter ized by an in verse-dis tance weight ma trix (IDW) with a cut-o of 300 km. The weight ma trix is set to be iden ti cal across all spa tial mod els, un der the prin ci ple of keep ing the weight ma trix sim ple (Lesage, 2014). We also use a squared IDW with the same cut-o for ro bust ness check. 6 As two conventional production air pollutants (as opposed to consumption air pollutants such as nitrogen dioxide and carbon monoxide), SO 2 and industrial soot are proved to be the major static sources of toxic smog in China, and are widely adopted to measure air pollution before more comprehensive and timely indices are introduced (see Yuan, Mu, & Wang, 2015; Poon, Irene Casas, & He, 2013 and Hu, 2006 for instance). 7 Total spending and the two spendings of interest are not highly correlated in our sample, with both correlation coe cients at the level of Therefore multicollinearity is not a major concern. 8 Total retail sales of consumer goods also captures the income e ect assuming a positive marginal propensity to consume. The city-level annual wage variable in our dataset is discarded since its descriptive statistics deviate substantially from the reality. 9 As indicated by the industrial classi cation codebook released by China's National Bureau of Statistics (GB/ T ), secondary sector includes industries of manufacturing, mining, construction, utility production and supplying (water, electricity, gas and heat). Tertiary sector incorporates industries of circulation, nance, consulting, tourism, technical service, public service, education, medical service and sports. 10 Air pollutants naturally move across geographical units without the presence of socioeconomic factors. This spatial autoregressive process leads us to select the SAR model as our point of departure. The SDM further considers anthropogenic in uences from neighboring units and overcomes the problem of xed ratio between indirect and direct e ects in the SAR setting. Given the nature of our data generating process, we do not consider models that leave out the endogenous dependence such as spatial error model (SEM), spatial error Durbin model (SDEM) and spatial lags of X (SLX) model. 6 (14) (15) 4

5 To es ti mate the dy namic panel mod els, we adopt the sys tem gen er al ized method of mo ments (GMM) pro posed by Blundell and 11 Bond (1998), which cor rects for po ten tial bi ases through in stru ment ing the en doge nous vari ables. Sys tem GMM is proved to be com pu ta tion ally ef cient, re silient to the weak in stru ment prob lem, and pro duces het eroskedas tic ity and au to cor re la tion-con sis tent (HAC) stan dard er rors. For the es ti ma tion of SAR and SDM, we do not de vi ate from sys tem GMM. Kukenova and Monteiro (2009) and Jacobs, Ligthart, and Vrijburg (2009) val i date that es ti ma tors pro duced by Spa tial Max i mum Like li hood Es ti ma tion (MLE), Spa tial Dy namic MLE, Spa tial Dy namic Quasi-MLE and the Arel lano-bond dif fer ence GMM are all out per formed by the sys tem GMM es ti ma tor when deal ing with dy namic spa tial panel mod els. While sys tem GMM takes care of the en do gene ity in the time-lagged terms of emis sion, two iden ti ca tion is sues re main un re solved. First, the processes of hu man-cap i tal ac cu mu la tion and clean-tech nol ogy adop tion can be time con sum ing, hence mak ing the im pact of s cal spend ing on pol lu tion emis sion into a lagged ef fect. Also, the ef fects of s cal spend ing on pol lu tion will no longer be in stan ta neous. We con se quently re place the two spend ing vari ables of in ter est and the to tal ex pen di ture term with their rst-or der lags and three-year mov ing av er age lags, to ac count for po ten tial lagged im pacts. Sec ond, ed u ca tion spend ing and R&D spend ing can 12 re main en doge nous mostly due to the omit ted vari able prob lem. We hence in stru ment these two vari ables with their time-lagged terms and re-es ti mate the dy namic mod els. Two in stru ment ing schemes are pro posed, the rst is to use only the rst-or der lag as sin gle in stru ments, and the sec ond is to use the set of rst-or der, sec ond-or der and third-or der lags as mul ti ple in stru ments, to cap ture rel a tive long-term im pacts Data The bulk of data used in this study are ex tracted from the China City Year book. In ter city dis tance data used to con struct the dis tance-based weight ma tri ces is ob tained from the Stata pack age CHI NA_S PAT DWM that pro vides great cir cle dis tance among Chi nese provinces and cities. The main sam ple of this study in cludes 4 di rect-con trolled cities and 280 pre fec ture-level cities in China 13 from 2003 to We sup ple mented the main sam ple with city-level Air Pol lu tion In dex (API) data pro vided by the Data Cen ter of China's Min istry of En vi ron men tal Pro tec tion (MEP). Table 1 pre sents de scrip tive sta tis tics of our full sam ple. The 284 cities act as ma jor emit ters by dis charg ing 18.2 mil lion tons of SO 2 and 8.5 mil lion tons of in dus trial soot each year on av er age within the stud ied decade, tak ing up 80% and 60%, re spec tively, of the na tion's av er age an nual emis sion of these two pol lu tants. While the sam pled to tal s cal ex pen di ture ac counts for 36% of the na tional level, the share of ed u ca tion spend ing in to tal ex pen di ture is 15% (equal to the na tional av er age), and the frac tion of R&D spend ing is spend ing is 2.8% (lower than the na tional av er age of 4%). The sam ple cov ers mil lion peo ple that makes up 50% of the ur ban pop u la tion and con tributes 48% of the to tal GDP. 4. Empirical results 4.1. Full- sample results Table 2 pre sents the es ti mates of SO 2 emis sion pro duced by dif fer ent model spec i ca tions. Plain OLS re sult in Col umn (1), al beit con t a m i nated by omit ted vari able bias, co in cides with our hy pothe ses that s cal spend ing on ed u ca tion and R&D are neg a tively as so ci ated with SO 2 emis sion. Fixed-ef fects re sult in Col umn (2) re port a weak com po si tion ef fect that shows SO 2 emis sion will re duce merely by 0.6% if s cal spend ing on ed u ca tion in creases by 10%, ce teris paribus. The tech nique ef fect mea sured by the elas tic ity of SO 2 emis sion with re spect to R&D spend ing ap pears to be even weaker ( 0.03), while the sta tis ti cal sig nif i cance van ishes in the xed-ef fects set ting. With time-lag of the in de pen dent vari able added, Col umn (3) show that a 10% in crease in ed u ca tion spend ing and R&D spend ing can lead to a 0.4% and 0.25% de crease in SO 2 emis sion, re spec tively. Over all we nd the com po si tion ef fect and tech nique ef fect are both small in mag ni tude, at least rel a tive to the ex ist ing coun try-level ev i dence. To tal spend ing is shown to have a co ef cient of 0.1 in the dy namic model, in di cat ing a (non-ro bust) pos i tive ef fect. The sign and sig nif i cance of GDP and its qua dratic term re ect the ex is tence of an En vi ron men tal Kuznets Curve (EKC) at city level. Share of green space seems to mit i gate SO 2 emis sion at a small mag ni tude, while past dis charge of SO 2 pos sesses non triv ial ex plana tory power. De mo graphic vari ables and sales of con sumer goods ap pear to be in signif i cant un der xed-ef fects spec i ca tions. Table 3 pre sents re sults of the in dus trial soot equa tions. Re sult of the dy namic panel set ting in Col umn (3) show that ed u ca tion spend ing, time-lag of soot emis sion, and two GDP vari ables are the only ones that pre serve sta tis ti cal sig nif i cance, with signs and mag ni tudes in line with their SO 2 coun ter parts. Pos si ble rea son for the es ti mates of the two air pol lu tants to ap proach each other is that SO 2 and in dus trial soot are largely emit ted si mul ta ne ously from iden ti cal sta tic sources such as ther mal power plants and oil re ner ies. 11 After killing the xed e ect by rst-di erencing the dynamic panel model, the di erenced residual and the di erenced lagged-depend variable are correlated since both of them are a function of the lagged residual (Nickell, 1981). 12 Reverse causation seems improbable to bias our estimates since pollution emission of a certain year is unlikely to a ect scal spending of the same year. Identical to most other countries, the scal spending scheme in China is predetermined by the governmental scal budget and can hardly be changed after being approved by the National People's Congress and the Ministry of Finance (see http: // www. mof. gov. cn for a compete scal budgeting process). 13 Chinese cities can be hierarchically divided into direct-controlled cities (mega city with signi cant political and economic importance), prefecture-level cities (large or median-size city) and county-level cities (small-size city). 5

6 Table 1 De scrip tive sta tis tics. Variable Mean Std. dev Min Max SO 2 emission (metric ton) 64, , , Industrial soot emission (metric ton) 30, , , Total scal spending (million yuan) , , Ed u ca tion spend ing , R&D spend ing , Urban population(million) Population per km (thousand) Per capita GDP (yuan) 39, , , Sales of consumer goods (million yuan) 25, , , % GDP from secondary sector % GDP from tertiary sector Green space coverage (%) Note: R&D spending incorporates scal expenditures on basic research, applied research and new product development. Population is measured by the number of registered urban residents (residents with urban Hukou) within municipal districts at the end of the year. Urban land area used to calculate population density is measured by urban built-up area which incorporates non-agricultural urban area and infrastructured suburban area. We leave out the non-built-up part of the prefecture to con ne our analysis within well-infrastructured and economically active urban areas, where most human-capital intensive activities and technical upgrading take place. Table 2 Re sults of SO 2 emis sion. (1) (2) (3) Education spending [0.061] [ 0.006] [0.004] R&D spending [0.028] [0.014] [0.014] Total spending [0.076] [0.095] [0.041] Population [0.054] [0.125] [0.133] Population density [0.023] [0.089] [0.049] GDP [0.339] [0.560] [0.461] 2 GDP [0.011] [0.018] [0.015] Sales [0.054] [0.039] [0.040] %GDP from industry [0.150] [0.131] [0.193] %Green space [0.052] [0.004] [0.006] Lagged SO [0.014] City xed-e ects No Yes Yes Year xed-e ects No Yes Yes Constant [2.631] [3.839] [3.228] 2 Adjusted R Observations Note: Variables are in logarithmic form except for the year dummies included in all models. Model (3) passes the Wald test for overall t, the AR(2) test for serial correlation, and the Hansen test for over-identi cation. Robust standard errors clustered at city level are in brackets Regional sub- sample results Ini ti ated in the late 1970s, the re form of s cal de cen tral iza tion in China en dowed lo cal gov ern ments with au ton omy on nan cial bud get ing. Given the long-ex ist ing so cioe co nomic di vide among east ern, cen tral and west ern China, un ob served re gion-spe ci c fac tors such as re gional pol icy of the cen tral gov ern men tal, so ci etal tol er ance to wards pol lu tion and lo cal bu reau cratic ef ciency can lead to dif fer ent s cal arrange ments. We split the sam ple by east ern (116 cities), cen tral (108 cities) and west ern (60 cities) area to ac count for po ten tial re gional het ero geneities. Figs. 1 and 2 show that east ern re gion ranks the high est in the level of s cal spend ing on ed u ca tion. Cen tral area, al beit close to the level of west ern re gion, ranks the low est. With a paucity of teach ing fa cil i ties, el i gi ble 6

7 Table 3 Re sults of in dus trial soot emis sion. (1) (2) (3) Education spending [0.065] [0.074] [0.019] R&D spending [0.029] [0.017] [0.131] Total spending [0.075] [0.081] [0.051] Population [0.058] [0.129] [0.159] Population density [0.026] [0.109] [0.053] GDP [0.362] [0.204] [0.508] 2 GDP [0.012] [0.008] [0.017] Sales [0.057] [0.071] [0.156] %GDP from industry [0.221] [0.210] [0.098] %Green space [0.061] [0.039] [0.015] Lagged soot [0.052] City xed-e ects No Yes Yes Year xed-e ects No Yes Yes Constant [2.779] [4.647] [3.917] 2 Adjusted R Observations Note: Variables are in logarithmic form except for the year dummies included in all models. Model (3) passes the Wald test for overall t, the AR(2) test for serial correlation, and the Hansen test for over-identi cation. Robust standard errors clustered at city level are in brackets. Fig. 1. Fis cal spend ing on ed u ca tion by re gion (mil lion yuan). teach ers and school man age r ial per son nel, west ern re gion has the high est share of spend ing on ed u ca tion in to tal gov ern ment ex pen di ture un der the guid ance and sup port of the cen tral gov ern ment. In terms of the R&D spend ing, Figs. 3 and 4 il lus trate that east ern re gion ranks the rst in both level and share. While only slight share-dif fer ences ex ist among the three re gions, the coastal-in land gap re gard ing lev els of R&D spend ing ap pears to be sig nif i cant and per sis tent. Table 4 pre sents the sub-sam ple re sults based on sys tem GMM es ti ma tion. For the east ern re gion a 10% in crease in ed u ca tion spend ing re duces SO 2 emis sion by 1.9% and soot emis sion by 1.4%. These es ti mates in di cate the ex is tence of a rel a tively larger com po si tion ef fect in this area com pared to the na tional av er age. As the pre cur sor of the Re form and Open-up in late 1970s, it is ev i dent that among the three re gions, the east ern part has the high est con cen tra tion of uni ver si ties, the best busi ness en vi ron ment for knowl edge-based start-ups, and the largest mar ket for hu man cap i tal-in ten sive prod ucts. Com po si tion ef fect tends to be aug mented and self-re in forc ing in cities since hu man cap i tal re ceives ex tra re turns from ag glom er a tion economies through la bor mar ket match ing and knowl edge spillover. Hence we ex pect the as so ci ated co ef cient to be smaller in ab solute value if county-level data are stud 7

8 Fig. 2. Share of ed u ca tion spend ing in to tal spend ing by re gion. Fig. 3. Fis cal spend ing on R&D by re gion (mil lion yuan). Fig. 4. Share of R&D spend ing in to tal spend ing by re gion. ied. Spend ing on R&D dis plays a slightly in creased tech nique ef fect of 0.07 than the full-sam ple re sult for SO 2 emis sion, while es ti mate of soot emis sion losses sta tis ti cal sig nif i cance. Un der ly ing forces that im pede the trans for ma tion from re search fund ing to the clean-tech nol ogy adop tion re main strong even in the more de vel oped east ern area. For the cen tral part of ur ban China, com po si tion ef fect is op er at ing only in abat ing in dus trial soot emis sion, at al most half the level of its east ern neigh bor. With most of the re search uni ver si ties lo cated in only a few provin cial cap i tal cities, the area still lacks the abil ity to fos ter cre ative class and in no v a tive in dus tries. Since eco nomic growth has long been re liant on tra di tional la bor-in ten sive or re source-based man u fac tur ing, it is un likely that hu man cap i tal-in ten sive rms can ef fec tively sub sti tute for pol lut ing rms, at least in the short run. Ex ist ing eco nomic struc ture along with less strin gent en vi ron men tal reg u la tion fur ther low ers the adop tion rate of clean tech nol ogy, which is re ected in the ig nor able tech nique ef fects. 8

9 Table 4 Re sults of SO 2 and in dus trial soot emis sion by re gion. Eastern Central Western SO2 Soot SO2 Soot SO2 Soot Education spending [0.056] [0.069] [0.056] [0.033] [0.201] [0.938] R&D spending [0.016] [0.019] [0.089] [0.004] [0.049] [0.374] Total spending [0.174] [0.056] [0.393] [0.062] [0.235] [0.962] Population [0.043] [0.074] [0.028] [0.077] [0.095] [0.576] Population Density [0.047] [0.475] [0.029] [0.047] [0.048] [0.645] GDP [0.604] [0.547] [0.118] [0.218] [0.934] [0.439] 2 GDP [0.203] [0.183] [0.246] [0.233] [0.382] [0.142] Sales [ 0.017] [0.047] [0.269] [0.041] [0.370] [0.749] %GDP from industry [0.032] [0.039] [0.144] [0.207] [0.174] [0.029] %Green Space [0.038] [0.046] [0.065] [0.049] [0.217] [0.101] Lagged SO 2/Soot [0.006] [0.009] [0.063] [0.030] [0.241] [0.016] Constant [4.538] [4.519] [2.481] [5.221] [9.186] [4.178] Observations Note: City xed-e ects and time xed-e ects are included in all models. Variables are in logarithmic form. All models pass the Wald test for overall t, the AR(2) test for serial correlation, and the Hansen test for over-identi cation. Robust standard errors clustered at city level are in brackets. The ca pac ity for west ern cities to pro vide higher ed u ca tion is the weak est among three re gions. Eco nomic growth dom i nates en vi ron men tal pro tec tion in this area and eco nomic re struc tur ing to wards hu man-cap i tal in ten sive in dus try might take decades to achieve. The re gion is not a fa vored place for high-tech in vest ments, in spite of po lit i cal en cour age ments. While we fail to iden tify any com po si tion or tech nique ef fects given the in signif i cant re sults with counter-in tu itive signs and mag ni tudes, it should be noted that air pol lu tion is not a triv ial prob lem in this area. Based on our data, the SO 2 emis sion is 13.5% higher than that of the cen tral area and only 7% lower than level in the east ern re gion. In dus trial soot emis sion ap pears to be sig nif i cantly lower com pared to the east ern area, but only 14% less than that of the cen tral area Results from spatial models Un like non-spa tial spec i ca tions, point es ti mates of SAR and SDM model do not pos sess a par tial de riv a tive in ter pre ta tion that re ects the true im pact of spa tially-au to cor re lated vari ables. Lesage and Pace (2009) pro poses two sum mary in di ca tors for the mea sure ment of di rect ef fect and in di rect ef fect to gen er ate more sur veyable re sults. Spa tial lags that ac count for po ten tial en doge nous and ex oge nous in ter ac tion ef fects are taken as con trols, while we fo cus on the vari ables of in ter est and re port only the di rect (mar ginal) ef fects. Table 5 shows that the re sults are ro bust when com pared with those we achieved from non-spa tial set tings. Fee ble com po si tion ef fects are iden ti ed across two air pol lu tants, and the mag ni tudes of R&D spend ing does not seem to in di cate any plau si ble tech nique ef fect. The nat ural drift ing process of the SO 2 is cap tured by the en doge nous lags with sta tis ti cal sig nif i cance at 1%, while we do not ob serve a sim i lar spillover ef fect for soot emis sion. These nd ings re main ro bust when we adopt an other spa tial 14 weight ma trix known as the squared in verse-dis tance weight ma trix. 14 Under the squared IDW arrangement, spatial dependence decays at a much faster rate than under the linear IDW. We also run regional sub-sample regressions for SDM and nd the results not deviating from what is presented in Table 4. 9

10 Table 5 Di rect-ef fect re sults of SAR and SDM. SO2 SAR SDM SAR SDM Education spending [0.012] [0.015] [0.013] [0.017] R&D spending [0.017] [0.020] [0.004] [0.003] Total spending [0.072] [0.077] [0.044] [0.050] Population [0.117] [0.130] [0.401] [0.433] Population density [0.102] [0.093] [0.072] [0.004] GDP [0.744] [0.751] [0.401] [0.333] 2 GDP [0.020] [0.013] [0.014] [0.011] Sales [0.039] [0.022] [0.105] [0.110] %GDP from industry [0.110] [0.059] [0.021] [0.030] %Green space [0.004] [0.004] [0.045] [0.049] Lagged SO 2/Soot [0.014] [0.011] [0.043] [0.049] W* SO 2/Soot [0.032] [0.071] [0.133] [0.029] Constant [3.043] 8.48 [2.564] [3.322] [3.110] Observations Note: City xed-e ects and time xed-e ects are included in all models. Variables are in logarithmic form. Estimates of the independent spatial lags in SDM models 2 are suppressed for brevity, with W*GDP, W*GDP and W*(Population density) signi cant at 5% for both pollutants. All models pass the Wald test for overall t, the AR(2) test for serial correlation, and the Hansen test for over-identi cation. Robust standard errors clustered at city level are in brackets Results using the Air Pollution Index The uni di men sional mea sure ment of air pol lu tion based on SO 2 and in dus trial soot leaves out the par ti cle pol lu tion gen er ated by 15 par tic u late mat ter (PM). While di rect panel data on city-level PM 10 and PM 2.5 are un avail able, the Data Cen ter of China's Min istry of En vi ron men tal Pro tec tion (MEP) recorded an Air Pol lu tion In dex (API) that mea sures air qual ity for ma jor Chi nese cities from to As only a lim ited frac tion of the cities have con sec u tive data from 2003 to 2012, we choose 26 provin cial cap i tal cities (ex clud ing Lhasa), 4 di rect-con trolled cities (Bei jing, Tian jin, Shang hai and Chongqing), and 5 cities specif i cally des ig nated in the state plan (Dalian, Ningbo, Qing dao, Xi a men and Shen zhen) to form an API sam ple. This sam ple is not meant to rep re sent the en tire pop u la tion of Chi nese cities, but to pro vide the strongest pos si ble ev i dence of com po si tion ef fect and tech nique ef fect, given the so cioe co nomic en dow ments and sta tus of these re gional cen tral cities. Table 6 re ports the full-sam ple re sults with our de pen dent vari able switched to the API of 35 ma jor cities. The two sta tis ti cally sig nif i cant es ti mates of ed u ca tion spend ing ( & 0.124) are in deed larger rel a tive to their main-sam ple coun ter parts, and the es ti mate of R&D spend ing un der sta tic xed-ef fects set ting ( 0.064) also in creased com pared to the SO 2 and soot re sults. The over all mag ni tude of the es ti mates sup ports our nd ing of weak com po si tion ef fect and tech nique ef fect, with the for mer re main slightly stronger than the later. 15 The U.S. Environmental Protection Agency de nes PM as a complex mixture of extremely small particles and liquid droplets that get into the air. Conventional PM includes inhalable particles with diameters equal to or below 10 m (known as PM 10), and ne inhalable particles with diameters equal to or below 2.5 m (known as PM 2.5). 16 The API is comprehensively calculated using SO 2, NO 2 and PM 10, based on the 1996 version of Ambient Air Quality Standard (GB ). Data Center of MEP reported API on a daily basis. We calculate the annual API by taking average of the monthly API, with monthly API calculated by taking average of daily API on three dates within one month: the 1st, the 15th and the 30th (28th/ 29th for February). In 2013, MEP replaced API with the Air Quality Index (AQI) by adding O 3, CO and PM 2.5 to the pollutant basket, based on the 2012 version of Ambient Air Quality Standard (GB ). The AQI has been recognized to outperform API with more stringent standard and higher monitoring frequency. We anticipate our results to be updated using AQI in future works. Soot 10

11 Table 6 Re sults of Air Pol lu tion In dex. (1) (2) (3) Education spending [0.202] [0.016] [0.012] R&D spending [0.305] [0.011] [0.079] Total spending [0.114] [0.126] [0.388] Population [0.135] [0.352] [0.580] Population density [0.135] [0.139] [0.205] GDP [1.232] [1.108] [2.126] 2 GDP [0.567] [0.049] [0.080] Sales [0.091] [0.202] [0.438] %GDP from industry [0.125] [0.098] [0.104] %Green space [0.112] [0.145] [0.151] Lagged API [0.103] Constant [6.754] [6.590] [1.366] 2 Adjusted R Observations Note: Variables are in logarithmic form except for the year dummies included in all models. Model (3) passes the Wald test for overall t, the AR(2) test for serial correlation, and the Hansen test for over-identi cation. Robust standard errors clustered at the city level are in brackets Results using lagged covariates and IVs The im pact of s cal spend ing on pol lu tion emis sion can be lagged if we ex am ine the chan nels of hu man-cap i tal ac cu mu la tion and tech no log i cal change. We adopt the rst-or der lags and three-year mov ing av er age lags of the spend ing vari ables to ac count for po ten tial lagged im pacts. We choose not to go fur ther with longer lags at the cost of sys tem atic sam ple at tri tion. Table 7 pre sents the re sults. While the over all sizes of co ef cients are not sub stan tially al tered, it can be seen that the ed u ca tion spend ing es ti mates re main to dom i nate the R&D spend ing es ti mates in terms of both mag ni tudes and sta tis ti cal sig nif i cance. The ef fects of ed u ca tion spend ing across all the three pol lu tion mea sure ments are also much stronger com pared to their non-lagged coun ter parts pre sented in Tables 2, 3 and 6, which val i dates the adop tion of lagged terms. Table 7 Re sults us ing lagged spend ing terms as re gres sors. SO2 Soot API 1-year 3-years 1-year 3-years 1-year 3-years Education spending [0.002] [0.069] [0.037] [0.032] [0.031] [0.038] R&D spending [0.019] [0.041] [0.054] [0.024] [0.042] [0.070] Total spending [0.102] [0.036] [0.155] [0.041] [0.331] [0.580] Lagged DV [0.185] [0.139] [0.079] [0.044] [0.140] [0.221] Observations Note: City xed-e ects and time xed-e ects are included in all models. Variables are in logarithmic form. All models pass the Wald test for overall t, the AR(2) test for serial correlation, and the Hansen test for over-identi cation. Estimates of non-lagged terms are suppressed for brevity. Robust standard errors clustered at city level are in brackets. 11

12 Table 8 pro ceeds to re port the re sults un der two in stru ment ing schemes. The sin gle-iv es ti mates and the mul ti ple-iv es ti mates do not seem to dif fer sub stan tially from each other. Iden ti cal to the re sults in Table 7, the strongest com po si tion ef fect is iden ti ed in the API mod els, while R&D spend ing re main no (does not make sense-per haps you mean ex hibits no sig nif i cant ef fect ) sig nif i cant ef fect on air pol lu tion. A joint ex am i na tion of Tables 7 and 8 sup ports our pre dic tion that the API sam ple of re gional cen tral cities would yield the high est pos si ble es ti mates of in ter est. 5. Discussion of mechanisms 5.1. Composition e ect The over all nd ing of the triv ial com po si tion ef fect in di cates that s cal spend ing on ed u ca tion does not trans late into suf cient pro-en vi ron men tal hu man-cap i tal in ten sive ac tiv i ties in ur ban China. While the in ten sity of hu man-cap i tal and its ac cu mu lat ing process are oner ous to mea sure, eco nomic re struc tur ing to wards the cleaner sec tor pro vides a tractable path for test ing the ef fect of hu man-cap i tal ac cu mu la tion. We hence regress two vari ables of ur ban eco nomic struc ture (per cent age of GDP gen er ated by the sec ondary and ter tiary sec tor) on ed u ca tion spend ing, to ver ify if such struc tural change took place within the stud ied pe riod. Table 9 pre sents the es ti mates of the dy namic mod els pro duced by sys tem-gmm. The only sta tis ti cally sig nif i cant main ef fect of in ter est arises in the ter tiary sec tor equa tion of cen tral re gion with counter-in tu itive di rec tion. The re sults re main largely un changed when we re place the spend ing terms with their 1-year and 3-year lags. It seems safe to con clude that ed u ca tion spend ing is not dri ving the vari a tion in ur ban eco nomic struc ture. Given the back ground that China re mains at the stage of up grad ing its in dus trial struc ture from la bor-in ten sive to phys i cal cap i tal-in ten sive, it is not im plau si ble that ed u ca tion spend ing fails to trig ger per va sive sub sti tu tion be tween phys i cal cap i tal-in ten sive in dus try and hu man cap i tal-in ten sive in dus try in cities. From the long-term per spec tive, the iden ti ed com po si tion ef fect could be un der es ti mated in our 10-year sam ple, since it might take much longer than a decade for ed u ca tion spend ing to af fect eco nomic struc ture through fos ter ing and in spir ing hu man-cap i tal in ten sive ac tiv i ties. At the mi cro level, spend ing on ed u ca tion might sup port an av er age ur ban worker ob tain ing a cer tain amount of higher ed u ca tion, but ap pears to be in suf cient to pro duce ad e quate num bers of in no v a tive en tre pre neurs and knowl edge-based rms that re shape the econ omy. The of cial 2016 China Higher Ed u ca tion Qual ity Re port is sued by the Min istry of Ed u ca tion an nounces that China owns the largest num ber of col lege stu dents (over 36 mil lion) and the sec ond largest num ber of uni ver si ties (2852) on the globe, with a 40% gross en roll ment rate. Nev er the less, the re port em pha sizes that the qual ity of Chi na's higher ed u ca tion re mains to be greatly im proved as it cur rently suf fers from the paucity of highly-cre ative fac ul ties and well-de signed train ing pro grams for col 17 lege stu dents. Suf cient amount of ed u ca tion spend ing, if not mis al lo cated, should di rectly con tribute to the hu man-cap i tal in ten sity of the econ omy by of fer ing stu dents with more cre ativ ity-based train ing, and in di rectly by im prov ing the in no v a tive ness of col lege teach ers. An al ter na tive de mand-side mech a nism pre dicts that in creased hu man-cap i tal ac cu mu la tion leads ur ban cit i zens to pay more at 18 ten tion to the non-pe cu niary as pect of their liv ing con di tions, re sult ing in stricter en force ment of en vi ron men tal reg u la tions. With out cred i ble data on city-level en vi ron men tal reg u la tion, the chan nel-block ing fac tor might come from the slug gish ness of lo cal gov ern ments in re sponse to the de mand shift. The im ple men ta tion of di rect en vi ron men tal reg u la tion and the Pigou vian tax scheme both re quire gov ern ment to ac quire ac cu rate knowl edge con cern ing the mar ginal dam age of pol lu tion and the mar ginal ben e t of con trol ling, which leads to a de ferred gov ern men tal re ac tion. This in for ma tion dis ad van tage can be fur ther am pli ed by in ef cient en vi ron men tal pol icy-mak ing processes, which ap pears to be per va sive among China's lo cal gov ern ments Technique e ect Anal o gous to the ef fect of ed u ca tion spend ing, R&D spend ing fails to make Chi nese cities cleaner, in di cat ing that rms are not re spond ing to R&D spend ing by adopt ing pro-en vi ron men tal tech nolo gies. We pro vide a tri-tier l ter in ter pre ta tion that as cribes the in ef fec tive ness of R&D spend ing to the com bi na tion of low out put-in put ra tio, low con ver sion rate from re search out come to ap plic a ble tech nol ogy, and low adop tion rate of clean tech nol ogy by pol lut ing rms. We ar gue that with the ex is tence of these l ter screens, only mi nor clean-tech nol ogy adop tion can be dri ven by R&D spend ing. 17 Data from the Ministry of Education shows that in 2015, only 18.7% of the full-time college teachers possess a Ph.D. degree, while 44.4% hold a bachelor's degree as their highest educational attainment. 18 It should be noted that we are not discussing the demand-side income e ect here, while income variables have been (partially) controlled for. The demand for a cleaner environment can be generated by education per se even without income variation. Using data from 14,000 Chinese households, a recent study based on the contingent valuation method (CVM) con rms that higher educational attainment is directly associated with signi cantly higher willingness-to-pay in terms of air pollution abatement (Zhang, Zhang, & Chen, 2017). 12