The Value of Centralization: Evidence from a Political Hierarchy Reform in China

The Value of Centralization: Evidence from a Political Hierarchy Reform in China Shiyu Bo This Version: July 12, 2015 Abstract Is regional development better under a reform that centralizes governments? This paper exploits an experiment in political hierarchy in China from the 1980s to investigate the effects of more centralized government institutions on urbanization and industrialization. To base my estimates on plausibly exogenous variation in the reform I conduct an event study to rule out the possibility of differential prereform trends. Empirical results reveal that government centralization improves the development of regions as a whole. But it will also produce distributional effects for different types of belonging counties. The counties which are capitals of their regions benefit significantly more relatively to other counties. To interpret the empirical results, I propose a theoretical framework arguing that gains from centralization are from internalizing spill-overs within a region and equalizing marginal revenues from government expenditures across counties. I use a firm-level dataset to validate its micro foundations. JEL Code: O18 R11 P25 H73 Department of Economics and STICERD, London School of Economics and Political Science. Email: s.bo@lse.ac.uk. I am grateful to my supervisor Robin Burgess for his guidance. I benefit from useful suggestions and comments from Gharad Bryan, Abhishek Chakravarty, Xiaoguang Shawn Chen, J. Vernon Henderson, Jason Garred, Yu-Hsiang Lei, Yatang Lin, Guy Michaels, Michael Olabisi, Rui Zhang and participants at various seminars. I also thank Cong Liu for her continued support. It is very preliminary and incomplete. Comments are all welcomed. All the errors are my own. 1

1 Introduction How centralized should a government be? This question has long been of interest to economists and policymakers. In the last three decades, decentralization experiments have been carried on the developing world, especially in many transition economies in Asia and Latin America (World Bank, 2000). There is a large theoretical body of literature proposing various stories on trade-offs behind centralized and decentralized institutions. The classical approach formalized by Oates (1972) assumes centralization can internalized spillovers across districts but accompanied uniformity will produce inefficiency under heterogeneous tastes. Recent works emphasized more on political process, such as conflicts of interests (Besley and Coate, 1998) and diminish accountability (Seabright, 1996). Empirically, however, quantitative evaluations on causal effects of (de)centralization are in a lack and existing evidence is mixed (Bardhan, 2002). This paper revisits such problem in a novel setting. I explore effects of a political hierarchy reform first launched in 1983 in China, after which political decisions were taken at a higher regional level instead of a lower local level. Figure 1 depicts such changes over time. I utilize variations in centralization brought by this reform to examine its outcomes in development, mainly including industrialization and urbanization. Such setting makes two contributions in the empirical literature of (de)centralization. First, most previous literature concentrated on national-provincial relations, referred to fiscal federalism (Qian and Weingast, 1997). For example, China s miracle since 1980s has been attributed a lot to the decentralization from the central government to provincial governments (see a survey by Xu (2011)). Unlike them, I focus on evolutions of political powers between two sub-provincial governments. They represent governance entities in China but are left as black boxes in the literature without thorough analysis neither theoretically nor empirically. lack evaluations on their power allocation. Second, previous settings care provisions of public goods most. But in developing countries, roles of governments are relatively more straightforward to foster economic development and growth. I investigate overall economic performance directly here, instead of common outcomes in public goods. [Figure 1 about here.] To guide my empirical analysis, I present a two-county spatial equilibrium model. Before the centralization reform, counties can make decisions on their own tax collections and subsidies to foster local urbanization and industrialization. After the reform, the municipality government will manage the developments in both counties. On one hand, the municipality government can internalize spillovers to improve the outcomes in both counties; on the other hand, it will equalize marginal revenues of subsidies, which may produce differential outcomes in two counties. I show that as a whole the municipality will benefit in this centralization reform, while at the same time there will be distributional effects on two belonging counties. To evaluate the impact of the centralization reform on aggregate regional development, I collect and digitize prefecture-level and county-level data on urbanization 2

and industrialization from 1983 to 2003 from various sources. Results show that capital counties where municipality governments locate benefited from the reform in which the governance was centralized into municipalities, while peripheral counties that lost decision rights were hurt. The overall effects on the whole region will enjoy improvements. Such heterogeneous response to this reform aggravate the disparities between capital counties and peripheral counties. Along with the results on overall economic performance, I also find that the reform was associated with enlarges in disparities in government sizes and expenditures between capital and peripheral counties. To validate the micro foundation, I document some patterns of industrial enterprises using a firm-level dataset. First, I show agglomeration can produce spill-overs on firms productivity within and across counties. Second, after the reform, resource misallocation within a region is mitigated, measured by a reduce in dispersions of revenue productivity as in Hsieh and Klenow (2009). Third, industrial sectors become more concentrated under the more centralized institution. Firms in counties with higher sector-specific productivity will produce more after the reform. A key insight from my paper is possible gains in efficiency from centralization. A common wisdom from classical works such as Oates (1972) is that the centralized system helps to internalize spill-overs across boundaries. In this paper, I show the existence of such externalities across different counties. After the reform, the externalities can be internalize under a centralized governance. That is the first source of values from centralization. Moreover, I also propose a novel channel from which centralization may bring benefits. In a decentralized institution, county governments collect tax and subsidize industrial development locally. When input factors are imperfect mobile as in China, marginal revenues brought by government spendings could be different across counties, which may produce resource misallocation within a region. A notorious consequence of such over decentralization is the duplication of industries across counties and the distortion of production away from patterns of comparative advantage as Young (2000) documented. However, when the power of governance is centralized at a higher municipality level, the municipality government can manage tax from counties and allocate them efficiently, then marginal revenues across counties can be equalized. A useful tool to show this pattern is to examine dispersions of Hsieh and Klenow (2009) s revenue productivity (TFPR). To illustrate the intuition, I plots distributions of TF- PR in a municipality before and after the centralization reform in Figure 2. Finally, the most common reason to prefer decentralization than centralization is better accountability in local governments since they will compete with each other. On the contrary, in the classical works, centralization is always assumed to happen in central governments, who act as monopolists and sources of accountability problem. But in this paper the centralization happen in a regional level (prefecture/municipality), not the central government. Prefecture/municipality government leaders also have to compete for promotion as county leaders. So in this arrangement centralization in such regional level can avoid traditional accountability problem. [Figure 2 about here.] 3

This paper also contributes to literature in urbanization and regional development. Urban economists have emphasized the importance of political institutions in determination of urbanization and urban primacy (Henderson and Becker, 2000; Henderson and Wang, 2007; Ades and Glaeser, 1995). This paper will contribute to offer a sub-national causal evidence on those topics. 2 Background 2.1 Political Hierarchy and Jurisdictions in China After 1978 In China, there are four levels of sub-national jurisdictions: province level, prefecture or municipality level, county level and township level. Among these four levels, township governments do not own enough decision rights on most economic affairs; their main function is to have policies conveyed to them by upper-level governments (Xu, 2003). Meanwhile, mainly due to large sizes of provinces, 1 it is difficult for provincial governments to administer local affairs heterogeneously; they have to rely on prefecture or municipality government to enact policies locally. As a result they are essential to the political and economic processes in China. This paper will concentrate on relations between them. The basic political hierarchy in China is sketched in Figure 3. [Figure 3 about here.] A key difference between a prefecture and a municipality is their relations with counties belonging to them. In a prefecture, all counties under it are almost autonomous and the county governments have decision rights on the development of their own counties. The prefecture government has no administration power on counties, since it is only a province government s resident agency in this prefecture s territory. Legally, leaders in the prefecture government cannot enact economic policies on their own; instead, they can only convey decisions by the province government to their belonged counties and oversee them. Moreover, as provinces in China are too large, many policy choices are left to lower levels of counties. As a result, counties under any prefecture can make decisions on their own affairs independently, including economic and political fields. For example, they can decide where public expenditures will be put into and what kinds of public goods will be offerer; local taxes and subsidies; locations of newly opened state-owned enterprises etc. To some extent, in this setting counties are of considerable autonomy. On the contrary, in a municipality, county governments lose the decision rights. The municipality government located in the capital county can administer the development of all its counties. The reason of for this difference is that the municipality government is not a province government s resident agency. Legally, counties belonging to it are under supervision of the municipality government itself, not of the province government. Under this situation, county governments will be manipulated by the municipality government: their leaders are nominated by the municipality government; their tax and public finance related affairs are not only supervised but also managed; they are not 1 The average size of a province in China is about 300,000 km 2, similar as the size of Italy. 4

allowed to set up new state-owned enterprises freely and such location choice will be coordinated by the municipality government. In general, counties in municipalities are not as autonomous as in prefectures. The municipality officials can give heterogeneous orders to specific counties. In Figure 3, the link between the prefecture government and its counties are denoted by dashed lines to indicate the fact that the prefecture government cannot administrate its counties; in contrast, the link between the municipality government and its counties are denoted by solid lines to indicate the fact that the prefecture government s actual power on its counties. Historically, prefectures were the main institutions above counties. For thousands of years, counties were the basic units of China s local entities. The jurisdiction above counties did not own many administration powers on them. No matter what its name was in any specific period, the jurisdiction above counties played a very similar role as prefecture does in modern China. Such tradition was challenged after the founding of the People s Republic of China in 1949. The main reason is the rise of cities. Cities in China emerged naturally in the process of modernization. These jurisdictions with large urban areas were named municipalities placed at the same level as prefectures in the political hierarchy system. After mess situations in the Cultural Revolution, the number of municipality and its percentage in jurisdictions of the same political level remained stable until the 1983 reform as Figure 1 shows. 2.2 The 1983 Reform: Turning prefectures into municipalities The reform starting from 1983 aimed to reform prefectures and turn them into municipalities. Primary goal of the reform was to accelerate urbanization in counties by utilizing industrial resources in municipalities to help peripheral counties whose territories were mostly rural areas. As supposed by the central government, after a prefecture was changed into a municipality, it would form a core-peripheral setting, with the more urbanized capital county in the core of the municipality and less developed peripheral counties around it. The central government expected this municipality setting to increase spillovers from capital counties to peripheral counties under the coordination of municipality governments. The actual process of the 1983 reform can be found in Figure 1. From 1983 to 1993, the number of municipality increase rapidly and the number of prefecture decrease in a similar pattern. Until 2003, except for a few special minority residences, almost all prefectures have turned to municipalities. Moreover, as the pattern of the number of total prefecture or municipality level jurisdiction suggests, such reform is an one-on-one transition between a municipality and a prefecture. Figure 4 presents an example of the reform. [Figure 4 about here.] I am going to make use of variations of this reform across space and timing to explore the heterogeneous development outcomes brought by centralization of governance. 5

As described above, under a prefecture, the powers of governance are decentralized to county governments and all counties behave independently; under a municipality, the powers are centralized to municipality government. The Turning prefectures into municipalities reform offers a great source of variations in the extent of centralization. Figure 5 depicts the variation of the reform in my sample. [Figure 5 about here.] 3 Conceptual Framework 3.1 Production The production side of this model is mainly adopted from the handbook chapter by Duranton and Puga (2004). In their work, Duranton and Puga model different microfoundations of urban agglomeration economies. Here I take their first source of agglomeration: sharing the gains from variety. I will focus on a two-region setting. In each region, there are two sectors, agriculture and manufacture. The agriculture sector employ labour only under constant return to scale. In the manufacturing sector, perfectly competitive firms produce goods for final consumption using intermediate goods following a constant elasticity substitution (CES) technology. function of a representative final-good producer j in urban area of county i is Y ji = AL 1 α ji ni 0 The production [x h ] α dh (3.1) where n denotes the total number of intermediate goods used and x h denotes the amount of intermediate h. Duranton and Puga (2004) follow Ethier (1982) to assume that intermediate inputs are produced by monopolistically competitive firms. It enables us to get many classical results as Dixit and Stiglitz (1977) have done. Factor demands by final good producer is: x h (P h ) = [ αal1 α P h ] 1/(1 α) (3.2) Following Dixit and Stiglitz (1977), the profit-maximizing price P h is a fixed markup over marginal cost: Plugging back to the production function yields P h = 1/α (3.3) Y ji = A 2 α 1 α α 2 1 α ni L ji (3.4) As for the number of intermediate varieties n, I assume it will depend on the total number of urban workers. It suggests that human capital in the whole urban area will contribute to innovations of new varieties (Marshall, 1890). For example, some faceto-face learnings or everyday skill acquisition may be easier in cities. Such idea of knowledge spillovers is developed by Romer (1990) and Grossman and Helpman (1991) 6

to sustain a dynamic economy s long-run growth. But in this paper I apply a static setting for simplicity. whole country (n N ): I assume available n i is derived from existing varieties in the n i = Lθ i Lφ i i L θ+φ n i N (3.5) N where θ (0, 1), φ A (0, 1), φ B = 0. Substituting the quantity of intermediate goods x(h) and the number of intermediate varieties n into the production function of final goods yields aggregate production in manufacture section as Y i = A 2 α 1 α α 2 1 α n N L θ+φ i N L φ i i L1+θ i (3.6) We can observe an aggregate increasing returns to scale in the manufacturing sector. It is noteworthy that in China, most innovations in the manufacturing sector were not born in private enterprises until the twenty-first century, so it is not an impossible task for firms to access and imitate the other innovations without patent protections. 3.2 Roles of Local Governments In this part I am going to model behaviours of local governments and their roles in the urbanization and industrialization. Since the production of final goods depends on the amount of labour in the manufacturing sector, it is necessary to consider determinants of labour supply. I assume that all workers employed in the manufacturing sector live in cities, and peasants in the agriculture sector live in rural areas. There are many ways to model a sustainable monocentric city. A key part in these urban models is to limit sizes of cities which is driven to infinity by the economies of scale as above. I assume that motivation of local governments is to increase manufacturing production for promotion, while facing a cost of moving peasant to cities, financed by local tax: U = f(y ) c(t ). A function M(L) measures the cost to sustain L workers in urban areas. I assume that f = log( ), c is convex to the tax collected locally and M is linear. Firstly, I will discuss the situation before the 1983 reform. Suppose that in a prefecture there is a capital county A and a peripheral county B. Before the 1983 Turning prefectures into municipalities reform, decision powers were decentralized to county governments. So both A and B can decide their optimal level of workers L i, i {A, B}, to maximize the production in the manufacturing sector Y i for their promotions. At this setting, each county can only make use of the tax collected locally. It is obvious to see in above setting that when the capital county attracts more city residences from its rural areas, it will produce a positive externalities to the peripheral county due to an expansion in varieties of intermediate goods in the whole economy. However, in this decentralized case before the 1983 reform, the local government cannot collude to internalize this source of positive externality. So each government i will choose its own optimal effort level ˆL i taking the other s efforts as given and they will reach a Nash equilibrium {ˆL A, ˆL B } in the end (Besley and Coate, 1998). To summarize, the 7

maximization problem faced by the local governments can be written as: max L i U i (L i ) = (1 + θ)log(l i ) + φ i log(l i ) c(t i ) (3.7) where facing the constraint T i = M(L i ). Now I turn to the situation after the reform. In this case, a prefecture was turned into a municipality and counties belonging to it lost their decision rights. The municipality government is in charge of its counties now. It can choose the optimal level of {L A, L B } at the same time to maximize its utility. At the same time, the municipality government can determine how much spent on each county. The municipality government s maximization problem is: max U(L A, L B ) = log(y A ) + log(y B ) c(t A ) c(t B ) (3.8) L A,L B with the constraint T A + T B = M(L A ) + M(L B ). I denote the equlibrium as L i Intuitively, after centralization, the externalities will be internalized and both counties will benefit, but some of them also may be hurt by equalizing marginal utility across county boundaries. From the expressions of ˆL i and L i above, some predictions can be shown for empirical tests: log( L A + L B ) > log( ˆ L A + ˆ L B ) (3.9) log( L A ) log( ˆ L A ) > log( L B ) log( ˆ L B ) (3.10) For manufacture outputs, results are similar. 4 Data 4.1 County-level Statistics on Population and Production The outcomes I will investigate are urbanization and industrialization in counties of China. To achieve these goals, I digitize some county-level variables from different sources. The main sources are a series of statistics yearbooks published by each province every year. These yearbooks report some primary statistics of counties in the province collected by local branches of the National Bureau of Statistics. Data on population and production is available in most counties. As for years before the series of yearbooks were published, I digitize these variables from local gazettes, where the data sources are also from local branches of the National Bureau of Statistics. 4.2 County-level Statistics on Public Finance Besides the outcomes in population and production, variables associated with public finance are also significant for confirming the expected mechanism of centralization of governance. I construct the public finance data from The Prefecture, Municipality and County Public Finance Statistics Yearbook, which reports government revenue and 8

government expenditure at county and prefecture levels annually from 1993. 4.3 Firm-level Statistics To explore more implications on the effects on industrial sectors and on enterprise, I utilize the firm-level data from China Annual Survey of Industrial Production. The survey is also conducted by National Bureau of Statistics (NBS) of China. It is an annual census of all non-state industrial firms with more than 5 million Yuan in revenue plus all state-owned firms. I will use the data from 1998 to 2003. It consists over 100,000 firms in 1998 and nearly 200,000 in 2003. 5 Impacts on Regional Development In this section I am going to evaluate the overall impacts on economic developments using aggregate level data from various sources. Summary statistics of main variables are presented in Table 1. [Table 1 about here.] The table shows that both capital and peripheral counties experience significant growth in urbanization and industrialization after the reform. But of course the comparison is preliminary and a more comprehensive analysis is needed. 5.1 Overall Impacts on Prefectures I first estimate the overall effects on prefectures as a whole. The baseline regression is: Y jpt = βt reat jpt 1 + φw jpt + α j + δ pt + ɛ jpt (5.1) where Y jpt is an outcome variable in prefecture j, province p and year t; T reat jpt 1 is the main independent variable, indicating whether prefecture j in province p received treatment by year t 1; α j is a prefecture fixed effect; δ pt is a year province fixed effect; w ijpt are control variables. The outcomes of interest in this paper are population in urban areas and industrial outputs. In this prefecture-level regression, the main variables I use in practice are log value of non-agricultural population, urbanization rate (measured by ratio of non-agricultural population to total population), log value of industrial outputs and industrialization rate (measured by ratio of industrial outputs to total outputs). For the interested independent variable T reat jpt 1, I use one-period lag value of treatment status, since reforms to most prefecture did not happen at the beginning of one year and many related documents suggest that the transitions rarely took place immediately throughout rest of that year. Nevertheless, the results are robust to different lags and one of them is presented in Section 5.3. The prefecture fixed-effects α j captures prefecture specific characteristics which are time-invariant, such geography and culture. 9

The year province fixed-effects δ pt captures province-level common shocks, for example, policies and regulations implemented by province or central governments. Table 2 presents the results from the baseline model. [Table 2 about here.] The left hand side variable in column (1) is the log value of non-agricultural population. After a prefecture become a municipality, the whole prefecture will experience an 9.0% increase in non-agricultural population. Column (2) reports the effects on urbanization rate. The positive effect on the urbanization rate is 4.2% and significantly different from zero at the 1 percent level. As for industrial production, the reform is associated with a 7.2% increase in industrial output. This estimation is significant at the 5 percent level. The industrialization rate will increase by 2.6% but only marginally significant. These results are reasonable in the sense of previous theoretical and literal expectations. When a prefecture is turn to a municipality, the decision powers are taken into hands of the municipality government. The centralization of governance will help to internalize the positive spill-over effects among counties. A natural challenge about validity of the baseline empirical strategy is non-random reform of prefectures. To address this concern, I propose an event study design. To be specific, I test the identification assumption of the baseline regression by estimating a set of twelve yearly treatment effects beginning five year before the reform event and continuing for seven years after the reform. It will enable me to check pre-trends in these yearly treatment effects to secure the validity of the identification assumption. Virtually it is a flexible baseline regression allowing the effect to vary by year relative to the reform. The event study specification can be written as follow: 7 Y jpt = β τ I(Y earssincet reat jt = τ) + φw ijpt + α j + δ pt + ɛ jpt (5.2) τ= 5 where outcome variables include log values of non-agricultural population and industrial output, which are probably most suspicious for endogeneity during policy making; I( ) is an indicator function, and Y earssincet reat jt counts the years at time t since prefecture j was treated. Then Y earssincet reat jt takes negative values counting years before the treatment, positive values after the treatment and zero when t is the year it got treated. [Figure 6 about here.] Figure 6 plots the results for the event studies on non-agricultural population and industrial output in prefectures respectively. There are not any significant pre-existing differential trends in the growth of non-agricultural population and industrial output. The absence of evidence on differential pre-trends suggests that the central or provincial level government did not choose targets of the reform by their growth rate of urban population or industrial output. 10

5.2 Distributional Effects on Counties Apart from the overall effects of the reform, we will expect there exist some distributional effects on different types of counties according to the conceptual framework. In this subsection, I will explore capital and peripheral counties heterogeneous responses to the reform respectively. The baseline regression on county level is a difference-in-differences model: Y ijpt = βt reat jpt 1 + γt reat jpt 1 Capital i + φw ijpt + α i + δ pt Capital i + ɛ ijpt (5.3) where Y ijpt is an outcome variable in county i, prefecture j, province p and year t; T reat jpt 1 is the main independent variable, indicating whether prefecture j in province p received treatment by year t 1; Capital i is a dummy variable indicating whether this county i is a capital county or not; α i is a county fixed effect; δ pt is a year province fixed effect; w ijpt are control variables. Besides the four outcome variables in the prefecturelevel regressions, I construct two more variables on county development in this subsection: county s share of non-agricultural population in prefecture and county s share of industrial output in prefecture. The regression results are presented in Table 3. [Table 3 about here.] The first three columns report the effects of the reform on outcomes related to urbanization. Column (1) reports the effects of the reform on non-agricultural population. It suggests that when a prefecture is reformed to a municipality, the capital county of it will experience an 14.1% increase in non-agricultural population relative to those capital counties remaining in prefectures. On the contrary, the effect on peripheral counties is a negative 0.5% but not significant. Column (2) and (3) report the effects on urbanization rate and county s share of non-agricultural population in prefecture or municipality. The positive effect on urbanization rate is not significantly different from zero, and capital counties non-agricultural population shares 3.2% more in municipalities than in prefectures. For peripheral counties, these two impacts are small and not significant. Turning to industrial production, column (4) to (6) show the results for log value of industrial outputs, industrialization rate and county s share of industrial outputs in prefecture or municipality respectively. In column (5), we can see that the reform increase industrial outputs in capital counties by 17.0% while decrease peripheral counties by 6.9%. Both effects are significantly different from zero at the 5 percent level. The industrialization rate does not appear to be correlated with the reform either in capital or in peripheral counties. Since the reform is conducted in prefectures, the county level regressions will not suffer serious endogeneity problem. The identification assumption in the difference-indifferences model is that in the absence of the reform, capital counties and peripheral counties should have evolved in parallel. So I also estimate the event study to nonagricultural population and industrial output in county-level regressions. The results are plotted in Figure 7. We can observe that the outcomes in capital counties diverge 11

significantly from their pre-existing trends and the trends in peripheral counties are nearly flat in contrast. Those two graphs are in support of the identification assumption and validate the difference-in-differences regression. [Figure 7 about here.] 5.3 Robustness In this subsection, I am going to run various robustness checks towards the baseline model. 5.3.1 Evidence from employment statistics from census data The data I use in baseline regressions are reported by the National Bureau of Statistics in provincial statistic yearbooks. The population-related variables are based on Household Registration (hukou) system. There is a concern about the use of such variables is that whether they really measure the economic activity of a resident. It is possible that an agricultural resident and a non-agricultural resident differ only in the literal hukou status. Then the increases in non-agricultural population after the reform we find can be due to changes in definition of hukou and do not reflect any improvements in urbanization process. To address this concern, I utilize the population census data containing information on agricultural and non-agricultural employment to do a similar analysis as in baseline regressions. There are three waves of the population census data suitable during or closed to my analysis: 1982, 1990 and 2000. The aggregate data from the population census reports population employed in different sectors aggregated to county level. It allows me to examine the changes in employments in agricultural and non-agricultural sectors, which is a more precise measurement of population engaged in rural and urban economic activities. Results are presented in Table 4. [Table 4 about here.] 5.3.2 Geographic conditions Geographic conditions are believed to have huge influences on developments. Although in my baseline regressions all time-invariant impacts brought by heterogeneous geographic conditions are absorbed by individual fixed effects, it is still possible that they will be related with regions developments at different timings. Specifically, it may affect the timing of the reform. From Figure 5, we can easily observe some reform patterns related with geographies. So it is necessary to check the robustness when we controlling for the impacts of geographies flexible for different years. Empirically, I include the interactions between geographic conditions and year dummies in the baseline model. I calculate a prefecture or county s distance from nearest river, coastal line and the corresponding provincial capital city and take those three as proxies of geographic conditions. Table 5 lists the prefecture-level regression results and Table 6 lists the county-level regression results. 12

[Table 5 about here.] [Table 6 about here.] We can see that all results remain almost same as in the baseline regressions. Furthermore, the estimated coefficients of three geographic variables are insignificant at most years. For simplicity I do not report them in the paper. Above results suggest that the possible caveats brought by geographies do not exist and help to make casual arguments from my results. 5.3.3 Individual-specific time trends In this part, I add individual-specific time trends to the baseline regression. The individual-specific time trends in a fixed-effect model will allow treatment and control groups to follow different trends. If the estimated coefficients do not change much, it will support the identification strategy of the baseline model much. The results are presented in Table 7 and Table 8. The magnitudes of coefficients are quite similar while the significances are gone, which may be partly due to reduction in the degree of freedom. The fact that all coefficients in these regressions lie in the confidence intervals in the baseline estimations is encouraging to the validity of the baseline model. [Table 7 about here.] [Table 8 about here.] 5.3.4 Using other lags of main independent variable Another robustness check is related to the definition of our interested independent variable. In the baseline regression, I use the one-period lagged treatment status T reat jpt 1 as the main independent variable. Here I replace it with the treatment status at current period T reat jpt to check the robustness. Since reforms often happened in the middle of one year, I use the proportion of treated months as the value of T reat jpt if t is the year when the reform took place at county i and prefecture j. For example, if county i in prefecture j was treated in September of year T, then T reat jpt = 0.75. The results are shown in Table 9 and Table 10. Almost all estimated coefficients are similar in magnitude, while some of the significances drop due to increased standard error. Furthermore, the results will not change much by choosing other lags. I do not list the results for conciseness. [Table 9 about here.] [Table 10 about here.] 13

5.3.5 Possible heterogeneity across different periods There are some reasons to believe that the effect of the centralization reform would be heterogeneous in different periods. I divide the examined time period in half: 1983-1993 and 1994-2003, and run the baseline regression based on them respectively. One of the main reasons to do that is in 1994, China took a significant fiscal centralization reform from local governments to the central government. Although there is no evidence that the reform is interacted with my political hierarchy reform between different levels of local governments, it is possible that the effect I found in this paper is smaller after 1993 due to the limitation of fiscal capability in local governments. Meanwhile, changes in population mobility along time may also be responsible for the possible heterogeneity. [Table 11 about here.] [Table 12 about here.] 5.4 Placebo Test In this subsection I am going to explore possibilities that the above effects of the reform are not brought by the centralized institution by estimating the effects of the placebo reform in Zhejiang Province. Zhejiang Province was in this reform and all its prefectures were turn into municipalities during 1983 and 2000 nominally. However, the reform in Zhejiang is only a de jure experiment. Instead, Zhejiang never took the decision rights of county governments to municipality governments. County governments could maintain their powers on public finance and enacting economy policy on their own territories. The de facto independence of counties have been confirmed by the province government from the start of the 1983 reform (The People s Government of Zhejiang Province, 1983). Such special setting in Zhejiang Province provides me an opportunity to run a placebo test to see whether the results in last two sections reflect the effects through the centralization of governance. If it were not the centralized institution, but some other simultaneous changes in the reform instead that are responsible for above results, we could observe similar effects in the placebo reform in Zhejiang. Otherwise, such effects would not appear. To be specific, I estimate the baseline models, but only use the data in Zhejiang Province. The first four columns in Table 13 presents prefecture-level results. Table 14 presents county-level results. [Table 13 about here.] [Table 14 about here.] From the results it can be found that there are not any effects no matter on the prefecture level or on the county level. The coefficients on various outcome variables are all close to zero and not significantly different from zero. It strongly support the story of centralization of governance to explain the heterogeneous effects we found in previous sections. 14

There are some concerns about Zhejiang s specialities. For example, since it is on the coast opposite to Taiwan, there are not so many state-owned enterprises in manufacturing sector as in inland provinces. Instead, it is famous for its developed private-owned enterprises. To address the concerns that such specialities might erode validity of my placebo test, I do a difference-in-differences style regression using Fujian province as a control group. Fujian province is adjacent to Zhejiang and also on the frontier opposite to Taiwan. Like Zhejiang, it is also characterised by numerous smallscale private-owned enterprises. To be specific, I regress a difference-in-differences model Y jpt = β 1 T reat jpt 1 + β 2 T reat jpt 1 Zhejiang p + φw jpt + α j + δ pt + ɛ jpt (5.4) where T reat jpt 1 Zhejiang p is an interactive term between lag treatment variable and a dummy Zhejiang p which equals to one when the county is from Zhejiang province. We expect the coefficient of the interaction negative when regressing in prefectures. The last four columns in Table 13 list the results. For the county-level regression, I run a tripledifference regression and expect the coefficients of T reat jpt 1 Capital i Zhejiang p is negative. Table 15 confirms it. [Table 15 about here.] 5.5 Government Sizes and Expenditures Mechanism of the whole story lies in the changes in behaviours of local governments. To test the channel, it is necessary to check whether there are any changes on dimensions of local governments associated with the reform, such as sizes of local governments, fiscal expenditure etc. The data is from The Prefecture, Municipality and County Public Finance Statistics Yearbook from 1993. I run regressions using the baseline model and government related outcomes including log value of government employment, log value of government administrative expenditure, county share of expenditure in the prefecture and expenditure ratio to revenue. The results are listed in Table 16. As expected, the treatment is strongly associated with increases in local government sizes and public expenditure in capital counties, and in peripheral counties it is a slight negative relation. The results is in support of the whole story strongly. [Table 16 about here.] Besides that, we can explore further what kind of expenditure is affected by the reform. In Table 17, I divide total expenditure into four main categories and regress them on the lagged treatment status respectively. The results show that expenditures on infrastructures respond strongly to the reform. [Table 17 about here.] 15

6 Patterns of Industrial Enterprises To validate the micro foundation, in this section I develop some patterns of productions of industrial enterprises before and after the reform. 6.1 Agglomeration Spillovers With firm-level data, it is possible to check some key assumptions in the theoretical part. First, we can observe economies of scale from urban population agglomeration in my simple model. Second, agglomerations in capital counties can not only produce spillovers on enterprises located in themselves, but also benefit those in their peripheral counterparts. Third, the centralization of governance affects individual productivity through increasing urban workers and strengthening the agglomeration forces. Following Henderson (2003), to test above assumptions and validate the whole story, I estimate firm level production functions and test whether each firm s total factor productivity benefits from scale externalities of local or neighbourhood economic activities. To be specific, in order to obtain a measure of firm-level productivity, I estimate an augmented Cobb-Douglas production function in which the total factor productivity is calculated from residuals from the regression for simplicity. Then I regress this measure of productivity as an outcome variable in a firm-level fixed-effect model. Main independent variables are total employments in various regions. If the assumptions in the conceptual framework are reasonable, we will observe the positive coefficients in total employments and the effects from the reform will be absorbed by employment variables. Table 18 presents the results. [Table 18 about here.] Results of the above table can confirm the three key assumptions in my model. From the first two columns, we can see that increases in employments in urban areas have significantly positive effects on individual enterprises productivity. Moreover, agglomerations in capital counties can produce externalities on peripheral enterprises, which is the source of inefficiency under a prefecture institution. On the contrary, enterprises in capital counties are not benefited from increases in peripheral county employments. The last four columns test the channel of our treatment. Column (3) and (5) reveal that the 1983 reform has a sizeable impact on capital county enterprises productivity, but such impact will be totally absorbed by including the employments in the whole county. It suggests that the reform takes effects through agglomeration forces as I modelled. 6.2 Misallocation Another possible consequence of centralization is mitigations in resource misallocation. The problem of the resource misallocation in developing countries draws attention in recent years. Its significance in explaining low aggregate output per worker and total factor productivity (TFP) in developing countries has been confirmed (Hsieh and Klenow, 2009). In the paper, they document dispersions of revenue productivity (TFPR) as 16

proxies of resource misallocation, and then measure how much aggregate manufacturing output in China and India could benefit from if equalizing marginal products of labour and capital to the extent observed in the United States. They also show that the extent of misallocation can be affected by various policy distortions like licensing and size restrictions. In this part, I am going to investigate whether the problem of resource misallocation will be mitigated by the more centralized institution. Intuitively,the decentralized institution before the reform would encourage county governments to protect local enterprises and distort efficient resource allocation within a prefecture. So we may expect after the centralization reform, the municipality government may reallocate the resources to promote the efficiency in the whole region. The regression equation used in this part is: Y jt = βt reat jt 1 + α j + δ pt + ɛ jst (6.1) where Y jst is the dispersion in TFPR within prefecture j and year t; α j is prefecture fix effects; δ pt is year province fixed effects. Here I use standard deviations, ratios of 75th to 25th percentiles and ratios of 90th to 10th percentiles in TFPR respectively to measure dispersions as Hsieh and Klenow (2009) do. All of those dispersion measures are standardized. Regression results are in Table 19. [Table 19 about here.] From above results we can find that the dispersions in TFPR in centralized municipalities are relatively smaller than decentralized prefectures. A counterfactual analysis similar as Hsieh and Klenow (2009) suggests that moving the dispersions in decentralized prefectures to the relative efficient centralized level will lead to a gain of 32% in TFP in 1998. 6.3 Concentration Another natural implication on a more centralized institution is that a specific sector may become more concentrated geographically. Literature (Young, 2000; Poncet, 2003) has shown that local protectionism and decentralization slow down industrial concentration and agglomeration after 1980. We can expect that after the municipality government take over the decision power from the county governments, the local protectionism should be mitigated and it is not necessary for each county to own every sector. So that a specific sector will concentrate in less places. To measure the geographic concentration of industries, I calculate the index developed by Ellison and Glaeser (1997). In their paper, the authors construct a model-based index of geographic concentration of economic activities (γ index). I revise their γ index to fit into my county-prefecture scenario as follows: γ sj = G s (1 i x2 i )H s (1 x 2 )(1 H s ) (6.2) 17

where γ sj is the Ellison-Glaeser index in sector s, prefecture j; G sj = i (x i s si ) 2 is the spatial Gini coefficient, where i is any county belonging to prefecture j, x i is i s share of total employment or output of all industries in j, s si is sector s s share of employment or output for region r in county i; H s is the Herfindahl index of sector s. The greater EllisonGlaeser index, the higher the geographic concentration. It equals to zero if all the firms randomly pick their location. In this part, I calculate the Ellison-Glaeser index γ sjt for every combination of year, sector and prefecture or municipality, and regress them on the treatment variable treat in a fixed-effect model: γ sjt = βt reat jt + α j + δ pt + θ st + ɛ jst (6.3) where γ sjt is the Ellison-Glaeser index in prefecture j, sector s and year t; α j is prefecture fix effects; δ pt is year province fixed effects; θ st is sector province fixed effects. Results are given in Table 20 [Table 20 about here.] The first column is the result on the Ellison-Glaeser Index measured in outputs and the second is the one in employment. From the results we can find that after the centralization reform, the extent of industrial concentration significantly raised. Although we can see that sectors will become more concentrated within a prefecture, it is still not clear about a specific sector will concentrate in which kind of counties. One plausible thinking is that firms in a sector will agglomerate in places with high average sector-specific TFP. To solve this problem, I do a firm-level regression: Y uijst = β 1 T reat jt 1 + β 2 T reat jt 1 T F P is + η u + δ pt + θ st + ɛ uijst (6.4) where Y uijst is log output of firm u in sector s, county i, prefecture j and year t; T F P is is average TFP in sector s and county i. To avoid endogeneity, I use the average TFP in the first available year (1998). If a sector will agglomerate in relative high-productive places, we should expect the coefficient of the interaction between the treatment variable and the average county-sector TFP. The results are in Table 21. [Table 21 about here.] From results above, we can see that firms in counties with higher sector-specific productivity will produce more after the reform. It suggests that the hierarchy reform will be helpful to sort firms into places with comparative advantages. Results in the later two columns show that such effects do not differ between capital counties and peripheral counties. The heterogeneous effects on overall industrial outputs shown earlier may be due to the reason that capital counties owns more sectors with comparative advantages. 7 Concluding Remarks This paper exploits a particular type of political reforms in China to understand the distributional effects of centralization on regional development. Prefectures in China 18

were reformed to be municipalities from 1983, which made counties belonging to them less autonomous and the hierarchy more centralized. After the reform, the municipality governments are responsible for economic policies and developments of belonging counties. Such centralization setting can internalize the positive spill-overs across counties as Oates (1972) describes. We can expect a positive overall effect on the whole prefecture. But at the same time, when the municipality governments can equalize marginal revenue across counties, capital counties may get more than peripheral counties. Then the distribution effects come from above two opposite forces. To quantify the effects, I collect prefecture-level and county-level data on population and production in China. In the baseline regression, the fixed-effect model reveals that capital counties benefit from centralization and peripheral counties suffer a slight loss, while prefectures as whole experience an overall positive growth. The baseline results are robust to various settings. Above treatmen effects do not appear to be due to differential pre-existing trends. An event study model confirms the validity of the identification strategy of my baseline model. Controlling for geographic conditions and individual-specific trends do not change the baseline results. Linking the results on urbanization and industrialization to those of similar regression on government sizes and public finance suggest the plausibility of the story. Besides that, a placebo test involving Zhejiang Province which took the reform only de jure also indicates the changes in the political hierarchy is the basic mechanism. Finally, I use the firm-level data on Chinese industrial enterprises to explore more micro evidence beneath the whole macro picture. To interpret the empirical results better, I construct a theoretical model involving a spatial equilibrium in urban and rural areas as a conceptual framework. All the empirical results of this paper describe significant impacts brought by the political hierarchy reform. They indicate an essential role the political factor play in the process of regional development. Moreover, the differential patterns in two types of counties remind policy makers the distributional effects from political institution changes. 19

References Ades, A. F. and E. L. Glaeser (1995). Trade and Circuses: Explaining Urban Giants. The Quarterly Journal of Economics 110 (1), 195 227. Bardhan, P. (2002). Decentralization of Governance and Development. Journal of Economic Perspectives, 185 205. Besley, T. and S. Coate (1998). Sources of Inefficiency in a Representative Democracy: a Dynamic Analysis. American Economic Review, 139 156. Dixit, A. K. and J. E. Stiglitz (1977). Monopolistic Competition and Optimum Product Diversity. The American Economic Review, 297 308. Duranton, G. and D. Puga (2004). Micro-Foundations of Urban Agglomeration E- conomies. Handbook of Regional and Urban Economics 4, 2063 2117. Ellison, G. and E. Glaeser (1997). Geographic Concentration in US Manufacturing industries: a Dartboard Approach. Journal of Political Economy 105 (5), 889 927. Ethier, W. J. (1982). National and International Returns to Scale in the Modern Theory of International Trade. The American Economic Review, 389 405. Grossman, G. M. and E. Helpman (1991). Innovation and growth in the global economy. MIT Press. Henderson, J. V. and R. Becker (2000). Political Economy of City Sizes and Formation. Journal of Urban Economics 48 (3), 453 484. Henderson, J. V. and H. G. Wang (2007). Urbanization and City Growth: the Role of institutions. Regional Science and Urban Economics 37 (3), 283 313. Hsieh, C.-T. and P. J. Klenow (2009). Misallocation and manufacturing tfp in china and india. The Quarterly Journal of Economics 124 (4), 1403 1448. Marshall, A. (1890). Principles of Political Economy. Maxmillan, New York. Oates, W. E. (1972). Fiscal Federalism. New York: Harcourt Brace Jovanovich. Poncet, S. (2003). Measuring Chinese Domestic and International Integration. China Economic Review 14 (1), 1 21. Qian, Y. and B. R. Weingast (1997). Federalism as a Commitment to Preserving Market Incentives. Journal of Economic Perspectives 14 (4), 83 92. Romer, P. M. (1990). Endogenous technological change. Journal of Political Economy 98, 71 102. Seabright, P. (1996). Accountability and Decentralization in Government: An Incomplete Contract Model. European Economic Review 40 (1), 61 89. 20

World Bank (2000). Entering the Twenty-First Century. Oxford and New York: Oxford University Press. Xu, C. (2011). The Fundamental institutions of China s Reforms and Development. Journal of Economic Literature, 1076 1151. Young, A. (2000). The Razor s Edge: Distortions and Incremental Reform in the People s Republic of China. Quarterly Journal of Economics, 1091 1135. 21

Figure 1: Changes in the number of prefecture or municipality level jurisdictions Note: This figure plots the numbers of prefecture/municipality level jurisdictions in China over years. The red line is the number of prefectures. The green line is the number of municipalities. The orange line is the number of other special types of jurisdictions such as minority autonomous zones. The blue line is the total number. 22

Figure 2: Changes in the number of prefecture or municipality level jurisdictions Note: This figure plots dispersions of TFPR in Ningde, Fujian Province, in 1998 and 2003. Ningde prefecture received the treatment to become a municipality in 1999. Distributions are for deviations of log(tfpr) from sector means. TFPR is calculated as in Hsieh and Klenow (2009). Figure 3: hierarchy of local governments in China Note: This figure graphs the basic structure of local governance in China. Dashed lines between different levels of local governments suggest that upper level governments do not have administrative powers on lower level governments jurisdictions. Solid lines suggest that upper level governments have administrative powers on lowers jurisdictions. 23

Figure 4: An example of the Reform Note: This figure visualizes an example of the reform. The blue prefecture is treated in the reform and turned to a municipality. The red prefecture is not treated and remains prefecture status. The dark-color zones are capital counties and the light-color zones are peripheral counties. Figure 5: Timing of the reform across space Note: This figure graphs the nationwide reform timing. The darker the colour is, the later the treatment happens. The white zones are those excluded in the sample. 24

(a) Event study on log non-agricultural population (b) Event study on log industrial output Figure 6: Event study on outcomes in prefectures Note: This figure reports estimates of the dynamic effect of the Turning prefectures into municipalities reform on developments in prefectures derived from an event study regression. Estimates are constructed by regressing the log of non-agricultural (a) or the log of industrial output (b) on a series of dummy variables indicating whether the year of observation falls in a given relative year as measured from the year of the reform happened. Relative year -6 is the omitted category so that all estimates should be interpreted as relative to the sixth year prior to the reform. All years beyond the relative year 6 are grouped into the effects of relative 7. The series in blue circle plots the estimate and 95 percent confidence interval for the relative year main effects. The red dashed line plots the fit from a regression of the estimates on a linear term of relative year using only the pre-reform years. 25

(a) Event study on log non-agricultural population (b) Event study on log industrial output Figure 7: Event study on outcomes in counties Note: This figure reports estimates of the dynamic effect of the Turning prefectures into municipalities reform on developments in capital and peripheral counties derived from an event study regression. Estimates are constructed by regressing the log of non-agricultural (a) or the log of industrial output (b) on a series of dummy variables indicating whether the year of observation falls in a given relative year as measured from the year of the reform happened and their interactions with a dummy variable indicating whether a county is a capital county or not. Relative year -6 is the omitted category so that all estimates should be interpreted as relative to the sixth year prior to the reform. All years beyond the relative year 6 are grouped into the effects of relative 7. The series in red triangles plots the coefficient estimates of the relative year main effects, representing the trend among peripheral counties. The series in blue circle plots the estimate and 95 percent confidence interval for the sum of the relative year main effects and the interaction with the capital county indicator, representing the trend among capital counties. The orange and green dashed line plot the fit from a regression of the estimates of capital and peripheral counties respectively on a linear term of relative year using only the pre-reform years. 26