Decentralization, Collusion and CoalmineDeaths in China∗

Ruixue Jia†

IIES, Stockholm UniversityHuihua Nie‡

Renmin University of China

August 24, 2011

Abstract

We propose a political-economics model of decentralization andcollusion to explain high death rates in the coalmine industry in China,where the local government allows the firms to choose dangerous buthigh-payoff production technology when collusion is feasible. In termsof death rates, our model predicts a positive effect of decentralization,a negative effect of media exposure and a positive effect of collusion.We collect a provincial level panel dataset from 1995 to 2005 andcompare the decentralization period (1998 to 2000) with the adjacentcentralization periods. We find that decentralization increases deathrates and that media exposure (proxied by newspapers published per1000 individuals) decreases death rates. Moreover, given decentral-ization, death rates are about 100% higher in provinces where thegovernor in charge of coalmine safety is a native or has been in offi celonger than average.Keywords: decentralization, collusion, public bads

∗Acknowledgement to be added.†Institute for International Econonomic Studies, Stockholm University. Email:

ruixue.jia@iies.su.se.‡School of Economics, Renmin University of China. Email: niehuihua@gmail.com

1

1 Introduction

Over the last few decades, China has experienced rapid industrialization and

economic growth. Meanwhile, China has become notorious internationally

for weak occupational health and safety protection. Fatalities from coalmine

accidents are among the most important health and safety issues in China.

Coal mining accounts for less than 4 percent of the broadly defined industrial

workforce, but over 45 percent of industrial fatalities (Wright 2004).

China’s coal mining industry is indeed the deadliest in the world in terms

of human safety where thousands of people die every year in the coal pits.

Coalmine fatalities in China account for 80 percent of the world total1. In

2007, China produced 41.1% and United States produced 18.7% of coal in

the world. At least 3598 people died in coal mine mishap in China, compared

with 34 in the United States. In fact, the death rate in China is not only much

higher than that of developed countries such as the US, Japan and Germany

but also more than ten times higher than that of many other developing

economies such as India and African countries. As mentioned in Wright

(2004), the head of China’s safety bureaucracy admitted in 2001 that China’s

coalmine death rate was 11 times higher than that in Russia, and 15 times

higher than that in India. Wang (2006) provides a cross-country comparison

of coalmine death rates from history to today.

China’s appalling record on coal-mine safety has drawn worldwide atten-

tion, especially in the media (The Economist 2002; The Economist 2005;

The Economist 2007). Some hypotheses about causes have been put foward,

such as insuffi cient safety input (Tang and Guo 2006) and lack of safety ed-

ucation (Asian Development Bank 2007), insecure property rights (Tan and

Tang 2004) and possible collusion between governments and coalmines (Nie

and Jiang 2011). However, to the best of our knowledge, all these existing

1“Coal Mine Fatalities in China Account for 80 percent of World Total,”BBC, Nov.13, 2004.See:http://news.bbc.co.uk/chinese/simp/hi/newsid_4000000/newsid_4009000/4009043.stm.

1

hypotheses are mainly suggestive and lack a clear theoretical framwork as

well as convincing empirical strategy. In this paper, we propose a model of

decentralization and collusion to explain high coalmine death rates. The lo-

cal government allows firms to choose dangerous but high-payoff technologies

when collusion between the local government and the firms is feasible. As

discussed below, our assumptions on technology and collusion are based on

existing investigations of deadly coalmine mishaps.2 Our model predicts a

positive effect of decentralization, a negative effect of media supervision and

a negative effect of transaction costs of collusion (given decentralization) on

death rates.

To bring our theory to data, we collect provincial-level panel data from

1995 to 2005 and explore an exogenous institutional change of the manage-

ment power of key state coalmines. In March 1998, the management of the

key state coalmines was delegated to provincial government and this made

collusion possible, or at least much easier. This decentralization period lasted

until Feburary 2001, when the State Administration of Coal Mine Safety was

established and re-centralized coalmine safety supervision. Hence, we have

one decentralization period (1998-2000) when collusion is possible and two

centralization periods (1995-1997 and 2001-2005) as comparison.

We find much higher death rates were increased during the decentralized

period. Using the number of newspapers published per 1000 poluation as

a proxy for media exposure, we find that higher exposure decreases death

rates. Since transaction costs of collusion are not observable, we approach

them by exploring biographical information on 71 governors in charge of

coalmine safety between 1995 and 2005. We assume that being a native of the

province or having been in offi ce for longer time, lowers the transaction costs

of collusion. Within the decentralized period, we find that the death rate

is about 100% higher in provinces where the governor in charge of coalmine

2There are so many anecdotal evidences on collusion that a phrase was coined for thisphenomenon "guan-mei goujie", which literally means "offi cial-coal collusion".

2

safety is a native or has been in offi ce longer than average.

Although the decentralization decision is a national policy and exogenous

to difference provinces, there still might be a concern that natives are ap-

pointed to provinces with higher death rates. We do not find that this is

true in our data. However, as a robustness check, we conduct a placebo test

with respect to being native using data before 1998 and do not find being

native matter for death rates. We also control for governor fixed effects and

explore within-governor variation. Further, We limit our sample by dropping

a subset of governors that might have been endogenously appointed.

Our study contributes to a few lines of literature. Many fruitful theories

in the organizational literature have focused the costs and benefits of decen-

tralization/delegation. Mookherjee (2006) discusses the existing theoretical

literature and some empirical studies of industrial organizations. Delegation

theory has been widely applied to the organizational forms of governments,

especially to the trade-off between decentralization and centralization. Del-

egation and collusion have been paid a lot of attention in the theoretical

literature (Tirole 1986; Kofman and Lawarree 1993; Baliga and Sjostrom

1998; Laffont and Martimort 1998; Mookherjee and Tsumagari 2004). Our

model emphasizes how decentralization makes the collusion between the su-

pervisor (the Local government) and the agent (the Firm) more likely. As a

result, the principal (the Central government) faces a trade-off between high

payoff and high probability of accidents.

Although the theorectical incentive-based literature focuses more on the

cost side of decentralization, existing empirical evidence has been celebrating

the virtues of decentralization in terms of economic performance and govern-

ment accountability. Bardhan and Mookherjee (2006) provides an excellent

survey of existing research along this line. Within the context of China, fis-

cal decentralization since 1994 is often taken as one of the driving forces of

China’s growth miracle (Jin, Qian, and Weingast 2005; Lin and Liu 2000),

despite sometimes opposite findings (Zhang and Zou 1998). A possible rea-

3

son for the more positive findings in existing empirical studies is that the

cost side is often less visible than economic growth rates and expenditure on

public goods. We focus on coalmine mishap as it is one of the most visible

health and safety problems. Our findings on the downside impacts of decen-

tralization share some flavor with the findings on regional protectionism in

Young (2000), who argues that increased autonomy and incentives induced

local governments to engage in provincial protectionism. Our finding of bad

outcomes due to the combination of decentralization and collusion remind

of the views inBlanchard and Shleifer (2000), who argue that political cen-

tralization and economic decentralization together reduce the risk of local

capture.

Our study is also related to the literature evaluating political connections.

This line of literature has documented the benefits captured by those with

stronger political connections (Fisman 2001; Faccio 2006). It is conceivable

that these political benefits might be captured at the costs of some public

interest. Our study provides new evidences on the cost side of political

connections, and the measures are closely related to descriptions of Chinese

networks ("guanxi") in political science (Guo 2001).

Moreover, our study contributes to an emerging strand of literature which

examines the impacts of media in nondemocratic countries (Reinikka and

Svensson 2005; Egorov, Guriev, and Sonin 2009). Why do dictators need

media? Our findings support the logic in Egorov, Guriev, and Sonin (2009),

who argue that media allows dictators to monitor bureaucrats.

Although we focus on a very specific example of coalmine deaths, our

theory has external validity to other "public bads" such as ecology destruction

and air pollution. Our theory is consistent with the current dilemma of

China: dramatic growth but overwhelming problems with workplace safety

and the environment. Moreover, our study also serves the first attempt to

provide both a theory and empirical evidences on high coalmine death rates

in China. To explore an exogenous change in our identification, we only focus

4

on key sate coalmines in this paper. This perspective also sheds possible light

on the death rates in smaller coalmines. For example, Wright (2004) mentions

that township and village coalmines are closely related to local governments

in a nexus of "local state corporatism" (Oi 1999).

The organization of the paper is as follows. Section 2 provides a model of

collusion and decentralization and the predictions derived from this model.

Section 3 describes the data. Section 4 presents the main estimation results,

where the predictions are confronted with the data. Section 5 provides var-

ious robustness checks to deal with concerns of endogeneity, heterogeneity

and alternative explanations. Section 6 colludes.

2 A Model of Decentralization and Collusion

Our model is based on the three-tier principal-supervisor-agent (PSA) frame-

work (Tirole 1986), where the Central government is the principal, the Local

government serves as the supervisor and the Firm is the agent. Decentraliza-

tion is characterized by the share of resources assigned to the Local govern-

ment in a grand contract. Collusion is characterized by a side contract signed

by the Local government and the Firm. Below we describe the environment,

the timing of the game, motivating evidences for our assumptions and the

solution.

2.1 Environment

There are three risk-neutral players: the Principal/Central government (C),

the Surpervisor/Local government (L), and the Agent/Firm (F ). To produce

y units of coal, F can use a good way or a bad way. The good way costs cy2

and there is no accident. The bad way costs cy2 and an accident A happens

with probability p. Parameters c and c are known to F and L but cannot be

observed by C.

5

When an accident happens and is detected, the local government (L)

and the firm (F ) get punished. We assume that both of them have limited

liability. Without loss of generality, assume that L gets 0 and F gets a

fine of f in the case of punishment. The probability of being punished is

mp, where m denotes the detection probability; especially, this is related to

media exposure.

Production y is taxed at the rate of t and income ty is divided between

C and L. L gets a share of s, which measures the extent of decentralization.

However, L and F can collude by signing a side contract3, although they

face some transaction costs to agree with this arrangement. For simplicity,

we assume transaction costs are proportional to output, i.e., the costs are are

φy(c). L gets a bribe b if he allows F to choose c, where b is proportional to

the payoff of F , i.e., b = k(1− t)(1−φ)y(c). This way, k can be thought as aresult of bargaining between L and F . In our solution, we take k as a choice

variable of L that satisfies F’s participation constraint in the side contract.

2.2 Timing

The timing of the game is as follows:

1. C decides the policy of centralization or decentralization characterized

by s. C also specifies the publishment of of discovering an accident. The

punishment is known to F and L , and denoted by (0, f), where 0 is the

payoff to the local government and f is the fine of the firm.

2. Given (s, 0, f), F and L decide to collude or not by signing a side

contract. The side contract specifies the share k, a choice of L that satisfies

F’s participation constraint in the side contract. There are some transaction

costs in this agreement given by φy.

3. Given (s, 0, f) and k, F chooses c or c. c or c is not observed by C.

4. The side contract, if it exists, is implemented.

3In collusion, F chooses c. Note that we assume away the case that F chooses c andthere is no collusion. This is because the information is symmetric between F and L.

6

5. Output (y) is realized. Accidents are observed, if they are detected.

The punishment schedule gets implemented in the case of detection.

2.3 Discussion of assumptions

Before turning to the solution of the game, we present some qualitative evi-

dences in support of the model’s assumptions. Specially, we clarify what the

technologies are, how collusion works and why there might be transaction

costs of collusion.

The technology in the model broadly refers to coal production technology

and safety monitoring technology. We focus on the short-run decisions of

firms such as where to mine, how many hours to work and how closely to

monitor the mine safety. In the real world, most coalmine accidents occur

because of these short-run decisions. For instance, one of the most common

types of mining accidents in China is gas explosion. Some coalmine managers

disregard the monitoring system for the sake of production. After 32 miners

died in a gas explosion in Henan Province on October 18, 2010, it was revealed

that gas emissions had exceeded the standard 22 hours prior to the disaster

but this warning was disregarded by the mine management.4 Similarly, 108

miners died of gas explosion in Heilongjiang Province on November 21, 2010

because the managers decided to mine at places that did not have qualified

gas drainage system.5 Because short-run decisions of firms are essential in

affecting coalmine safety, centralization or decentralization of management

and supervion may play an important role.

As for collusion, government offi cials allow mine managers to employ

dangerous technology in exchange for money or shares in the enterprises.

When accidents occur, offi cials also help the owners cover them up as they

will be punished together. For instance, when 81 miners died of mine flooding

in Nandan County (Guangxi Province) on July 17, 2001, local offi cials quickly

4Source: http://society.people.com.cn/GB/12986786.html5Source: http://society.people.com.cn/GB/8217/174547/index.html

7

teamed up with the coalmine owner to cover up the accident. Once the

catastrophe was disclosed to the public, it was revealed that the head of

Nandang County had received 3.21 million yuan ($410,000) in bribes during

the prior two years (Tu 2007). Wright (2008) provides more information on

rent seeking in the coal industry.

Since collusion is illegal, trust is essential in the agreement of collusion.

Li and Wu (2010) provide some qualitative evidences on the transactions

between bribers and bribees in China and the Philippines. They show that

personal networks play an important role in facilitating corruption, especially

in a relation-based governance system such as the Chinese case. Empirically,

we will argue that a common background and a long period of interaction

facilitates trust-based collusion.

2.4 Solution

We solve the game by backward induction.

First, consider the firm’s choice between the production technologies char-

acterized by c and c, given (s, 0, f) and k. If it chooses the good c−technology,its problem is,

maxy(1− t)y − c

2y2.

The solution gives y(c) = 1−tcand π(c) = (1−t)2

2c. Instead, if it chooses the

bad c−technology, the firm’s problem becomes,

maxy(1− t)(1− φ)y − c

2y2 − k(1− t)(1− φ)y −mpf.

Now, we get y(c) = (1−t)(1−k)(1−φ)c

and π(c) = (1−t)2(1−k)2(1−φ)22c

−mpf .

Therefore, it is optimal to choose the c−technology iff π(c) ≥ π(c) i.e.,

(1− t)2(1− k)2(1− φ)2

2c− (1− t)2

2c≥ mfp. (FPC)

8

Now consider the local government’s choice of k. Its problem is as follows,

maxk(1−mp)[(1− φ)sty(c) + k(1− t)(1− φ)y(c)],

s.t.

(1−mp)[(1− φ)sty(c) + k(1− t)(1− φ)y(c)] ≥ sty(c) (LPC)

where y(c) = (1−t)(1−k)(1−φ)c

, as solved above.

First, we ignore the participation contraint of L and solve its maximiza-

tion problem. Then we check when the participation constraint is satisfied.

The first-order condition with respect to k gives:

k∗ =1− t− st

2(1− t). (LIC)

Subsitute k∗ into (LPC) and get a suffi cient condition cc≥ 4t(1−t)

(1−mp)(1−φ)2

(see appendix). We assume that (LPC) is satisfied and focus on whether F

proposes the side contract.

Obviously,∂k∗

∂s= − t

2(1− t)< 0.

Substituting k∗ into the firm’s participation constraint (FPC) for collu-

sion, we get the following equivalent condition (FPC′):

(1− t)2(1− k∗)2(1− φ)2

2c−mfp ≥ (1− t)2

2c. (FPC’)

Given (LIC) and (FPC′), we can characterize the likelihood of collusion

(Collusion) by the following proposition.

Proposition 1 Given that cc≥ 4t(1−t)

(1−mp)(1−φ)2 ,collusion is more likely for higher

decentralization (s) and less likely for higher media exposure (m): ∂Collusion∂s

>

0, ∂Collusion∂m

< 0. Besides, higher transaction costs coeffi cient (φ) makes the

9

effect of decentralization small: ∂2Collusion∂s∂φ

< 0. The predictions for death

rates have the same signs: ∂Deathrate∂s

> 0, ∂Deathrate∂m

< 0, ∂2Deathrate∂s∂φ

< 0.

Proof. Consider the comparative statistics with respect to the left hand side(LHS) of (FC

′).

First,

∂LHS

∂s= −(1− t)2(1− k∗)(1− φ)2

c

∂k∗

∂s

=(1− t)2(1− k∗)(1− φ)2

c

t

2(1− t)> 0

Second,∂LHS

∂m= −fp < 0

Moreover,

∂2LHS

∂s∂φ= −(1− t)2(1− k∗)(1− φ)

c

t

(1− t)< 0

Given that Collusion is more likely for a higher LHS, we get ∂Collusion∂s

>

0, ∂Collusion∂m

< 0, ∂2Collusion∂s∂φ

< 0.

Since the c−technology is chosen underCollusion, we will also get ∂Deathrate∂s

>

0, ∂Deathrate∂m

< 0, ∂2Deathrate∂s∂φ

< 0.

Note that our model takes the decentralization decision of the Central

government as given rather than provide an optimal contract from the per-

spective of the Central government. In a more complete model, the Central

government would face a trade-offbetween high output y and dangerous tech-

nology c. Its choice of decentralization would depend on the relative utility

of output and safety concerns. However, we do not attempt to endogenize

its choice of decentralization in this paper. Instead we bring the predictions

of our theory to data, given the decentralization, or centralization, decisions

of the Central government.

10

In terms of empirical predictions, we can observe s since the centralization

or decentralization policy of the Central government is public. We can also

observe m to some extent. For example, number of newspapers would be

a reasonable measure for m. But we cannot observe the technology device

by the coalmining firms (c or c) nor can we observe the transaction cost

coeffi cient φ directly. Instead, we consider some proxies for φ. For instance,

being a native and having more local network connections will lower the

transaction costs of collusion. Similarly, being in offi ce for a couple of year will

also enchance connections and hence decrease the transaction costs. Hence

we have the following corollary.

Corollary 1 Being native (N) and having long experience in offi ce (E)increases the effect of decentralization on collusion: ∂2Collusion

∂s∂N> 0 and

∂2Collusion∂s∂E

> 0. We also have ∂2Deathrate∂s∂N

> 0 and ∂2Deathrate∂s∂E

> 0.

This is simply because ∂φ∂N

< 0 , ∂φ∂E

< 0 and ∂2Collusion∂s∂φ

< 0. As mentioned

above, we will also get ∂2Deathrate∂s∂N

> 0 and ∂2Deathrate∂s∂E

> 0 as c−technology ischosen under Collusion.

In sum, our theory has four testable predictions:

(1) Higher decentralization leads to higher death rates: ∂Deathrate∂s

> 0;

(2) More media supervision leads to lower death rates: ∂Deathrate∂m

< 0;

(3) Having a native local leader increases the effect of decentralization on

death rates: ∂2Deathrate∂s∂N

> 0;

(4) Having a local leader in offi ce for longer time increases the effect of

decentralization on death rates: ∂2Deathrate∂s∂E

> 0.

3 Data

We collect panel data on key state coalmines, governors in charge of coalmine

safety, as well as provincial characteristics for 22 provinces across China be-

tween 1995 and 2005. Among china’s remaining provinces, Tianjin, Shanghai,

11

Hainan and Tibet do not produce any coals, while there are no key coalmines

in Fujian, Hubei, Gungdong, Guangxi and Qinghai. Figure 1 maps the dis-

tribution of key coalmines across China in terms of production. Table 1

presents some summary statistics.

[Figure 1 here]

[Table 1 here]

3.1 Decentralization/delegation of key state coalmines

According to their ownership, Chinese coalmines can be divided into three

types: key state coalmines, local state coalmines and township and village

coalmines. There are 94 key state coalmines, which in 2003 produced 47.81%

of the total coals in China whereas local state coalmines produced 16.94%

and and township and village coalmines produced 35.25% (The State Ad-

ministration of Coal Mine Safety 2004).

In this paper, we focus on the management of the key state coalmines.

Before 1998, the coal mining industry was overseen by the Ministry of Coal

Industry in the Central government. Due to many policy burdens for state-

owned enterprises and the competition of small coalmines, the profitability

of the key state mines was negative in the 1990s. To provide more incen-

tives for profitability, the management of the key state mines was shifted to

provincial governments in 1998 (The State Council 1998). This delegation

decision was passed at China’s Ninth National People’s Congress in March

1998. This delegation was also associated with 206 enterprises affi liated with

these coalmines, an asset of 237.9 billion yuan ($30 billion) and 4.35 mil-

lion employees. After delegation, the taxes from the key state mines were

handed over to the provincial government and the profits were left to the

coalmines. Meanwhile, the powers of safety supervision were also shifted to

the provincial governments.

This decentralization period of key state coalmines lasted until Feburary

12

2001, when the State Administration of Coal Mine Safety (SACMS) was

established and re-centralized safety supervision powers. The centralization

was further increased in 2003, when SACMS became part of the general

offi ces of the State Council.

Hence, we take the period between 1998 and 2000 as the decentralized

period when collusion became possible or at least much easier, whereas 1995-

1997 and 2001-2005 are centralized comparison periods. Note that the two

centralizated periods differ slightly. Before 1998, the key state coalmines were

standard state-owned enterprises, whose management power and supervision

power all belonged to the central government. These powers were all dele-

gated to the provincial governments in 1998. Only the supervision power got

centralized since 2001. We group the two centralized period together in our

baseline estimations and separate them as a robustness check.

3.2 Transaction costs of collusion and media exposure

As mentioned above, we gauge the transaction costs of collusion (φ in the

model) by examining the biographical information of the provincial governor

in charge of coalmine safety. The governor in charge of coalmine safety is

one of the vice governors in a province and has a tenure of at most 10 years.

Besides coalmine safety, the vice governor is also in charge of safety of other

industries such as construction-site safety and road-traffi c safety. We trace

the careers of all the 71 governors in charge of coalmine safety across 22

provinces between 1995 and 2005. These data come from yearly provincial

government reports and the CVs on People’s Daily online. The average time

in offi ce for these 71 governorsis is about 4 years.

We proxy the transaction costs of collusion by (1) whether the governor

in charge of coalmine safety is a native, i.e. born in the same province that

he governs and (2) whether he has been in offi ce for longer than average, i.e.

for more than 4 years. Our results are robust to whether we use four, five

or six years as a cutoff. The motivation is that being native and being in

13

offi ce for longer facilitates collusion as they are proxies for political network

connections.

About 43% of the vice governors in charge of coalmine safety are native.

This share is very similar to the share of natives among the major provincial

governors. Figure 2 shows the share of native governments over time. We

do not see any change along with decentralization and only a slight increase

after 2001.

[Figure 2 here]

We realize that these measures might also capture something else besides

the possibility of collusion. For example, being a native might make it easier

to under-report death numbers whereas being in offi ce fora long time is also

a proxy for learning experience. However, these concerns would only bias our

findings in the opposite direction from that of our main hypotheses.

We measure media exposure (m in the model) by newspapers published

per 1000 individuals between 1996 and 2005. Numbers of newspapers pub-

lished come from China Statistical Data of Press and Publication published

since 1997. We divide them by provincial population sizes from China Sta-

tistical Yearbook. The number of newspapers published per 1000 individuals

varies from 0.36 to 8.58.

3.3 Death rates

Number of deaths and the yearly production of key state coalmines come

from China Coal Industry Yearbook and China Statistical Yearbook. Death

rates are measured by number of deaths per million tons of coal production.

As shown in Table 1, the mean of death rate is about 2.5 people per million

tons of production.

As for measurement error in deaths, the death rates of key state coalmines

are believed to be more reliable than those of smaller coalmines. This is

14

another reason why we focus on them in this paper. On top of this, sup-

pose the provincial governments have incentives to under-report deaths, they

would have better ability of under-reporting in the decentralized period. This

should again bias our findings in the opposite direction from the predictions

of our model.

Figure 3 plots the changes of death rates over time by whether the gover-

nor in charge of coalmine safety is a native. Clearly, the death rate jumped

dramatically between 1998 and 2000, especially for the provinces with native

observations.

[Figure 3 here]

We also observe a higher jump of death rates in 1998. This might be

related to the findings in the literature of political business cycles in China

(Tao 2006; RIETI 2004). This line of literature finds that there is a strong

tendency for the economic growth rate to reach the peak in the year fol-

lowing the National Congress as provincial leaders have stronger incentives

to promote economic growth and production. The year of 1998 happened

to be after the 15th Party Congress in 1997. In our estimations below, we

take the three years of decentralization together. But our basic results on

media exposure and transaction costs still hold after dropping 1998 though

the magnitude become smaller.

3.4 Other controls

We also control for characteristics of coalmines, other information on the

governors and other provincial characteristics.

Data on coalmines includes average yearly wages and size of fixed assets.

The average yearly wage is a proxy for labor quality, whereas size of fixed

assets is a proxy for the scale of coalmines. These data also come from China

Coal Industry Yearbook and China Statistical Yearbook.

15

Other information on the governors in charge of coalmines includes their

age and whether they have some experience of safety management before

coming to offi ce. We use these two variables to control for the career concerns

and experience of the governors. Like the proxies for transaction costs of

collusion, these data come from yearly provincial government reports.

Provincial characteristics include real GDP per capita, whether the main

provincial governor is a native, and the distance of the provincial capital

to Beijing. GDP per capita comes from China Statistical Yearbook. The

biographical information of the main governors is from yearly provincial gov-

ernment reports. The distance of the provincial capital to Beijing is calcu-

lated with ArcGIS and works as a proxy for the supervision intensity by the

Central government. In practice, we use log(1+distance to beijing) to take

account of the 0 observations of Beijing. This transformation will not change

our results.

4 Baseline Estimates

In our baseline estimation, we explore within-province variation by running

a fixed-effect model including province fixed effects and year fixed effects. To

test our basic prediction, we run the following specification:

Deathrateit = λi + γt + βDecentralized+ εit, (1)

where λi denotes province and year fixed effects. Decentralized is a dummy

for year 1998-2000.

Our theory predicts that β > 0. Column (1) and (2) present the results

from OLS and fixed effects regressions. The results are consistent with our

theory.

16

To test predictions (2)-(4), we run the following specification:

Deathrateit = λi + γt + β1Nativeit ∗Decentralized+ β2Experit ∗Decentralized+β3News_perit + β4Nativeit + β5Experit (2)

+γ ′Xit ∗Decentralized+ υ′Xit +22∑i=2

δiprovi ∗ year + εit,

where Nativeit is a binary indicator whether the governor is a native in

province i and year t. Similarly, Experit is a binary indicator whether the

governor has been in offi ce for more than 4 years in i and year t. News_peritis the number of newspapers published per 1000 individuals. We do not

include the dummy of Decentralized itself as we have already controlled for

year fixed effects. Including it does not change our results.

Xit is a vector of controls, including some other characteristics of the

governor in charge of coalmine safety, such as whether he has work experi-

ence in safety supervision departments and his age. We control for coalmine

characteristics including the log amount of fixed assets and log of the average

yearly wage of mine workers. Moreover, we also control for other provincial

characteristics such as GDP per capita, whether the major provincial gover-

nor is a native and the log distance of the provincial capital to Beijing. We

also control for the variables mentioned above as well as their interactions

with decentralization.

Finally, we also control for provincial specific trends by including the

interaction of provincial dummies and year. To be cautious about the auto-

correlation concern, we report AR (1) standard errors for all the specifications

to follow.6

The results are presented in Table 2. Column (3) and (4) are the results

6We also tried with other ways of dealing with standard errors. FGLS (xtgls) givesus similar magnitudes and higher significance. Clustering at the provincial level gives thesame magnitudes and lower significance levels (around 10% in the baseline estimations)as we only have 22 provinces in our sample.

17

from the fixed effects model without and with controls. Column (5) presents

the results including provincial specific trends. Similarly, Column (6)-(8) are

the results after considering the data of newspapers per 1000 individuals.

Our theory predicts that β1 > 0, β2 > 0 and β3 < 0. The results are

in line with these predictions and stable over different specifications. In the

decentralized period, the mean of the dependent variable is about 3.3 and its

standard deviation is about 6.6. Hence, within the decentralized period, the

death rate is about 100% higher in provinces where the governor in charge

of coalmine safety has collusion possibilities. On the other hand, a standard

deviation of newspaper per 1000 individuals (1.36) decreases the death rate

by 100%.

[Table 2 here]

5 Robustness Checks

In this section, we conduct a couple of robustness checks, to tackle endo-

geneity concerns and possible heterogeneity. Although the decentralization

decision is exogenous, there still might be a concern that natives are ap-

pointed to provinces with higher death rates. We do not find this true in our

data. We conduct a placebo test with data before 1998 to check that being a

native did not matter before the decentralization. As a robustness check, we

also control for governor fixed effects and explore within-governor variation.

Besides, we limit our sample by excluding possibly endogenous governors.

Given that we have two periods of centralization (1995-1997 and 2001-

2005) and one period of decentralization (1998-2000), we can test our theory

separately to see whether the effects are heterogeneous over time. We find

the effects are generally larger for the period between 1995 and 2000.

Our empirical results also speak to some of prospective alternative expla-

nations. But we do not find support for them.

18

5.1 A placebo test for having a native governor

To check whether there were already some different trends for the provinces

with native vice governors before the decentralization decision in 1998, we

conduct a placebo test using data before 1998 and control for province fixed

effects and year fixed effects. The specifications is as follows:

Deathrateit = λi + γt + α1Nativeit + α2Experit + υ′Xit + εit.

The results are presented in Table 3. Column (1)-(3) are results from

OLS and fixed effects without and with controls. These results imply that

having a native vice governor did not matter before the decentralization.

[Table 3 here]

5.2 Within-governor estimation

Analogous to the baseline estimations in (1) and (2), we run fixed-effect mod-

els including governor fixed effects and year fixed effects to explore within-

governor variation. The specifications are as follows:

Deathrateit = µi + γt + β′Decentralized+ εit,

and

Deathrateit = µi + γt + β′1Nativeit ∗Decentralized+ β′2Experit ∗Decentralized+β′3News_perit + γ

′Xit ∗Decentralized

+υ′Xit +

22∑i=2

δiprovi ∗ year + εit.

where λi denotes governor fixed effects. Since we include governor fixed

effects, we do not include whether the governor is a native or whether he has

been in offi ce for more than four years in the specification.

19

The estimation results are presented in Table 4. The impact of newspa-

pers is still negative but not significant in this specification. This is because

there is not enough variation in terms of this variable within governors. The

rest of the results are very similar to what we find in the within-province

estimation. This provides further evidence on decentralization and collusion,

since we are now looking at the effect within the role of the same governor.

[Table 4 here]

5.3 Excluding possibly endogenous governors

We trace the career paths of the governors and define a subset of governors

whose predecessor was politically demolished as possibly endogenous gover-

nors. In our definition, a governor is demolished if he did not move to a

higher level or a similar level of offi ce and did not retire. In our data, 20%

of this subset are native whereas 50% of the rest are native. These data do

not suggest that natives are more likely to be appointed to provinces with

higher death rates.

As a robustness check, we drop the subset of governors whose predecessors

were demolished, which is about 25% of our observations. We then run the

same regression as (1) and the results are presented in Table 5. The results

from column (1) to column (8) are generally close to our baseline estimates.

[Table 5 here]

5.4 Comparing 1995-2000 and 1997-2005

We know from government offi cial documents that the decentralization deci-

sion in 1998 was due to the consideration of profitability (The State Council

1998). However, we do not know the precise motives behind the centraliza-

tion decision in 2001. Besides, the centralization refers to the supervision

power particularly. Our model predicts that we should see an effect, both

20

when decentralization is introduced and when it is taken away. To check

whether the effects are heterogenous between the two decisions, we can esti-

mate separate regressions for subperiods.

Now we replicate the same regressions in (1) and (2) but separate the time

into two sample periods: 1995-2000 and 1998-2005. In the first subsample

1995-2000, 1995-1997 is the centralization period whereas 1998-2000 is the

decentralization period. In the second subsample 1998-2005, 1998-2000 is

the decentralization period where 2001-2005 is the centralization period.

The results are presented in Table 6. The upper panel is for 1995-2000 and

the lower panel is for 1998-2005. We find the effects of the decentralization

decision are generally somewhat larger. But the conclusions from the baseline

estimations hold for both periods.

[Table 6 here]

5.5 Alternative explanations

As mentioned in the introduction, both media and academia have provided

quite a few hypotheses to explain China’s high coalmine death rate. We

categorize the alternative hypotheses into three groups. The first hypothesis

is insuffi cient safety input (Tang and Guo 2006). This argues that insuffi cient

inputs are the main reason behind the high accident rates in China, especially

for small local coalmines. For example, gas drainage systems have not been

set up in 40% of the coalmines.

A second hypothesis is insecure property right (Tan and Tang 2004).

This would argue that the ownership of mines is not stable and often subject

to policies of the Central government. The Central government sometimes

closes coalmines without noticing owners of the mines when safety becomes

politically sensitive. Insecure property rights makes owners short-sighted.

Hence, the owners under-invests in safety facilities.

A third hypothesis is low quality of coalmine workers. This would argue

21

that the coalmine workers are usually uneducated, lack knowledge on safety

and do not have strong demands of improvements in their working conditions

(Whyte 1999).

Insuffi cient safety inputs no doubtedly matter for the high death rates.

But the next question is to answer why the safety inputs are not suffi cient,

which is what we try to do in this paper. The property right hypothesis may

provide a possible explanation. However, this hypothesis would generate op-

posite results from our findings, since decentralization should provide better

property rights protection.

It is not easy to measure the demand for workplace safety. We can use

yearly wages of mine workers as proxies for labor quality. This proxy does

not matter after controlling for province fixed effects and year fixed effects.

Hence, the hypothesis of low quality of coalmine workers is not supported.

The coeffi cients of the controls we use are presented in Table A1.

[Table A1 here]

Our theory provides a political-economics explanation about why safety

inputs might be insuffi cient. Our study contributes to the understanding of

high coalmine death rates by providing empirical evidence in support of this

theory.

6 Conclusion

We propose a theory of decentralization and collusion to explain China’s

abysmal record on coalmine safety. In this theory, the local government

allows firms to choose dangerous but high payoffproduction technology when

collusion is feasible. We explore an institutional change in the management

of key coalmines to test our theory, using a provincial level panel dataset from

1995 and 2005. Although our theory is tested with a very specific example

of coalmine deaths, it has external validity to other "public bads" such as

22

ecology destruction and air pollution. The bottom line of our argument is

that the local government and firms might sacrifice the benefits of the public

to achieve growth.

Our theory is consistent with the current dilemma of China: dramatic

growth and overwhelming problems with workplace safety and the environ-

ment. While decentralization has led to growth achievements, it has also

brought many unexpected problems. More careful pondering is needed to

evaluate the consequences of public-sector decentralization.

We use aggregate data at the provincial level in this paper. Another

perspective to investigate the death rates in coalmines would be to study

data at the level of individual coalmines. Coalmine level data may also help

us better analyze the welfare implications to coalmine workers. This task is

left for future work.

7 Appendix

The participation constraint of the local government (LPC) is as follows:

(1−mp)[(1− φ)st+ k∗(1− t)(1− φ)](1− t)(1− k∗)(1− φ)

c≥ st

1− t

c

Given k∗ = 1−t−st2(1−t) , the participation constraint is equivalent to:

c

c≥ 4st(1− t)

(1−mp)(1− t+ st)2(1− φ)2

The maximum of the right hand side is 4t(1−t)(1−mp)(1−φ)2 . Hence, we get a

suffi cient condition for (LPC) to hold: cc≥ 4t(1−t)

(1−mp)(1−φ)2 .

23

References

Asian Development Bank (2007). People’s Republic of China: Coal Mine

Safety Study. www.adb.org/Documents/Reports/Consultant/39657-

PRC/39657-PRC-TACR.pdf .

Baliga, S. and T. Sjostrom (1998). Decentralization and Collusion. Journal

of Economic Theory 83 (2), 196—232.

Bardhan, P. and D. Mookherjee (2006). Decentralization, Corruption and

Government Accountability. International Handbook on the Economics

of Corruption, 161—188.

Blanchard, O. and A. Shleifer (2000). Federalism Wth and Without Polit-

ical Centralization: China versus Russia. NBER.

Egorov, G., S. Guriev, and K. Sonin (2009). Why Resource-Poor Dicta-

tors Allow Freer Media: A Theory and Evidence From Panel Data.

American Political Science Review 103 (04), 645—668.

Faccio, M. (2006). Politically Connected Firms. The American Economic

Reviewe 96 (1), 369—386.

Fisman, R. (2001). Estimating the Value of Political Connections. The

American Economic Review 91 (4), 1095—1102.

Guo, X. (2001). Dimensions of Guanxi in Chinese Elite Politics. The China

Journal (46), 69—90.

Jin, H., Y. Qian, and B. Weingast (2005). Regional Decentralization and

Fiscal Incentives: Federalism, Chinese style. Journal of Public Eco-

nomics 89 (9-10), 1719—1742.

Kofman, F. and J. Lawarree (1993). Collusion in Hierarchical Agency.

Econometrica 61 (3), 629—656.

Laffont, J. and D. Martimort (1998). Collusion and Delegation. The Rand

Journal of Economics 29 (2), 280—305.

24

Li, S. and J. Wu (2010). Why Some Countries Thrive Despite Corruption:

The Role of Trust in the Corruption—Effi ciency Relationship. Review

of International Political Economy 17 (1), 129—154.

Lin, J. and Z. Liu (2000). Fiscal Decentralization and Economic Growth

in China. Economic Development and Cultural Change, 1—21.

Mookherjee, D. (2006). Decentralization, Hierarchies, and Incentives: A

Mechanism Design Perspective. Journal of Economic Literature 44 (2),

367—390.

Mookherjee, D. and M. Tsumagari (2004). The Organization of Sup-

plier Networks: Effects of Delegation and Intermediation. Economet-

rica 72 (4), 1179—1219.

Nie, H. and M. Jiang (2011). Zhengqi hemou yu kuangnan(Coal Mine Acci-

dents and Collusion between Local Governments and Firms). Jingjiyan-

jiu (Economic Research Journal) Forthcoming.

Oi, J. (1999). Rural China Takes off: Institutional Foundations of Eco-

nomic Reform. University of California Press.

Reinikka, R. and J. Svensson (2005). Fighting Corruption to Improve

Schooling: Evidence from a Newspaper Campaign in Uganda. Jour-

nal of the European Economic Association 3 (2-3), 259—267.

RIETI (2004). The Political Aspect of China’s Business

Cycle- Synchronization with the communist party congress.

http://www.rieti.go.jp/en/china/04071401.html .

Tan, M. and X. Tang (2004). Chanquan niuqu: kuangnandeshencengsikao

(Property Rights and Coalmine Accidents). Meitanxuebao (Journal of

Coal Mine) 29 (6), 756—759.

Tang, L. and X. Guo (2006). Woguo xianjieduan kuangnan pinfa cheny-

injiqiduice: jiyu anquan touru de shjiao (The Reasons for Coalmine

25

Mishap in China and Policy Implications). Zhongguo gongye jingji

(China Industrial Economy) 12.

Tao, Y. (2006). The Evolution of Political Business Cycle in Post-Mao

China. Issues & Studies 42 (1), 163—194.

The Economist (2002). Pits of Death: Counting the Cost of Coal.

http://www.economist.com/node/1284235 .

The Economist (2005). Going Under.

http://www.economist.com/node/3966158 .

The Economist (2007). Bottoming Out.

http://www.economist.com/node/9687898 .

The State Administration of Coal Mine Safety (Ed.) (2004). China Coal

Industry Yearbook. The Coal Industry Press.

The State Council (1998). Guanyu guoyou zhongdian meikuang guanli

tizhi youguan wenti de tongzhi (Notice on the Reform of the Key State

Coalmines).

Tirole, J. (1986). Hierarchies and Bureaucracies: On the Role of Collusion

in Organizations. Journal of Law, Economics, and Organization 2 (2),

181.

Tu, J. (2007). Coal Mining Safety: China’s achilles’heel. China Security 3,

36—63.

Wang, S. (2006). Regulating Death at Coalmines: Changing Mode of Gov-

ernance in China. Journal of Contemporary China 15 (46), 1—30.

Whyte, M. (1999). The Paradox of China’s Post-Mao Reforms. Harvard

Uni.

Wright, T. (2004). The Political Economy of Coal Mine Disasters in China:

Your Rice Bowl or Your Life. The China Quarterly 179 (-1), 629—646.

Wright, T. (2008). Rents and Rent Seeking in the Coal Industry. Taylor &

Francis.

26

Young, A. (2000). The Razor’s Edge: Distortions and Incremental Re-

form in the People’s Republic of China. Quarterly Journal of Eco-

nomics 115 (4), 1091—1135.

Zhang, T. and H. Zou (1998). Fiscal Decentralization, Public Spending,

and Economic Growth in China. Journal of Public Economics 67 (2),

221—240.

27

Figure 1: Average production of key state coalmines between 1995 and 2005 (million tons)

Figure 2: Share of native governors in charge of coalmine safety between 1995 and 2005

decentralized in 1998 centralized in 2001

0.2

.4.6

.81

Sha

re o

f nat

ive

gove

rnor

s

1995 2000 2005year

Figure 3: Coalmine death rates between 1995 and 2005

decentralized in 1998 centralized in 2001

02

46

810

Dea

ths

per

mill

ion

tons

of

coal

s

1995 2000 2005year

deathrate_native deathrate_meandeathrate_nonnative

Table 1: Summary statistics for the province-year data

Variables Mean S.D. Min Max NumberTransaction costs of collusion and media exposure

Native 0.43 0.50 0 1 239 Whether has been in office more than 4 years 0.41 0.49 0 1 283 Published newspapers per 1000 individuals 1.58 1.36 0.36 8.58 218

Coalmine characteristics Death rates (million tons) 2.45 4.10 0 46.25 233 Production (1,0000 tons) 2976 3815 42 25621 238 Yearly wage (RMB) 10292 4782 4321 25827 240 Fixed assets (100 million RMB) 36.99 59.69 0.02 329.84 238

Other leader’s characteristics Years in office 3.6 2.3 1 10 239 Age 52.6 4.67 36 61 238 Safety Work Experience 0.31 0.46 0 1 242

Other provincial characteristics GDP per capita 7438 4926 1826 28050 241 Whether the main governor is native 0.42 0.49 0 1 242 Distance to Beijing (Km) 1075 674 0 2559 282

Table 2: Within-province estimation results

Dependent var. Deaths per million tons of coal production Models OLS Fixed

effects Fixed effects

Fixed effects

Fixed effects

Fixed effects

Fixed effects

Fixed effects

(1) (2) (3) (4) (5) (6) (7) (8) Periods 1995-2005 from 1995 to 2005 from 1995 to 2005 Decentralized 1.182** 2.819** (0.595) (1.092) Native*Decentralized 4.305*** 4.056*** 4.174** 3.720*** 3.561** 4.370** (1.256) (1.534) (1.730) (1.318) (1.597) (1.783)Exper*Decentralized 3.422*** 3.091** 2.821* 3.018** 2.610* 2.267 (1.270) (1.391) (1.576) (1.318) (1.480) (1.647)Publish. newspapers -2.556*** -2.228** -2.238*per 1000 individuals (0.977) (1.085) (1.219)Native 0.038 0.271 -0.088 0.520 0.497 -0.280 (0.856) (1.059) (1.445) (0.940) (1.142) (1.575)Exper 0.352 0.229 1.015 0.767 0.815 1.532 (0.804) (0.887) (1.066) (0.868) (0.993) (1.196)Controls Y Y Y Y Year fixed effects Y Y Y Y Y Y Province fixed effects

Y Y Y Y Y Y

Province time trends Y Y Observations 233 233 209 206 Y 189 187 187

Notes: (a) AR(1) standard error are reported. *** indicates p<0.01, ** indicates p<0.05, and * indicates p<0.1. (b) Controls include (1) some other characteristics of the governor in charge of coalmine safety: whether he has work experience in safety supervision departments and his age; (2) coalmine characteristics including log of the amount of fixed assets and log of the average yearly wage of mine workers and (3) other provincial characteristics: log of GDP per capita, whether the major provincial governor is a native and log of distance to Beijing as well as (4) their interactions with the dummy of decentralization.

Table 3: A placebo test using data between 1995 and 1997 Dependent variable Deaths per million tons of coal production Models OLS Fixed Effects Fixed Effects (1) (2) (3) Native 0.677 -0.190 1.298 (0.422) (1.411) (1.599) Exper 0.024 -0.668 -0.236 (0.498) (1.268) (0.860) Safety 1.607 (2.265) Age 0.356* (0.180) Log (Fixedasset) 1.628 (1.596) Log (Wage) 2.570 (3.190) Log(GDP_Per) -0.530 (7.627) Whether the main governor 6.039*** is native (1.415) Controls Y Year fixed effects Y Y Leader fixed effects Y Y Observations 58 58 57 Notes: (a) The placebo best is to check whether having a native vice governor matters before the decentralization. (b) AR(1) standard error are reported. *** indicates p<0.01, ** indicates p<0.05, and * indicates p<0.1.

Table 4: Within-governor estimation results

Dependent var. Deaths per million tons of coal production Models OLS Fixed

effects Fixed effects

Fixed effects

Fixed effects

Fixed effects

Fixed effects

Fixed effects

(1) (2) (3) (4) (5) (6) (7) (8) Periods from 1995 to 2005 from 1995 to 2005 from 1996 to 2005 Decentralized 1.182** 3.466*** (0.595) (1.292) Native*Decentralized 5.243*** 3.414** 2.985 4.971*** 3.216* 3.138 (1.268) (1.629) (1.936) (1.330) (1.719) (2.087)Exper*Decentralized 2.737** 2.500** 2.322* 2.755** 2.448** 2.264 (1.119) (1.188) (1.276) (1.148) (1.233) (1.368)Publish. newspapers -1.080 0.008 -0.857per 1000 individuals (0.947) (0.983) (1.238)Controls Y Y Y Y Year fixed effects Y Y Y Y Y Y Province fixed effects

Y Y Y Y Y Y

Province time trends Y Y Observations 233 231 231 228 228 211 209 209

Notes: (a) *** indicates p<0.01, ** indicates p<0.05, and * indicates p<0.1. (b) Controls include (1) some other characteristics of the governor in charge of coalmine safety: whether he has work experience in safety supervision departments and his age; (2) coalmine characteristics including log of the amount of fixed assets and log of the average yearly wage of mine workers and (3) other provincial characteristics: log of GDP per capita, whether the major provincial governor is a native and log of distance to Beijing as well as (4) their interactions with the dummy of decentralization.

Table 5: Dropping possibly endogenous leaders Dependent var. Deaths per million tons of coal production Models OLS Fixed

effectsFixed effects

Fixed effects

Fixed effects

Fixed effects

Fixed effects

Fixed effects

(1) (2) (3) (4) (5) (6) (7) (8) Periods 95-05 from 1995 to 2005 from 1996 to 2005 Decentralized 1.807** 2.291 (0.742) (1.397) Native*Decentralized 3.653** 3.685* 3.933 2.916* 2.400 3.030 (1.639) (2.164) (2.486) (1.709) (2.298) (2.625) Exper*Decentralized 4.886*** 4.108* 3.224 4.026** 3.718 2.387 (1.671) (2.285) (2.601) (1.746) (2.350) (2.690) Publish. newspapers -3.024*** -3.025** -3.205**per 1000 individuals (1.129) (1.280) (1.498) Native 0.742 0.697 1.328 1.245 0.964 0.404 (1.245) (1.445) (2.106) (1.335) (1.584) (2.294) Exper 0.408 0.626 1.677 1.263 1.828 3.645** (1.017) (1.189) (1.464) (1.109) (1.335) (1.677) Controls Y Y Y Y Year fixed effects Y Y Y Y Y Y Province fixed effects

Y Y Y Y Y Y

Province time trends Y Y Observations 178 178 157 154 154 141 139 139

Notes: (a) AR(1) standard error are reported. *** indicates p<0.01, ** indicates p<0.05, and * indicates p<0.1. (b) Controls include (1) some other characteristics of the governor in charge of coalmine safety: whether he has work experience in safety supervision departments and his age; (2) coalmine characteristics including log of the amount of fixed assets and log of the average yearly wage of mine workers and (3) other provincial characteristics: log of GDP per capita, whether the major provincial governor is a native and log of distance to Beijing as well as (4) their interactions with the dummy of decentralization

Table 6: Comparing two periods

Dependent Variable Deaths per million tons of coal production From 1995 to 2000

Models OLS Fixed effects

Fixed effects

Fixed effects

Fixed effects

Fixed effects

(1) (2) (3) (4) (5) (6) Periods 1995-2000 1995-2000 1996-2000 Decentralized 1.776** 1.862** (0.876) (0.790) Native*Decentralized 5.464*** 6.938** 5.519*** 5.921* Exper*Decentralized (1.700) (2.647) (2.054) (3.416) 2.647 3.268 3.329 2.746 (2.289) (2.611) (2.644) (3.223) Published newspapers -1.863 -1.816 per 1000 individuals (2.140) (2.239) Native -2.807 -2.362 -2.717 -0.378 (2.044) (3.019) (2.511) (3.779) Exper 1.393 0.885 1.069 1.703 (1.799) (2.107) (2.195) (2.719) Controls Y Y Y Year fixed effects Y Y Y Y Y Province fixed effects Y Y Y Y Y Observations 125 125 123 121 103 102

From 1998 to 2005 Periods 1998-2005 1998-2005 1998-2005 Decentralized 0.852 0.665 (0.727) (0.624) Native*Decentralized 3.878*** 3.826** 3.534** 3.487** Exper*Decentralized (1.394) (1.697) (1.382) (1.693) 3.419** 3.192* 2.958** 2.694 (1.463) (1.694) (1.456) (1.702) Publish. newspapers -2.148** -1.919* per 1000 individuals (0.949) (1.063) Native 0.570 1.056 0.849 1.032 (1.038) (1.292) (1.031) (1.280) Exper 0.269 -0.012 0.732 0.495 (0.985) (1.209) (0.992) (1.231) Controls Y Y Y Year fixed effects Y Y Y Y Y Province fixed effects Y Y Y Y Y Observations 173 173 173 171 173 171

Table A1. Results on controls Models Fixed

Effects Fixed

Effects Fixed

Effects Fixed

Effects Column number in Table 2 (4) (5) (7) (8) Native*Decentralized 4.056*** 4.460** 3.561** 4.370** (1.534) (1.723) (1.597) (1.783) Exper*Decentralized 3.091** 3.041* 2.610* 2.267 (1.391) (1.573) (1.480) (1.647) Published newspapers -2.228** -2.238* per 1000 individuals (1.085) (1.219) Native 0.271 -0.002 0.497 -0.280 (1.059) (1.448) (1.142) (1.575) Exper 0.229 0.667 0.815 1.532 (0.887) (1.039) (0.993) (1.196) Safety*Decentralized -1.593 -1.385 -1.517 -1.315 (1.622) (1.805) (1.681) (1.906) Age*Decentralized 0.216 0.253 0.204 0.294 (0.222) (0.249) (0.233) (0.267) Log (Fixedasset)* -0.140 0.135 -0.032 0.361 Decentralized (0.585) (0.682) (0.603) (0.693) Log (Wage)*Decentralized 2.697 0.130 2.344 -1.257 (3.331) (3.652) (3.476) (3.866) Log(GDP_Per)*Decentralized 0.814 1.087 -0.170 0.875 (1.733) (1.883) (1.855) (2.045) Prov_native*Decentralized 1.111 0.763 1.524 1.052 (1.514) (1.634) (1.601) (1.715) Log(Dis_Bejing)*Decentralized 0.592 0.516 0.272 0.130 (0.678) (0.796) (0.752) (1.020) Safety -0.074 -0.528 -0.266 -1.301 (1.332) (1.680) (1.460) (1.785) Age 0.098 0.065 0.093 0.032 (0.086) (0.102) (0.094) (0.127) Log (Fixedasset) -0.162 -0.420 -0.165 -0.370 (0.419) (0.497) (0.449) (0.513) Log (Wage) 2.404 1.815 3.176 1.195 (3.861) (4.932) (4.157) (5.179) Log(GDP_Per) -3.811 0.097 -4.776 -0.253 (4.214) (7.071) (4.642) (13.350)

Whether the main governor -1.187 -0.545 -1.520 -0.967 is native (0.932) (1.202) (1.035) (1.225) Year fixed effects Y Y Y Y Province fixed effects Y Y Y Y Province-specific trends Y Y Observations 206 206 187 187