Economic Modelling xxx (2018) 1–8

Contents lists available at ScienceDirect

Economic Modelling
journal homepage: www.journals.elsevier.com/economic-modelling

On the reverse causality between output and infrastructure: The case of

China☆
Qu Feng * , Guiying Laura Wu
School of Social Sciences, Nanyang Technological University, Singapore

A R T I C L E I N F O A B S T R A C T

JEL Classification: After the 2008 global financial crisis, promoting public infrastructure investment as a growth engine has been
C23 revived by economists. China has been considered as such a successful example of enhancing economic growth
E22 by massive infrastructure investments in the past decades. However, the literature has provided conflicting
H54
empirical results on the productivity effect of public infrastructure using aggregate data, mainly due to reverse
O40
causality. Thus, the estimated productivity effect could be either upward or downward biased. In this paper we
O50
rely on the institutional background of infrastructure investment in China, and explore several alternative ways
to mitigate the reverse causality. Using China’s provincial-level data over 1996–2015 and within the frame-
Keywords:
work of an aggregate production function estimation, we find that an upward bias dominates when estimating
Infrastructure
output elasticity of public infrastructure, and that weak evidence is found on the productivity effect of public
Productivity
Chinese economy infrastructure. This finding highlights the necessity of using alternative identification strategies or data types.

1. Introduction telecommunications infrastructure in OECD countries, and the recent

works surveyed in Redding and Turner (2015) for transport infrastruc-
After the 2008 global financial crisis, promoting public infrastruc- ture. In China, rates of return to railroad and road are found over 10%
ture investment as a growth engine has been revived by economists and 20%, respectively (Li and Li, 2013; Li and Chen, 2013).
and policy makers. For example, a 4 trillion Chinese Yuan (equivalent To address whether public infrastructure investment as a whole
to 600 billion US dollars) fiscal stimulus package was introduced by the enhances the growth of the whole economy, we take a macro view
Chinese government to invest mainly in the infrastructure in its western and focus on the productivity and return of the total public infras-
provinces in 2008 (Ouyang and Peng, 2015). Recently, as Chinese econ- tructure investment. For this purpose, following the literature starting
omy started to slow down in 2015, 1 trillion Chinese Yuan was further from Aschauer (1989), we estimate the output elasticity with respect to
proposed to invest in infrastructure (Financial Times, August 5, 2015). public infrastructure in an aggregate production function using China’s
For a specific project on infrastructure investment, e.g., building an provincial panel data over 1996–2015.
airport, it is straightforward to calculate its economic return if the ben- The importance of studying China’s case is in two folds. First, it is
efits and costs of the project are well defined and recorded. However, well known that China is considered as an investment-driven economy
its social return may not be fully captured in a financial evaluation with the investment-to-GDP ratio above 45% since 2009, far exceed-
framework. For a specific type of infrastructure, the literature has also ing other developing countries and advanced economies.1 As a major
developed various ways to identify its productivity effect, for example, component of the total investment, public infrastructure investment
Fernald (1999) for road in the US, Röller and Waverman (2001) for

☆ We would like to thank the editor Paresh Narayan, associate editor Russell Smyth and three anonymous referees for their constructive comments and suggestions.

We are grateful to Zhao Chen, Shigeyuki Hamori, Shaoqing Huang, Hao Shi, Guang-zhen Sun, Yi-Chan Tsai, and the audiences at TED conference 2015, Fudan
University, the 1st Annual International Conference on Applied Econometrics in Hawaii 2015, and University of Macau for their helpful suggestions. Zhifeng Wang
provided excellent research assistance. Financial support from the MOE AcRF Tier 1 Grant M4011642 is gratefully acknowledged. Division of Economics, School of
Social Sciences, Nanyang Technological University. Address: 14 Nanyang Drive, Singapore, 637332. Emails: qfeng@ntu.edu.sg (Q. Feng), guiying.wu@ntu.edu.sg
(G. Wu).
* Corresponding author.
E-mail address: qfeng@ntu.edu.sg (Q. Feng).
1
See the World Bank website https://data.worldbank.org/indicator/NE.GDI.TOTL.ZS?locations=CN-TH-VN-IN.

https://doi.org/10.1016/j.econmod.2018.05.006
Received 1 January 2018; Received in revised form 23 April 2018; Accepted 5 May 2018
Available online XXX
0264-9993/© 2018 Elsevier B.V. All rights reserved.

Please cite this article in press as: Feng, Q., Wu, G.L., On the reverse causality between output and infrastructure: The case of China, Economic
Modelling (2018), https://doi.org/10.1016/j.econmod.2018.05.006
Q. Feng and G.L. Wu Economic Modelling xxx (2018) 1–8

accounts for an average rate of 9.3% of China’s GDP during 1996- with the reverse causality are discussed in Section 3. Section 4 presents
2015.2 Thus, it is of policy significance to evaluate the productivity the data and reports the empirical findings. Section 5 concludes.
and return of public infrastructure investment in China. Second, China’s
institutional context may provide unique identification strategies for 2. Empirical model
the endogeneity problem due to the reverse causality between output
and public infrastructure when estimating its elasticity. To model the general idea that public infrastructure investment pro-
Using the framework of an aggregate production function estima- motes economic growth, following literature we introduce an aggregate
tion, the literature has provided conflicting empirical results, mainly production function:
due to reverse causality. As surveyed in Bom and Ligthart (2014), the
output elasticity of public capital varies from the highest estimate of Y = AK 𝛾k L𝛾l ,
2.04 for Australia in one research to the lowest one of −1.7 for New where Y is the total output; L is the total labor force; and K is the
Zealand in another research. In between, many estimates are statisti- stock of non-infrastructure capital. The public infrastructure capital B,
cally not different from zero. The output elasticity of public infrastruc- measuring the stock of public infrastructure investment, enters the pro-
ture capital could be overestimated when a growth in output facilitates duction function as a contributing component to the total productivity
an increase in public infrastructure investment. That is, public infras- factor (TFP) A, i.e., A = A0 B𝛾b , where A0 is the component of TFP that
tructure investment could be induced by economic growth, instead of is unrelated to public infrastructure. Thus, the aggregate production
driving economic growth. Alternatively, the output elasticity of pub- function becomes
lic infrastructure capital could be underestimated when public infras-
tructure investment is used as a countercyclical tool to boost economic Y = A0 B𝛾b K 𝛾k L𝛾l . (1)
growth during economic recession.
The stock variables, B and K, accumulate according to the following
In a recent study with a focus on the investment efficiency in China,
laws of motion:
Shi and Huang (2014) argue that a downward bias is more likely in
China’s case. This is because the Chinese government tends to use Bt = (1 − 𝛿b )Bt −1 + Gt (2)
infrastructure investment as a choice for stimulating its economy when
a negative productivity shock is expected. Consistent with this logic, and
they find that the output elasticity using a proxy approach developed by Kt = (1 − 𝛿k )Kt −1 + It . (3)
Ackerberg et al. (2015) is even larger than that from the OLS approach.
Using China’s provincial panels over 1995–2011, they obtain a big and Here Gt measures the infrastructure investment in industries with exter-
positive output elasticity of public infrastructure, with a magnitude nalities, such as electricity, gas, water, transport, information transmis-
around 0.22 to 0.29. This implies a rate of return more than 50%.3 sion, and I t is the investment in non-infrastructure sectors. 𝛿 b and 𝛿 k
In this paper we rely on the institutional background of infrastruc- are depreciation rates of B and K, respectively.
ture investment in China, and explore several alternative ways to miti- Under the assumption of constant returns to scale (CRS),4
gate the reverse causality between aggregate output and public infras- 𝛾 b + 𝛾 k + 𝛾 l = 1, so that (1) becomes Y ∕L = A0 (B∕L)𝛾b (K ∕L)𝛾k . Thus
tructure. Using different approaches we find that an upward bias dom- the aggregate production function in the intensive form can be written
inates when estimating output elasticity of public infrastructure using as
China’s provincial-level data over 1996–2015. Within the framework of y = 𝛾 0 + 𝛾 b b + 𝛾 k k,
an aggregate production function estimation, weak evidence is found
on the productivity effect of public infrastructure in China. This finding where y = log(Y/L), b = log(B/L), k = log(K/L) and 𝛾 0 = log(A0 ). In
suggests the necessity of using alternative identification strategies or this equation, 𝛾 b and 𝛾 k are the output of elasticities of public infras-
data types, e.g., a disaggregation approach using firm-level data, such tructure and non-infrastructure capital. The economic return of public
as Fisher-Vanden et al. (2015); Li et al. (2017); and Wu et al. (2017). infrastructure, or the marginal output of public infrastructure, can be
The rest of the paper is organized as follows. Section 2 introduces a measured as
macroeconometric model using an aggregate production function, aug-
𝜕 Y ∕𝜕 B = 𝛾b Y ∕B.
mented with public infrastructure capital. Various strategies of dealing
To estimate the coefficients 𝛾 b , 𝛾 k , a panel data model based on the
aggregate production function above is used
2
This rate is calculated using the data from the website of National Bureau
of Statistics of China. Also see Fig. 14.3 of Naughton (2007) for the ratios of yit = 𝛾0 + 𝛾b bit + 𝛾k kit + 𝜇i + Tt +𝜀it , (4)
physical infrastructure investment to GDP during 1981–2004.
3
where y it is the logarithm of GDP per labor in province i in year t, and
There are several other studies on China’s infrastructure in the literature.
bit is the logarithm of public infrastructure stock per labor, and kit is
Shi et al. (2017) incorporate a CES production function in Mankiw et al.
(1992) model, and estimate the relationship between infrastructure and eco-
the logarithm of non-infrastructure capital stock per labor. 𝜇 i denotes
nomic growth in a vector error correction model using a panel data set of province specific factors, such as different land area, location, weather,
China’s 30 provinces over 1990–2013. Lin and Song (2002) obtain a signifi- endowments of raw materials and myriad other factors. Time effects T t
cant OLS estimate of output elasticity of city infrastructure above 0.102 in a can be used to control for national-level macro shocks, including busi-
cross-section regression of the relationship between per capita GDP growth and ness cycles and counter-cyclic policies. 𝜀it denotes idiosyncratic shocks
investment, foreign direct investment, labor force growth, government expen- or measurement error in output. To deal with the non-stationarity in
diture and urban infrastructure using a data set of 189 large and medium-sized macroeconomic variables, first-differencing Eq. (4) gives our estimat-
Chinese cities for the period 1991–1998. Ward and Zheng (2016) estimate the ing equation:
contribution of telecommunications services to economic growth using a panel
data set of 31 Chinese provinces over the period from 1991 to 2010. To address Δyit = 𝛾b Δbit + 𝛾k Δkit + ΔTt + Δ𝜀it . (5)
the concern of reverse causality between telecommunications and per capita
growth, system GMM estimators combined with external instruments are used
in a dynamic panel data model. For a detailed survey on the effect of infras-
4
tructure on economic growth in China using aggregate level data, see Shi et Results without the CRS restriction are not reported here for the sake of
al. (2017). Wu et al. (2017) also provide an extensive discussion on the litera- space but are available upon request. Despite the small variations in the output
ture on the relationship between public infrastructure and economic growth in elasticities with and without the CRS restriction across various models, the main
China using disaggregate data. message obtained under the CRS restriction remains unchanged.

2
Q. Feng and G.L. Wu Economic Modelling xxx (2018) 1–8

3. Dealing with reverse causality as their response the GDP yardstick competition. Hence a province with
better growth prospects could expect to produce higher output and col-
When we write down Eq. (4) or (5), our aim is to identify the causal lect more fiscal revenue in the future, which in turn may allow the
effect of public infrastructure on output. However, as pointed out, e.g., province to invest more in current infrastructure via various financing
by Gramlich (1994), the causality could go from output to public infras- schemes. This would also imply a positive correlation between Δbit and
tructure. Higher output may mean greater demand for the services from Δ𝜀it in Eq. (5).
public infrastructure; higher output may also mean more income for It is a well-known fact that the 30 provinces in China are at dif-
expenditure on public infrastructure. Hence, a positive estimated elas- ferent levels of economic development, varying substantially in GDP
ticity could be mainly driven by this reverse causality. Thus, the OLS per capita, public facilities and fiscal budget (Naughton, 2007). Hence,
estimator of 𝛾 b in (5) (i.e., the first difference (FD) estimator of (4)) over a relative long span of time, such positive correlation generated by
could be biased upward. Alternatively, in the literature as summarized financing abilities cross provinces could overpower the negative corre-
by Bom and Ligthart (2014), due to the Keynesian multiplier effect, lation between output and public infrastructure due to the short-run
public infrastructure investment is often used to boost economic growth countercyclical story or national policies to reduce regional disparity.
during the period of economic recession. In this case, output and public Therefore, after including time effects in Eq. (5) to mitigate the effect
infrastructure investment could be negatively correlated. Thus, the OLS of national-level countercyclical policies, we conject that the upward
estimator of 𝛾 b in (5) (i.e., the first difference (FD) estimator of (4)) bias due to the reverse causality is more likely when estimating output
could be biased downward. elasticity of public infrastructure 𝛾 b in (5).
In the literature, there are several ways to deal with this endogene- In this paper, we employ several ways to deal with or mitigate the
ity issue due to reverse causality. The first and general approach is the endogeneity issue due to reverse causality. The first approach is to use
instrumental variable (IV) estimation, e.g., Holtz-Eakin (1994), Baltagi an alternative measure of investment in fixed assets reported by the
and Pinnoi (1995) and the more recent literature surveyed in Redding National Bureau of Statistics of China (NBS): Newly Increased Fixed
and Turner (2015). An alternative way to address the reverse causality Assets (NIFA hereafter) (xinzeng guding zichan touzi in Chinese). Differ-
is the simultaneous-equations approach, explicitly modeling the rela- ent from the usual measure of investment to construct public infras-
tionship between y and b in an additional equation, such as Röller tructure capital and non-infrastructure capital in (4), Total Investment
and Waverman (2001) and Cadot et al. (2006). Another approach is in Fixed Assets (TIFA hereafter) (quanshehui guding zichan touzi in Chi-
to explore the heterogeneity of output effect from disaggregated data. nese), which measures total cost spent on constructing and purchasing
A leading example is Fernald (1999). Recently, Calderon et al. (2015) fixed assets, NIFA measures investment in fixed assets that have been
use a panel cointegration approach to deal with the nonstationarity used for production after the process of construction and purchase is
and establish only one cointegrating relation to address concerns with completed.6 Due to the time to build, NIFA is less likely to be affected
reverse causality in a panel data set with a long span of time periods. by the current output. Thus, the reverse causality between output and
In the Chinese context, Shi and Huang (2014) claim that the reverse public infrastructure (or non-infrastructure) capital is mitigated. 7
causality could lead to a negative correlation between output and pub- We also make use of a measure of bit in the level Eq. (4) (or Δbit
lic infrastructure since “Chinese government tends to use infrastruc- in the differenced Eq. (5)) that is less likely to be affected by y it (or
ture investment as a choice for reviving its economy when it expects Δyit ). A natural candidate in the literature is the lagged value of bit (or
a large negative TFP shock”, which will bias downward the estimated Δbit ). Different from bit (or Δbit ), bit−1 (or Δbit−1 ) is less likely to be
output elasticity of infrastructure. In their paper, the endogeneity due affected by y it (or Δy it ) under the assumption that the current output
to reverse causality is interpreted as the negative correlation between only affects the current and future, instead of the past, values of public
Δbit and Δ𝜀it , where this correlation is dealt with by the proxy approach infrastructure. As a stock variable accumulating all past public infras-
developed by Ackerberg et al. (2015). tructure investments, bit −1 still provides service to future production.
Different from Shi and Huang (2014), we argue that regarding the As a general approach to deal with endogeneity, instrumental vari-
feedback effect of output on public infrastructure, a positive correlation able estimation is also used to consistently estimate 𝛾 b . In this paper,
is more likely to dominate in the case of China. Bai and Qian (2010) pro- three different sets of instruments are explored. First, as in Holtz-Eakin
vide an interesting survey on the specific institutional background for (1994), twice-lagged variables Δbit −2 and Δkit −2 are employed as inter-
infrastructure investment in China. Two stylized facts can be summa- nal instruments for Δbit and Δkit in Eq. (5).8 Second, as widely doc-
rized from the survey. First, most infrastructure investment are made by umented in the literature one of distinctive institutional features of
state-owned enterprises with funds from both the central and the local China’s economic miracle is that under the so-called “GDP tournament”
governments. Second, among various jurisdiction levels, the provincial scheme local governments have been playing an active role in pro-
governments play a key role in infrastructure investment decision. Wu moting economic growth, including investing in infrastructure (Li and
et al. (2017) survey several hypotheses on the investment incentives of Zhou, 2005; Jin et al., 2005; Wang et al., 2017). Under this scheme,
the Chinese governments that have been discussed in the literature. In local governments compete with each other on GDP growth, and their
short, for the central government, first, infrastructure development is investment behavior could affect each other. Thus, Δbit in neighboring
needed to fight against the worsening regional inequality by promoting
the catch-up of lagging inland provinces with coastal provinces. This
would imply a negative correlation between bit and 𝜇 i in Eq. (4) and
can be eliminated by first differencing as in Eq. (5).5 Second, infrastruc- 6
Ozyurt (2009) uses NIFA as a measure of effective investment in a study of
ture development is necessary to support the rapid economic growth of estimating China’s aggregate production function using time series data cover-
the country that fuels an ever-increasing demand for infrastructure ser- ing 1952–2005.
7
vices. This would imply a positive correlation between Δbit and Δ𝜀it in NIFA is not a formal measure of investment reported by NBS. It is reported
Eq. (5). Finally, for the provincial governments, under China’s region- to show the extent of how investment process in fixed assets has been com-
ally decentralized authoritarian system, infrastructure investment has pleted in some years and some sectors. Since the data on NIFA are not available
been adopted as the most effective instrument by the local governments before 2002, TIFA is used as a formal measure of investment throughout the
paper. We construct the data of NIFA before 2002 by using the components of
basic construction and renovations of NIFA and their ratios in provinces and
industries in China Statistics Yearbooks.
5 8
When infrastructure investment is used to reduce regional inequality at the Δbit −1 and Δkit −1 could be correlated with Δ𝜀it . It is worth noting that this
growth of output, instead of the level of output, Δbit and Δ𝜀it could be nega- IV approach is different from using Δbit −1 and Δkit −1 as regressors in the FD
tively correlated, as in Shi and Huang (2014). regression above.

3
Q. Feng and G.L. Wu Economic Modelling xxx (2018) 1–8

Table 1
Summary statistics of variables.
Symbol Definition Unit Mean Std. D. Form in regression Data sources
y real output per labor 10,000 yuan 2.38 1.79 log China NBS Website
b real infrastructure capital per labor 10,000 yuan 1.17 0.92 log China NBS Website
k real non-infrastructure capital per labor 10,000 yuan 5.03 4.84 log China NBS Website
newb real infrastructure capital per labor based on NIFA 10,000 yuan 0.69 0.51 log China NBS Website
nb real infrastructure capital per labor in neighboring provinces 1.04 0.72 log authors’ calculation
G infrastructure investment flow 100 million yuan 674 624 China NBS Website
L number of labor force 10,000 3080 1847 China NBS Website
age 1 age of provincial governor 57.9 4.0 level Wikipedia, baike.baidu.com
age 2 age of provincial party leader 59.7 4.1 level Wikipedia, baike.baidu.com
Notes.
1. All variables are measured in provincial level.
2. Units and summary statistics of all variables are reported before taking log.

provinces, denoted as Δnbit , can serve as an instrument for Δbit .9 A assets by industry and by province. Infrastructure investment G is mea-
recent study by Zheng et al. (2015) finds that infrastructure spending sured by the sum of investments in the 3 industries: (1) production and
in a province is positively correlated with infrastructure spending in its supply of electricity, gas and water; (2) transport, storage and post;
neighboring provinces. In addition, since Δy it is only affected by Δbit (3) information transmission, computer services and software.11 I is
and Δkit conditional on time dummies in Eq. (5), instruments of Δbit −2 defined as total investment minus G. Stock variables of B and K are
and Δnbit have no direct effect on Δy it . They affect Δy it only through constructed as in (2) and (3) using depreciation rates 𝛿 b = 𝛿 k = 10%.12
Δbit . Table 1 reports the summary statistics for the variables used in
Third, we use the ages of provincial governors and party leaders as the analysis. GDP, public infrastructure investment, non-infrastructure
external instruments for public infrastructure in (5). In China’s current investment are deflated by the province-specific price indices of invest-
political system, provincial governors and party leaders retire at an age ment in fixed assets.13 The unit, mean and standard deviation for the
of 65 if they are not promoted to top-level officials in Chinese central real output per labor, real public infrastructure and non-infrastructure
government. Given that GDP growth is the most important key perfor- capital stocks per labor and other variables before taking logarithms are
mance indicator and that investment is one of the major contributing reported. These variables are used in the log form in regressions, so that
factors of GDP growth, provincial governors and party leaders are less the corresponding coefficients can be interpreted as elasticities.
motivated to invest when their ages are closer to 65.10 In this case, We first report estimation results on elasticities 𝛾 b and 𝛾 k without
the ages of provincial governors and party leaders could be negatively dealing with reverse causality. Column (1) of Table 2 reports fixed
correlated with public infrastructure investment. In terms of exclusion effects (FE) estimates of 𝛾 b and 𝛾 k , which are 0.057 and 0.303, respec-
restriction, like instruments of twice-lagged variables and neighboring tively. To eliminate unit roots and common trends in the macro data,
public infrastructure, the ages of provincial governors and party leaders first-differencing is needed. Column (2) presents FD estimates, showing
are considered to be irrelevant to output (or growth) in the aggregate that the estimated elasticity of public infrastructure capital is 0.127 and
production function (4) (or (5)). significant at 1% level.14 Considering that the return of public infras-
The empirical results using the identification strategies above are
reported in Section 4 below. Using a Chinese provincial panel data 11 The definition of G here is consistent with the description of physical infras-
set during 1996–2015, we show that after dealing with the endogene-
tructure in Fig. 14.3 of Naughton (2007) for China, and the literature in general,
ity issue due to reverse causality, the estimated output elasticities are
e.g., Calderon et al. (2015). Shi and Huang (2014) also include investment in
notably smaller than the FD estimates, suggesting that an upward bias management of water conservancy, environment, and public facilities as part
due to reverse causality is prevalent in China’s case. of public infrastructure investment. When we broaden the definition of infras-
tructure as in Shi and Huang (2014) in robustness checks, we obtain similar
4. Data and empirical results findings as in our benchmark results.
12
The choice of depreciation rate in the literature typically varies between
Data on GDP (Y) are obtained from the website of National Bureau 3% and 16%. Thus we set 10% as our benchmark depreciation rate and con-
ducted robustness checks using other rates as alternatives. The main finding of
of Statistics of China. We collect data for 30 provinces excluding Tibet
our empirical exercise turns out to be not sensitive to the depreciation rate. To
over years 1996–2015. As in Shi and Huang (2014), the size of labor
implement the perpetual inventory method, one has to start with an initial value
force (L) is calculated by number of residents multiplied by the ratio of for Bit and Kit . In our application, we assume that Bi1996 = Gi1996 /(𝛿 b + g) and
age cohort of 16–65. For the key variables public infrastructure invest- Ki1996 = I i1996 /(𝛿 k + g), where g = 10%, the average long-run growth rate dur-
ment (G) and non-infrastructure investment (I), we collect data on the ing our sample period. This assumption is based on the property of a balanced-
total investment in fixed assets (TIFA) from Statistical Yearbooks of growth-path model, in which new investment is made to compensate deprecia-
The Chinese Investment in Fixed Assets and China Statistical Yearbooks. tion and guarantee a constant growth in capital stock.
13
These two series of statistics yearbooks report total investment in fixed According to The Chinese Statistic Yearbook, the investment in fixed assets
consists of three components, namely the investment in construction and instal-
lation, the investment in purchases of equipment and instrument, and the
investment in other items. Price indices of investment in fixed assets are cal-
9
We define a province as a neighboring province of i if it shares com- culated as the weighted arithmetic mean of the price indices of the three com-
mon border of province i. For examples, the neighboring provinces of Shang- ponents of investment in fixed assets. Under our definition, both infrastruc-
hai are Jiangsu and Zhejiang, and Jiangxi’s neighbors are Zhejiang, Anhui, ture and non-infrastructure investment contain investment in all three com-
Hubei, Hunan, Fujian and Guangdong provinces. nbit is defined as the log (sum ponents. Without knowing the exact proportion of each component, we apply
of infrastructure stock in neighboring provinces/sum of labor in neighboring the price indices of investment in fixed assets to both infrastructure and non-
provinces). The instrument used is its first difference. infrastructure investment.
10 14
A similar argument can be seen in Li and Zhou (2005), Wang et al. (2017), Clustered standard errors are reported in parenthesis below estimates,
in which age is an important factor for the career concerns of provincial leaders. adjusted for 30 clusters in province.

4
Q. Feng and G.L. Wu Economic Modelling xxx (2018) 1–8

Table 2
Output elasticities: Fixed-effects and first-differenced estimates.
Dependent variable: Output per labor
Independent variables: FE FD FDnew FDlag
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Infrastructure capital per labor 0.057 0.127∗∗∗ 0.144∗∗∗ 0.088∗∗ 0.037∗ 0.033 0.035 0.005 −0.030
(0.07) (0.03) (0.03) (0.04) (0.02) (0.02) (0.03) (0.04) (0.03)
Non-infrastructure capital per labor 0.303∗∗∗ 0.324∗∗∗ 0.340∗∗∗ 0.315∗∗∗ 0.228∗∗∗ 0.250∗∗∗ 0.210∗∗∗ 0.215∗∗∗ 0.402∗∗∗
−0.04 −0.03 −0.04 −0.02 −0.02 −0.03 −0.02 −0.04 −0.03
Periods All All 1996–2007 2008–2015 All 1996–2007 2008–2015 All All
Year effects Yes Yes Yes Yes Yes Yes Yes Yes Yes
Overall R2 0.84 0.73 0.76 0.67 0.60 0.63 0.54 0.45 0.71
No. of observations 599 569 329 240 569 329 240 539 539
Notes.
1. FE and FD in columns (1)–(4) denote fixed effects regression and first difference regression, respectively.
2. FDnew in columns (5)–(7) refer to the first-difference estimates using data based on newly increased fixed asset investment.
3. FDlag in columns (8) refer to the first difference estimates using the lags of both public infrastructure and non-infrastructure capital. In column (9) only the
lagged value of public infrastructure capital is used.
4. Standard errors are reported in parentheses. The stars, ∗, ∗∗ and ∗∗∗ indicate the significance level at 10%, 5% and 1%, respectively.
5. Standard errors are adjusted for 30 clusters in province.
6. Depreciation rate 10% is used to calculate public infrastructure and non-infrastructure capital stocks.
7. For the definition, unit of variables and data sources, please refer to Table 1.

Table 3
Output elasticities: Instrumental variable estimates.
Independent variables: FD IV1 FD IV2 FD IV3
(1) (2) (3) (4) (5) (6)
Infrastructure capital per labor −0.095 −0.050 −0.063 −0.098 0.059 0.140
(0.09) (0.14) (0.10) (0.19) (0.21) (0.15)
Non-infrastructure capital per labor 0.332∗∗∗ 0.210∗∗∗ 0.370∗∗∗ 0.333∗∗∗ 0.258∗∗∗ 0.220∗∗∗
(0.04) (0.08) (0.05) (0.11) (0.09) (0.06)
Periods All 1996–2007 2008–2015 All All All
Year effects Yes Yes Yes Yes Yes Yes

Overall R2 0.62 0.60 0.63 0.62 0.70 0.71

No. of observations 509 269 240 509 509 509

Instruments Δbt-2 , Δkt-2 𝛥nbt , Δkt-2 age1, Δkt-2 age1, age2, Δkt-2
1st-stage regression coefficient 0.337 0.356 0.345 0.272 −0.001 −0.001
1st-stage t-ratio (6.94) (4.93) (5.36) (3.67) (-2.25) (-2.11)
Sargan test (p-value) 0.46
Notes.
1. FD IV denotes first difference instrumental variable regression.
2. Depreciation rate 10% is used to calculate the capital stocks.
3. Standard errors are reported in parentheses. The stars, ∗, ∗∗ and ∗∗∗ indicate the significance level at 10%, 5% and 1%, respectively.
4. Standard errors are adjusted for 30 clusters in province in columns (1)–(6).

tructure capital is 𝜕 Y/𝜕 B = 𝛾 b Y/B and Y/B = 2.254 averaging over big drop in estimated output elasticity of infrastructure suggests that
1996–2015 for depreciate rates 𝛿 b = 𝛿 k = 10% in the sample, this elas- an upward bias is more likely than a downward bias in the FD esti-
ticity indicates a return rate of 28.6%. This means that investment in mate in column (2), and that a positive productivity effect of public
public sectors is very productive and profitable. To examine the change infrastructure capital could be driven in part by the positive feedback
of return over time, FD estimates using subsamples are also reported effect of output on public infrastructure. In the subsample estimates of
in columns (3) and (4), 0.144 and 0.088 for periods of 1996–2007 and columns (6) and (7) of Table 2, similar estimated elasticities 𝛾 b and 𝛾 k
2008–2015, respectively. This implies rates of return to public infras- are shown.
tructure capital of 0.144 × 2.394 = 34.5% and 0.088 × 2.043 = 18.0%, We also use the lagged values of Δbit (and Δkit ), instead of the cur-
respectively. rent values, to reduce the feedback effect of y on b. The resulting FD
However, due to the reverse causality discussed above, FD estimates estimates using the lagged values of Δbit and Δkit , labelled by FDlag,
could be upward or downward biased. To mitigate this issue, first, we are reported in column (8) of Table 2. Completely different from FD
use an alternative measure of public infrastructure capital based on estimate of 𝛾 b in column (2), after mitigating reverse causality, the
NIFA, which is less likely affected by y. Column (5) of Table 2 displays FDlag estimate of 𝛾 b drops to 0.005 and insignificant. Though using
the FD estimates using this new measure, labelled by FDnew. Consistent the lagged value may weaken the direct impact of infrastructure on
with the discussion above, after weakening the positive linkage from y output, the sharp difference in estimated 𝛾 b between columns (2) and
to b (and k), the estimated elasticity of public infrastructure capital of (8) suggests that the big positive elasticity of public infrastructure cap-
𝛾 b becomes less significant and falls markedly to 0.037 from 0.127 in ital in column (2) could be overestimated due to the positive feedback
column (2), with a rate of return of 0.037 × 3.707 = 13.7%. Such a effect of output on public infrastructure. By contrast, the FDlag estimate

5
 Q. Feng and G.L. Wu
Table 4
Output elasticities: Robustness checks.
Dependent variable: Output per labor
Independent variables: A: Depreciation rates 𝛿 b = 4%, 𝛿 k = 10% B: year-end employment C: FE on Differenced data
FD FDnew FDlag FDIV1 FDIV2 FD FDnew FDlag FDIV1 FDIV2 FE FEnew FElag FE IV1 FE IV2
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15)
Infrastructure capital 0.182∗∗∗ 0.075∗∗ −0.003 −0.089 −0.144 0.156∗∗∗ 0.047∗ 0.051 0.096 −0.242 0.167∗∗∗ 0.039∗ −0.009 −0.227 0.228
per labor (0.031) (0.030) (0.046) (0.10) (0.22) (0.04) (0.03) (0.05) (0.08) (0.41) (0.02) (0.02) (0.02) (0.14) (0.18)
Non-infrastructure 0.301∗∗∗ 0.211∗∗∗ 0.219∗∗∗ 0.326∗∗∗ 0.351∗∗∗ 0.357∗∗∗ 0.257∗∗∗ 0.175∗∗∗ 0.252∗∗∗ 0.399∗ 0.378∗∗∗ 0.238∗∗∗ 0.137∗∗∗ 0.245∗∗∗ 0.188∗∗∗
capital per labor (0.028) (0.024) (0.035) (0.04) (0.12) (0.05) (0.04) (0.03) (0.04) (0.20) (0.04) (0.03) (0.04) (0.07) (0.06)
Regions All All All All All All All All All All All All All All All
Periods All All All All All All All All All All All All All All All
Year effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

Overall R2 0.74 0.61 0.45 0.63 0.60 0.69 0.55 0.40 0.68 0.43 0.72 0.60 0.43 0.35 0.69
6

No. of observations 569 569 539 509 509 569 569 539 509 509 569 569 539 509 509

Instruments Δbt-2 ,Δkt-2 𝛥nbt ,𝛥kt-2 Δbt-2 , Δkt-2 𝛥nbt ,𝛥kt-2 Δbt-2 , Δkt-2 𝛥nbt ,𝛥kt-2
1st-stage regression 0.360 0.274 0.397 0.193 0.202 0.210
coefficient
1st-stage t-ratio (7.14) (3.81) (8.47) (2.70) (4.08) (2.42)
Notes.
1. Panel A: depreciation rates of 4% and 10% are used to construct the capital stocks. Definitions of FD, FDnew, FDlag, FDIV1 and FDIV2 remain as in Tables 2
and 3
2. Panel B: year-end employment is used to measure the labor force. Depreciate rates of 10% remain as in Tables 2 and 3
3. Panel C: FE, FEnew, FElag, FEIV1 and FEIV2 refer to fixed effects estimates using differenced data and those using newly increased fixed asset investment,
lags of public infrastructure and private capital stocks, instruments of lagged values and neighboring public infrastructure, respectively.
4. Standard errors are reported in parentheses. The stars, ∗, ∗∗ and ∗∗∗ indicate the significance level at 10%, 5% and 1%, respectively.
5. Standard errors are adjusted for 30 clusters in province in columns (1)–(15).

Economic Modelling xxx (2018) 1–8

Q. Feng and G.L. Wu Economic Modelling xxx (2018) 1–8

Table 5
Output elasticities: Additional robustness checks.
Dependent variable: Output per labor
Independent variables: D: Alternative IV2 E: Broad Definition of Infrastrstructure F: Subsample of Eastern Region
FD IV2 FD FDnew FDlag FDIV1 FDIV2 FD FDnew FDlag FDIV1 FDIV2
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Infrastructure capital 0.077 0.140∗∗∗ 0.058∗∗ 0.03 −0.055 −0.160 0.107∗∗ 0.034 −0.021 −0.409 −0.125
per labor (0.15) (0.03) (0.02) (0.04) (0.08) (0.46) (0.038) (0.031) (0.053) (0.27) (0.31)
Non-infrastructure 0.250∗∗∗ 0.306∗∗∗ 0.208∗∗∗ 0.201∗∗∗ 0.337∗∗∗ 0.412 0.321∗∗∗ 0.215∗∗∗ 0.261∗∗∗ 0.500∗∗∗ 0.363∗
capital per labor (0.08) (0.03) (0.02) (0.04) (0.05) (0.33) (0.033) (0.027) (0.033) (0.12) (0.17)
Regions All All All All All All Eastern Eastern Easter Eastern Eastern
Periods Yes All All All All All All All All All All
Year effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

Overall R2 0.70 0.73 0.60 0.45 0.66 0.59 0.76 0.64 0.52 0.34 0.67
No. of observations 509 569 569 539 509 509 209 209 198 187 187

Instruments 𝛥nbt ,Δkt-2 Δbt-2 ,Δkt-2 𝛥nbt ,𝛥kt-2 Δbt-2 ,Δkt-2 𝛥nbt ,𝛥kt-2
1st-stage regression 0.256 0.381 0.185 0.182 0.287
coefficient
1st-stage t-ratio (4.61) (7.30) (2.54) (2.18) (2.40)
Notes.
1. Panel D: an instrument based on a new measure of infrastructure investments in neighboring provinces, defined as their GDP competitors instead of their
geographic neighbors.
2. Panel E: an alternative definition of infrastructure by including investments in industries related to management of water conservancy, environment, and
public facilities.
3. Panel F: subsample of eastern region is used.
4. Definitions of FD, FDnew, FDlag, FDIV1 and FDIV2 remain as in Tables 2–4
5. Standard errors are reported in parentheses. The stars, ∗, ∗∗ and ∗∗∗ indicate the significance level at 10%, 5% and 1%, respectively.
6. Standard errors are adjusted for 30 clusters in province in columns (1)–(6). Robust standard errors are used in columns (7)–(11).

of non-infrastructure capital elasticity 𝛾 k is still of a big magnitude of and insignificant.

0.215 and significant, though decreasing from 0.324 in column (2). To Unlike 𝛾 b , the corresponding estimates of 𝛾 k in columns (1), (4) and
further confirm the effect of reverse causality on estimating 𝛾 b , Column (5) in Table 3, 0.332, 0.333 and 0.258, are still positive and significant,
(9) gives FD estimates using the lagged value of Δbit and current value and are comparable with the FD estimates in Table 2.17 Thus, the dif-
of Δkit . Same pattern remains as in column (8). ference between 𝛾 b and 𝛾 k indicates the different roles that the public
Table 3 reports IV estimates of elasticities using instruments infrastructure and non-infrastructure capital play in the aggregate pro-
of twice-lagged variables (FDIV1), neighboring public infrastructure duction function. Public infrastructure is more likely positively affected
(FDIV2) and the ages of provincial governors and party leaders by the output than non-infrastructure capital.
(FDIV3), respectively. Non-infrastructure capital Δkit is also considered Three robustness checks are reported in Table 4: using depreciate
as endogenous and instrumented by Δkit −2 . The estimates of public cap- rates 𝛿 b = 4%, 𝛿 k = 10% in panel A,18 replacing calculated labor force
ital elasticity using the full sample are −0.095 and −0.098 in columns with year-end employment reported by NBS in panel B, and running
(1) and (4), respectively. Similar to FDnew and FDlag estimates Table 2, fixed effects estimation on differenced data instead of pooled OLS in
after dealing with the reverse causality between y and b (and k), the FD panel C. In panels A and B, we report FD and 4 estimates using NIFA
IV estimates of output elasticity public infrastructure drop to small neg- (FDnew), lagged variables (FDlag) and two internal instruments Δbit −2 ,
ative numbers, and are no longer statistically significant from 0. The FD Δkit−2 (FDIV1) and Δbit in neighboring provinces (FDIV2), correspond-
IV estimates of 𝛾 b using external instruments of the ages of provincial ing to columns (5), (8) of Table 2 and columns (1), (4) of Table 3,
governors alone and both ages of provincial governors and party leaders respectively. In panel C, FE and FE estimates using NIFA (FEnew),
are 0.059 and 0.140 in columns (5) and (6), respectively. Both are pos- lagged variables (FElag) and 2 sets of instruments (FEIV1, FEIV2) are
itive and of a big magnitude, but statistically insignificant.15 Columns presented in columns (11)–(15), respectively. Consistent with the mes-
(2)–(3) also give FDIV1 estimates using subsamples in the periods of sage delivered by Tables 2 and 3, estimates of 𝛾 b in columns (2)–(5),
1996–2007 and 2008–2015.16 The estimates of 𝛾 b are small, negative, (7)–(10), (12)–(14) decrease substantially after reverse causality is
taken into consideration. In column (15), using differenced data the
fixed effects IV estimate of 𝛾 b is 0.228 but insignificant. No robust pat-
tern of a big positive and significant estimates of 𝛾 b are found in various
15 cases, sharply constrasted with the estimates of 𝛾 k .
The first-stage regression results of regressing instruments for Δb on exoge-
nous variables in Eq. (5) are reported in the last three rows in Tables 3 and 4
Table 5 shows the results of three additional robustness checks. First,
For instruments of instruments of twice-lagged variables and neighboring public
infrastructure, both are very informative. The magnitude of instrument of age
17
of provincial governors (age1) is small but still statistically significant. Sargan The mean value for the ratio Y/K is 0.624 during our sample period. Thus
test for overidentification is conducted in column (6) of Table 3. No evidence the output elasticities of non-infrastructure capital from Tables 2 and 3 indicate
shows that instruments of age1 and age2 are invalid. a rate of return around 20%. This number is close to the results reported by Bai
16
The year of 2008 as the cutoff point is used because a 4 trillion Chinese and Zhang (2014).
18
Yuan fiscal stimulus package was introduced by the Chinese government to We also conduct robustness checks using other different depreciate rates,
invest mainly in the infrastructure in its western provinces in 2008. This event including combinations of i) 𝛿 b = 5%, 𝛿 k = 10%; ii) 𝛿 b = 15%, 𝛿 k = 15%; iii)
could lead to different output elasticities of infrastructure before and after 2008. 𝛿 b = 10%, 𝛿 k = 15%; iv) 𝛿 b = 15%, 𝛿 k = 10%. Main results remain.

7
Q. Feng and G.L. Wu Economic Modelling xxx (2018) 1–8

panel D employs an alternative measure of infrastructure investments Appendix A. Supplementary data

in neighboring provinces, defined as their GDP competitors instead of
their geographic neighbors used in column (4) of Table 3.19 The esti- Supplementary data related to this article can be found at https://
mated output elasticity of infrastructure becomes 0.077 and statistically doi.org/10.1016/j.econmod.2018.05.006.
insignificant. Second, in Panel E we consider an alternative definition
of infrastructure by including investments in industries related to man-
agement of water conservancy, environment, and public facilities, i.e.,
the fourth category of infrastructure investment considered in Shi and References
Huang (2014). As in Table 4, FD, FDnew, FDlag, FDIV1 and FDIV2
estimates are reported in columns (2)–(6). Third, considering China’s Ackerberg, D.A., Caves, K., Frazer, G., 2015. Identification properties of recent
geographic heterogeneity and different economic development across production function estimators. Econometrica 83 (6), 2411–2451.
Aschauer, D., 1989. Is public expenditure productive? J. Monetary Econ. 23, 177–200.
regions, we split the sample into 3 groups: eastern, central and western Bai, C., Qian, Y., 2010. Infrastructure development in China: the cases of electricity,
regions. Panel F presents 5 estimates as in panel E. As in Table 4, the highways, and railways. J. Comp. Econ. 38 (1), 34–51.
same pattern emerges. Once the reverse causality between the output Bai, C., Zhang, Q., 2014. Returns to capital and their determinants in China. World
Econ. 10, 2–30.
and infrastructure is mitigated, the estimates of 𝛾 b decrease remark- Baltagi, B., Pinnoi, N., 1995. Public capital stock and state productivity growth: further
ably and become statistically insignificantly in most cases. This evi- evidence from an error components model. Empir. Econ. 20, 351–359.
dence suggests that the reverse causality may lead to an upward bias. Bom, P., Ligthart, J., 2014. What have we learned from three decades of research on the
productivity of public capital? J. Econ. Surv. 28 (5), 889–916.
Cadot, O., Röller, L., Stephan, A., 2006. Contribution to productivity or pork-barrel? The
5. Conclusion two faces of infrastructure investment. J. Publ. Econ. 90, 1133–1153.
Calderon, C., Moral-Benito, E., Serven, L., 2015. Is infrastructure capital productive? A
This paper is motivated by the question whether infrastructure dynamic heterogeneous approach. J. Appl. Econom. 30, 177–198.
Fernald, J., 1999. Roads to prosperity? Assessing the link between public capital and
investment contributes to productivity gains and economic growth in productivity. Am. Econ. Rev. 89 (3), 619–638.
China. We address this issue in the framework of an aggregate produc- Fisher-Vanden, K., Mansur, E.T., Wang, Q., 2015. Electricity shortages and firm
tion function, in which public infrastructure capital is modelled as a productivity: evidence from China’s industrial firms. J. Dev. Econ. 114, 172–188.
Gramlich, E., 1994. Infrastructure investment: a review essay. J. Econ. Lit. 32,
contributing factor of TFP, and a panel data set of 30 Chinese provinces 1176–1196.
during 1996–2015 is used to estimate the output elasticities of public Holtz-Eakin, D., 1994. Public-sector capital and the productivity puzzle. Rev. Econ. Stat.
infrastructure and non-infrastructure capital stocks. In such a frame- 76, 12–21.
Jin, H., Qian, Y., Weingast, B., 2005. Regional decentralization and fiscal incentives:
work, the main identification problem is the reverse causality between federalism, Chinese style. J. Publ. Econ. 89, 1719–1742.
the output and public infrastructure investment, which could lead to an Li, Z., Chen, Y., 2013. Estimating the social return to transport infrastructure: a
upward or downward bias. price-difference approach applied to a quasi-experiment. J. Comp. Econ. 41,
669–683.
In this empirical study, we proposed several different ways to mit-
Li, H., Li, Z., 2013. Road investments and inventory reduction: firm level evidence from
igate the reverse causality. Unlike Shi and Huang (2014), we find that China. J. Urban Econ. 76, 43–52.
an upward bias dominates when estimating output elasticity of pub- Li, H., Zhou, L.-A., 2005. Political turnover and economic performance: the incentive
role of personnel control in China. J. Publ. Econ. 89, 1743–1762.
lic infrastructure in China’s context. After controlling for the reverse
Li, Z., Wu, M., Chen, B., 2017. Is road infrastructure investment in China excessive?
causality between the GDP growth and public investment, we find weak Evidence from productivity of firms. Reg. Sci. Urban Econ. 65, 116–126.
evidence of a big positive productivity effect of public infrastructure Lin, S., Song, S., 2002. Urban economic growth in China: theory and evidence. Urban
within the framework of an aggregate production function. Stud. 39, 2251–2266.
Mankiw, N.G., Romer, D., Weil, David N., 1992. A contribution to the empirics of
This, of course, does not deny the possibility that public infrastruc- economic growth. Q. J. Econ. 107 (2), 407–437.
ture investment may play an important role in economic growth and Naughton, B., 2007. The Chinese Economy: Transitions and Growth. MIT Press.
development. As surveyed by Gramlich (1994), Shi and Huang (2014) Ouyang, M., Peng, Y., 2015. The treatment-effect estimation: a case study of the 2008
economic stimulus package of China. J. Econom. 188, 545–557.
and Calderon et al. (2015), there are other econometric issues that are Ozyurt, S., 2009. Total factor productivity growth in Chinese industry: 1952-2005. Oxf.
not discussed in the short note. Instead, what we want to highlight here Dev. Stud. 37 (1), 1–17.
is the challenge of identifying the productivity effect of public infras- Redding, S.J., Turner, M.A., 2015. Transportation costs and the spatial organization of
economic activity. Handb. Reg. Urban Econ. 5, 1339–1398.
tructure investment in the aggregate production function estimation Röller, L., Waverman, L., 2001. Telecommunications infrastructure and economic
development: a simultaneous approach. Am. Econ. Rev. 91 (4), 909–923.
19 For example, the neighbors of Jiangsu, the ranked 2nd in 2016, are Guang-
Shi, H., Huang, S., 2014. How much infrastructure is too much? A new approach and
evidence from China. World Dev. 56, 272–286.
dong and Shandong. The information on Chinese provinces GDP ranking 2016 is Shi, Y., Guo, S., Sun, P., 2017. The role of infrastructure in China’s regional economic
from Wikipedia: https://en.wikipedia.org/wiki/List_of_Chinese_administrative_ growth. J. Asian Econ. 49, 26–41.
divisions_by_GDP Wang, Z., Zhang, Q., Zhou, L.-A., 2017. To Build Outward or Upward: the Spatial
Pattern of Urban Land Development in China. Working Paper.
framework. Dealing with reverse causality is of the first order impor- Ward, M., Zheng, S., 2016. Mobile telecommunications service and economic growth:
tance, and it is difficult to find good external instruments due the nature evidence from China. Telecommun. Pol. 40 (2–3), 89–101.
Wu, G.L., Feng, Q., Wang, Z., 2017. Estimating Productivity of Public Infrastructure
of aggregate data. This difficulty suggests the unique value of using Investment. Working Paper.
alternative identification strategies or data types, e.g., a disaggregation Zheng, X., Song, F., Yu, Y., Song, S., 2015. Search of fiscal interactions: a spatial analysis
approach using firm-level data such as Fisher-Vanden et al. (2015); Li of Chinese provincial infrastructure spending. Rev. Dev. Econ. 19, 860–876.
et al. (2017); and Wu et al. (2017).