Industrial coal demand in China: A provincial analysis

Industrial Coal Demand in China: a Provincial Analysis

Cristina Cattaneoa , Matteo Manerab, Elisa Scarpac a Fondazione Eni Enrico Mattei, Italy, and University of Sussex, UK bUniversity of Milan-Bicocca, Italy, and IEFE-Bocconi, Milan, Italy cEdison Trading, Milan, Italy

Revised: April 2009

Abstract. The general concern on the environmental implications of the rising demand for coal registered in China during the last few years has induced considerable research effort to produce accurate forecasts of China’s energy requirements. Nevertheless, no previous study has modelled the coal demand in China at provincial level. The aim of this paper is twofold. First, we estimate and forecast the Chinese demand for coal using panel data disaggregated by provinces and accounting for spatial heterogeneity. Second, given the spatial nature of the data, we explicitly capture the spatial autocorrelation among provinces using spatial econometrics. In particular, we specify the Chinese industrial coal demand at provincial level with a fixed-effect spatial lag model and a fixed-effect spatial error model. The fixedeffect spatial lag model seems to better capture the existing interdependence between provinces. This model forecasts an average annual increase in coal demand to 2010 of 4 percent. Keywords. Energy demand in China; Coal demand in China; Chinese provinces; Panel data; Spatial econometrics; Forecasting. JEL classification. C23; E6; Q31; Q41. Acknowledgements. This paper is part of the research work being carried out by the International Energy Markets Unit at the Fondazione Eni Enrico Mattei. Previous versions were presented at the First International Workshop on Empirical Methods in Energy Economics, organised by ETH Zurich, Switzerland, August 29-30, 2008, at the Fondazione Eni Enrico Mattei, Milan, and at the Department of Statistics, University of Milan-Bicocca. We are very grateful to Andrea Bigano, Massimo Filippini, Marzio Galeotti, Alessandro Lanza, Anil Markandya, and Michael McAleer for helpful comments and suggestions. Corresponding author. Matteo Manera, Professor of Economics (Econometrics), Department of Statistics, University of Milan-Bicocca, Via Bicocca degli Arcimboldi 8, I-20126 Milano, Italy, e-mail: [email protected].

1

Industrial Coal Demand in China: a Provincial Analysis 1. Introduction The exceptional economic development of China has been traditionally the central issue for many economists, interested in understanding the causes of such an astonishing growth. The key ingredients of the Chinese economic performance have been deeply analysed, with the aim to evaluate China’s growth potential and its future directions. Only recently, however, has the interest shifted towards the analysis of the implications of such spectacular development, with particular attention to the impacts on the world energy economy. Since the onset of economic reforms, China has shown sustained economic growth at an average annual rate of 9.5 percent. This extraordinary growth rate has enabled the country to increase, by more than 10 times, both total GDP and per capita GDP during the last 25 years (WDI, 2006). This surge emerged after the Central Government set in 1978 the basis for shifting the system towards a socialist modernisation and towards a revitalised domestic economy, open up to outside word. Not surprisingly, this expansion, which came along with massive industrialisation, urbanisation and motorisation, has induced a substantial increase in energy consumption. In 2004, for example, the rate of growth of energy consumption reached 18.3 percent (APERC, 2007), while the current levels of energy consumption make China the largest energy user in Asia and the second largest in the world, second only to US. China’s energy structure has been traditionally dominated by coal, given the abundance of coal mines located in the country. For instance, around the half of 20th Century, coal was nearly the sole fuel for energy production in China, while nowadays other fuels enter the energy mix. Nevertheless, the share of coal remains extremely high and largely dominates the Chinese energy scenario. A substantial share of energy demand is absorbed by the heavy industry, which accounts for more than 80 percent of total final energy consumption. Actually, China is characterised by an exceptionally large share of valued added in industry and a low share in services. Therefore, a major challenge for China is the rebalance of the economy through a relative shift from industry-led growth to services-led growth and labour-intensive urban growth. Not only will this change allow to balance the path of development, aligning China to the economic structure of the industrialised countries, but also it will help to save energy and resources and to protect the environment. On this respect, the continued industry-led expansion requires more energy and puts heavy burden on the 2

environment, predicting a non-sustainable growth over the longer horizon. Although significant effort is placed to forge a sustainable energy path, which eventually substitutes more efficient fuels for coal, nevertheless coal will remain the greatest energy input for many years to come. In the recent past, concern regarding the environmental implications of the rising Chinese energy demand has induced considerable research efforts to produce long-term forecasts of China’s coal requirements. The numerous economic analyses, which are generally conducted at an aggregate country level, are almost unanimous in projecting increasing trends in coal demand for the near future. To our knowledge, however, no previous study has modelled the coal demand in China at provincial level and focussed on coal demand forecasts. The availability of data at provincial level allows us to fill this gap. The aim of this paper is twofold. First, we analyse the Chinese industrial coal demand using panel data disaggregated by provinces and taking into account spatial heterogeneity. Second, given the spatial nature of the data, we explicitly capture autocorrelation among provinces using spatial econometrics. In particular, we specify the Chinese industrial coal demand at provincial level with a fixed-effect spatial lag model and a fixed-effect spatial error model. The remainder of the paper is organized as follows. Section 2 presents a brief review of the literature on energy consumption in China. Section 3 outlines the methodology adopted. Section 4 describes the data set used and offers an overview of China’s major energy trends. Section 5 identifies potential sources of spatial interaction among Chinese provinces. Section 6 presents model specifications and empirical results. Section 7 illustrates the forecasts of industrial coal demand in China at provincial level. Finally, Section 8 contains a summary of the empirical findings and concludes. 2. Literature review The empirical literature on modelling energy demand in China can be classified in two broad groups, according to its two distinct goals. The first one is the estimation of long-run income elasticities of energy demand, generally based on time series techniques such as cointegration and vector error correction (VECM), as well as the provision of energy demand forecasts. The second aim is the detailed analysis of the factors which are responsible for the energy intensity decline experienced by China during its ongoing extraordinary economic development. This latter goal is generally achieved through the use of econometric models which decompose the contributions of variables such as sectoral shifts or subsector productivity changes to total energy use. Among the contributions which belong to the first group is the work by Masih and Masih (1996), who employ the dynamic OLS procedure developed by Stock and Watson to estimate the 3

demand for coal in China using annual data from 1953 to 1992. The advantage of this methodology is that it accounts for the endogenous nature of the regressors, as well as for the non-stationarity of the variables involved. Controlling for the simultaneity bias between coal demand and its determinants, the authors find that absolute long-run price and income elasticities are close to unity for China. Modelling and forecasting coal demand in China is also at the heart of the contribution by Chan and Lee (1997). Applying the Engle-Granger’s two-step procedure, they estimate an error correction model, which includes price, income, and a structural shift as regressors. The authors present long-run elasticities during the period 1953-1994 which are slightly lower than one, suggesting a less-than-proportional reaction of coal demand to income and price changes. They predict a rise for coal consumption from around 1.2 billion tons (Bt) in 1994 to a level of 1.48 Bt in 2000, as a consequence of a 10 percent average increase in national income, together with a 4 percent annual increase in real coal price and a 1 percent drop in the share of heavy industry’s output on national income. A value less than one of the GDP elasticity for electricity demand is reported in Lin (2003), who employs a VECM for the period 1975-2001. It should be noted that, extending the sample period prior to the Chinese reforms phase, this elasticity increases but still remains lower than one. Moreover, the electricity demand rates are expected to grow by 5.8 percent on average between 2002 and 2010. This prediction indicates a remarkable decline in the consumption growth rate if compared with its historical trend, which averaged 9 percent in the period 1978-2001. Lee and Chang (2008) apply a VECM to a panel of 16 Asian countries, including China, during the period 1971-2002, in order to estimate the causal relationship between GDP and energy consumption, controlling for capital and labour input. The authors find that energy Granger-causes GDP in the long-run. Conversely, neither short-run, nor long-run causal effects from GDP to energy are present. The empirical evidence on the existence of Granger-causality between energy and income in China is quite mixed. Yuan et al. (2008) test the causality between GDP and energy consumption at aggregated and disaggregated levels for different fuels. While Granger-causality running from electricity and oil consumption to GDP is documented, the causal effect from coal and total energy consumption to GDP does not emerge. GDP is also found to Granger-cause total energy, coal and oil consumption, whereas GDP does not Granger-cause electricity consumption. Those results support the empirical evidence produced by Shiu and Lam (2004) and Yuan et al. (2007), where electricity consumption is found to Granger-cause GDP, but not vice-versa. 4

Cointegration analysis and Granger-causality tests are also conducted by Soytas and Sari (2006), for energy consumption and economic growth. The authors, however, do not find any cointegration relationship between the two variables, and they conclude that sustainable growth can be achieved without putting too much pressure on the environment. Zou and Chau (2006) test the relationship between economic growth and oil consumption, finding support for the existence of bidirectional Granger causality in the long-run. Crompton and Wu (2005) propose an alternative methodology to forecast energy consumption in China, which uses a Bayesian vector autoregression (VAR). This methodology avoids the problem of overparameterisation, which generally affects standard VAR models with a large number of lags. With an estimation sample period running from 1956 to 2003, they forecast over the period 2004-2010 an average 3.8 percent rise in total energy consumption and a 3.3 percent increase in coal demand. Adams and Shachmurove (2007) apply an energy balance system to model energy functions in China. This methodology is traditionally used as an accounting framework to investigate energy supply and utilisation, although the authors produce energy demand prospects for 2010 and 2020. Econometric estimation of the linkage parameters suggests an income elasticity for fuel consumption which varies between 0.5 and 0.6. The size of the parameter, which is substantially lower than one, eventually reflects some ongoing improvements in the use of energy, or, alternatively, a shift towards products which are less energy-intensive. The authors conclude that there must be a substitution of more efficient fuels for coal, since they predict a decrease in coal demand of 6 percent, despite the increase in final energy consumption amounts to 2.6 percent. Tang and La Croix (1993) analyse the impact of economic activity on total energy consumption in China, using pooled cross sectional data at a provincial level for the period 19851989. They report a value for the income elasticity of total energy demand which is around one, lower than the coefficient estimated with a similar methodology for developing countries by Zilberfarb and Adams (1981), and Reister (1987). Tang and La Croix suggest that a unit elasticity implies a constant energy intensity for Chinese provinces along the process of economic development. A different view over the trends of energy intensity in China is shared by a number of empirical contributions, which are almost unanimous in finding a decline in energy consumption per unit of GDP, despite the increase in total energy consumption. The main objective of those papers, which include, among others, Kambara (1992), Garbaccio et al. (1999), Chu at al. (2006), is to offer sound explanations for such a decline in energy intensity. The general conclusion is that 5

technical and structural changes are responsible for the improved efficiency in China. Using disaggregated firm-level data, Fisher-Vanden et al. (2004) show that nearly half of the decline in energy intensity is due to gains in efficiency at a firm level. Liu and Ang (2007) find that the aggregated industrial energy intensity decline is due to changes in sectorial energy intensity and structure change. Conversely, Ma et al. (2008) point out that energy intensity in China has slightly increased, as a consequence of the adoption of more energy-intensive technologies. 3. Modelling industrial coal demand in China: a spatial econometrics approach The main feature of spatial models is the introduction of some forms of interactions within agents, who are influenced in their choices by other agents, rather than behaving in isolation. The existence of such interactions generate the so-called “spatial dependence”: in general, this term refers to the presence of a functional relationship between what happens at one point in space and what happens elsewhere (Anselin, 1988). In the modern economic theory, a spatial approach can be found in the new economic geography, as embodied, among others, in the works by Krugman (1991, 1996). These contributions explain the concentration of production by means of increasing returns, path dependence and imperfect competition, which determine spatial externalities and spillovers. Given the existence of such spatial interactions, an empirical analysis of these models requires a spatial econometric approach. A classical economic framework, where total costs are minimized subject to a production constraint, generates the following industrial coal demand: C it = f ( Z it , β )

(1)

where C it is the coal demand in province i (i=1,…,N) at time t (t=1,…,T), Zit is a vector containing K standard factors that influence demand, such as input prices and output, and β is a vector of K unknown parameters. However, it may be the case that the demand function depends also on a term capturing strategic interaction, represented by the coal demand in other provinces. It is well known that the oldest coal-fired power stations in China were built close to mines to reduce transportation costs. New energy firms have decided to locate their plants in the same area, in order to benefit from pecuniary externalities generated by the existence of coal-specific infrastructure, such as installed grids for power transfer. This form of development implies that nowadays the coal-based energy firms are concentrated where other energy producing firms are located, that is in areas with good access to fuel inputs, rather than in more industrialized areas. Therefore, standard coal demand 6

determinants cannot fully explain the industrial coal demand, and spatial interaction terms should be considered in the analysis. Spatial dependence can be incorporated in a regression model through a spatial lag or a spatial error term (Anselin, 1988). Using panel data, the spatial lag model can be written in structural form as: N

K

j =1

r =1

(2a)

C it = ρ ∑ wij C it + ∑ β r z rit + μ i + u it

which corresponds to the reduced form:

N ⎛ ⎞ C it = ⎜⎜1 − ρ ∑ wij ⎟⎟ j =1 ⎝ ⎠

−1

K

∑ β r z rit r =1

N ⎛ ⎞ + ⎜⎜1 − ρ ∑ wij ⎟⎟ j =1 ⎝ ⎠

−1

(μ i

+ u it )

(2b)

In equations (2a)-(2b), i and j indicate provinces (i.e. i,j=1,...,N); t indicates time (i.e. t=1,...,T); ρ is the spatial autoregressive parameter; wij is the (i,j)-th element of the spatial weight matrix W of order N, with wij=wji for any i≠j, and wii=0 for any i; zrit is the r-th element of vector Zit, which may or may not include a common constant term; µi is the i-th individual fixed effect; uit is a classical error term. The spatial weight matrix W is crucial in the specification, since it defines the structure of dependence among provinces. It should be noted that this structure is maintained invariant with respect to time, since it is based on the geographical location of coal mines. If we stack variables across provinces and years, model (2a) can be rewritten as: C = ( I - ρΨ ) -1 Zβ + ( I - ρΨ ) -1 ( M + U )

(2c)

In equation (2b), C is the NT·1 vector of stacked observations industrial coal demand, I is the identity matrix of order NT, Ψ = W ⊗ I , Z is the NT·K matrix of stacked observations on the explanatory variables, β is the K·1 vector of parameters, µ is the N·1 vector of individual fixed effects, J is a T·1 vector of ones; M = μ ⊗ J ; U is the NT·1 vector of stacked error terms uit. The inverse matrix ( I − ρΨ ) −1 is a full matrix which induces error terms in all locations (Anselin and Bera, 1998). Since the OLS estimator would be biased and inconsistent in this case, maximum 7

likelihood (ML) estimation or instrumental variables should be used. The spatial error model has the following structure: K

C it = ∑ β r z rit + μ i + ε it

(3a)

r =1

N

ε it = λ ∑ wij ε it + vit j =1

where λ is the spatial autoregressive coefficient. Equation (3a) can be represented in matrix form as:

C = Zβ + M + E

(3b)

E = λΨ E + V

where E (V) is the NT·1 vector of stacked error terms εit (vit). This regression fails to meet the standard condition of spherical error terms, as the covariance matrix of errors E contains nonconstant diagonal elements. Therefore, the existence of spatial dependence in the disturbances induces heteroskedasticity in Ε, irrespective of the covariance structure of the errors V. The nuisance related to the spatial error dependence introduces a problem of inefficiency in the estimated parameters1, and again ML estimation can be used. 4. Data description Two sources of data are used for our analysis, namely the CEIC dataset and the China Energy Databook (2004), published by the China Energy Group of the Lawrence Berkeley National Laboratory. In particular, the latter source of data collects important statistical information, drawn from national and provincial energy balances on energy production and consumption, disaggregated by energy types, sectors and uses. Of particular relevance for our empirical investigation is the total industrial coal demand, which embodies the industrial energy consumption, the industrial non-energy use, as well as consumption lost during the process of energy conversion. For China, the demand for coal in the industrial sector absorbs a considerable share of aggregate coal demand, although this share has 1

Anselin and Bera (1998) interpret the spatial error dependence as a nuisance induced by spatial autocorrelation in

measurement errors, or in variables that are not crucial to the model, such as ignored spillovers across the observational units.

8

slightly increased over time, as reported by the China Energy Databook (2004). Specifically, the industrial sector is responsible for 85 percent of total demand in 1995 and more than 90 percent in 2002. Therefore, industry is the primary driver for coal demand. The industrial coal consumption is mainly concentrated in four sectors, namely power and heat generation, metallurgy, chemical production and material building. Among these, power and heat generation is the main contributor of coal consumption, with a share of 75 percent in 2002. The metallurgic industry mainly employs coking coal, fuel coal and coal injected into blast furnace. The chemical industry demands lump anthracite for nitrogen fertilizer and fuel coal for heat supply. The building industry uses coal for cement and brick production. A second important feature is the relevance of coal as a source of energy. Coal is extremely abundant in China, which makes this country the larger producer and consumer of coal in the world. The China Energy Databook (2004) reports the composition of the Chinese primary energy demand, by fuel source. Coal represents the key component of energy demand, accounting for over 60% of the primary energy consumption mix, followed by oil, hydro-power and natural gas. The importance of coal has slightly decreased over time, nevertheless coal still remains the undisguised driver of primary energy. China shows large unbalances in the composition of the output value, since it is characterized by an exceptionally high share of value added in industry and an exceptionally low share in services. This feature holds also when China is compared to US and Latin America. Regarding its administrative structure, China is composed by 22 provinces, 5 autonomous regions, 4 municipalities, and 2 special administrative regions. Unfortunately, for some of the territories, our data set offers incomplete information. Therefore, in order to form a balanced structure, only 27 administrative units have been selected. Moreover, since the provincial energy balances are accessible from 1995 to 2002 with annual frequency, the total number of observations used in our empirical investigation is 216. 5. Spatial nature of China’s industrial coal demand data

The structure of the generating capacity in China varies enormously from region to region, and it is influenced by the location of the major sources of energy. Hydropower plans, for example, are mainly located in Central and Southern China, whereas thermal-power plans can be found in the North and North-Eastern regions, where coal is abundant (IEA, 2006). FIGURE 1: ABOUT HERE 9

Although China owns substantial reserves of conventional energy resources, coal largely dominates the other energy inputs. Actually, China possesses extensive coal reserves, which make the country the world’s largest coal producer. Moreover, the coal reserves are highly geographically concentrated, as two-thirds of total coal reserves are located in the Northern provinces. The black circles in Figure 1 indicate the location of the top major coal mines, which supply more than 10 million tonnes (Mt) per annum. Figure 1 shows that the largest coal mines are concentrated in 8 provinces, all located in the North-Eastern and Northern regions. Table 1 reports the exact production of coal mines in 1996: among others, Daton in Shanxi produces more than 30 Mt alone, while Kailuan in Hebei and Pingdingshan in Henan supply more than 17 Mt. TABLE 1: ABOUT HERE It is worth noticing that the 10 major energy producing provinces, which use coal as an energy source, are at close distance from the major coal mines zones. According to the China Energy Databook (2004), the major energy producing provinces are: Shanxi, Shandong, Inner Mongolia, Henan, Hebei, Anhui, Heilongjiang, Shaanxi, Sichuan, and Liaoning. This evidence indicates that the power generation industry is concentrated in few inland provinces. Coal-fired power stations are built at close distance from the major mines, despite the largest centres for energy demand are located in the industrialised coastal areas. This correspondence is eventually due to transportation constraints, as the Chinese inland coal transport infrastructures, which move more than one half of total coal, are inefficient and highly congested. Conversely, there has been a substantial effort to increase the energy transfer capacity from the country’s resource-rich areas to the high energy-demanding regions: about 40 percent of total investments in the power sector is represented by transmission investments. Table 2 reports the first 10 provinces with the highest demand of industrial coal. Again, the location of those provinces reflect the position of coal mines. This fact suggests that the location of the energy sources eventually plays an important role in explaining the industrial coal demand. Therefore, interdependence among close provinces can arise2.

2

Despite the extensive coal reserves which are available, China has increased the volume of imported coal, in particular

to meet the energy demand raised by the coastal provinces. The existence of a large fraction of coal imports in the coastal area could weaken the hypothesis of pecuniary externalities which can be exploited in the coal reserve areas. However, the provincial energy balances show that the relevance of imports was minimal during the period 1995-2002,

10

In particular, this fact may justify a form of interdependence, where coal demand of province i is a function of coal demand in province j. TABLE 2: ABOUT HERE It is important to note that, although the demand for industrial coal includes both energy and non-energy use of coal, the former is far more important than the latter. The fraction of coal employed for energy production over total industrial use is large and increasing during the sample period. In 2002, the power generation accounted for 75 percent of total industrial consumption. This evidence allows us to incorporate in modelling the industrial coal demand the spatial structure described above, which is largely based on the link between coal mines, power generation and demand for coal. As a first step towards the investigation of the existence of spatial autocorrelation in the industrial coal demand, a simple representation of the quintile distribution of this variable is reported. As Figure 2 clearly shows, a cluster of high coal demand values can be identified in the North-Eastern regions, in proximity to the coal mines zone, whereas a cluster of low demand is located in the Western area.3 The 5 provinces whose coal demand is, on average, more than 70 Mt are indicated in dark black and are concentrated in a bounded zone. Conversely, the 5 provinces whose demand for coal is less than 20 Mt appear to be slightly more dispersed, though 3 of them are neighbours. These provinces are identified in white. FIGURE 2: ABOUT HERE The presence of spatial correlation between observational units can be further detected by means of formal tests, which capture the extent to which values similarity matches with locations similarity. In particular, the demand for coal can show positive spatial correlation if likewise values tend to cluster in space, or it can display negative correlation whenever the locations are surrounded by neighbours with dissimilar values. Finally, a zero spatial correlation implies that it is not possible to identify a specific spatial pattern among coal values. There are a number of widely used statistics which detect sample spatial autocorrelation, which are divided into two groups, namely global and local statistics. Among the global measures of spatial autocorrelations are the Moran’s I tests: as well as that some provinces with significant coal imports were not located in the coastal region. 3

Given the panel dimension of our data set, the variable employed is a 8-years average of coal demand.

11

I =

N S0

∑∑w i

ij

j

z i z j / ∑ z i2

(4)

i

and the Geary’s c tests:

c = ( N − 1)∑∑ wij ( x i − x j ) 2 / 2( S 0 )∑ z i2 i

j

(5)

i

with z i = x i − μ , S 0 = ∑∑ wij . i

j

In our analysis two different weight matrices are employed. The first matrix (W1) defines the neighbourhood set according to common borders, i.e. wij = 1 if provinces i and j have common

border and wij = 0 otherwise. It should be noted that, in order to facilitate interpretation, the weight matrix is standardized by imposing that the row elements sum up to one. The second matrix (W2) is based on a distance decay function. In this case, wij = 1 / Dist ij2 if Distij is less than 600 Km, where Distij is the Euclidean distance between province i and j, while wij = 0 otherwise. A spatial clustering of high/high or low/low values results from Moran’s I>-1/(N-1) and Geary’s c 10 Mt per annum)

 

Source: EPA (1996).

Figure 2. Industrial coal demand clusters (Mt, 8-years average)

27

Table 1. Coal production for major coal mines Coal production Coal mine

Province

Datong Kailuan Pingdingshan Huaibei Xishan Hegang Xuzhou Fuxin Longkou Yanzhuo Jixi Huainan Yangquan Fengfeng Jincheng Tiefa Xinwen Qitaihe Shuangyashan

Shanxi Hebei Henan Anhui Shanxi Heilongjiang Jiangsu Liaoning Shandong Shandong Heilongjiang Anhui Shanxi Hebei Shanxi Liaoning Shandong Heilongjiang Heilongjiang

(thousand tonnes per annum)

31754.6 17604.8 17147.8 14232.1 14127.7 13130.7 13103.1 12698.7 12003.1 12003.1 11644.2 11498.5 10476.9 10370.0 10320.6 10241.0 10089.0 10060.1 10015.5

Source: EPA (1996)

Table 2. Industrial coal demand for selected provinces (Mt) Province

1995

1996

1997

1998

1999

2000

2001

2002

Shanxi Hebei Shandong Jiangsu Henan Liaoning Guangdong Inner Mongolia Anhui Zhejiang

140.0 90.8 86.6 83.1 66.7 81.5 45.3 37.7 44.8 39.5

143.8 92.0 91.6 83.2 69.6 80.9 47.7 40.6 47.5 43.2

137.7 92.3 91.5 79.9 65.4 79.7 48.6 48.6 49.0 45.1

143.5 94.8 87.1 81.8 68.1 76.4 48.1 44.1 48.0 44.3

128.4 96.7 86.1 83.4 71.0 73.9 50.8 48.3 50.0 45.8

133.2 101.6 75.3 84.5 76.9 86.0 57.3 53.4 53.2 48.3

138.8 106.5 94.7 87.1 83.1 81.2 58.9 57.1 57.2 53.5

170.1 117.8 110.9 94.0 93.7 84.0 64.8 63.4 59.7 58.6

Source: Authors’ computations from China Energy Databook (2004).

Table 3. Global Moran’s I test on average industrial coal Moran’s I

I

E(I)

sd(I)

z

W1 0.305 -0.038 0.122 2.811 W2 0.321 -0.038 0.156 2.303 Notes: The test statistic z is asymptotically normally distibuted.

P-value 0.005 0.021

Table 4. Global Geary’s c test on average industrial coal Geary’s c

c

E(c)

sd(c)

z

W1 0.624 1.000 0.141 -2.665 W2 0.728 1 0.185 -1.469 Notes: The test statistic z is asymptotically normally distibuted.

28

P-value 0.008 0.142

Table 5. Local Moran’s Ii test on average industrial coal for selected provinces Province

P-value

Ii

W1 Beijing Tianjin Hebei Shanxi Inner Mongolia Shaanxi Gansu Qinghai Xinjiang Liaoning Jilin Heilongjiang Shanghai Jiangsu Zhejiang Anhui Fujian Jiangxi Shandong Henan Hubei Hunan Guangdong Guangxi Sichuan Guizhou Yunnan

-0.37 -0.35 1.17 1.59 0.04 -0.33 0.51 0.84 1.18 0.54 -0.13 -0.03 -0.14 0.40 0.00 0.06 0.19 0.12 1.39 0.87 -0.01 0.10 -0.13 0.30 -0.09 0.21 0.30

P-value

Ii W2

0.61 0.63 0.00 0.00 0.80 0.40 0.16 0.09 0.06 0.27 0.87 0.99 0.88 0.32 0.92 0.77 0.66 0.66 0.00 0.01 0.94 0.69 0.83 0.45 0.88 0.57 0.52

0.10 0.07 1.84 2.42 0.10 -0.64 1.04 1.29 -0.15 -0.16 -0.01 -0.05 0.10 0.00 0.08 0.33 0.00 0.62 0.91 -0.01 0.07 -0.11 0.17 -0.10 0.26 0.19

0.84 0.86 0.00 0.00 0.78 0.25 0.14 0.16 0.88 0.86 0.97 0.99 0.77 0.94 0.80 0.49 0.92 0.07 0.00 0.95 0.83 0.91 0.75 0.91 0.57 0.73

Notes: i) The null hypothesis is no spatial autocorrelation; ii) For some provinces the local statistics is not computed, as there is no neighbour associated to these provinces.

Table 6. Variables description Variable

Description

COAL RGVA POP HEAVY

Industrial coal consumption (Mt). Gross value added (Billion Yuan, constant 1995). Millions. Heavy industry dummy variable. HEAVY=1 for Shanxi, Gansu, Liaoning, Jilin and Heilongjiang; HEAVY = 0 otherwise. Dummy variable for municipalities. MUNIC=1 for Beijing, Shanghai and Tianjin; MUNIC=0 otherwise. Coastal dummy variable. COAST=1 for Tianjin, Liaoning, Shanghai, Jiangsu, Zhejiang, Fujian, Shandong, Guangdong, Guangxi; COAST=0 otherwise. Trend variable. TREND=3 for 2002; TREND=2 for 2001; TREND=1 otherwise.

MUNIC COAST TREND

29

Table 7. Descriptive statistics Variable COAL RGVA POP

N. 216 216 216

Mean 47.21 147.47 44.48

Std. Dev. 31.19 107.13 25.07

Min. 3.11 8.21 4.81

Table 8. Estimated models of industrial coal demand No spatial RGVA lnTREND*RGVA lnTREND POP HEAVY MUNIC COAST Constant Observations Log-likelihood Rho (ρ)

-0.017 (0.015) 0.027 (0.009)** 1.201 (1.824) 0.970 (0.086)** 57.872 (3.834)** 9.612 (1.479)** -15.368 (6.826)* -0.983 (0.610)

W1

W2

Lag model -0.016 (0.014) 0.028 (0.008)** 0.755 (1.623) 0.975 (0.079)** 60.571 (3.560)** 9.475 (1.462)** -17.808 (6.318)** -3.525 (0.988)**

Error model -0.014 (0.015) 0.027 (0.008)** 1.288 (1.762) 0.960 (0.082)** 58.042 (3.580)** 9.738 (1.440)** -15.546 (6.387)* -1.059 (0.584)

Lag model -0.016 (0.014) 0.028 (0.008)** 0.910 (1.660) 0.982 (0.078)** 59.573 (3.612)** 9.583 (1.371)** -17.186 (6.222)** -2.108 (0.984)*

Error model -0.016 (0.014) 0.027 (0.008)** 1.193 (1.685) 0.968 (0.083)** 57.894 (3.550)** 9.640 (1.417)** -15.333 (6.368)* -0.997 (0.583)

216 -602.8 0.063 (0.019)***

216 -607.1

216 -606.3 0.024 (0.017)

216 -607.4

Lambda (λ)

0.087 0.015 (0.125) (0.110) Joint test for F(23, 185)=204 Chi2(23)=4844 Chi2(23)=5109 Chi2(23)=5402 Chi2(23)=5414 provincial dummies P-value=0.000 P-value=0.000 P-value=0.000 P-value=0.000 P-value=0.000 Notes: i) Maximum likelihood estimation is used in the spatial regressions; ii) Robust standard errors are reported in parenthesis; iii) * (**) [***] indicates 10% (5%) [1%] statistical significance; iv) ln(TREND) takes the value of zero between 1995 and 2000. It equals 0.69 in 2001 and 1.10 in 2002; v) The spatial weight matrix W1 is computed setting ω ij = 1 if provinces i and j share common borders, and

ωij = 0

otherwise; vi) The spatial weight matrix W2 is computed setting

ω ij = 1 / Dist ij2 , where

Distij is the Euclidean distance between province i and j, if Distij is ≤ 600 Km, and ωij = 0 otherwise.

30

Max. 170.10 545.20 114.30

Table 9. The income (RGVA) elasticity of coal (COAL) (year 2002) Region North

Northeast

East

South-Central

Southwest

Northwest

Overall

Model Province Beijing Tianjin Hebei Shanxi Inner Mongolia Liaoning Jilin Heilongjiang Shanghai Jiangsu Zhejiang Anhui Fujian Jiangxi Shandong Henan Hubei Hunan Guangdong Guangxi Sichuan Guizhou Yunnan Shaanxi Gansu Qinghai Xinjiang Mean

No spatial 0.09 0.04 0.03 0.01 0.02 0.04 0.03 0.04 0.07 0.07 0.08 0.04 0.11 0.06 0.06 0.04 0.05 0.07 0.11 0.07 0.06 0.02 0.05 0.04 0.03 0.04 0.05 0.05

W1 Error Lag 0.11 0.11 0.05 0.06 0.04 0.04 0.01 0.01 0.02 0.02 0.05 0.05 0.04 0.04 0.05 0.05 0.09 0.09 0.08 0.08 0.10 0.10 0.04 0.04 0.13 0.13 0.08 0.08 0.07 0.07 0.05 0.05 0.06 0.06 0.09 0.09 0.13 0.13 0.09 0.09 0.07 0.07 0.02 0.02 0.06 0.06 0.05 0.05 0.04 0.04 0.05 0.05 0.06 0.06 0.06 0.06

W2 Error 0.10 0.05 0.03 0.01 0.02 0.04 0.04 0.05 0.08 0.07 0.08 0.04 0.12 0.07 0.06 0.04 0.06 0.07 0.12 0.07 0.06 0.02 0.05 0.04 0.03 0.04 0.05 0.06

Lag 0.11 0.05 0.04 0.01 0.02 0.05 0.04 0.05 0.09 0.08 0.09 0.04 0.13 0.08 0.07 0.05 0.06 0.08 0.13 0.08 0.07 0.02 0.06 0.05 0.04 0.05 0.06 0.06

Table 10. Spatial tests Test

P-value

Test

W1 Moran's I

P-value W2

0.815

0.415

0.181

0.856

0.525 0.004

0.469 0.951

0.010 0.042

0.919 0.838

9.242 8.720

0.002 0.003

2.329 2.360

0.127 0.124

Spatial error: LM-err Robust LM Spatial lag: LM-lag Robust LM

31

Table 11. Forecasts of industrial coal demand (Mt, years 2005, 2010) Drivers specification A

B

C

D

E

F

Model

No spatial Lag Error No spatial Lag Error No spatial Lag Error No spatial Lag Error No spatial Lag Error No spatial Lag Error

Aggregate industrial coal demand Prediction 2005 1630 1643 1634 1617 1629 1621 1611 1621 1615 1624 1635 1627 1610 1621 1614 1623 1635 1627

Prediction 2010 1975 2021 1983 1951 1994 1958 1975 1910 1880 1899 1936 1904 1874 1909 1880 1899 1935 1904

Average annual percent change 2002-2010 3.98 4.33 4.04 3.80 4.12 3.85 3.98 3.48 3.25 3.40 3.68 3.44 3.21 3.47 3.25 3.40 3.67 3.44

Notes: i) The AR model for RGVA is applied in drivers specifications A and B; ii) The Holt-Winters filter for RGVA is used in drivers specifications C and D, whereas the double-smoothing filter is used in drivers specifications E and F; iii) As far as POP is concerned, specifications A, D and F use the double-smoothing model, while specifications B, C and E apply the trend model; iv) Forecasted values refer to total China industrial coal demand, although only the 27 provinces - out of 30 - considered in our analysis contributed to the aggregation; v) The percentage change is computed with reference to year 2002, where the actual value for industrial coal demand is Mt 1454.

32

Table 12. Forecasts of industrial coal demand by provinces (Mt, years 2005, 2010) Region North

Province

Beijing Tianjin Hebei Shanxi Inner Mongolia Northeast Liaoning Jilin Heilongjiang East Shanghai Jiangsu Zhejiang Anhui Fujian Jiangxi Shandong South-Central Henan Hubei Hunan Guangdong Guangxi Southwest Sichuan Guizhou Yunnan Northwest Shaanxi Gansu Qinghai Xinjiang

Prediction 2005 33.54 27.26 116.79 151.89 56.83 98.55 46.48 64.42 56.74 114.49 74.13 65.38 32.71 30.19 120.38 93.38 64.47 47.73 88.03 32.77 58.13 40.50 33.09 36.27 23.40 8.43 26.92

Prediction 2010 51.35 37.07 136.09 158.21 61.93 111.04 53.69 74.42 70.64 154.66 117.47 72.67 45.09 35.31 159.29 114.21 73.60 57.43 122.93 44.34 64.95 48.45 38.28 44.70 27.40 12.81 32.82

Notes: The reported predictions at provincial level refer to driver specification A and to the spatial lag model, which uses the weight matrix W1.

33