AN ABSTRACT OF THE DISSERTATION OF Man Li for the degree of Doctor of Philosophy in Agricultural and Resource Economics presented on June 16, 2010. Title: Essays on Land-use Change, Carbon Sequestration and Emissions in China. Abstract approved: JunJie Wu Jeffrey J. Reimer China has experienced rapid economic growth in the last twenty years, accompanied by large-scale land conversion, severe environmental degradation, and rising carbon dioxide (CO2) emissions. Designing policies for sustainable development requires a comprehensive understanding of the relationship between economic growth, land-use change, carbon sequestration and emissions in China. This dissertation consists of three essays that address several relevant issues from an economic perspective. The first essay presents an empirical analysis to identify the major drivers of land-use change in China for the period 1988-2000 by using highly-disaggregated, national-scale GIS land use data and a state-of-art econometric method. Results indicate that GDP growth and agricultural investment had relatively larger impacts on farmland conversion, while population growth and agricultural investment were more influential in grassland loss. Implications of the results for the design of farmland protection policies are discussed. The second essay examines the relationship between land-use change and soil carbon sequestration in China. Results indicate that farmland and grassland loss, deforestation, and land idleness, driven by GDP growth, accelerated soil carbon runoff. Implementation of the green growth policy could generate up to 0.7-1.1 million Mg SOC and result in 22.2-37.4 million CNY welfare losses annually from 2001 to 2050. The marginal welfare loss is approximately ¥15.3/Mg (equivalent to $2.25/Mg) for sequestering about 1 million Mg SOC per year. The third essay presents a new method to examine the sources of change in CO2 emissions in China between 1991 and 2006. Results indicate that GDP scale effect accounted for the majority of emission increments. The emission index associated with capital was a dominant contributor to emission abatement. The effects of technical change in production and change in the GDP-composition by sector played positive roles in curtailing emissions. ©Copyright by Man Li June 16, 2010 All Right Reserved ESSAYS ON LAND-USE CHANGE, CARBON SEQUESTRATION AND EMISSIONS IN CHINA by MAN LI A DISSERTATION submitted to Oregon State University in partial fulfillment of the requirements of the degree of DOCTOR OF PHILOSOPHY Presented June 16, 2010 Commencement June 2011 Doctor of Philosophy dissertation of Man Li presented on June 16, 2010. APPROVED: Co-Major Professor, representing Agricultural and Resource Economics Co-Major Professor, representing Agricultural and Resource Economics Head of the Department of Agricultural and Resource Economics Dean of the Graduate School I understand that my dissertation will become part of the permanent collection of Oregon State University libraries. My signature below authorizes release of my dissertation to any reader upon request. Man Li, Author AKNOWLEDGEMENTS I have been fortunate to have a number of excellent people who have assisted my degree and without whom this dissertation would have not been possible. I wish to express sincere gratitude to all members of my PhD program committee, Dr. JunJie Wu, Dr. Jeffrey J. Reimer, Dr. Andrew J. Plantinga, Dr. Lan Xue, and Dr. Russell E. Ingham, for their invaluable comments and suggestions on this dissertation. I particularly appreciate the consistent support, guidance, and encouragement of my co-major professor, Dr. JunJie Wu. All of his comments and suggestions have left mark on this work. A special thanks is extended to Dr. Xiangzheng Deng, who has been helpful in collecting most data used in this dissertation. I would also like to thank my family for their patient support and encouragement throughout these years. CONTRIBUTION OF AUTHORS Dr. JunJie Wu was involved in the design, analysis, and writing of Chapters 2-3. Dr. Xiangzheng Deng assisted with data collection of Chapters 2-3. TABLE OF CONTENTS Page 1. Introduction ...................................................................................................................1 References ................................................................................................................6 2. Indentifying Drivers of Land-use Change in China: A Spatial Multinomial Logit Model Analysis .............................................................................................................7 Abstract ....................................................................................................................8 Introduction ..............................................................................................................9 The Model ..............................................................................................................13 Empirical Specification .....................................................................................13 Econometric Issues ............................................................................................15 Data ........................................................................................................................19 results .....................................................................................................................23 Drivers of Land-use Changes .................................................................................35 Public Agricultural Investment and Farmland Protection ......................................38 Conclusions ............................................................................................................41 Endnotes .................................................................................................................43 Acknowledgements ................................................................................................45 References ..............................................................................................................46 Appendices .............................................................................................................49 3. An Empirical Economic Analysis of Land-use Change and Soil Carbon Sequestration in China ................................................................................................64 Abstract ..................................................................................................................65 TABLE OF CONTENTS (CONTINUED) Page Introduction ............................................................................................................66 The SOC Density Model ........................................................................................69 Data ........................................................................................................................73 results .....................................................................................................................75 SOC Density Model ..........................................................................................75 Land Use Change Model ...................................................................................77 Green Growth, Carbon Sequestration, and Welfare Loss ......................................79 Conclusions ............................................................................................................88 Endnotes .................................................................................................................90 Acknowledgements ................................................................................................91 References ..............................................................................................................92 4. Decomposing the Change of CO2 Emissions in China: A Distance Function Approach .....................................................................................................................96 Abstract ..................................................................................................................97 Introduction ............................................................................................................98 Methodology ........................................................................................................100 Application ...........................................................................................................106 Data .................................................................................................................106 Result and Discussions ....................................................................................107 Concluding Comments .........................................................................................113 Endnotes ...............................................................................................................114 TABLE OF CONTENTS (CONTINUED) Page Acknowledgements ..............................................................................................115 References ............................................................................................................116 Appendices ..........................................................................................................118 5. Conclusions...............................................................................................................123 Bibliography .............................................................................................................126 LIST OF FIGURES Figure Page 2.1. Distribution of Land-Use Change on Farmland, 1988–2000......................................10 2.2. Distribution of Land-Use Change on Grassland, 1988–2000 .....................................11 2.3. Marginal Costs of Preserving Farmland in Counties with Different Percent Farmland Coverage (Study Area: Huang-Huai-Hai Plain, Yangtze River Delta, and Sichuan Basin) ..........................................................................................................................39 3.1. Histogram of R-square for all GWR‟s in the SOC density model ..............................75 3.2. Box plot of land-use dummy estimators for all GWR‟s in the SOC density model ...76 3.3. The area of land by use in the baseline scenario .........................................................81 3.4. The area of land by use in the scenario of initial 10% GDP growth rate ...................82 3.5. The flow of soil organic carbon relative to the baseline for different GDP growth rates .............................................................................................................................83 3.6. Time series plot of expected logsum welfare gains for different GDP growth rates relative to the baseline ................................................................................................85 3.7. The marginal welfare losses of soil carbon sequestration by discount rate under the Green Growth policy scenario ....................................................................................86 3.8. The average welfare losses of soil carbon sequestration by discount rate under the Green Growth policy scenario ....................................................................................87 4.1. National Carbon Dioxide Emissions in China, 1991-2006 .......................................108 LIST OF TABLES Table Page 2.1. Summary Statistics of Explanatory Variables ...........................................................19 2.2a. Land-use Transitions from 1988 to 1995 ..................................................................20 2.2b. Land-use Transitions from 1995 to 2000 ..................................................................20 2.3a. Coefficient Estimates for the Standard Multinomial Logit Model of Land-use Change on Farmland, 1988-1995................................................................................25 2.3b. Coefficient Estimates for the Standard Multinomial Logit Model of Land-use Change on Farmland, 1995-2000................................................................................26 2.4a. Coefficient Estimates for the Spatial Multinomial Logit Model of Land-use Change on Farmland, 1988-1995 .............................................................................................27 2.4b. Coefficient Estimates for the Spatial Multinomial Logit Model of Land-use Change on Farmland, 1995-2000 .............................................................................................28 2.5a. Coefficient Estimates for the Standard Multinomial Logit Model of Land-use Change on Grassland, 1988-1995 ...............................................................................31 2.5b. Coefficient Estimates for the Standard Multinomial Logit Model of Land-use Change on Grassland, 1995-2000 ...............................................................................32 2.6a. Coefficient Estimates for the Spatial Multinomial Logit Model of Land-use Change on Grassland, 1988-1995 ............................................................................................33 2.6b. Coefficient Estimates for the Spatial Multinomial Logit Model of Land-use Change on Grassland, 1995-2000 ............................................................................................34 2.7a. Description of Simulation Scenarios ........................................................................35 2.7b. Simulated Changes in Land Supplies of Six Major Uses, 1988-2000 ......................37 3.1. Summary Statistics of Explanatory Variables ...........................................................73 3.2. Description of Scenarios with Different Annual GDP Growth Rates .......................80 4.1. Change in CO2 Emissions and its Seven Decomposing Indices, 1991-2006 ...........109 LIST OF TABLES (CONTINUED) Table Page 4.2. Geometric Means of Annual Changes for Each Consecutive Two-year Period, 19912006 ..........................................................................................................................111 LIST OF APPENDIX TABLES Table Page A2.1. Coefficient Estimates for the Standard Multinomial Logit Model of Land-use Change on Forestland, 1988-1995 ..............................................................................52 A2.2. Coefficient Estimates for the Standard Multinomial Logit Model of Land-use Change on Forestland, 1995-2000 ..............................................................................53 A2.3. Coefficient Estimates for the Spatial Multinomial Logit Model of Land-use Change on Forestland, 1988-1995 ...........................................................................................54 A2.4. Coefficient Estimates for the Spatial Multinomial Logit Model of Land-use Change on Forestland, 1995-2000 ...........................................................................................55 A2.5. Coefficient Estimates for the Standard Multinomial Logit Model of Land-use Change on Water Area, 1988-1995 ............................................................................56 A2.6. Coefficient Estimates for the Standard Multinomial Logit Model of Land-use Change on Water Area, 1995-2000 ............................................................................57 A2.7. Coefficient Estimates for the Spatial Multinomial Logit Model of Land-use Change on Water Area, 1988-1995..........................................................................................58 A2.8. Coefficient Estimates for the Spatial Multinomial Logit Model of Land-use Change on Water Area, 1995-2000..........................................................................................59 A2.9. Coefficient Estimates for the Standard Multinomial Logit Model of Land-use Change on Unused Land, 1988-1995 .........................................................................60 A2.10. Coefficient Estimates for the Standard Multinomial Logit Model of Land-use Change on Unused Land, 1995-2000 .........................................................................61 A2.11. Coefficient Estimates for the Spatial Multinomial Logit Model of Land-use Change on Unused Land, 1988-1995 .........................................................................62 A2.12. Coefficient Estimates for the Spatial Multinomial Logit Model of Land-use Change on Unused Land, 1995-2000 .........................................................................63 ESSAYS ON LAND-USE CHANGE, CARBON SEQUESTRATION AND EMISSIONS IN CHINA CHAPTER 1 INTRODUCTION MAN LI 2 China has been at the forefront of the surge in global economic growth, with an approximate annual 10% increase in the gross domestic product (GDP) for the last twenty years (NBSC 2005; 2006-07). This rapid growth has been heavily affected by policy. For example, the growth rate reached its first peak in 1992 as Chinese reformer leader, Xiaoping Deng, made his famous southern tour of China using his travels as a method to reassert his economic policy, which was a milestone in the process of China‟s economic reform. In 1998, the Chinese government implemented industrial restructuring and shut down thousands of inefficient enterprises, including small-scale mines and power plants. Economic growth slowed somewhat during this period, but the economy kept growing strongly after China‟s join the World Trade Organization in 2001. Along with this economic success, CO2 emissions have been increasing rapidly in China since the 1990s. The emissions peaked for the first time in 1996 then dropped off and touched the bottom in 1998, which was concurrent with the industrial restructuring as just discussed. Again, after the entry to the WTO in 2001, the emissions went up at a high speed of 12% per year. In particular, China surpassed the United States to become the largest CO2 emitter of the world in 2006, releasing nearly 1.7 million petagrams of carbon (Pg C) into the atmosphere, which is 1.4 times more than the amount it emitted in 1991 (Marland et al. 2009). China has also experienced rapid urbanization in the last twenty years. From 1988 to 2000, the total developed area increased by 1.78 million hectares. The fraction of population residing in urban areas increased from 26% in 1990 to 46% in 2008 (Chinese Academy of Social Sciences 2009). The rapid urbanization has led to dramatic land-use changes in many parts of China, particularly in coastal regions and areas near major cities. For example, in traditional agricultural regions such as the Huang-Huai-Hai Plain, Yangtze River Delta and Sichuan Basin, rapid urban expansion to high-quality cultivated land is threatening the national food security, which is an extremely important issue in China where one-fifth of the world's population lives. The Chinese government has implemented the Basic Farmland Protection Regulation (1994) and revised the Land 3 Administration Law (1998) to slow this trend, but these laws and regulations have only achieved limited success. Despite the great loss of high-quality farmland in traditional agricultural regions, the total farmland in China increased by 2.8 million hectares from 1988 to 2000, due largely to farmland gained through grassland conversion and deforestation. Most grassland conversion to farmland occurred in the farming-pasture zones of East Inner Mongolia, North China Plain, and Loess plateau. Total grassland and forestland declined by 3.5 and 1.2 million hectares in China during this period, which reduced the soil carbon density and made China‟s ecological environment worse off. For example, agricultural development and desert expansion reduced the soil organic carbon (SOC) density by 1020 kg/m2 in the alpine meadow of the Southeast Tibet Plateau and by 2-3 kg/m2 in the grassland of the Mid-east Inner Mongolia (Wang et al. 2003). The connections between economic growth, CO2 emissions, land-use change, and carbon sequestration are complex. Designing policies for sustainable development requires a comprehensive understanding of these relationships. This dissertation consists of three essays that address several relevant issues from an economic perspective. The first essay (Chapter 2), Identifying Drivers of Land-use Change in China: A Spatial Multinomial Logit Model Analysis, presents an empirical analysis to identify the major drivers of land use change in China for the period 1988-2000. The analysis compiles highly disaggregated, national-scale GIS land-use data and develops a state-ofart method to estimate a spatial multinomial logit model that takes into account of spatial autocorrelation explicitly. Results indicate that both socioeconomic and geophysical variables affected land-use change in China. GDP growth and agricultural investment had larger impacts on farmland conversion, while population increase and agricultural investment were more influential in grassland loss. The model is used to evaluate the effectiveness of public agricultural investment as a policy instrument for farmland protection in traditional agricultural regions (Huang-Huai-Hai Plain, Yangtze River Delta, and Sichuan Basin). The results indicate targeting public agricultural investment in 4 counties with relatively low coverage of farmland (< 50%) is more cost-effective than spending money across regions. The second essay (Chapter 3), An Empirical Economic Analysis of Land-use Change and Soil Carbon Sequestration in China, examines the relationship between land-use change and soil carbon sequestration in China. A statistical Soil Organic Carbon (SOC) density model and an econometric land-use change model (i.e., a version of the multinomial logit model in the first essay) are developed to link the socioeconomic factors with the SOC density. The approach captures spatial autocorrelation and spatial heterogeneity simultaneously and can be applied to a large region. Results indicate that SOC density is generally highest in forest and grass lands, and lowest in unused land in China. GDP growth leads to farmland and grassland loss, deforestation, and idleness, which accelerates soil carbon runoff. The models are integrated to evaluate the welfare effects of China‟s green growth policy. The Chinese government has in 2006 announced six major measures, including legislation, industrial structure, technology, energy consumption management, incentive policies, and mechanism, in pursuit of green growth and a resources saving society in conjunction with soil carbon sequestration. The government listed resource reservation and environment protection as a major national policy in its “11th Five-Year Plan (2006 to 2010) for National Economic and Social Development”. The policy could generate up to 0.7-1.1 million Mg SOC and 22.2-37.4 million CNY welfare losses annually from 2001 to 2050. The marginal welfare loss curve is approximately ¥15.3/Mg (equivalent to $2.25/Mg) to sequester about 1 million Mg SOC per year. The third essay (Chapter 4), Decomposing the Change of CO2 Emissions in China: A Distance Function Approach, examines the sources of change in CO2 emissions in China between 1991 and 2006. It evaluates the relative contributions of the sources to emission abatement using a new empirical approach. The method uses the data envelopment analysis (DEA) technique to decompose emission changes into seven components based on the Shephard output distance function. The method accounts for factors that increase carbon emissions, as well as decrease them. It allows for cross- 5 sectional analysis under flexible data requirement. Results indicate that GDP scale effect (i.e., an expansion of the economy equivalent to the ratio of GDP between the two time periods) accounted for the majority of emission increments. The emission index associated with capital was a dominant contributor to emission abatement. The effects of technical change in production and change in the GDP-composition by sector played positive roles in reducing emissions. 6 REFERENCES Chinese Academy of Social Sciences. 2009. “中国城镇人口 2008 年末 6.07 亿.” http://www.cpirc.org.cn/news/rkxw_gn_detail.asp?id=10684 (accessed December 1, 2009). Marland, G., Boden, T., Andres, R.J., 2009. National CO2 emissions from fossil-fuel burning, cement manufacture, and gas flaring: 1751-2006, in: Trends online: A compendium of data on global change. Carbon Dioxide Information Analysis Center, O.R.N.L., U.S. Department of Energy, Oak Ridge Tennessee. National Bureau of Statistics of China, 2006-07. China Statistical Yearbook. China Statistical Press, Beijing. National Bureau of Statistics of China, 2005. Comprehensive Statistical Data and Materials on 55 Years of New China. China Statistical Press, Beijing. Wang, S., Tian, H., Liu, J., Pan, S., 2003. Pattern and change of soil organic carbon storage in China: 1960s-1980s. Tellus 55B, 416-427. 7 CHAPTER 2 INDENTIFYING DRIVERS OF LAND-USE CHANGE IN CHINA: A SPATIAL MULTINOMIAL LOGIT MODEL ANALYSIS MAN LI 8 ABSTRACT This essay presents an empirical analysis to identify the major drivers of land use change in China for the period 1988-2000. The analysis compiles highly disaggregated, national-scale land-use data and develops a state-of-art method to estimate a spatial multinomial logit model that takes into account of spatial autocorrelation explicitly. Results indicate that both socioeconomic and geophysical variables affected land-use change in China. GDP growth and agricultural investment had larger impacts on farmland conversion, while population increase and agricultural investment were more influential in grassland loss. The model is used to evaluate the effectiveness of public agricultural investment as a policy tool for farmland protection in traditional agricultural regions, including the Huang-Huai-Hai Plain, Yangtze River Delta, and Sichuan Basin. The results indicate targeting public agricultural investment in counties with relatively low coverage of farmland (< 50%) is more cost-effective than spending money across regions. 9 INTRODUCTION China has experienced rapid urbanization in the last twenty years. From 1988 to 2000, the total developed area increased by 1.78 million hectares. The fraction of population residing in urban areas increased from 26% in 1990 to 46% in 2008 (Chinese Academy of Social Sciences 2009). The rapid urbanization has led to dramatic land-use changes in many parts of China, particularly in coastal regions and areas near major cities. For example, in traditional agricultural regions such as the Huang-Huai-Hai Plain, Yangtze River Delta and Sichuan Basin, rapid urban expansion to high-quality cultivated land is threatening the national food security, which is an extremely important issue in China where one-fifth of the world's population lives. The Chinese government has implemented the Basic Farmland Protection Regulation (1994) and revised the Land Administration Law (1998)1 to slow this trend, but these laws and regulations have only achieved limited success. Despite the great loss of high-quality farmland in traditional agricultural regions, the total farmland in China increased by 2.8 million hectares from 1988 to 2000, due largely to farmland gained through grassland conversion and deforestation. Most grassland conversion to farmland occurred in the farming-pasture zones of East Inner Mongolia, North China Plain, and Loess plateau. Total grassland and forestland declined by 3.5 and 1.2 million hectares in China during this period, which made China‟s ecological environment worse off. Understanding the drivers of land-use change in China is useful for designing efficient agricultural, environmental, land use policies. Many previous studies have made efforts to obtain such knowledge: some of them focused on relatively small geographic areas (Deng et al. 2002; Deng et al. 2008b; Long et al. 2007; Long et al. 2008; Ostwald and Chen 2006), others concerned land development to urban built-up use (Deng et al. 2008a; Seto and Kaufmann 2003). Few have conducted a comprehensive, systematic analysis of land-use change at the national scale. One exception is Liu et al. (2003), who documented the spatial pattern of land-use change in China from 1995 to 2000, but did not quantify the major drivers of land-use change. A majority of finding from these 10 previous studies is that both geophysical and socioeconomic factors may affect land-use in China. Figure 2.1. Distribution of Land-use Change on Farmland, 1988–2000. The purposes of this essay are 1) to conduct a national-scale analysis to identify the major drivers of land use conversions in China; 2) to assess the relative importance of socioeconomic drivers; and 3) to design a cost-effective scheme to preserve farmland from development in traditional agricultural regions. To achieve these objectives, we compile a unique dataset that covers Mainland China, and develop a standard and a spatial multinomial logit models to analyze land use choice among six major uses (i.e., farmland, forestland, grassland, water area, urban area, and unused land). Other data provided by the Chinese Academy of Sciences, include terrain, climate, and socioeconomic variables, which are measured at a scale of 10 by 10 kilometers, except for socioeconomic data that are measured at county level – the most disaggregated unit available. 11 Figure 2.2. Distribution of Land-use Change on Grassland, 1988–2000. Discrete dependent variable models have been widely applied to the studies of land use and land-use change (Carrión-Flores and Irwin 2004; Lewis and Plantinga 2007; Lubowski et al. 2006; Nelson et al. 2001; Nelson and Hellerstein 1997; Polyakow and Zhang 2008; Wu et al. 2004; Wu and Cho 2007), where spatial autocorrelation are an important econometric concern because land uses are spatially distributed. The cost of not correcting for spatial dependence is inefficient and/or inconsistent estimates if the error structure or land-use choice is correlated over space (Anselin 2006). However, in the context of limited dependent variable model, it is technically challenging to overcome computational burdens when dataset is large. Some studies employ spatial sampling technique to solve this problem (Carrión-Flores and Irwin 2004); others construct spatial lags as instrumental variables for the right hand side of equation (Nelson et al. 2001; Nelson and Hellerstein 1997). But most of the literature ignores the potential spatial interdependence.2 In this study, we adopt two approaches, including an explicitly spatial 12 multinomial logit model, to correct for the potential endogeneity resulting from spatial autocorrelation in the dependent variable. This essay departs from previous studies in two aspects. First, we use unusually detailed, national-scale land use data, which were developed based on the US Landsat image with a spatial resolution of 30 by 30 meters. To the best of our knowledge, it is the first economic application in land use literature. Previous econometric studies on China typically aggregate land use data from pixel level to county level. The aggregation provides contiguous coverage of land conversion in a region but does not provide information on the spatial pattern of land-use change within a county. In contrast, this study uses data at highly disaggregated level, which helps understand the emergence of a collective pattern of land-use change where interdependence between land use choices is an important element. Second, we apply a new econometric technique in multinomial logit regression, which allows for modeling spatial autocorrelation explicitly with a large dataset. The remainder of this essay is organized as follows. Section 2 describes the land-use change model. Section 3 discusses data. Section 4 reports the estimation results. Section 5 and Section 6 present simulation results. The final section concludes. 13 THE MODEL To model land-use change in China, we must fully understand China‟s landownership. Unlike the United States and many European countries, China has no private land. Land can be owned by the state or by village collective, depending on land use types. For example, all urban land and most forest, pasture, water area, and unused land belong to the state; and all farmland is collectively owned by villagers. Land use is also heavily regulated by the government. The state retains the right to requisition farmland and other collectively owned land for urban construction, industrial development, and transport infrastructure by paying subsidies to villagers based on the original use of the land. Land requisition is the single type of land ownership transaction.3 Empirical Specification In this context of landownership, land use decision can be made by two types of agents – government (county-level or above) and village collective. They have different concerns: government officials are interested in their political and economic achievements to get more promotion opportunities, whereas individual villagers care the net returns to land. We assume that each type of (risk-neutral) agent makes land use decision to maximize utility. Based on their concerns, the utility of government includes the level of local GDP and image-building projects; while the utility of villagers comprise household income and employment opportunity. There are six alternative uses for each parcel of land: farmland, grassland, forestland, water area, urban area, and unused land. Let k and s be initial and final land use, respectively. We assume that urban development is irreversible, i.e., urban area will never be converted to nonurban uses. Therefore k can be any of five nonurban uses and s can be any of all six uses. Let U is|k denote the agent‟s utility from converting land grid i from use k to use s. U is|k can be decomposed into a deterministic component and an unobserved random component: Uis|k Vis|k is|k . The key variables affecting the deterministic component 14 Vis|k are identified based on urban and land economics theory. In a monocentric open city, the market land curve equals the upper envelope of the equilibrium bid rent curve of household at each location: (2.1) R t, r I t T r , where r is the distance from CBD, T r is the transport cost at r , and I t is the household income at time t . I 0, T 0, and Tr 0 . Although the assumption of competitive land market does not hold in China, Deng et al. (2008b) show that the monocentric city model has fairly high explanatory power when applied to China. So we adopt equation (2.1) to motivate the empirical specification of urban land rent. In a perfectly competitive market, rent in agricultural land equals revenue minus costs of other inputs per unit land given long-run profits is zero. Von Thünen‟s theory on agricultural land rent (Hall 1966) serves the theoretical basis of farmland bid rent. Based on the economic theories, we use five pixel-level geophysical variables and four county-level socioeconomic variables to construct Vis|k . The geophysical variables are land productivity, precipitation, temperature, the temporal variations in precipitation and temperature, respectively. These variables, discussed in details in the data section, measure agricultural yield potentials. For example, land productivity is estimates of crop yield and climate variables are supplements to land productivity.4 Three more pixel-level geophysical variables designed to capture spatial effects are discussed below. The socioeconomic variables are county GDP, population, public agricultural investment, and highway density. Specifically, population captures the effect of household income together with county GDP, highway density measures transport costs for conveying agricultural products, and public agricultural investment contributes to improving agricultural productivity in the long run. The unobserved random component is|k is assumed to follow a type-I extreme value distribution. Under this assumption, the probability of converting land grid i from use k to use l is: 15 Pil|k Pr U il|k U is|k , l s Pr Vil|k il|k Vis|k is|k , l s . Pr is|k il|k Vil|k Vis|k , l s Vil|k e Vis|k se (2.2) Equation (2.2) defines a multinomial logit regression model for each starting use k . To avoid redundant parameters, we set the initial use k as reference and normalize the corresponding coefficients to zero‟s such that Vik |k 0 . Hence there are five probability equations in the regression for each starting use k . We use maximum likelihood method to maximize the joint probability of multiple land-use choices based on equation (2.2). Econometric Issues Spatial autocorrelation is an important econometric concern when applying contiguous geographic data for empirical analysis. The cost of not correcting for spatial dependence is inefficient and/or inconsistent estimates if the error structure or land-use choice is correlated over space. But in practice it is technically challenging to distinguish between two types of spatial autocorrelation. Further, true residuals are unobservable in a limited dependent variable model, which makes it more difficult to test against spatial autocorrelation. Kelejian and Prucha (2001) develop a generalized Moran‟s I statistic (asymptotically equivalent to a Lagrange Multiplier statistic) that can be used to examine the existence of spatial error correlation. However, the econometric theory of testing for spatial interdependency of discrete LHS variable is still in its infancy in the literature.5 In this essay we ignore the potential for spatial dependence in error term because the estimates would be asymptotically efficient when the data sets used in estimation are extremely large.6 To correct for the potential endogeneity resulting from spatial autocorrelation in the dependent variable, we experiment with the following two approaches. In the first approach, we add three geophysical variables – terrain slope, elevation, and neighborhood index – as instruments to the right hand side (RHS) of the utility equation. We adopt a regular structure (i.e., an unlagged form) of terrain slope and 16 elevation instruments, which differs from the previous studies which use RHS spatial lags in the spatial analysis (Nelson et al. 2001; Nelson and Hellerstein 1997). Terrain slope and elevation used in this essay are able to capture the information from grids adjacent to the original location because they are generated from China‟s digital elevation model (DEM). DEM has taken spatial effects into account when estimating or retrieving the values of other locations during the interpolation process. Neighborhood index is constructed based on a six-dimensional vector, measuring the average of the percent land coverage of the eight cells surrounding the original location for each use. It is of theoretical significance to include neighborhood index in the utility equation. For example, in the local jurisdictional models, the land rents in the same community are correlated because they are affected by local public services providing by the local jurisdiction. The surrounding urban use coverage may serve as a proxy for the neighborhood effects. Hence the deterministic component of utility Vis|k can be written as (2.3) Vil|k V xil , y i , z m l|k xill|k y i βl|k z mγ l|k , where lk is transition-specific constant capturing conversion costs.; xil is the neighborhood index; y i is a vector of variables describing the locational characteristics of grid i, such as soil quality, topographic features, and weather conditions; and z m is a set of socioeconomic variables indexed by county m in respect that county is the most disaggregated unit available for measuring socioeconomic data. Since the absolute magnitude of coefficient in a multinomial logit model has no economic interpretation, we set initial use in k as reference and normalize the coefficients so that k |k 0 , k |k 0 , β k |k 0 , and γ k |k 0 , as we discussed in the previous section. The normalization avoids an overidentification issue in the regression. Building on the first approach, the second approach models the spatial autocorrelation explicitly. Specifically, we develop a spatial multinomial logit model by assuming that agents‟ utilities are spatially dependent.7 Having surrounding land in the 17 same use could help lower maintenance costs, encourage government to invest in infrastructure; agents may also benefit from knowledge spillover. In these situations, net returns to adjacent land parcels are correlated. Therefore, we add a spatially lagged utility to the RHS of the utility equation such that (2.4) U WU V ε , where is a spatial autoregressive parameter ( 1 ). The magnitude of represents the extent to which an element of LHS variable ui is affected by the remaining elements u j for j i . Thus the standard model is a special case of the spatial model when 0 . W is a row-standardized n n weight matrix such that wii 0 and n j 1 wij 1 . We specify the (i,j)th entry of the weight matrix W as a Gaussian function of geographical distance from location j to location i as equation (2.5) shows.8 (2.5) wij exp dij2 h2 n j 1 exp dij2 h2 , i, j 1, , n, and i j , where d ij measures the Euclidean distance between location i and j , and h is referred to as the bandwidth. The reduced form of equation (2.4) is given by (2.6) U I W V I W ε . 1 1 For notational convenience, let V* I W V . Now the expression of probability of 1 converting grid i from land-use k to land-use l is (2.7) Pil|k Vil*|k e s e is|k V* . In practice it is infeasible to conduct estimation in the context of limited dependent variable regression when dataset is large, because evaluating a log-likelihood function needs an n-dimensional integration where n is sample size. To overcome this problem, Pinkse and Slade developed a generalized method of moments, which works for a spatial error model. But it is still technically impracticable to apply this approach in a spatial lag model. Recently, Klier and McMillen revised Pinkse and Slade‟s method by developing a linearized9 logit version in a spatial lag framework, which simplifies the 18 algorithm to only two steps – a standard logit estimation followed by a (linear) two-stage least squares regression. Hence it is feasible to estimate a logit model with spatially lagged dependent variables in a large dataset. This study follows this strand and extends Klier and McMillen‟s approach to the context of multinomial logit regression. To the best of our knowledge, no published studies have done this before. However, as we have discussed, there are no formal results in the literature so far that can test for the existence of spatial interdependency of discrete LHS variable. Therefore we directly test against the null hypothesis of 0 using Student‟s t statistics reported in the two-stage least squares regression. If the test rejects the null, then spatial endogeneity exists in the dependent variable. Another difficulty with the spatial regression is that the estimated parameters are, in part, functions of the weighting function. As the bandwidth h tends to infinity, the weighting function exp dij2 h2 is close to one for all pairs of points so that wij n 1 1 j i . Equivalently, the weight becomes uniform for every point j no matter how far it is from location i . Conversely, as h becomes smaller, utility will increasingly depend on observations in close proximity to i . In particular, the weighting function exp dij2 h2 tends to zero when the distance d ij exceeds approximately 2.15 times as long as the bandwidth h . The problem hence becomes how to select an appropriate bandwidth or decay function in regression. In this study we assume a uniform h for all models and choose h on a criterion of minimum Predicted Residual Error Sum of Squares (PRESS), where the fitted value with the point i omitted from the calibration process. In the remaining of this essay, we refer to the model developed based on the first correction as the standard multinomial logit model, and refer to the model estimated by using the second approach as the spatial multinomial logit model.10 19 DATA Our study covers Mainland China. Most data used in this essay were provided by the Chinese Academy of Sciences (CAS), including land-use type, terrain, climate, and socioeconomic data. They are contiguous data measured at a scale of 10 by 10 square kilometers, except for socioeconomic data, which are measured at county level. Contiguous data are more desirable than dispersed sample plots in the prediction of land conversions. Table 2.1 provides a detailed summary of the data. Table 2.1. Summary Statistics of Explanatory Variables Variable M easurement Unit 10-km-gird level Land productivity Terrain slope Elevation Precipitation, 1991-1995 Precipitation, 1996-2000 Std. of precipitation, 1991-1995 Std. of precipitation, 1996-2000 Temperature, 1991-1995 Temperature, 1996-2000 Std. of temperature, 1991-1995 Std. of temperature, 1996-2000 g/ha. degree km 1000 mm 1000 mm 1000 mm 1000 mm degree Celsius degree Celsius degree Celsius degree Celsius county level highway GDP, 1989 GDP, 1996 GDP, 2000 Population, 1989 Population, 1996 Population, 2000 Agricultural investment, 1994 Agricultural investment, 1995 Agricultural investment, 1999 Agricultural investment, 2000 m/10000 ha. billion RM B yuan billion RM B yuan billion RM B yuan million people million people million people million RM B yuan million RM B yuan million RM B yuan million RM B yuan N M ean Std. Dev. M inimum M aximum 93902 94662 94612 94173 94173 94173 94173 94173 94173 94173 94173 1.413 3.555 1.837 0.468 0.478 0.081 0.081 6.298 6.677 0.385 0.599 2.632 5.010 1.742 0.421 0.436 0.077 0.067 8.021 8.045 0.105 0.153 0.000 0.000 -0.153 0.006 0.006 0.002 0.002 -17.740 -17.000 0.084 0.239 14.168 72.790 7.040 1.877 1.824 0.402 0.368 31.420 31.620 0.904 1.693 2331 2236 2247 2251 2331 2332 2333 2138 2137 2143 2140 1.022 1.351 2.593 3.956 0.468 0.510 0.529 0.073 0.076 0.077 0.096 3.794 3.579 6.518 11.121 0.456 0.499 0.514 0.379 0.410 0.423 0.526 0.000 0.016 0.021 0.041 0.005 0.006 0.006 0.000 0.000 0.000 0.000 155.708 116.195 202.418 364.877 10.228 10.616 10.817 11.783 13.055 13.653 17.057 Land-use data are generated from a unique land cover and land use database, which was developed based on the US Landsat TM/ETM images with a spatial resolution of 30 by 30 meters (Deng et al. 2008a; Liu et al. 2003). The data are available for three years – the late 1980s, the mid-1990s, and the late 1990s, denoted as 1988, 1995, and 20 2000, respectively. CAS made visual interpretation and digitization of TM images to generate thematic maps of land cover, and sorted the data with a hierarchical classification system of 25 land cover classes. Further, CAS grouped 25 classes of land cover into 6 aggregated classes of land use, i.e., farmland, forestland, grassland, water area, urban area11, and unused land. Deng et al. (2006) provides a detailed explanation of the six land-use types. Table 2.2a. Land-use Transitions from 1988 to 1995 Initial land-use Farm Forest Grass Water Urban Unused Freq Prob Freq Prob Freq Prob Freq Prob Freq Prob Freq Prob Total Farm 11,131 0.662 2,787 0.125 1,931 0.064 415 0.155 106 0.321 246 0.012 16,616 Forest 2,952 0.176 15,976 0.719 2,974 0.099 179 0.067 29 0.088 312 0.016 22,422 Final land-use Grass Water 1,947 386 0.116 0.023 2,997 161 0.135 0.007 21,333 336 0.709 0.011 400 1,353 0.150 0.506 16 10 0.048 0.030 3,142 329 0.157 0.016 29,835 2,575 Urban 178 0.011 36 0.002 16 0.001 28 0.010 160 0.485 11 0.001 429 Unused 212 0.013 272 0.012 3,518 0.117 298 0.111 9 0.027 16,026 0.799 20,335 Total 16,806 1 22,229 1 30,108 1 2,673 1 330 1 20,066 1 92,212 Urban 100 0.006 24 0.001 11 0.000 12 0.005 305 0.696 7 0.000 459 Unused 152 0.009 253 0.011 2,630 0.088 268 0.103 4 0.009 16,665 0.819 19,972 Total 16,636 1 22,469 1 29,861 1 2,596 1 438 1 20,359 1 92,359 Table 2.2b. Land-use Transitions from 1995 to 2000 Initial land-use Farm Forest Grass Water Urban Unused Total Freq Prob Freq Prob Freq Prob Freq Prob Freq Prob Freq Prob Farm 12,531 0.753 2,344 0.104 1,720 0.058 235 0.091 85 0.194 188 0.009 17,103 Forest 2,122 0.128 17,422 0.775 2,261 0.076 97 0.037 15 0.034 204 0.010 22,121 Final land-use Grass Water 1,478 253 0.089 0.015 2,281 145 0.102 0.006 22,937 302 0.768 0.010 248 1,736 0.096 0.669 9 20 0.021 0.046 3,025 270 0.149 0.013 29,978 2,726 21 Table 2.2a-2.2b describe land transition matrices of six land-use classes for the time intervals of 1988-1995 and 1995-2000. Land-use exchanges mainly occur between farmland, forestland, and grassland, as well as between grassland and unused land. Urban area expansion is not as significant as anticipated when viewed from a national perspective. Data on geophysical variables are generated from a geographical information system (GIS) database, including cross-sectional data of land productivity, terrain slope, and elevation. Land productivity is a pixel-specific (5-kilometer-grid) variable, originally estimated by a research team from Institute of Geographical Sciences and Natural Resources Research, CAS by using standalone software of Estimation System for the Agricultural Productivity (Deng et al. 2006). Terrain slope and elevation are generated from China‟s digital elevation model as part of the basic CAS database. Climate panel data are initially collected from over 400 weather stations and organized by the Meteorological Observation Bureau of China. The dataset includes mean annual precipitation and mean annual temperature from 1991 to 2000, CAS interpolated the point climate data into surface data with the method of thin plate smoothing spline (Hartkamp et al. 1999) to get more disaggregated information for each pixel. We calculate the standard deviations of mean annual precipitation and mean annual temperature along time, and use them as measures of temporal variations in climate. Socioeconomic variables, such as county GDP and population are gathered from several versions of statistical yearbooks and population yearbooks for China‟s counties and cities for three years (1989, 1996, and 2000). A common suggestion is that county GDP and population are “endogenous.” We use lagged county GDP and population measures (1989, 1996), so that statistical endogeneity seems less likely. Data on public agricultural investment are collected from province and county level statistical yearbooks for four years (1994, 1995, 1999, and 2000). The investment comes from fiscal budget of the state and local government. It is mainly used for developing agriculture infrastructure like seeds, fertilizers, and irrigation. Data on highway density are available for one year. Based on a digital map of transportation networks in the mid-1990s, highway density are 22 calculated as the total length of all highways in a county divided by land area of that county. Data in value terms are measured at the 2000 real Chinese yuan (hereafter CNY or ¥). All of these variables are county-level data. 23 RESULTS We estimate the standard and the spatial multinomial logit models separately with a dataset composed of observations at a 10-km-land-grid scale. There are two transition periods, 1988-1995 and 1995-2000, for the analysis. During each period there are five initial land uses (farmland, forestland, grassland, water area, and unused land) and six final uses (farmland, forestland, grassland, water area, urban area, and unused land). So we estimate twenty separate models in total.12 We apply maximum likelihood method to estimate ten standard multinomial logit models. The pseudo R2 (McFadden's likelihood ratio index) ranges between 0.546 and 0.825. Estimation of ten spatial multinomial models is conducted using the linearized generalized method of moments. Appendix A provides a detailed description of the algorithm. The estimated value of the bandwidth h is 200 km. The essential idea behind the bandwidth is that for each location i there is a circle centered at i with a radius of 430 km ( 430km 200km 2.15 ). Within the circle points around i have “bump of influence” on i ; beyond the circle the influence of points are negligible. To save space, we will report and discuss the estimation results of land-use change on farmland and grassland for two transition periods, presenting the remaining results in the Appendix B. Table 2.3a-2.3b report coefficient estimates for the standard model of land conversion on farmland, respectively, from 1988 to 1995 and from 1995 to 2000. Estimates and standard errors of parameters in equation (2.3) are presented in each column by land-use choice. It shows that the sign, magnitude, and statistical significance of estimates are consistent in transition period and are in line with the economic interpretation on the whole. For example, all transition-specific constants have negative estimates and almost all of them are statistically different from zero at the 1% level, indicating that conversion cost deters land conversion on farmland. Likewise, the estimates of land productivity are negatively significant, implying that a patch of farmland with higher crop yield potential is less likely to be changed to other uses. There is evidence that the odds of farmland conversion are associated with climate, e.g., a patch 24 of high-rainfall farmland is more likely to be afforested and a patch of low-rainfall and low-temperature farmland is less likely to be abandoned. The results provide strong evidence for the association between the odds of farmland conversion and county-level socioeconomic factors, such as county GDP, population, agricultural investment, and highway density. It is particularly significant during the transition period of 1988-1995. As shown in Table 2.3a and 2.3b, land is more likely to be changed out of farming in a county with higher level of GDP. With higher GDP, the demand for residential development and industrial and commercial uses increases. Farming is generally a low-paying job in China, so farmers are willing to be engaged in other higher-returned activities rather than farming. Conversely, farmers are more likely to farm in a county with low GDP for the lack of high-paying jobs. Increased population may increase labor supply. In a county with large population, farmland is less likely to be converted to other uses because farming is still the main way to make a living if the industry is underdeveloped. Likewise, public agricultural investment contributes to improving agricultural productivity in the long run, which decreases the probabilities of farmland conversions. In the empirical model of this study, highway density measures freights of conveying agricultural products. Dense highway tends to lower transport costs, which decreases the probabilities of farmland conversions. The GDP coefficient on urban expansion is not as statistically significant as anticipated, especially in the second period. For one reason, urban development mainly occurred in East China. Therefore GDP only has moderate effects on urban expansion if viewed from the whole country. For another, the neighborhood effect, as a proxy for accessibility (or location) rent is so large that it outperforms the role of GDP in encouraging urban development. Table 2.4a-2.4b report estimated parameters of the spatial multinomial model of land conversion on farmland for the periods of 1988-1995 and 1995-2000.13 Likewise, the sign, magnitude, and statistical significance of estimates are generally consistent for the two periods. The spatial autoregressive parameter ( ) is estimated to be 0.0760 and 0.3308 and is statistically significant at 5% and 1% levels for two transition periods, respectively. Positive estimates imply positive spatial externalities of land-use choices Table 2.3a. Coefficient Estimates for the Standard M ultinomial Logit M odel of Land-use Change on Farmland, 1988-1995 Indep. Variable Forestland Grassland Water area Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -2.4944*** (0.2027) -2.4109*** (0.2505) -4.0919*** (0.4224) -6.1098*** (0.7112) Land productivity -0.0570*** (0.0122) -0.0959*** (0.0144) -0.0577*** (0.0207) -0.0457* (0.0270) County GDP 0.0479*** (0.0167) 0.0623*** (0.0202) 0.0706** (0.0293) 0.0944*** (0.0332) Population -0.3910*** (0.1031) -0.6637*** (0.1354) -0.2232* (0.1271) -0.1864 (0.1854) Agricultural investment -0.1021 (0.1508) -0.0822 (0.1196) -0.5175 (0.4002) -0.7971** (0.3207) Highway density -0.1156*** (0.0296) -0.1249*** (0.0464) 0.0040 (0.1078) 0.0429 (0.0534) Terrain slope 0.0541*** (0.0104) 0.0799*** (0.0108) 0.0169 (0.0338) -0.1900 (0.1186) Elevation 0.0950* (0.0576) 0.1758*** (0.0574) -0.2773 (0.1733) 0.3259 (0.2554) Precipitation 0.9175*** (0.1924) -0.0651 (0.2312) -0.7200* (0.4004) 0.4239 (0.5523) Temperature -0.0065 (0.0099) 0.0142 (0.0105) 0.0889*** (0.0244) 0.1137** (0.0452) Std Err of precipitation -2.4573*** (0.6117) -0.6045 (0.9496) 0.0051 (1.1433) -0.2298 (1.7038) Std Err of temperature -0.9903*** (0.3680) -0.2820 (0.4576) -0.5643 (0.8490) -2.6193** (1.1561) Neighborhood index 0.0511*** (0.0012) 0.0519*** (0.0015) 0.0869*** (0.0032) 0.1032*** (0.0048) Number of observations 15012 M cFadden's LRI 0.6436 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -0.9629 (0.7185) -0.1269*** (0.0320) -0.0004 (0.0905) -0.2542 (0.2555) -1.0209 (1.2504) 0.0370 (0.1292) -0.0129 (0.0818) -0.7721*** (0.2108) -3.0655*** (0.8473) -0.0753*** (0.0223) -0.3552 (4.8060) -0.7981 (1.3608) 0.0530*** (0.0032) 25 Table 2.3b. Coefficient Estimates for the Standard M ultinomial Logit M odel of Land-use Change on Farmland, 1995-2000 Indep. Variable Forestland Grassland Water area Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -3.8590*** (0.2606) -3.2968*** (0.2424) -4.9410*** (0.6075) -7.2280*** (0.9423) Land productivity -0.1027*** (0.0136) -0.1318*** (0.0157) -0.0510** (0.0233) -0.0463 (0.0368) County GDP 0.0026 (0.0081) -0.0045 (0.0146) 0.0315* (0.0166) 0.0114 (0.0226) Population 0.0379 (0.0947) -0.3445*** (0.1274) -0.2059 (0.2344) 0.0661 (0.2134) Agricultural investment -0.1981* (0.1019) -0.0524 (0.1141) -0.2718 (0.2381) -0.1176 (0.2517) Highway density -0.1610*** (0.0504) -0.1808*** (0.0635) -0.1645 (0.1513) -0.0802 (0.1757) Terrain slope 0.0726*** (0.0091) 0.0498*** (0.0110) -0.2091*** (0.0585) -0.0871 (0.1163) Elevation 0.0780 (0.0591) 0.4166*** (0.0643) -0.0817 (0.1958) 0.4502* (0.2462) Precipitation 0.6961*** (0.1513) -0.3161* (0.1863) 0.8743** (0.3844) 0.5525 (0.5124) Temperature -0.0020 (0.0101) 0.0221* (0.0116) 0.0040 (0.0290) 0.0392 (0.0422) Std Err of precipitation 0.3500 (0.5999) 0.3058 (0.9794) -0.6821 (1.7468) -1.5711 (2.9525) Std Err of temperature 1.0990*** (0.3650) 1.0649*** (0.2832) 1.2361 (0.8399) 1.1539 (1.1732) Neighborhood index 0.0419*** (0.0013) 0.0383*** (0.0015) 0.0567*** (0.0037) 0.0980*** (0.0064) Number of observations 14794 M cFadden's LRI 0.6662 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -0.4472 (0.8303) -0.1482*** (0.0411) 0.0429** (0.0208) -0.7696 (0.5879) 0.1167 (0.3628) -0.1332 (0.1735) -0.2446* (0.1289) -0.3200 (0.2467) -4.3673*** (1.0255) -0.1388*** (0.0324) 1.3712 (3.3825) -0.6066 (1.1337) 0.0438*** (0.0040) 26 Table 2.4a. Coefficient Estimates for the Spatial M ultinomial Logit M odel of Land-use Change on Farmland, 1988-1995 Indep. Variable Forestland Grassland Water area Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -2.4164*** (0.1622) -2.6617*** (0.2012) -3.8213*** (0.7071) -5.4724*** (1.6243) Land productivity -0.0680*** (0.0077) -0.1186*** (0.0090) -0.0152 (0.0962) 0.2871 (0.6642) County GDP -0.0997** (0.0467) -0.1016** (0.0494) -0.0788 (0.4152) 0.1159 (0.8281) Population 0.3707*** (0.0099) -0.6283*** (0.0132) -0.2020*** (0.0314) -0.1826*** (0.0599) Agricultural investment -0.1800*** (0.0404) -0.0529 (0.0437) -0.5279*** (0.1216) -0.8027*** (0.0700) Highway density -0.1230*** (0.0120) -0.1316*** (0.0186) -0.0056 (0.0372) 0.0135 (0.0405) Terrain slope 0.1243 (0.0860) 0.0883 (0.1428) 0.0815 (0.1703) -0.0204 (0.3133) Elevation 0.1127 (0.0860) 0.2020* (0.1081) -0.1642 (0.5866) 0.4496 (0.4085) Precipitation 0.7794*** (0.1449) 0.2288 (0.1828) -0.4695 (0.6419) 1.2320 (1.3165) Temperature -0.0053 (0.0070) 0.0034 (0.0078) 0.0743* (0.0420) 0.0277 (0.1018) Std Err of precipitation -2.7082*** (0.4249) -1.6864** (0.8138) -0.7702 (1.6734) -1.1668 (3.5181) Std Err of temperature -0.5423** (0.2733) 0.7490** (0.3310) -0.5931 (1.2747) -1.9306 (2.4402) Neighborhood index 0.0520*** (0.0009) 0.0534*** (0.0011) 0.0866*** (0.0040) 0.1036*** (0.0093) Spatial parameter (ρ) 0.0760** (0.0333) Number of observations 15012 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err 1.1830 (1.1756) 0.0027 (0.2380) 0.0701 (0.4799) -0.2442*** (0.0819) -1.0304*** (0.1476) 0.1581 (0.1877) 0.4488 (0.9424) -2.2122 (2.6257) -1.9152 (2.3880) -0.0731 (0.0491) -11.4784 (9.9609) -3.0681* (1.8193) 0.0520*** (0.0032) 27 Table 2.4b. Coefficient Estimates for the Spatial M ultinomial Logit M odel of Land-use Change on Farmland, 1995-2000 Indep. Variable Forestland Grassland Water area Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -2.4895*** (0.2394) -2.1248*** (0.1949) -3.0390*** (1.103) -6.4762** (2.6449) Land productivity -0.1100*** (0.0056) -0.1419*** (0.0073) -0.2288 (0.2453) 0.1070 (0.5066) County GDP -0.2073*** (0.0414) -0.3589*** (0.0493) 0.4184 (0.4735) 0.2532 (0.8937) Population 0.0745*** (0.0103) -0.2951*** (0.0145) -0.2072*** (0.0390) 0.0573 (0.0833) Agricultural investment -0.3675*** (0.0509) -0.1681*** (0.0530) -0.2051 (0.3024) -0.2037 (0.2658) Highway density -0.1764*** (0.0079) -0.2601*** (0.0281) -0.1750*** (0.0180) -0.0995*** (0.0365) Terrain slope 0.1755** (0.0717) 0.2501* (0.1455) -0.0581 (0.2787) 0.1157 (0.4991) Elevation -0.1064 (0.1267) 0.3102*** (0.1131) -0.0685 (0.2742) 0.5453 (0.3399) Precipitation 0.1399 (0.1271) -0.0453 (0.1428) 0.9428 (0.6093) 1.0594 (1.5738) Temperature -0.0109 (0.0069) 0.0035 (0.0078) -0.0113 (0.0436) 0.0553 (0.1321) Std Err of precipitation 0.5744 (0.3746) 0.8760 (0.7963) -0.7223 (2.6062) -4.5177 (7.4682) Std Err of temperature 1.1702*** (0.2649) 1.0190*** (0.1764) 0.3942 (1.6958) 2.4109 (3.3430) Neighborhood index 0.0405*** (0.0008) 0.0353*** (0.0009) 0.0548*** (0.0034) 0.1013*** (0.0132) Spatial parameter (ρ) 0.3308*** (0.0357) Number of observations 14794 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -1.9842* (1.1742) -0.8823* (0.5176) 1.0089* (0.5360) -0.7455*** (0.1264) 0.3147 (0.2937) -0.1344*** (0.0383) -0.4215 (1.2291) -0.3368 (0.8959) -0.4517 (2.5570) -0.0354 (0.0758) 10.6778 (4.7202) -0.4945 (1.7233) 0.0421*** (0.0032) 28 29 between neighboring locations. Particularly, it demonstrates an increasing trend of spatial dependence over time. A comparison of results in Table 2.4a-2.4b and Table 2.3a-2.3b shows that the signs and relative magnitudes of most estimates, such as land productivity, population, agricultural investment, highway density, neighborhood index, etc., are robust to the spatial lag specification. Nevertheless, we find that in the utility equation of converting farmland to unused land, the absolute values of estimated coefficients on standard error of precipitation are extremely high, indicating that results of spatial multinomial model are sensitive to some explanatory variables especially when a choice is least likely to be selected. In addition to land conversion on farmland, we report coefficient estimates of land-use change on grassland, for the standard and spatial model, and for two transition periods in Tables 2.5a-2.6b. Similar to the foregoing outcome presented in Tables 2.3a2.4b, the results are generally consistent in time and robust to the spatial lag specification, including sign, magnitude, and statistical significance of estimates. For example, transition-specific constants are estimated to be significantly negative and neighborhood indices are estimated to be significantly positive at the 1% level. Estimates of land productivity, population, and highway density are statistically positive in the utility equations of farmland and forestland, implying that cultivation and afforestation are more likely to take place on a patch of grassland that possesses higher productivity, larger population, or denser highway. By contrast, a patch of grassland with lower productivity or smaller population is more likely to be converted to unused land. Compared with county GDP, population is a more stable socioeconomic factor in driving grassland change. In the context of grassland conversion, the magnitude of spatial dependence is not as time-sensitive as that of change on farmland. Spatial autoregressive parameters ( ) in two periods are respectively estimated to be 0.3709 and 0.3542, which are statistically significant at the 1% level. Again, in the utility equation of small probability event (e.g., urban development in this case), we find unusual coefficients estimates of some explanatory variables (e.g., intercept, terrain slope, standard error of precipitation and 30 temperature). Hence caution should be exercised when applying the spatial multinomial model, where the estimates are not robust to small probability events. Table 2.5a. Coefficient Estimates for the Standard M ultinomial Logit M odel of Land-use Change on Grassland, 1988-1995 Indep. Variable Farmland Forestland Water area Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -2.5272*** (0.2438) -3.2139*** (0.2120) -4.5855*** (0.4887) -8.1604*** (2.9041) Land productivity 0.0782*** (0.0171) 0.1272*** (0.0197) 0.1898*** (0.0508) 0.3214* (0.1786) County GDP 0.0527** (0.0250) -0.0053 (0.0366) 0.1095* (0.0574) 0.0575 (0.6718) Population 0.4507*** (0.1550) 0.8189*** (0.1669) 0.0486 (0.5327) -0.8714 (3.6464) Agricultural investment -0.1943 (0.1893) 0.3088** (0.1298) 0.4021 (0.4338) -0.1310 (4.6872) Highway density 0.1609*** (0.0428) 0.0957 (0.0642) 0.1711 (0.1471) -0.4573 (0.7783) Terrain slope 0.0140 (0.0095) 0.0380*** (0.0069) -0.0609*** (0.0228) -0.0157 (0.1109) Elevation -0.4717*** (0.0510) -0.1437*** (0.0335) -0.1110 (0.0848) 0.5892 (0.7513) Precipitation 0.5282** (0.2319) 0.7794*** (0.2024) 0.5972 (0.6803) -0.9467 (3.8030) Temperature -0.0039 (0.0105) -0.0163** (0.0080) -0.0669*** (0.0221) 0.2157 (0.2476) Std Err of precipitation 0.5156 (1.0097) -1.3059 (0.8522) 1.2753 (2.7736) -0.2273 (20.728) Std Err of temperature -1.0408** (0.4464) -0.8175** (0.4057) -0.6959 (0.9127) -3.7185 (8.2997) Neighborhood index 0.0556*** (0.0015) 0.0583*** (0.0013) 0.0997*** (0.0038) 0.1662*** (0.0339) Number of observations 17893 M cFadden's LRI 0.6611 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -2.4478*** (0.2899) -0.1199** (0.0549) 0.1456*** (0.0426) -0.5820 (0.4588) -1.1709*** (0.3279) 0.0610 (0.0844) -0.0324*** (0.0083) 0.0267 (0.0481) -1.1148*** (0.4158) -0.0166 (0.0112) 1.5814 (2.4712) -1.3545*** (0.4911) 0.0559*** (0.0013) 31 Table 2.5b. Coefficient Estimates for the Standard M ultinomial Logit M odel of Land-use Change on Grassland, 1995-2000 Indep. Variable Farmland Forestland Water area Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -3.1325*** (0.2143) -3.4663*** (0.1859) -5.7489*** (0.7341) -5.0566 (9.2609) Land productivity 0.1507*** (0.0147) 0.0755*** (0.0183) 0.2003*** (0.0652) 0.1022 (0.7749) County GDP -0.0741*** (0.0170) -0.0139 (0.0153) 0.0166 (0.0731) -0.1279 (0.7389) Population 0.6967*** (0.1522) 0.7112*** (0.1441) 0.3270 (0.8176) -0.7448 (10.550) Agricultural investment 0.2416** (0.1163) -0.1530 (0.1156) 0.0391 (0.8088) -0.3704 (31.015) Highway density 0.2093*** (0.0612) 0.1865*** (0.0625) 0.1130 (0.1575) 0.6222 (0.5172) Terrain slope 0.0051 (0.0099) 0.0459*** (0.0063) -0.0908*** (0.0240) -0.1523 (0.6673) Elevation -0.3827*** (0.0568) -0.1004*** (0.0341) 0.2322 (0.1546) -2.2314 (8.8000) Precipitation -0.0689 (0.1657) 0.0342 (0.1496) -0.1111 (0.8861) 0.0435 (13.290) Temperature 0.0363*** (0.0094) -0.0001 (0.0073) -0.0044 (0.0313) 0.0735 (0.6435) Std Err of precipitation -0.3487 (0.9213) 3.1369*** (0.7336) 0.5856 (4.7753) 0.1013 (65.121) Std Err of temperature 0.1826 (0.2723) -0.0810 (0.2547) 0.5449 (0.7414) -1.6131 (12.739) Neighborhood index 0.0435*** (0.0015) 0.0398*** (0.0012) 0.0819*** (0.0047) 0.0927 (0.1070) Number of observations 18116 M cFadden's LRI 0.6157 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -3.8559*** (0.2912) -0.0911** (0.0425) 0.1583*** (0.0175) -1.8065*** (0.3724) -1.4602*** (0.2927) -0.7420*** (0.0853) -0.0359*** (0.0080) 0.3706*** (0.0561) -2.4384*** (0.3838) 0.1005*** (0.0121) 0.1737 (1.93550 1.9762*** (0.3033) 0.0287*** (0.0012) 32 Table 2.6a. Coefficient Estimates for the Spatial M ultinomial Logit M odel of Land-use Change on Grassland, 1988-1995 Indep. Variable Farmland Forestland Water area Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -1.2389*** (0.2120) -1.7452*** (0.1945) -4.1508*** (1.3296) -15.369 (17.646) Land productivity 0.0752*** (0.0091) 0.1233*** (0.0048) 0.2215*** (0.0550) 1.2020* (0.6366) County GDP 0.4807*** (0.0455) 0.1561*** (0.0258) 0.2479 (0.1884) -1.9941 (4.0108) Population 0.4411*** (0.0134) 0.7872*** (0.0162) 0.0936 (0.1232) -0.8311* (0.4721) Agricultural investment -0.2927*** (0.0363) -0.4496*** (0.0493) 0.4143*** (0.0878) -3.1804 (2.4861) Highway density 0.1770*** (0.0185) 0.1045*** (0.0317) 0.1350** (0.0683) 0.1944 (1.1951) Terrain slope -0.5795*** (0.1284) 0.4184*** (0.1292) 0.8213 (1.3910) -20.545* (7.9721) Elevation -0.3153** (0.1375) -0.0700 (0.1820) -0.1089 (0.2945) -0.8731 (11.818) Precipitation -0.3434* (0.1903) -1.0597*** (0.1809) -1.1070 (1.6367) 5.9941 (12.212) Temperature -0.0129 (0.0087) 0.0103* (0.0061) -0.1032* (0.0583) 1.0981 (0.9376) Std Err of precipitation 2.0066** (0.7889) 0.4617 (0.6011) 11.845* (6.8454) -74.123 (55.410) Std Err of temperature -1.8063*** (0.3588) -1.0802*** (0.3070) 0.9165 (2.4233) 23.377 (30.939) Neighborhood index 0.0522*** (0.0011) 0.0556*** (0.0010) 0.0976 (0.0068) 0.2504 (0.1243) Spatial parameter (ρ) 0.3709*** (0.0257) Number of observations 17893 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -2.2998*** (0.2716) -0.1152*** (0.0104) 0.2036*** (0.0470) -0.4763*** (0.0978) -1.1561*** (0.0755) 0.0004 (0.0384) 0.5781 (0.5507) 0.6049* (0.3200) -0.3755 (0.4613) 0.0040 (0.0103) 2.8330 (3.1432) -0.2967 (0.4525) 0.0531 (0.0010) 33 Table 2.6b. Coefficient Estimates for the Spatial M ultinomial Logit M odel of Land-use Change on Grassland, 1995-2000 Indep. Variable Farmland Forestland Water area Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -2.4354*** (0.2097) -2.2088*** (0.1806) -4.0181** (1.7919) -18.003 (27.936) Land productivity 0.1509*** (0.0098) 0.0657*** (0.0044) 0.1694* (0.0947) 1.1325 (2.4068) County GDP 0.3276*** (0.0574) 0.0264 (0.0266) -0.0061 (0.3465) -4.7848 (17.159) Population 0.6756*** (0.0120) 0.6873*** (0.0157) 0.3905*** (0.1222) -2.0046** (0.8319) Agricultural investment 0.1693*** (0.0494) -0.3135*** (0.0476) 0.0468 (0.4883) 0.2128 (0.8250) Highway density 0.2077*** (0.0184) 0.1971*** (0.0140) 0.0411 (0.0957) 1.1293 (1.9200) Terrain slope -0.3060** (0.1320) -0.2399* (0.1229) 0.5193 (1.4622) 12.592 (17.755) Elevation -0.4028*** (0.0733) -0.0310 (0.1199) 0.2544 (1.2613) -45.132 (58.018) Precipitation -0.3070** (0.1377) -0.7090*** (0.1323) 0.4320 (1.9937) -12.234 (12.927) Temperature 0.0136 (0.0090) 0.0052 (0.0058) -0.0433 (0.0825) 0.7652 (0.9008) Std Err of precipitation 1.1634 (0.7682) 1.8629*** (0.5421) 0.5036 (11.469) 61.710 (59.686) Std Err of temperature 0.4735** (0.2226) 0.4435* (0.2338) 0.3880 (1.9510) 17.952 (54.206) Neighborhood index 0.0385*** (0.0011) 0.0367*** (0.0009) 0.0791*** (0.0067) -0.1580 (0.1477) Spatial parameter (ρ) 0.3542*** (0.0330) Number of observations 18116 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -3.5409*** (0.2751) -0.0991*** (0.0102) 0.0412 (0.0601) -1.9724*** (0.0732) -0.9308*** (0.1529) -0.7098*** (0.0153) -0.6200 (0.4491) 0.7635** (0.3426) -0.9096** (0.4186) 0.0794*** (0.0119) -0.6759 (1.9727) 2.7153*** (0.2338) 0.0281*** (0.0008) 34 35 DRIVERS OF LAND-USE CHANGE Although the results in Tables 2.3a-2.6b demonstrate the significance of explanatory variables in land-use change decisions, they say little about the relative importance of their influences. Due to the nonlinear, multinomial form of the model, answer to the second question – what are the relative importance of socioeconomic drivers – can be discerned only through a series of simulations. We use the empirical standard multinomial logit models14 to predict national land-use change by uses from 1988 to 2000 under a factual and four counterfactual scenarios. As described in Table 2.7a, the factual simulation uses actually historical observations; it provides a benchmark to measure land use changes under counterfactual scenarios; the counterfactual scenarios respectively hold county GDP, population, agricultural investment, and neighborhood index at a hypothetical level and keeps the remaining variables at their historically observed values. Table 2.7a. Description of Simulation Scenarios Scenario Description Factual All variables at actual values No change in GDP Fix county GDP at 1989 values No change in population Fix population at 1989 values No agricultural investment Restrict agricultural investment coefficients to be zero No neighborhood effects Restrict neighborhood index coefficients to be zero Simulations are run at a 10-km-land-grid scale (equivalent to 10,000 hectares). Change in the total area for each use between 1988 and 2000 is estimated in the following five steps: 1) using the models for the period of 1988-1995 to predict the probabilities of land-use choice of every gird in 1995, given the historical use in 1988;15 2) using the models for the period of 1995-2000 to estimate probabilities of land-use choice of each grid in 2000, conditional on six uses in 1995; 3) multiplying the probabilities predicted in the first step by the conditional probabilities predicted in the second step, and hence obtaining the joint probabilities of land-use choice in 2000 for every individual land cell; 36 4) summing the choice probabilities by land-use type across individuals and multiplying the summations by 10,000 hectares; 5) calculating the difference between aggregate hectare of each use estimated in the fourth step and the historical land-use hectare in 1988. The procedure is applied to every scenario. For each use, we divide the outcome of the factual simulation, obtained from the fifth step, by the actual land area in 1988 and then use these ratios to evaluate the performance of the simulation model. On the whole, the simulation model performs moderately well in regenerating the direction and relative magnitudes of land-use changes from 1988 to 2000. Except for a bit underestimation of unchanged land area by -13.68% to -5.71%, errors of the factually-simulated changed area are within a range of 0.03-5.72% for all uses. Table 2.7b reports the simulation results, where change in hectare is the total simulated land area change for each use between 1988 and 2000, and percent change is calculated by subtracting the factually-simulated hectare change from the counterfactually-simulated hectare change, and divided the difference by the factually-simulated hectare change. It measures the direction and relative magnitudes of counterfactual scenario to the factual scenario. In particular, positive (negative) value of percent change indicates that the factor contributes to decreasing (increasing) the land hectare for that use. The simulation results reveal that socioeconomic factors have large impacts on land-use change. As we have discussed in the previous section, GDP growth leads to farmland loss (i.e., positive percent change) while population growth and agricultural investment contribute to farmland expansion (i.e., negative percent change). Further, the magnitudes of GDP and agricultural investment are greater than population, implying that GDP growth and agricultural investment are more important than population growth in driving farmland change. In contrast, grassland change tells a different story. Specially, population growth plays the most important role among three socioeconomic factors, indicating that it is human activity (particularly farming activity) rather than economic growth that caused the loss of grassland. One reason is that farming is still a main way for people to make a living in many underdeveloped regions (e.g., Horqin area, Inner Mongolia (Zhang and Zhao 2003)). 37 Table 2.7b. Simulated Changes in Land Supplies of Six Major Uses, 1988-2000a No change in No change in Change in major land use Factual county GDP population Farmland (1,000 ha.) 1,274.5 1,847.8 991.8 % 0.0% 45.0% -22.2% Forestland (1,000 ha.) -649.9 -518.9 -966.9 % 0.0% 20.2% -48.8% Grassland (1,000 ha.) -2,320.2 -2,317.2 -2,095.1 % 0.0% 0.1% 9.7% Water area (1,000 ha.) -217.1 -445.3 -218.7 % 0.0% -105.1% -0.7% Urban area (1,000 ha.) 2,526.7 2,545.4 2,532.6 % 0.0% 0.7% 0.2% Unused land (1,000 ha.) -613.8 -1,111.7 -243.8 % 0.0% -81.1% 60.3% No agricultural investment 722.8 -43.3% -300.4 53.8% -2,442.2 -5.3% -220.0 -1.3% 2,604.4 3.1% -364.6 40.6% No neighborhood effects 2,306.0 80.9% -9,301.7 -1331.2% -700.5 69.8% 4,310.4 2085.1% 1,597.3 -36.8% 1,788.5 391.4% a Change in hectare is the total land area change for each use between 1988 and 2000. Percent change is the counterfactuallysimulated net hectare change relative to the factually-simulated hectare change. Positive (negative) value of percent change indicates that the factor contributes to decreasing (increasing) the land hectare for that use. We also find that GDP growth and agricultural investment result in deforestation and population growth contribute to afforestation. In particular, the effect of GDP growth on forestland is less than effects of other two factors. With regard to change in unused land, effect of socioeconomic factors generally has opposite tendency against the foregoing farmland, grassland, and forestland. Specifically, GDP growth tends to expand unused land because agricultural activities such as cultivation, pasturing, and afforestation is a relatively low-paying activity as discussed before; while population growth can shrink unused land since job market is more competitive in a densely populated area and hence agricultural land will be more valuable; likewise, investment in agriculture infrastructure such as irrigation can also increase the value of agricultural land. Therefore, the opportunity cost of converting land out of agricultural use is raised, which leads to decreases in unused and urban area. Again, the last column of no neighborhood effects in Table 2.7b confirms the overwhelming effects of spatial externality on land-use change. Specially, the historical land-use patterns in 1988 and 1995 encourage afforestation and urban development. This result reveals that the accessibility (or location) rent, though unobservable, predominates over other rents such as rent on household income in the equilibrium bid rent structure of China. Hence urban expansion is more likely to occur in areas with more developed land. 38 PUBLIC AGRICULTURAL INVESTMENT AND FARMLAND PROTECTION As we have discussed above, public agricultural investment can help protect farmland from conversion, which is the third question we are going to address – could we design a cost-effective scheme to preserve farmland from development in traditional agricultural zones? To answer this question, we select three major agricultural production regions for the policy analysis: the Huang-Huai-Hai Plain, Yangtze River Delta, and Sichuan Basin. The regions have 128.08 million hectares of land, covering 700 counties across 7 provinces and four provincial-level metropolis in China. The three regions are highly productive agricultural areas, containing 34% of China‟s farmland. From 1988 to 2000, these regions lost 852.77 thousand hectares of farmland and gained 959.35 thousand hectares of developed land. The rapid urban development on high-quality farmland is threatening food security in China. To protect farmland and food security, the central government successively constituted and revised the Basic Farmland Protection Regulation (1994, 1998) and the Land Administration Law (1998, 2004). Apparently, these policies had only limited success in protecting farmland in these regions. It is of practical significance to adopt economic incentive as instruments to supplement the obligatory laws and regulations in farmland preservation. Public agricultural investment can serve for this purpose. In this section, we design a farmland protection scheme by increasing public agricultural investment in counties with different percent of farmland coverage. To this end, we calculate the marginal cost of farmland preservation for counties with different levels of farmland coverage. Specifically, we divide county-level farmland coverage percentage from 5% to 75% into seven intervals uniformly (i.e., 5-15%, 15-25%, …, 65-75%) and use the every midpoint to represent the interval. We estimate the marginal cost of farmland protection for each interval of farmland coverage. The formula of marginal cost is given by equation (2.8): (2.8) MCI lim I I I I 0 A ( I I ) A ( I ) I I I A ( I I ) A ( I ) for small I , where A( I ) is the total acreage of farmland in the study area when agricultural investment is I . We set I and I at ¥20,000/county and ¥5,000, respectively. 39 Figure 2.3. Marginal costs of preserving farmland in counties with different percent farmland coverage (Study area: the Huang-Huai-Hai Plain, Yangtze River Delta, and Sichuan Basin). † Marginal costs are measured at an investment level of ¥20,000 per county. Figure 2.3 illustrates the estimated marginal costs, where the horizontal axis represents the percent of farmland coverage and the vertical axis denotes the estimated marginal cost at the investment level of ¥20,000. As is shown there are very large variations in the costs, with the highest value (¥1282/ha.) being almost six times larger than the lowest value (¥190/ha.). The marginal cost curve is increasing and convex, implying that it is more cost-effective to protect farmland in counties with lower farmland coverage. In particular, 50% (or the range of 45-55%) is a cutoff level, below which marginal costs are moderately low. The result seems some counterintuitive. Intuitively, if a county has high farmland coverage, farmland will be less likely to be converted. Public spending on farmland protection should be more cost-effective in those counties. Then how to explain this counterintuitive result? Basically, in counties with high farmland coverage, the propensity to maintain land for agricultural use is already high. In this situation, even if there were no agricultural investment, farmland would still have a high tendency to be preserved. Therefore the marginal effect of agricultural 40 investment is relatively small. This result implies that for farmland protection, it is more cost-effective to make agricultural investment in counties with lower farmland coverage. The result has important policy implications. 41 CONCLUSIONS In this essay, we explored three questions: 1) What are the major drivers of landuse changes in China? 2) What is the relative importance of socioeconomic drivers? 3) How to design a cost-effective scheme to preserve farmland from development? To this end, we compiled a unique, national-scale dataset that includes high-quality land use data, and developed two set of multinomial logit models to analyze land-use choice among six major uses for two time intervals of 1988-1995 and 1995-2000. We answered these questions by estimating two set of models and generating a series of simulations with the estimation results. The estimation results show that both socioeconomic factors (e.g., county GDP, population, agricultural investment, and highway density) and geophysical variables (e.g., land productivity, terrain slope, elevation, and climate-related variables) affect land-use change on farmland and grassland, while the effect of geophysical variables on deforestation is more significant than that of socioeconomic factors.16 The spatial analysis provides convincing evidence for the existence of spatial dependence, where spatial autoregressive parameters are significantly positive in a range of 0.0104-0.8865. A comparison of results in the standard and spatial models reveals robust estimates of most parameters of spatial models with respect to the signs and relative magnitudes. We found that GDP growth play an influential role in the conversion of farmland, forestland, and unused land. In particular, GDP growth causes farmland loss, deforestation, and idleness. On the contrary, population growth shrinks unused land but enlarges land area for cultivation and forestation. Population growth is also major factor driving grassland degradation. Public agricultural investment affects all uses except water area. Specifically, it encourages land use for farming and pasturing. We also evaluated the effects of alternative variable on land use. GDP growth and agricultural investment have larger impacts on farmland conversion, while population growth and agricultural investment are more influential in deforestation and grassland conversion. Urban development is more sensitive to agricultural investment rather than county GDP. Nevertheless, access to developed areas dominates the effects of 42 agricultural investment and county GDP on urban expansion, which confirms that spatial externality is a major determinant of land-use change. It is more cost-effective for agricultural investment to larger counties with lower farmland coverage (<50%) when the objective is to reduce farmland conversion. This study contributes to the literature in two aspects. First, we use unusually detailed Chinese data in the analysis. To the best of our knowledge, no one has used such data for economic research in the land use literature. Second, we apply a new econometric method in multinomial logit regression. The method takes into account spatial autocorrelation in the dependent variable explicitly, while the traditional approach does not work. Caution should be exercised when applying the spatial multinomial logit model in the empirical work, considering the sensitivity of results to small probability events and to the specification of weight matrix. 43 ENDNOTES 1 There are relevant clauses to protect arable land in the revised Land Management Law. 2 Some studies use sample plot data from the National Resources Inventory (NRI) database, which are generated by a stratified sampling routine that ensures plots geographically dispersed. 3 China‟s land market is generally referred to as land-use right market, which emerged since the amended Constitution legalized land-use right transaction in 1988. It contains conveyance market and transfer market of land-use right. Conveyance market is a primary land market, in which transactions occur between government and land users; transfer market is a secondary land market, in which transactions occur between land users. 4 We did not use data on net returns to farmland for two reasons: 1) Data quality is doubtable. 2) Data on net return are measured at county level; while data on land productivity and climate variables are available at a more disaggregate level. 5 It is because that the test procedure needs to estimate coefficients and spatial autoregressive parameter simultaneously. 6 The sizes of all datasets are in a range of 1499-19488 observations. 7 This spatial model may have a potential overidentification problem, which is possible to generate inefficient estimates. Again, this issue is ignored because of large datasets used in the study. Alternatively, we can develop a spatial multinomial logit regression without including three instrumental variables in the model. A potential risk is that the alternative model may suffer from an underidentification problem and result in inconsistent estimates. 8 The weight matrix can be specified in several forms: the nearest-neighbor matrix, the binary block matrix, the distance inverse matrix, the negative exponential function, the Gaussian function, and the Spherical function (Dubin 1998). We adopt a Gaussian functional form because 1) the entries are continuous and 2) we can estimate the optimal bandwidth rather than choosing each entry a priori. 9 Linearization of generalized residuals is implemented around the initial estimates of the standard model, i.e., ρ=0. 10 Another econometric consideration pertains to the IIA property of the standard multinomial logit model, i.e., the relative odds of choosing l over k are independent of the other alternatives. Some studies appeal for more general models (e.g., nested logit model 44 and mixed logit model) to relax IIA assumption (Lubowski et al. 2006; Polyakov and Zhang 2008). But this approach may be infeasible for a large sample, or may lead to misspecification when a particular nested or mixed logit model is specified. An alternative approach is to employ Hausman specification test to examine IIA property. But even in a well-specified model, Hausman test of IIA often reject the assumption when alternatives seem distinct (Cheng and Long 2007). In our study, it is unsatisfactory to apply Hausman test given six land-use alternatives, which requires 15 essential tests for every initial land use (6!/[2!(6-2)!] = 15). In addition, some applications to land use have demonstrated that IIA assumption is not a serious problem for empirical work (Lewis and Plantinga; Lubowski et al. 2006; Polyakov and Zhang 2008). 11 Urban area consists of urban core and other built-up area such as roads, mines, and development zones that are not contiguous with the urban core. 12 Twenty models are two sets of five starting uses in two transition periods for both standard and spatial multinomial logit specifications. 13 R2 is not reported because it does not have a statistical interpretation in a two-stage least squares regression. 14 We adopt standard multinomial logit models rather than spatial models for four reasons: 1) The standard model is able to capture spatial lagged effect to some extent by involving three instrumental variables. 2) The standard model is more robust in the estimation of small probability events. 3) The spatial model might be sensitive to the form of weight matrix. 4) Generating probabilities of land-use change in every spatial model needs an inversion of n by n matrix, which is technically infeasible; alternatively, we may use a polynomial expansion as a proxy for the inverse operation, but it would increase prediction errors. Thus we expect that the standard model perform better than the spatial model in prediction. 15 For any land grid starting in urban uses, the probabilities of converting to other uses equal zero provided the assumption of irreversible urbanization. 16 We report the estimation results of land-use change on forestland in the Appendix B (Tables B1-B4). 45 ACKNOWLEDGMENTS We thank the National Science Foundation of China (70873118) and the Chinese Academy of Sciences (KZCX2-YW-305-2) for the financial support to generate the dataset used in this study. 46 REFERENCES Anselin, L.. 2006. Spatial econometrics. In T.C. Mills and K. Patterson (Eds.), Palgrave Handbook of Econometrics: Volume 1, Econometric Theory: 901-969. Palgrave Macmillan, Basingstoke. Carrión-Flores, C., and E. G. Irwin. 2004. Determinants of residential land-use conversion and sprawl at the rural-urban fringe. American Journal of Agricultural Economics 86:889-904. Cheng, S., and J. S. Long. 2007. Testing for IIA in the multinomial logit model. Sociological methods and Research 35:583-600. Chinese Academy of Social Sciences. 2009. “ 中国城镇人口 2008 年末 6.07 亿.” http://www.cpirc.org.cn/news/rkxw_gn_detail.asp?id=10684 (accessed December 1, 2009). Deng, X., J. Huang, S. Rozelle, and E. Uchida. 2006. Cultivated land conversion and potential agricultural productivity in China. Land Use Policy 23:372-384. Deng, X., J. Huang, S. Rozelle, and E. Uchida. 2008a. Growth, population and industrialization, and urban land expansion of China. Journal of Urban Economics 63:96-115. Deng, X., J. Liu, D. Zhuang, J. Zhan, and T. Zhao. 2002. Modeling the relationship of land use change and some geophysical indicators for the interlock area of farming and pasturing in China. Journal of Geographical Sciences 12:397-404. Deng, X., H. Su, and J. Zhan. 2008b. Integration of multiple data sources to simulate the dynamics of land systems. Sensors 8:620-634. Dubin, R.A., 1998. Spatial autocorrelation: A primer. Journal of Housing Economics 7, 304-327. Hall, P., editor. 1966. Von Thünen's Isolated State: An English edition of Der Isolierte Staat by Johann Heinrich Von Thünen. First edition. Pergamon Press. Hartkamp, A. D., K. D. Beurs, A. Stein, and J. W. White. 1999. Interpolation techniques for climate variables. CIMMYT, Mexico, D.F. Kelejian, H. H., and I. R. Prucha. 2001. On the asymptotic distribution of the Moran I test statistic with applications. Journal of Econometrics 104:219-257. 47 Klier, T., and D. P. McMillen. 2008. Clustering of auto supplier plants in the United states: generalized method of moments spatial logit for large samples. Journal of Business and Economic Statistics 26:460-471. Lewis, D. J., and A. J. Plantinga. 2007. Policies for habitat fragmentation: combining econometrics with GIS-based landscape simulations. Land Economics 83:109-127. Liu, J., M. Liu, D. Zhuang, Z. Zhang, and X. Deng. 2003. Study on spatial pattern of land-use change in China during 1995-2000. Science in China (Series D) 46:373384. Long, H., G. K. Heilig, X. Li, and M. Zhang. 2007. Socio-economic development and land-use change: analysis of rural housing land transition in the transect of Yangtse River, China. Land Use Policy 24:141-153. Long, H., X. Wu, W. Wang, and G. Dong. 2008. Analysis of urban-rural land-use change during 1995-2006 and its policy dimensional driving forces in Chongqing, China. Sensors 8:681-699. Lubowski, R. N., A. J. Plantinga, and R. N. Stavins. 2006. Land-use change and carbon sinks: econometric estimation of the carbon sequestration supply function. Journal of Environmental Economics and Management 51:135-152. Nelson, G. C., V. Harris, and S. W. Stone. 2001. Deforestation, land use, and property rights: empirical evidence from Darién, Panama. Land Economics 77:187-205. Nelson, G.C., Hellerstein, D., 1997. Do roads cause deforestation? Using satellite images in econometric analysis of land use. American Journal of Agricultural Economics 79, 80-88. Ostwald, M., and D. Chen. 2006. Land-use change: impacts of climate variations and policies among small-scale farmers in the Loess Plateau, China. Land Use Policy 23:361-371. Pinkse, J., and M. E. Slade. 1998. Contracting in space: an application of spatial statistics to discrete-choice models. Journal of Econometrics 85:125-154. Polyakov, M., and D. Zhang. 2008. Property tax policy and land-use change. Land Economics 84:396-408. Seto, K. C., and R. K. Kaufmann. 2003. Modeling the drivers of urban land use change in the Pearl River Delta, China: intergating remote sensing with socioeconomic data. Land Economics 79:106-121. 48 Wu, J., R. M. Adams, C. L. Kling, and K. Tanaka. 2004. From microlevel decisions to landscape changes: an assessment of agricultural conservation policies. American Journal of Agricultural Economics 86:26-41. Wu, J., and S.-H. Cho. 2007. The effect of local land use regulations on urban development in the Western United States. Reginal Science and Urban Economics 37:69-86. Zhang, Y., and S. Zhao. 2003. Analysis of land use change in Horqin Sandy land and its surrounding area during the past 15 years (In Chinese). Journal of Natural Resources 18:174-181. 49 APPENDICES 50 APPENDIX A Let Vi V xi , y i , z m X i θ , where X 1, x, y, z , θ μ, α, β, γ . Let dP parameter Γ θ, and gradient G . Conversion probability of land grid i can be dΓ rewritten as: Pil|k (A2.1) exp Vil**|k exp X il**|k θ exp V exp X s where X** I W X* and X il*|k 1 ** sl |k X il|k i s ** sl |k θ . Let Dil|k is an indicator variable, equal to one if latent variable Uil|k Uis|k , l s and zero otherwise. The generalized residual uil|k equals: uil|k Dil|k Pil|k (A2.2) When the spatial autocorrelation parameter 0 , for any l , s k , gradient terms can be expanded as: Pil|k (A2.3) θl|k Pil|k (A2.4) θ s|k Pil|k (A2.5) k Pil|k 1 Pil|k Xi Pil|k Pis|k Xi if s l Pil|k WX i θl|k s k Pis|k WX i θs|k There are four steps in the estimation procedure. Step 1: Estimate the standard multinomial logit model and assume initial values for Γ 0 0 are Γ θˆ , 0 Step 2: Calculate generalized residuals uil 0|k Dil|k Pˆil|k and the gradient terms G 0 based on Eq. (A2.3)-(A2.5). Step 3: Regress G on instruments Z X, WX, WWX, WWWX and predict Ĝ . 51 0 0 0 0 0 Step 4: Regress uil |k Gil|k Γ on Ĝ . Note that uil |k Gil|k Γ uil |k Gil|k θˆ when 0 . The coefficients are estimated values of parameters Γ θ, . So no large matrices have to be inverted in the algorithm. All it requires is standard logit followed by linear two-stage least square regression. A more detailed discussion of the procedure is provided by Klier and McMillen (2008). Table A2.1. Coefficient Estimates for the Standard M ultinomial Logit M odel of Land-use Change on Forestland, 1988-1995 Indep. Variable Farmland Grassland Water area Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -2.7835*** (0.1834) -3.0779*** (0.1945) -4.6856*** (0.7387) -4.4508*** (1.6021) Land productivity -0.0065 (0.0130) 0.0566*** (0.0170) 0.0691 (0.0430) -0.1788 (0.1189) County GDP 0.0027 (0.0147) 0.0182 (0.0225) -0.1406 (0.0966) 0.0157 (0.0574) Population -0.0550 (0.0930) -0.0021 (0.1321) 0.8228** (0.4039) 0.2045 (0.4704) Agricultural investment 0.0491 (0.0886) -0.1197 (0.1150) 0.2975 (0.6051) 1.1349* (0.6332) Highway density 0.1676*** (0.0498) 0.1716*** (0.0567) 0.1136 (0.2419) -0.3792 (0.8857) Terrain slope -0.0243*** (0.0072) -0.0095 (0.0062) -0.2402*** (0.0435) -0.2248* (0.1173) Elevation -0.1688*** (0.0500) 0.1285*** (0.0322) -0.5179*** (0.1666) -2.8689*** (0.6227) Precipitation -0.4353*** (0.1574) -1.0997*** (0.1715) -0.7107 (0.6314) -1.0955 (1.3883) Temperature 0.0507*** (0.0084) 0.0301*** (0.0068) 0.0771** (0.0323) 0.0790 (0.0784) Std Err of precipitation -1.6625*** (0.4919) 1.3443** (0.6553) -0.6897 (1.9371) -1.0067 (4.6230) Std Err of temperature -0.1248 (0.3061) 0.5206 (0.3607) -0.3588 (1.3495) -0.4862 (3.2093) Neighborhood index 0.0576*** (0.0013) 0.0583*** (0.0012) 0.1132*** (0.0056) 0.1507*** (0.0138) Number of observations 19345 M cFadden's LRI 0.6769 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. APPENDIX B Unused land Estimate Std Err -3.8801*** (0.8209) 0.0345 (0.0934) 0.0414 (0.0781) 0.2703 (0.6201) 0.3852 (0.4356) 0.3911** (0.1587) -0.0552* (0.0322) 0.1158 (0.1030) -2.3025*** (0.8236) -0.1096*** (0.0279) -0.4653 (4.4423) 0.3350 (1.5463) 0.0793*** (0.0040) 52 Table A2.2. Coefficient Estimates for the Standard M ultinomial Logit M odel of Land-use Change on Forestland, 1995-2000 Indep. Variable Farmland Grassland Water area Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -3.1282*** (0.2178) -2.8665*** (0.1752) -4.8250*** (0.7760) -6.0391 (4.1793) Land productivity 0.0990*** (0.0117) 0.0545*** (0.0171) 0.0116 (0.0441) 0.0249 (0.1358) County GDP -0.0041 (0.0067) 0.0150 (0.0092) -0.0010 (0.0491) -0.0115 (0.0348) Population -0.0752 (0.0822) -0.1991* (0.1049) 0.2150 (0.4086) -0.4762 (1.3089) Agricultural investment 0.1755** (0.0827) 0.1137 (0.0966) 0.2246 (1.0142) 0.3394 (4.7189) Highway density 0.0393 (0.0459) 0.0786 (0.0605) 0.0395 (0.2035) -0.2168 (0.5256) Terrain slope -0.0391*** (0.0077) -0.0175*** (0.0063) -0.1895*** (0.0505) -0.1908 (0.1509) Elevation -0.2066*** (0.0483) 0.1420*** (0.0328) -1.2181*** (0.1924) -0.4573 (0.6115) Precipitation -0.4585*** (0.1338) -1.1202*** (0.1378) -0.6447 (0.5126) -1.6371 (1.8932) Temperature 0.0480*** (0.0085) 0.0391*** (0.0069) 0.0831*** (0.0300) 0.1561 (0.1177) Std Err of precipitation -1.1567** (0.5049) 0.0428 (0.6458) -2.6334 (1.9783) -0.9844 (6.5859) Std Err of temperature 1.0541*** (0.3198) 0.7215*** (0.2633) 1.3924 (1.1542) -0.4434 (4.5842) Neighborhood index 0.0411*** (0.0012) 0.0387*** (0.0012) 0.0813*** (0.0064) 0.1319*** (0.0160) Number of observations 19488 M cFadden's LRI 0.6771 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -3.9343*** (0.7437) 0.1021 (0.0832) -0.1897 (0.2619) 0.6480 (0.9809) -2.1817** (1.1069) -0.8677* (0.4527) -0.1638*** (0.0411) 0.3046* (0.1579) -4.4378*** (0.7968) 0.0179 (0.0293) 0.1497 (3.0380) 2.6569** (1.1504) 0.0536*** (0.0050) 53 Table A2.3. Coefficient Estimates for the Spatial M ultinomial Logit M odel of Land-use Change on Forestland, 1988-1995 Indep. Variable Farmland Grassland Water area Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -2.4914*** (0.1422) -2.8930*** (0.1445) -2.2327 (1.8396) -16.320** (6.7621) Land productivity 0.0004 (0.0060) 0.0610*** (0.0043) 0.0540 (0.2249) 1.0111 (0.6792) County GDP -0.0175 (0.0402) -0.0804*** (0.0250) -1.5031 (1.0247) 6.2719 (8.7911) Population -0.0616*** (0.0086) -0.0203 (0.0126) 0.7417*** (0.0982) 0.2533 (0.5593) Agricultural investment 0.0505* (0.0302) -0.1347*** (0.0363) 0.4098 (0.5779) -0.1737 (3.4311) Highway density 0.1655*** (0.0091) 0.1646*** (0.0174) 0.0018 (0.2189) -0.2317 (0.1466) Terrain slope -0.0258 (0.0639) 0.0975 (0.0951) 0.0353 (0.8526) -2.0554** (0.9444) Elevation -0.1918*** (0.0501) 0.0719 (0.0978) -0.1823 (0.7130) -4.6428** (2.0376) Precipitation -0.2156** (0.1098) -0.6717*** (0.1542) -2.5938* (1.5742) 2.8612 (5.6797) Temperature 0.0306* (0.0062) 0.0184*** (0.0049) 0.2139** (0.0846) 0.4209 (0.3030) Std Err of precipitation -1.6763*** (0.3249) 0.1280 (0.4899) -4.2647 (4.2123) -36.9332* (21.974) Std Err of temperature -0.1208 (0.2061) 0.6758*** (0.2598) -2.5237 (3.1073) 26.6217** (12.267) Neighborhood index 0.0569*** (0.0008) 0.0582*** (0.0008) 0.1121*** (0.0104) 0.0858* (0.0439) Spatial parameter (ρ) 0.1113*** (0.0326) Number of observations 19345 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -3.9696*** (1.3840) -0.0607 (0.0865) 0.0989 (0.2123) 0.1812 (0.3189) 0.5776*** (0.1934) 0.3594*** (0.0835) 0.1893 (0.6361) 0.2502 (0.5595) -0.9389 (1.8741) -0.1563*** (0.0530) -2.7104 (12.691) 0.8181 (3.0545) 0.0821*** (0.0066) 54 Table A2.4. Coefficient Estimates for the Spatial M ultinomial Logit M odel of Land-use Change on Forestland, 1995-2000 Indep. Variable Farmland Grassland Water area Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -2.9879*** (0.1911) -2.8763*** (0.1509) -2.8032 (2.4545) -4.7163 (13.748) Land productivity 0.1003*** (0.0065) 0.0572*** (0.0044) -0.0609 (0.2285) -0.4921 (1.2736) County GDP 0.0113 (0.0403) -0.0220 (0.0249) -1.1205 (1.4875) 5.2055 (4.7932) Population -0.0815*** (0.0071) -0.2041*** (0.0128) 0.0915 (0.0948) -0.9660** (0.4532) Agricultural investment 0.1801*** (0.0235) 0.1091*** (0.0295) 0.1969 (0.2802) 0.3395 (2.8244) Highway density 0.0335*** (0.0056) 0.0761*** (0.0072) 0.0252 (0.0280) -0.2228*** (0.0546) Terrain slope 0.0226 (0.0617) 0.0713 (0.0922) 0.0083 (0.4085) 1.0567 (1.7127) Elevation -0.2216*** (0.0417) 0.1314** (0.0571) -0.963 (1.3877) -0.9937 (3.6407) Precipitation -0.3677*** (0.0849) -1.1385*** (0.1214) -0.7238 (1.2673) -2.0367 (6.1076) Temperature 0.0371*** (0.0059) 0.0299*** (0.0051) 0.0764 (0.0804) 0.1456 (0.4834) Std Err of precipitation -1.1818*** (0.3182) 0.9039* (0.4905) -4.6062 (4.6128) 11.289 (30.353) Std Err of temperature 0.9609*** (0.2185) 0.8341*** (0.1663) -0.1402 (4.1137) -7.5278 (21.492) Neighborhood index 0.0401*** (0.0007) 0.0385*** (0.0007) 0.0745*** (0.0088) 0.1406** (0.0628) Spatial parameter (ρ) 0.0104** (0.0400) Number of observations 19488 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -3.0628*** (1.1302) -0.2339 (0.1837) 0.0208 (0.3188) 0.6964** (0.2779) -2.3272** (1.1077) 0.2435 (0.9153) -4.3353 (3.3343) -3.6362 (2.6858) -4.3051** (2.0884) -0.1182* (0.0643) -0.3077 (7.1770) 4.2925** (1.7857) 0.0439*** (0.0062) 55 Table A2.5. Coefficient Estimates for the Standard M ultinomial Logit M odel of Land-use Change on Water Area, 1988-1995 Indep. Variable Farmland Forestland Grassland Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -1.2987** (0.6354) -0.4612 (0.8319) -1.9364** (0.7971) -5.0766** (2.1037) Land productivity 0.0178 (0.0344) 0.0748 (0.0545) 0.0569 (0.0557) -0.1077 (0.1391) County GDP 0.0038 (0.0473) -0.0312 (0.0638) 0.0156 (0.0862) 0.0526 (0.0670) Population -0.2362 (0.3010) 0.2117 (0.3147) -1.6273** (0.6962) 0.5125* (0.3031) Agricultural investment -0.8632 (0.5013) 0.2582 (0.8848) 0.2368 (0.9636) -0.5197 (1.5840) Highway density 0.5379 (0.2291) 0.1928 (0.2839) -0.1738 (0.2443) 0.8766** (0.3614) Terrain slope 0.0400 (0.0708) 0.0480 (0.0555) 0.0241 (0.0251) 0.0197 (0.5225) Elevation -0.5889** (0.2781) -0.6459* (0.3621) 0.0361 (0.1345) 0.0383 (0.7558) Precipitation 0.2545 (0.6478) -1.0663 (0.8095) 0.0482 (0.7722) -0.6825 (2.4874) Temperature -0.0187 (0.0397) 0.0304 (0.0490) 0.0728** (0.0333) 0.2135 (0.2106) Std Err of precipitation -0.5331 (1.9227) -1.8934 (2.6294) -3.3747 (3.5569) -1.9733 (5.4332) Std Err of temperature -3.1317** (1.3006) -5.4057*** (1.6528) -1.6622 (1.4569) -3.9594 (4.2369) Neighborhood index 0.0607*** (0.0044) 0.0703*** (0.0062) 0.0556*** (0.0043) 0.0863*** (0.0199) Number of observations 1520 M cFadden's LRI 0.6046 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -1.9380* (1.0214) -0.0152 (0.1280) -0.0827 (0.1738) -0.8808 (1.0169) -0.1456 (1.0513) -0.0280 (0.3184) 0.0304 (0.0194) -0.2351 (0.1673) 0.4337 (1.4136) -0.0112 (0.0414) -2.5342 (10.649) -1.5564 (1.8052) 0.0622*** (0.0053) 56 Table A2.6. Coefficient Estimates for the Standard M ultinomial Logit M odel of Land-use Change on Water Area, 1995-2000 Indep. Variable Farmland Forestland Grassland Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -3.5146*** (0.8073) -3.7134*** (1.2267) -2.6763*** (0.7753) -5.0281 (8.8008) Land productivity 0.0510* (0.0299) -0.1498** (0.0721) -0.0780 (0.0640) -0.0793 (0.3040) County GDP -0.0125 (0.0164) -0.1913** (0.0935) -0.0766 (0.1137) -0.0531 (1.0614) Population 0.1826 (0.1634) 0.8162* (0.4889) 0.0832 (0.8059) -1.0301 (10.113) Agricultural investment 0.1811 (0.5320) 0.6879 (0.8779) -0.0209 (0.8614) -2.9744 (49.987) Highway density 0.3167** (0.1523) 0.2871 (0.3881) 0.2984 (0.1967) 0.5073 (2.6891) Terrain slope 0.0332 (0.0660) -0.0219 (0.0720) -0.0088 (0.0299) 0.2408 (0.4943) Elevation -0.3081 (0.2287) -0.3472 (0.2199) 0.0328 (0.1515) -2.1311 (3.7833) Precipitation -0.7302 (0.4833) 0.6039 (0.7174) 0.2418 (0.6534) -4.0765 (7.4583) Temperature 0.0416 (0.0346) 0.0286 (0.0432) 0.0272 (0.0345) 0.3731 (0.5964) Std Err of precipitation 2.3506 (2.4991) -5.3109* (2.9318) -4.1736 (4.2787) -0.6482 (39.566) Std Err of temperature 0.5533 (1.1182) 0.0807 (1.9814) -0.3213 (0.9870) -1.7576 (14.561) Neighborhood index 0.0357*** (0.0037) 0.0563*** (0.0065) 0.0425*** (0.0044) 0.0660*** (0.0256) Number of observations 1499 M cFadden's LRI 0.5461 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -0.4896 (0.9003) -0.0385 (0.0663) 0.0635* (0.0381) -1.1591* (0.5957) 0.4033 (0.6715) 0.0319 (0.3327) -0.0215 (0.0273) -0.1232 (0.1482) -0.2975 (0.8468) -0.0782 (0.0363) 2.3921 (4.6187) -2.5544** (1.2470) 0.0358*** (0.0041) 57 Table A2.7. Coefficient Estimates for the Spatial M ultinomial Logit M odel of Land-use Change on Water Area, 1988-1995 Indep. Variable Farmland Forestland Grassland Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -1.8994*** (0.6160) -1.1767 (1.0315) -2.2193*** (0.8342) -0.6541 (5.2873) Land productivity -0.0494 (0.0591) 0.0171 (0.0536) 0.0508** (0.0204) -0.2956 (0.5526) County GDP 0.6754** (0.2655) 0.6877* (0.3978) -0.0283 (0.1480) -1.6757 (2.1664) Population -0.2335*** (0.0319) 0.1949*** (0.0690) -1.6255*** (0.0760) 0.4767*** (0.2184) Agricultural investment -1.1074 (0.2532) -0.2126 (0.4021) 0.1824 (0.2889) -0.676** (0.3372) Highway density 0.5194*** (0.0409) 0.2285*** (0.0745) -0.1432 (0.1364) 0.8811*** (0.0951) Terrain slope 0.0825 (0.2287) 0.0103 (0.2487) 0.6613 (1.2947) -0.1976 (0.4324) Elevation -0.4621 (0.5539) 0.0522 (0.9268) -0.1363 (0.8458) 0.4193 (1.0127) Precipitation 0.8679 (0.6391) -0.7604 (1.0214) 1.3498 (0.9425) -1.3304 (3.9337) Temperature -0.0860** (0.0384) 0.0052 (0.0526) 0.0374 (0.0343) 0.0846 (0.3042) Std Err of precipitation 1.1571 (1.7994) -1.0477 (3.4869) -7.1965 (5.4740) -2.1646 (11.204) Std Err of temperature -1.1016 (1.2709) -2.0640 (2.3261) 0.2087 (1.3601) -3.3985 (7.9236) Neighborhood index 0.0607*** (0.0042) 0.0687*** (0.0085) 0.0541*** (0.0043) 0.0706** (0.0291) Spatial parameter (ρ) 0.4308*** (0.0971) Number of observations 1520 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -2.0830** (0.9170) -0.0195 (0.0235) -0.1322 (0.2058) -0.6868** (0.2014) -0.1408 (0.4422) 0.0288 (0.2602) -0.6756 (1.4122) -0.0677 (1.4673) 1.4839 (1.5847) -0.0402 (0.0508) -2.7134 (11.342) 0.0644 (1.6562) 0.0673*** (0.0067) 58 Table A2.8. Coefficient Estimates for the Spatial M ultinomial Logit M odel of Land-use Change on Water Area, 1995-2000 Indep. Variable Farmland Forestland Grassland Urban area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -1.8363*** (0.7096) -0.5158 (1.7580) -1.3938* (0.8229) -1.1221 (11.327) Land productivity 0.0786 (0.0589) -0.2172** (0.1016) -0.0619** (0.0264) -0.4175 (0.4205) County GDP 0.2139 (0.2277) -0.1685 (0.4783) -0.2219 (0.1493) 5.9953 (7.2205) Population 0.1964*** (0.0208) 0.8974*** (0.1059) 0.1628** (0.0772) -0.8920*** (0.3459) Agricultural investment 0.1339 (0.1366) 0.6698** (0.3189) -0.1140 (0.1779) -3.3167** (1.4755) Highway density 0.3284*** (0.0111) 0.5044*** (0.1725) 0.2482* (0.1433) 1.9945 (1.4160) Terrain slope -0.0392 (0.1098) -0.5313 (0.5956) 0.4134 (0.8284) -9.6580 (7.3380) Elevation -0.3343 (0.2929) -0.8531 (0.6399) -0.5860 (0.8721) -1.0027 (49.686) Precipitation -0.2450 (0.4097) -0.5815 (0.8611) 0.3456 (0.7523) 3.8893 (7.1175) Temperature -0.0222 (0.0308) -0.0047 (0.0544) 0.0007 (0.0312) -0.3088 (0.6331) Std Err of precipitation 0.8176 (1.8157) -2.8022 (3.2156) -5.9253 (4.5718) 16.355 (46.403) Std Err of temperature -0.2159 (0.9671) -1.3158 (2.5181) 0.0060 (0.8827) 2.8721 (17.650) Neighborhood index 0.0336*** (0.0027) 0.0506*** (0.0085) 0.0373*** (0.0036) 0.0235 (0.0355) Spatial parameter (ρ) 0.4108*** (0.0864) Number of observations 1499 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Unused land Estimate Std Err -0.7791 (0.6841) -0.0344 (0.0212) 0.1879 (0.1170) -1.1285*** (0.1123) 0.6434* (0.3804) 0.0578 (0.0385) 0.4459 (0.8933) -0.4297 (0.4684) -1.2845 (0.8413) -0.0518 (0.0306) 9.6359 (4.6155) -1.9914 (1.0886) 0.0354*** (0.0030) 59 Table A2.9. Coefficient Estimates for the Standard M ultinomial Logit M odel of Land-use Change on Unused Land, 1988-1995 Indep. Variable Farmland Forestland Grassland Water area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -2.9605*** (0.6643) -7.6077*** (0.8299) -5.1937*** (0.2789) -6.9052*** (0.7162) Land productivity 0.0889* (0.0497) -0.0770 (0.1078) 0.0118 (0.0479) 0.1109 (0.0729) County GDP 0.3721*** (0.1439) -0.1357 (0.3696) 0.2932*** (0.0720) 0.3161* (0.1675) Population -0.3381 (0.7308) 2.0092* (1.0686) -0.7116* (0.4122) 0.0264 (0.9001) Agricultural investment -2.4682 (2.2017) -0.0468 (0.7565) -1.686*** (0.2841) -1.8844** (0.8420) Highway density 0.4908*** (0.1322) 0.2492 (0.2135) -0.3221*** (0.0912) 0.4114* (0.2484) Terrain slope -0.1487 (0.1185) -0.0228 (0.0357) -0.0079 (0.0094) -0.0187 (0.0239) Elevation -0.9107*** (0.2407) 0.2993** (0.1354) 0.2632*** (0.0491) -0.2121* (0.1147) Precipitation 2.3108** (0.9521) 2.9074*** (1.0460) 2.7848*** (0.4213) 3.9858*** (1.0558) Temperature -0.1522*** (0.0318) 0.0641** (0.0322) 0.0424*** (0.0117) -0.1531*** (0.0293) Std Err of precipitation 0.4155 (5.8175) 1.4269 (5.6772) -2.0200 (2.7633) 0.9080 (7.8870) Std Err of temperature -0.9815 (1.2870) 1.8069 (1.5764) 4.2513*** (0.4904) 5.6654*** (1.3302) Neighborhood index 0.0669*** (0.0038) 0.0917*** (0.0047) 0.0483*** (0.0012) 0.0880*** (0.0046) Number of observations 13748 M cFadden's LRI 0.7996 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Urban area Estimate Std Err -5.2945 (207.12) 0.6917 (7.3289) 0.2265 (9.8671) -2.5527 (173.50) 0.3002 (157.35) -1.3511 (81.250) -2.6121 (126.67) -0.8593 (44.950) -3.0495 (330.89) -0.0963 (5.4338) -0.7762 (1191.0) -1.6975 (595.47) 0.1133 (0.6806) 60 Table A2.10. Coefficient Estimates for the Standard M ultinomial Logit M odel of Land-use Change on Unused Land, 1995-2000 Indep. Variable Farmland Forestland Grassland Water area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -4.8764*** (0.6026) -4.8348*** (0.8487) -3.7899*** (0.2968) -5.1084*** (0.7763) Land productivity 0.2270*** (0.0407) 0.1060 (0.1063) 0.0917* (0.0488) 0.0971 (0.0969) County GDP -0.0274 (0.0691) -0.0516 (0.2770) -0.0033 (0.0976) -0.0500 (0.2248) Population 0.6492 (0.5992) 0.5469 (1.4022) -0.3835 (0.4998) 0.1833 (1.0677) Agricultural investment -0.9822 (0.9698) 0.8982 (0.8526) -1.0307*** (0.3874) 0.7685 (0.7521) Highway density 0.6207*** (0.084) 0.5508*** (0.1871) 0.3246*** (0.0734) 0.2749 (0.2886) Terrain slope 0.0499 (0.0659) 0.0540 (0.0356) -0.0133 (0.0083) 0.0050 (0.0227) Elevation -1.3933*** (0.2063) -0.1362 (0.1450) 0.1523** (0.0615) -0.4027*** (0.1527) Precipitation 1.0789 (0.7369) 2.9483*** (0.8192) 0.2822 (0.4103) 2.3348*** (0.9070) Temperature -0.1027*** (0.0256) -0.0449 (0.0338) -0.0430*** (0.0127) -0.1669*** (0.0325) Std Err of precipitation 7.2734** (3.0348) 0.0589 (3.2958) 10.155*** (2.0429) 3.7363 (5.4435) Std Err of temperature 2.5676*** (0.9076) -2.0130 (1.4603) 0.3636 (0.3312) 1.4047 (0.9426) Neighborhood index 0.0365*** (0.0038) 0.0749*** (0.0053) 0.0367*** (0.0014) 0.0627*** (0.0049) Number of observations 13276 M cFadden's LRI 0.8246 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Urban area Estimate Std Err -2.1247 (66.415) -2.4456 (25.816) -0.7431 (6.5123) -2.4836 (53.576) 5.6398 (9.5645) 0.1292 (7.7573) -0.4287 (8.5267) -4.6396 (21.063) -0.2501 (94.280) -0.4150 (1.9989) 0.2473 (239.16) 0.2657 (100.50) 0.2189 (0.3466) 61 Table A2.11. Coefficient Estimates for the Spatial M ultinomial Logit M odel of Land-use Change on Unused Land, 1988-1995 Indep. Variable Farmland Forestland Grassland Water area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept -1.1580* (0.6435) -7.9209*** (1.1678) -3.6298*** (0.2195) -6.6046*** (0.9006) Land productivity -0.1035 (0.1709) -0.1823** (0.0741) 0.0152*** (0.0057) 0.1399*** (0.0294) County GDP 1.6850*** (0.2944) 0.8448*** (0.1946) 0.2583*** (0.0301) 0.5824*** (0.1362) Population -0.2937*** (0.0375) 1.9131*** (0.1477) -0.6797*** (0.0328) -0.1433* (0.0777) Agricultural investment -2.6681*** (0.0874) -0.2365 (0.1974) -1.5420*** (0.0743) -2.0739*** (0.1480) Highway density 0.5411** (0.2144) -0.0435 (0.5343) -0.2848*** (0.0884) 0.6723*** (0.1384) Terrain slope 1.7388** (0.8552) 4.3721*** (1.3380) 1.2258*** (0.2737) 1.0015 (0.7198) Elevation -1.7953 (3.2603) -0.3226 (1.5399) -0.2791 (0.2395) -5.5762*** (1.6775) Precipitation -0.5085 (1.0432) -2.0745* (1.1036) 0.7017** (0.3022) -0.0357 (1.0667) Temperature -0.1062*** (0.0319) 0.1725*** (0.0413) 0.0426*** (0.0066) -0.0954** (0.0423) Std Err of precipitation 1.4730 (5.6052) 15.299** (6.2364) -4.9435*** (1.4370) 11.046 (8.1613) Std Err of temperature -5.4274*** (1.3844) -0.4240 (2.0870) 2.5256*** (0.2920) 8.0209*** (1.6261) Neighborhood index 0.0623*** (0.0043) 0.1065*** (0.0073) 0.0463*** (0.0006) 0.0845*** (0.0054) Spatial parameter (ρ) 0.2384*** (0.0288) Number of observations 13748 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Urban area Estimate Std Err -22.558 (80.246) -500.30*** (133.80) -18.673 (17.437) 5.1809 (5.0428) -113.84** (52.227) -5.0450 (4.1769) 4.1919 (114.21) 76.936 (59.914) -667.13*** (168.07) -0.5171 (2.2688) 2195.1*** (585.71) 374.50* (211.82) -0.6196* (0.3189) 62 Table A2.12. Coefficient Estimates for the Spatial M ultinomial Logit M odel of Land-use Change on Unused Land, 1995-2000 Indep. Variable Farmland Forestland Grassland Water area Estimate Std Err Estimate Std Err Estimate Std Err Estimate Std Err Intercept 0.7936 (0.5268) 0.7583 (0.8980) 0.0516 (0.1835) -0.4944 (0.8034) Land productivity 0.5044*** (0.0840) 0.0781*** (0.0237) 0.1144*** (0.0044) 0.1153*** (0.0196) County GDP 1.6538*** (0.2510) 0.6779*** (0.1273) -0.1979*** (0.0313) -0.1728 (0.1456) Population 0.5643*** (0.0253) 0.5666*** (0.0943) -0.3894*** (0.0244) 0.0930 (0.1113) Agricultural investment -1.2783*** (0.0592) 0.6952*** (0.1269) -1.1846*** (0.0321) 0.2614 (0.3409) Highway density 0.7369*** (0.0400) 0.6680*** (0.1924) 0.2841*** (0.0340) 0.4318*** (0.1241) Terrain slope -0.8374*** (0.2840) -0.9899 (0.9374) 0.8731*** (0.2207) -0.5967 (0.6022) Elevation -6.3319*** (1.4206) 0.0722 (0.9794) -0.1784 (0.2544) -1.1809 (0.8196) Precipitation -4.0402*** (0.3995) -2.8814*** (0.6699) -2.6201*** (0.2151) -0.6586 (0.8663) Temperature 0.1626*** (0.0191) 0.0715** (0.0343) -0.0361*** (0.0063) -0.1491*** (0.0390) Std Err of precipitation 2.9214 (1.8005) 3.1733 (2.2808) -0.4682 (1.0251) -2.9577 (4.6923) Std Err of temperature -2.8901*** (0.8245) -3.9449** (1.6225) 0.0504 (0.1563) 3.0074*** (1.1234) Neighborhood index 0.0364*** (0.0021) 0.0631*** (0.0055) 0.0303*** (0.0006) 0.0624*** (0.0036) Spatial parameter (ρ) 0.8865*** (0.0315) Number of observations 13276 Note: *, **, and *** indicate statistical significance at 10, 5, and 1% levels, respectively. Urban area Estimate Std Err 27.642 (71.712) -25.291 (19.476) -8.6139 (28.553) 48.085 (35.433) 1.4877 (8.7185) 4.0688 (5.0529) -56.152 (42.321) 2.0491 (11.939) -3.256 (100.47) -1.6407 (2.4767) -103.76 (159.85) 14.897 (70.029) 0.6825 (0.4681) 63 64 CHAPTER 3 AN EMPIRICAL ECONOMIC ANALYSIS OF LAND-USE CHANGE AND SOIL CARBON SEQUESTRATION IN CHINA MAN LI 65 ABSTRACT This essay presents an empirical analysis to examine the relationship between land-use change and soil carbon sequestration in China. The analysis compiles highquality, national-scale land use data with a spatial resolution of 30 by 30 meters and soil profile data (10 by 10 kilometers). A statistical soil organic carbon (SOC) density model and an econometric land-use change model are developed to link the socioeconomic factors with the SOC density. The approach captures spatial autocorrelation and spatial heterogeneity simultaneously and can be applied to a large region. Results indicate that the SOC density is highest in forest and grass lands and lowest in unused land in China. GDP growth leads to farmland and grassland loss, deforestation, and idleness, which accelerates soil carbon runoff. The models are integrated to evaluate the welfare effects of China‟s green growth policy, which targets resource reservation and environment protection in conjunction with soil carbon sequestration. Results indicate that the policy could generate up to 0.7-1.1 million Mg SOC and 22.2-37.4 million CNY welfare losses annually throughout the simulation period (2001-2050). The marginal welfare loss is approximately ¥15.3/Mg (equivalent to $2.25/Mg) for sequestering about 1 million Mg SOC per year. A comparison of the results reveals that soil-based carbon sequestration, oriented by the green growth policy, merits consideration. 66 INTRODUCTION Carbon dioxide (CO2) emissions are going up at a rapid speed in China. Increased energy demand, driven by fast economic development and unprecedented urbanization, makes it a limited strategy for China to rely on energy-based abatement alone. As a supplement, biological carbon sequestration has attracted more attention because of its environmental benefits for soil conservation and sustainable economic development. Biological sequestration involves managing land in ways that enhance the natural absorption of atmospheric carbon by soil and vegetation. Economic efficiency is a major criterion for evaluating the feasibility of alternative carbon sequestration strategies. Many American and European researchers have addressed this issue by estimating the sequestration costs in the United States and European countries. Their primary interests are forest sequestration costs (Adams et al. 1993; Alig et al. 1997; Lubowski et al. 2006; Parks and Hardie 1995; Plantinga et al. 1999; Stavins 1999) and economic potential for agricultural soil sequestration (Pautsch et al. 2001; Antle and Diagana 2003; Capalbo et al. 2004; Feng et al. 2006; Antle et al. 2007). It shows that cost estimation of conservation tillage is sensitive to the choice of baseline and the spatial heterogeneity of the area. Richards and Stokes (2004) reviewed the literature on this subject. They found that afforestation in the United States could sequester 250 to 500 million megagrams of carbon (Mg C) per year at a price ranging 10-150 USD per Mg C, whereas conservation tillage would generate 0.25 to 6.2 million Mg C in soil per year by spending 12-270 USD per Mg C. With the total land area of 946.8 million hectares, China, like the United States and Canada, has a large potential for soil carbon sequestration. Nevertheless China has witnessed a decreased storage of soil organic carbon (especially in grassland) resulting from dramatic land use change over the past decades. According to the first and second national soil surveys of China, the storage of soil organic carbon (SOC) was estimated to drop from 93 petagrams of carbon (Pg C) in the 1960s to 92 Pg C in the 1980s (Wang et al. 2003). Specially, extensive agricultural development and desert expansion reduced the SOC density by 10-20 kg/m2 in the alpine meadow of the Southeast Tibet Plateau and by 67 2-3 kg/m2 in the grassland of the Mid-east Inner Mongolia. It is of practical importance to estimate the spatial distribution and dynamic changes of soil carbon storage in China. Many Chinese scientists have made efforts to obtain such knowledge (Fang et al. 2001; Tang et al. 2006; Wang et al. 2003; Wang et al. 2004; Wu et al. 2003; Yang et al. 2007; Zhang et al. 2006). Huang and Sun (2006) surveyed the changes in topsoil carbon of croplands in China over the last two decades by selecting 132 representative articles from literature databases published since 1993. Canadian and Chinese researchers also implemented a four-year (2002-2006) project of carbon sequestration in China‟s forest ecosystems and their achievements were published in a special issue of Journal of Environmental Management (2007). These studies have two limits in common: 1) the focus is mainly on scientific analysis without taking socioeconomic factors into account, and 2) the estimation of soil carbon typically relies on site-specific process models1 that were developed based on field-level inputs and hence are restricted to local analysis in a relatively small region. The objectives of this essay are to estimate the association between land-use change and SOC density in China and to evaluate the welfare effects of the green growth policy that targets co-benefits of soil carbon sequestration and environmentally sustainable development. For this purpose, we compiled a unique, national-scale dataset that includes high quality land use data and soil profile data. Using 10-kilometer-gird (10km-gird) as the analysis unit, we develop a statistical model to explore the association of SOC density with six major land uses, i.e., farmland, forestland, grassland, water area, urban area, and unused land. Besides land use dummy, the right-hand-side covariates include soil profile and climate variables. This approach can be applied to a large region and thus overcomes the limitation of a process model that typically works for the fieldlevel study. We adopt a version of multinomial logit model for land-use change from 1995 to 2000, as developed and discussed in the first essay (Chapter 2), to analyze the effects of socioeconomic factors on land conversion among six uses. Socioeconomic variables are measured at county level, including county gross domestic product (GDP), population, public agricultural investment, and highway density. By combining two 68 models, we are able to use the “logsum” formula to estimate welfare changes for different GDP growth rates (Small and Rosen 1981). This essay contributes to the literature in two aspects. First, it is the first economic research on soil carbon sequestration in China, whereas previous studies are typically conducted by scientists. Second, it applies the “logsum” formula to evaluate the welfare effects of policy-oriented land conversion. The “logsum” is a measure of consumer surplus in the context of (multinomial) logit models and has been used in some fields, such as transport demand, recreation demand, and residential location (Mansfield et al. 2008; Phaneuf et al. 2008). But to the best of our knowledge, there have not been such applications in the land use literature. Econometrically, this essay applies geographically weighted regress (GWR) technique to a spatial autoregressive (SAR) model to correct for spatial autocorrelation and spatial heterogeineity simultaneously, which is a novel application of spatial econometrics in empirical study. The remainder of this essay is arranged as follows. Section 2 discusses the SOC density model. Section 3 describes data. Section 4 reports the estimation and simulation results. Section 5 generates a discussion on green growth, carbon sequestration, and welfare losses. The final section concludes. 69 THE SOC DENSITY MODEL The dynamics of SOC flow are a complex process, where SOC storage is determined by the balance of carbon inputs from plant production and outputs through a decomposition process (Jobbágy and Jackson 2000; Parton et al. 1993; Schlesinger 1977) and soil temperature, moisture, and texture jointly control the decomposition rates of SOC in various carbon pools (Parton et al. 1993). Interactions among these factors generally demonstrate a complicated, nonlinear pattern. For example, the effects of soil temperature and soil moisture on the decomposition rates exhibit an inverted-U shape with a heavy left-tail. The decomposition rates of SOC in the active pool tend to increase with sand content and the decomposition rates in the slow carbon pool tends to decrease with clay content. SOC density is negatively correlated with soil bulk density (Wang et al. 2004; Wu et al. 2003; Yang et al. 2007). It is challenging to apply detailed site-specific process models for regional analysis with field-level inputs of soil profile and vegetation. To overcome this challenge, we develop a statistical model that relies on relatively flexible data and hence can be tailored for specific use and easily be applied to a large region. Of primary interest is the association of SOC density with six major land uses. The covariates2 specified in this model include soil profile and climate factors, which are discussed in details in the data section. Equation (3.1) presents the general formula of the model. (3.1) y = Xθ + ε , where the bold type denote a vector or a matrix, dependent variable y is the logarithm of SOC density; X represents independent variables, i.e., land use dummy, soil profile, and climate variables; θ is the coefficient of X ; ε denotes error term. We adopt a quadratic polynomial functional form of soil profile and climate factors in the analysis, which allows capturing the potential nonlinear relationships within and between these variables. There are six land use groups: farmland, forestland, grassland, water area, urban area, and unused land. Soil profile variables include PH value, bulk density, and soil loam and sand contents.3 Climate variables include mean annual precipitation and mean annual 70 temperature. Previous study explained 84% of the variations in SOC storage in China using similar soil profile and climate variables (Yang et al. 2007). We refer to equation (3.1) as OLS model since it can be estimated using ordinary least squares (OLS) regression. When applying contiguous geographic data for the empirical analysis, OLS regression is inappropriate because of potentially spatial heterogeneity (spatial variation in parameters) and spatial autocorrelation (spatially correlated error term and/or spatially correlated dependent variable). The costs of not correcting for these issues are inconsistent/biased estimates if there is spatially interdependence or spatial heterogeneity. But in practice it is technically infeasible to distinguish between them. In this study, the OLS model is able to capture potential endogeneity resulting from spatial autocorrelation in the dependent variable because the disaggregated data on soil and climate are generated by interpolating the original point data into a surface respectively using Kriging algorithm (Kravchenko and Bullock 1999) and thin plate smoothing spline method (Hartkamp et al. 1999). These techniques have already taken spatial effects into account when estimating or retrieving the values of other locations during the interpolation process. To correct for the remaining problems, we extend OSL regression of equation (3.1) to a spatial autoregressive (SAR) model which allows for spatial error autocorrelation; meanwhile, we adopt geographically weighted regress (GWR) technique to capture spatial heterogeneity in coefficients (Fotheringham et al. 1998). The model is rewritten as (3.2) y = Xθ ui , vi + ε, ε = Wε + μ, where ui , vi denotes the coordinates of the i th point in space, θ ui , vi is a realization of the continuous function θ u, v at point i , is the spatial autoregressive coefficient, W is a row-standardized n n matrix with wii 0 and n j 1 wij 1 for each i , μ is 71 heteroscedastic noise such that E μμ 2M ui , vi 1 , and M ui , vi is an n n diagonal matrix. Hence the second moment of error ε is expressed as (3.3) E εε 2 I W M ui , vi 1 1 I W 1 , which is equivalent to its variance-covariance matrix. We refer to equations (3.2) and (3.3) as the spatial model. There are two weight matrices, W and M ui , vi , used for SAR and GWR, respectively. We assume a substantially identical weighting scheme in both matrices, where each non-zero entry is specified as a Gaussian function of geographical distance from location j to location i , as in (3.4) wij exp dij2 h2 n j 1 exp dij2 h2 , i, j 1, , n, and i j and (3.5) m jj ui , vi exp dij2 h2 , i, j 1, ,n . In equations (3.4) and (3.5), d ij measures the Euclidean distance between points i and j , and h is referred to as the bandwidth. One difficulty with spatial regression is that the estimated parameters are, in part, functions of the weighting function. As the bandwidth h tends to infinity, the weighting function exp dij2 h2 tend to one for all pairs of points so that wij n 1 1 j i and m jj ui , vi 1 i, j . Equivalently, the weights wij and m jj ui , vi becomes uniform for every point j no matter how far it is from location i , and GWR becomes equivalent to SAR. Conversely, as h becomes smaller, the parameter will increasingly depend on observations in close proximity to i . In particular, the weighting function exp dij2 h2 tends to zero when the distance d ij exceeds approximately 2.15 times as long as the bandwidth h . The problem hence becomes how to select an appropriate bandwidth or decay function in regression. In this study we choose h on a criterion of minimum Predicted Residual Error Sum of Squares (PRESS), where the fitted value with the point i omitted from the calibration process. 72 The essential idea of GWR is that for each point i there is a bump of influence around i corresponding to the weighting function so that sampled observations near to i have more influence in the estimation of the parameters of i than do sampled observations farther away. We perform weighted least squares regression for each point i in a SAR model. Equation (3.6) demonstrates the expression of the theoretical coefficient estimates. 1 (3.6) θˆ ui , vi X I W M ui , vi I W X X I W M ui , vi I W y . SAR models are typically estimated by the maximum likelihood (ML) method, where the likelihood function corresponds to the normal distribution. A practical difficulty with the ML method in SAR models is that the estimation of entails significant computational complexities. In this essay we adopt a generalized method of moments (GMM) estimator that is computationally simple irrespective of the sample size (Kelejian and Prucha 1999). Besides, we conduct Lagrange Multiplier (LM) test for every GWR as shown in equation (3.7). (3.7) LM ui , vi nμˆ ui , vi Wμˆ ui , vi μˆ ui , vi μˆ ui , vi 2 tr WW WW , where μˆ ui , vi is a GWR residuals based on estimation under the restricted model, i.e., 0 . The test statistic is asymptotically distributed as 1 (Burridge 1980). 73 DATA Our study covers Mainland China. Most data used in this essay were provided by the Chinese Academy of Sciences (CAS) including soil profile, climate, land use, terrain, and socioeconomic data. They are measured at a scale of 10 by 10 square kilometers, except for socioeconomic data, which are measured at county level. Table 3.1 provides a detailed summary of the data. Table 3.1. Summary Statistics of Explanatory Variables Variable Measurement Unit 10-km-gird level SOC density N Mean Std. Dev. Minimum Maximum kg/C m2 93,802 6.242 5.736 0.274 82.957 Soil PH value N/A 93,802 6.060 1.835 0.000 9.000 Bulk density g/cm3 93,802 1.340 0.149 1.013 2.544 Soil loam content percent 93,802 24.40 9.62 0.00 54.00 Soil sand content percent 93,802 56.06 17.71 19.00 100.00 Precipitation, long-term 1000 mm 93,802 0.579 0.496 0.009 2.498 Temperature, long-term degree Celsius 93,802 5.887 8.127 -20.900 26.000 Land productivity Terrain slope Elevation Precipitation, 1996-2000 Std. of precipitation, 1996-2000 Temperature, 1996-2000 Std. of temperature, 1996-2000 g/ha. degree km 1000 mm 1000 mm degree Celsius degree Celsius 93,902 94,662 94,612 94,173 94,173 94,173 94,173 1.413 3.555 1.837 0.478 0.081 6.677 0.599 2.632 5.010 1.742 0.436 0.067 8.045 0.153 0.000 0.000 -0.153 0.006 0.002 -17.000 0.239 14.168 72.790 7.040 1.824 0.368 31.620 1.693 m/10000 ha. billion CNY billion CNY million people million people million CNY million CNY 2,331 2,247 2,251 2,332 2,333 2,143 2,140 1.022 2.593 3.956 0.510 0.529 0.077 0.096 3.794 6.518 11.121 0.499 0.514 0.423 0.526 0.000 0.021 0.041 0.006 0.006 0.000 0.000 155.708 202.418 364.877 10.616 10.817 13.653 17.057 county level highway GDP, 1996 GDP, 2000 Population, 1996 Population, 2000 Agricultural investment, 1999 Agricultural investment, 2000 Data on soil profile were generated from a geographical information system (GIS) database and includes cross-sectional data of SOC density, PH value, bulk density, soil loam and sand content. They were initially collected by a special nationwide research and documentation project (the Second National Soil Survey of China) organized by the 74 State Council and run by a consortium of universities, research institutes and soils extension centers. CAS interpolated the soil information into surface data using Kriging algorithm, to get more disaggregated information for each pixel. Data on land-use type, terrain, climate, and socioeconomic variables have been discussed in the first essay (Chapter 2). 75 RESULTS We conduct cross-sectional estimation for the SOC density model (equations (3.2) - (3.3)) and the land-use change model (equations (2.2) - (2.3)), respectively, at the unit of 10-km-grid scale. Estimation results of two models are reported and discussed below SOC Density Model Figure 3.1. Histogram of R-square for all GWR‟s in the SOC density model. Under the criterion of minimum Predicted Residual Error Sum of Squares (PRESS), the bandwidth h is estimated to be 75 km. The essential idea behind the bandwidth is that for each location i there is a circle centered at i with a radius of 161.25 km ( 161.25 km 75 km 2.15 ). Within the circle points around i have “bump of influence” on i ; beyond the circle the influence of points are negligible. We perform GWR for each point i in the context of SAR model. GWR reports complete coefficient 76 estimates for 48830 observations out of 93820 in total.4 So the attributes of estimated parameters are discussed via their statistical distributions rather than any particular point estimates. 1.0 coefficient 0.5 0 -0.5 -1.0 farmland forest grass water urban unused landtype Figure 3.2. Box plot of land-use dummy estimators for all GWR‟s in the SOC density model. The SOC density model performs well. Figure 3.1 illustrates the distribution of R2, with the mean value equal to 0.633. For each SAR model, there is convincing evidence that error term is correlated over space (P-value<<0.001). The spatial autoregressive parameter ( ) is estimated to be 0.999 uniformly in all models. We remove the intercept from the regression to avoid collinearity caused by the land-use dummy. Thus the reference is referred to as the expected mean value of logarithm of SOC density for the pooled sample, which contains all observations of the six land-use categories. The absolute magnitude of coefficient of land-use dummy measures the difference in the expected mean of logarithm of SOC density for each separate land-use category relative to the reference. The P-value associated with the land-use dummy 77 variable ranges between 0.000 and 0.100. The average of the P-value is 0.032, presenting credible evidence for the statistically significant association of the logarithm of SOC density with land uses. We also report box plots of land-use dummy estimates by uses (Figure 3.2). It demonstrates large spatial variations in the regression results. Among estimates of six land-use dummies, forestland has the highest average (0.0068) and is followed by grassland (0.0030). By contrast, the average estimates of four remaining uses are found to be negative. In particular, unused land has the lowest mean (-0.0092). Like land-use dummy estimators, estimates of other covariates exhibit spatial nonstationarity. We do not generate a discussion on these covariates because they are not of our primary interest in the essay. Estimation results of these variables are available upon request. Land Use Change Model For each initial land use, we employ ML method to estimate multinomial logit models on land use changes from 1995 to 2000. There are five initial uses (farmland, forestland, grassland, water area, and unused land) and six final uses (farmland, forestland, grassland, water area, urban area, and unused land). So there are five separate models in total. These models perform well, where pseudo R2 (McFadden's likelihood ratio index) ranges between 0.546 and 0.825. We report the estimation results of land-use change on farmland, grassland, forestland, water area, and unused land in Tables 2.3b, 2.5b, A2.2, A2.6, and A2.10 (Chapter 2). Estimates and standard errors of parameters of equation (2.3) are reported in each column of these tables by land-use choice. It shows that the sign, magnitude, and statistical significance of estimates are in line with the economic interpretation. For example, all transition-specific constants have negative estimates and almost all of them are statistically different from zero at the 1% level, indicating that conversion cost deters land conversion on initial use. Likewise, the estimates of land productivity are negatively significant in all columns of Table 2.3b and positively significant in the first columns of Tables 2.5b, A2.2, and A2.10, implying that a land patch with higher crop yield potential is more likely to be used for farming. We 78 also found that the odds of land conversion are associated with climate, including mean annual rainfall, mean annual temperature, and their temporal variations. The results provide strong evidence for the association between the odds of land use change and socioeconomic factors, i.e., county GDP, population, public agricultural investment, and highway density. The sign and statistical significance of estimates are consistent in land conversion on different uses. As is shown, GDP estimates are reported statistically positive in the utility equations of unused land in Table 2.3b (land-use change on farmland) and statistically negative in the utility equations of farmland in Table 2.5b (land-use change on grassland), implying that land is less likely to be used for cultivation in a county with higher level of GDP. With higher GDP, the demand for residential development and industrial and commercial uses increases. Farming is generally a lowpaying job in China, so farmers are willing to be engaged in other higher-returned activities rather than farming. Conversely, farmers are more likely to farm in a county with low GDP for the lack of high-paying jobs. Estimates of population are statistically negative in the utility equations of grassland in Tables 2.3b and A2.2 (land-use change on farmland and forestland) and are statistically positive in the utility equations of farmland and forestland in Table 2.5b (land-use change on grassland), implying that cultivation and afforestation are more likely to take place than pasturing on a patch of land that possesses larger population. In addition, a patch of grassland with smaller population is more likely to be converted to unused land. The effects of highway density on land conversion are very similar to those of population. Highway density measures freights of conveying agricultural and forest products. Dense highway tends to lower transport costs, which decreases the probabilities of farmland conversions and deforestation. Public agricultural investment increases the probabilities of agricultural use (farmland and grassland), because investment in agriculture infrastructure such as irrigation can to improve agricultural productivity in the long run and hence increase the value of agricultural land. Consequently, the opportunity cost of converting land out of agricultural use is raised. 79 GREEN GROWTH, CARBON SEQUESTRATION, AND WELFARE LOSS It is impressive that China has been at the forefront of the surge in economic growth for the last twenty years. However, China‟s economic success was at the expense of degraded environment and diminishing natural resources. According to China Green National Accounting Study Report 2004 (SEPA and NBS 2006),5 the financial loss caused by environmental pollution was 511.8 billion CNY (equivalent to 66.3 billion USD), accounting for 3.05% of national GDP in 2004. The growing economy has particularly made a large impact on the SOC storage. From 1988 to 2000, GDP growth was estimated to induce farmland loss, deforestation, and idleness respectively by 573, 131, and 498 hectares in China (Li et al. 2010). These changes heavily lowered its SOC density. In this situation, it is imperative for Chinese policymakers to change development guidelines from „grow first, clean up later‟ to a more responsible long-term strategy. Green growth policy can serve for this purpose because it pursues environmentally sustainable economic development and generates co-benefits for carbon sequestration and emission abatement.6 However, it is costly to develop green growth strategies in the short run because part of the resources would be allocated to design and implement policy instruments, including taxes, regulations, incentives for clean technologies, and funding for the basic research. These instruments would inevitably result in an economic slowdown. In this section, we intend to evaluate the welfare losses associated with land conversion and soil carbon sequestration in the context of green growth policies. To this end, we assume the structure of land-use in the past would continue to work in the future and simulate land-use changes and SOC flows caused by economic growth for each five-year interval during the period 2001-2050 by coupling the empirical land-use change model and SOC density model. The baseline simulation uses actual observations in 2000 throughout the simulation period. It provides a benchmark to measure the effects of GDP growth on land-use changes and SOC flows. We design ten alternative scenarios (Table 3.2), in which the annual GDP growth rates range in 0.5% reduction from 10% to 5.5% 80 for the first decade and then decline by 1% per five-year. We assume that there was a growth rate floor (2.0%). The remaining variables are held at their historically observed values in 2000. Table 3.2. Description of Scenarios with different Annual GDP Growth Rates Time interval 2001-05 2005-10 2010-15 2015-20 2020-25 2025-30 Baseline 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% Scenario 1 10.0% 10.0% 9.0% 8.0% 7.0% 6.0% Scenario 2 9.5% 9.5% 8.5% 7.5% 6.5% 5.5% Scenario 3 9.0% 9.0% 8.0% 7.0% 6.0% 5.0% Scenario 4 8.5% 8.5% 7.5% 6.5% 5.5% 4.5% Scenario 5 8.0% 8.0% 7.0% 6.0% 5.0% 4.0% Scenario 6 7.5% 7.5% 6.5% 5.5% 4.5% 3.5% Scenario 7 7.0% 7.0% 6.0% 5.0% 4.0% 3.0% Scenario 8 6.5% 6.5% 5.5% 4.5% 3.5% 2.5% Scenario 9 6.0% 6.0% 5.0% 4.0% 3.0% 2.0% Scenario 10 5.5% 5.5% 4.5% 3.5% 2.5% 2.0% 2030-35 0.0% 5.0% 4.5% 4.0% 3.5% 3.0% 2.5% 2.0% 2.0% 2.0% 2.0% 2035-40 0.0% 4.0% 3.5% 3.0% 2.5% 2.0% 2.0% 2.0% 2.0% 2.0% 2.0% 2040-45 0.0% 3.0% 2.5% 2.0% 2.0% 2.0% 2.0% 2.0% 2.0% 2.0% 2.0% 2045-50 0.0% 2.0% 2.0% 2.0% 2.0% 2.0% 2.0% 2.0% 2.0% 2.0% 2.0% Simulations are run at a 10-km-grid scale (equivalent to 10,000 hectares). For each time interval, the probability of choosing use l for every grid i is a joint probability depending on the unconditional probabilities in the last period and the conditional (transition) probabilities between the last and current periods. The formula is presented in equation (3.8).7 (3.8) Pil t k 1 Pik t 5 Pil|k t , 6 t 2005, 2010, , 2050. The total area of each use for each time interval is calculated through summing the choice probabilities over all grids and multiplying the summations by 10,000 hectares (formula (3.9)). (3.9) Areal t 10,000 i 1 Pil t , n t 2005, 2010, , 2050. For each time interval, the SOC density of every grid i is estimated as the product of the SOC density in 2000 and the expected SOC density change due to land conversion, as shown in formula (3.10). (3.10) SOCDi t SOCDi 2000 exp 6 ˆ Pil t Pil 2000 , l 1 il t 2005, 2010, , 2050, 81 where ˆil is the coefficient estimate of land use dummies in SOC density model. Likewise, the total SOC at time period t is computed by summing SOC densities across all land grids and multiplying by 10,000 hectares (formula (3.11)). (3.11) SOC t 10,000 i 1 SOCDi t , n t 2005, 2010, , 2050. The procedure is applied to every scenario. We subtract the values obtained from formulas (3.9) and (3.11) under the baseline from the corresponding values under every GDP-growth scenario. The differences are land-use changes and SOC flows driven by economic growth. Figure 3.3. The area of land by use in the baseline scenario. Figure 3.3 illustrates the area of land by use in the baseline scenario. As is shown, the areas of unused land, farmland, grassland, and forestland are relatively large in China, ranging from 138 to 195 million hectares in 2000. In contrast, the urban and water areas are small, covering 3.8 and 15.7 million hectares, respectively. Under the 82 baseline scenario, the largest change occurs in urban land, with an increase from 3.8 to 14.7 during the 50-year simulation period. Meanwhile, the areas of land in cultivation, forest, and pasturing decline. Specifically, farmland increases in the first five years but begins to decline afterwards. Grassland experiences the greatest absolute loss by 7 million hectares. Generally speaking, except for urban area, changes in most uses are relatively small under the baseline simulation. Figure 3.4. The area of land by use in the scenario of initial 10% GDP growth rate. Once economic growth is considered, e.g., the annual GDP grows at 10% in the first decade followed by 1% reduction in the annual rate every five-year, unused land almost doubles from 138 to 239 million hectares throughout the simulation period (figure 3.4). Most of the increase is from declines in farmland, forestland, and grassland. Despite an increase in the first five years, farmland suffers the largest absolute loss by 55.5 million hectares, which is 13 times as many as the loss in the baseline. Following farmland, forestland and grassland experience the second and the third greatest loss by 83 31.4 and 27.3 million hectares, respectively. The losses accelerate temporally though GDP grows at a diminishing rate. The general pattern of land-use changes is similar at other levels of the annual GDP growth rate. Conversions of farmland, forestland, and grassland to idleness lead to SOC losses.8 Figure 3.5 pictures negative flows of SOC relative to the baseline for ten levels of annual GDP growth rates. It shows that more carbon runs off as the growth rate increases. The variation of carbon flows is large in different growth rates. For example, by the end of the simulation, the SOC runoff in the scenario of the fastest GDP growing (20.7 million Mg C) is nearly twice more than the outflow in the scenario of the slowest GDP growing (7.5 million Mg C). Figure 3.5. The flow of soil organic carbon relative to the baseline for different GDP growth rates. In the context of discrete land-use choice model, the usual techniques for measuring welfare effects must be modified since the welfare properties are more complicated. In this essay we apply the “logsum” formula derived by Small and Rosen (1981) to compute the expected compensating variation (CV) associated with land conversion for changes in GDP growth rate.9 As an approximate measure of consumer 84 surplus in the (multinomial) logit model framework, the “logsum” has been used in the transport literature (transport demand) and urban literature (residential location, see Phaneuf et al. 2008). But as far as we know, it has not been applied in land use studies. Equation (3.12) presents the expression of the “logsum” formula. For a land grid i with initial use k at time t , the compensating variation (CV) for a change in GDP growth rate is (3.12) 6 exp Vis|k t CVik t ln 6s 1 exp Vis|k t s 1 V I I t 2005, 2010, , , 2050 , where Vis|k t and Vis|k t are the deterministic components of utility of converting patch i from use k to s before and after the change in growth rate respectively, and I represents household income. Hence V I I expresses marginal utility of income (or money), which is assumed constant across agents and land-use choices and over time. The estimates of V I I is 1.49 105 . Considering the randomness of initial use in every time interval, the expected CV for patch i at time t is computed as (3.13) CVi t k 1 CVik t Pik t , 6 t 2005, 2010, , 2050 . Summing CV across all individual land grids, we obtain the total expected CV for a change in annual GDP growth rate at time t (formula (3.14)). (3.14) CV t i 1 CVi t , n t 2005, 2010, , 2050 . Figure 3.6 portrays the expected CV relative to the baseline for ten levels of annual GDP growth rates along time. We found that agents‟ welfare associated with land conversion is improved as the growth rate increases. Specially, relative to the baseline, the net welfare gain in the scenario of the fastest GDP growing (745 million CNY) is six times as many as the net gain in the scenario of the slowest GDP growing (122 million CNY) by the end of the simulation. We compute the present values of total welfare gain and total carbon loss for each scenario throughout the simulation period with three levels of social discount rate 85 Figure 3.6. Time series plot of expected logsum welfare gains for different GDP growth rates relative to the baseline. (2%, 4%, and 6%). We use annualized equivalents to express them (Richards and Stokes 2004; Stavins 1999). The annualization indicates that marginal damages caused by additional units of atmospheric carbon are constant and that benefits and costs are to be discounted at the same rate. Consider the Chinese government implemented a mix of policy instruments to achieve green development. These instruments might include pollution taxes, natural resource charges, emission standards, regulations in the production and use of toxic chemicals, motivating end-of-pipe technology innovations, funding for fundamental R&D, and so on. Implementation of such policies would slow down economy growth to different extents and we assume the economy decelerate from scenario 1 to scenario 10. Figures 3.7-3.8 illustrate the expected marginal and average welfare losses associated with soil carbon sequestration by discount rate, which are the main results of this essay. Moving from scenario 1 to scenario 10 could generate up to 0.7-1.1 million annualized Mg C and 22.2-37.4 million CNY losses of welfare for each year throughout 86 Figure 3.7. The marginal welfare losses of soil carbon sequestration by discount rate under the Green Growth policy scenario. the simulation period. Both marginal and average curves are downward-sloping, implying a concave total welfare curve. The marginal curves are convex at low levels of welfare losses (below 25 CNY/Mg) and approximately concave throughout most of its range. Empirically, we found that higher discount rates reduce the annualized carbon flows, the annualized marginal and average losses of welfare. Hence both curves shift inwards (or leftwards) as the discount rate goes up. For example, in sequestering 0.7 million Mg SOC per year, the annualized marginal (or average) welfare loss per Mg C declines from ¥34.3 (¥42.7) to ¥21.8 (¥37.8) and ¥13.1 (¥37.1) at discount rates of 2%, 4%, and 6%, respectively. We compare these results with the marginal costs of agricultural practice, including conservation tillage and reduction in fellow, estimated in previous studies of soil carbon sequestration. To generate about 1 million Mg C per year, the estimated marginal welfare loss is ¥15.3/Mg (equivalent to $2.25/Mg) in our study, which is lower than the costs estimated by previous literature in which the marginal costs are $11.6/Mg 87 (Feng et al. 2006), $50/Mg (Antle et al. 2007), $20-100/Mg (Capalbo et al. 2004), and $190-270/Mg (Pautsch et al. 2001). With these results, we see a promising strategy oriented by green growth to sequester SOC. In addition, the total social welfare loss would be much smaller if the co-benefits of green growth policy were considered. Figure 3.8. The average welfare losses of soil carbon sequestration by discount rate under the Green Growth policy scenario. 88 CONCLUSIONS In this essay, we explored two questions: 1) What is the association between land conversion and SOC density in China? 2) What is the welfare loss associated with soil-based carbon sequestration through implementing the green growth policy? For this purpose, we compiled a unique, national-scale dataset which includes high quality land use data and soil profile data. We developed two models for analysis. One is a statistical model to explore the association of SOC density with six major land uses; the other is a multinomial logit model to analyze the effects of socioeconomic factors on land conversion among six uses. We answered these questions by estimating the models and generating a series of simulations with the estimation results. In the SOC model, we found that the SOC density is higher in forest and grass lands and lower in unused land on average. The spatial analysis provides strong evidence for the existence of spatial nonstationarity in the parameter estimates and spatial autocorrelation in error term. In the land-use change model, we found that land-use changes were affected by socioeconomic factors, including county GDP, population, agricultural investment, and highway density. In a county with higher GDP, land is less likely to be used for cultivation. In a county possessing larger population or denser highway, cultivation and afforestation are more likely to take place than pasturing. Public agricultural investment increases the probabilities of agricultural use (farmland and grassland). Combining results of two models, we predicted the effects of GDP growth on land use, SOC storage, and social welfare. We found that GDP growth leads to farmland and grassland loss, deforestation, and idleness, which accelerates soil carbon runoff. We estimated the expected marginal and average welfare losses associated with soil carbon sequestration by generating green growth policy simulations from 2001 to 2050. We found that the policy could sequester up to 0.7-1.1 million annualized Mg SOC and 22.237.4 million CNY losses of welfare per year. The marginal and average welfare loss curves are downward-sloping. The marginal welfare loss is approximately ¥15.3/Mg (equivalent to $2.25/Mg) for sequestering about 1 million Mg SOC per year. A 89 comparison of the results reveals that soil-based carbon sequestration, oriented by the green growth policy, merits consideration. Caution should be exercised when interpreting the welfare effect. It measures the expected compensating variation for policy-induced land-use changes, which is not equivalent to the total social benefits and/or losses resulting from the green growth policy. The contribution of this essay to literature is twofold. First, it is the first economic application in soil carbon sequestration using detailed, national-scale Chinese data. By contrast, previous research is typically conducted by scientists. Second, it applies the “logsum” formula to evaluate the welfare effects of policy-oriented land conversion. The formula allows measuring consumer surplus in the context of multinomial logit models. To the best of our knowledge, no one has used this method for welfare analysis in the land use literature. 90 ENDNOTES 1 E.g., Denitrification-Decomposition (DNDC) model and Integrated Terrestrial Ecosystem C-budget (InTEC) model. 2 A covariate is a secondary variable that can affect the relationship between the dependent variable and other independent variables of primary interest. 3 We also have soil clay content. The summation of soil loam, sand, and clay content equals one. 4 In some circles, the number of parameters is greater than the number of observations which leads to missing estimates. We exclude those incomplete estimates in the following analysis. 5 As an experiment in national accounting, the Green GDP effort collapsed in failure in 2007, when it became clear that the adjustment for environmental damage had reduced the growth rate to politically unacceptable levels, nearly zero in some provinces. 6 Green growth can be seen as a way to pursue economic growth and development, while preventing environmental degradation, biodiversity loss, and unsustainable natural resource use. It builds on existing sustainable development initiatives in many countries and aims at identifying cleaner sources of growth, including seizing the opportunities to develop new green industries, job and technologies, while also managing the structural changes associated with the transition to a greener economy (OECD 2010). 7 For any land grid starting in urban uses, the probabilities of converting to other uses equal zero provided the assumption of irreversible urbanization. 8 The estimation results of SOC density model show that ceteris paribus, the SOC density is highest in forestland and grassland but lowest in unused land on average. 9 Small and Rosen (1981) extended the conventional methods of applied welfare economics to discrete choice models. They showed that the expected CV can be computed as an integral of the Hicksian choice probabilities. 91 ACKNOWLEDGMENTS We thank the National Science Foundation of China (70873118) and the Chinese Academy of Sciences (KZCX2-YW-305-2) for the financial support to generate the dataset used in this study. 92 REFERENCES Adams, R.M., Adams, D.M., Callaway, J.M., Chang, C.-c., McCarl, B.A., 1993. Sequestering carbon on agricultural land: social cost and impacts on timber markets. Contemporary Policy Issues 11, 76-87. Alig, R., Adams, D.M., McCarl, B.A., Callaway, J.M., Winnett, S., 1997. Assessing effects of mitigation strategies for global climate change with an intertemporal model of the U.S. forest and agriculture sectors. Environmental and Resource Economics 9, 259-274. Antle, J.M., Capalbo, S.M., Paustian, K., Ali, M.K., 2007. Estimating the economic potential for agricultural soil carbon sequestration in the Central United States using an aggregate econometric-process simulation model. Climatic Change 80, 145-171. Antle, J.M., Diagana, B., 2003. Creating incentives for the adoption of sustainable agricultural practices in developing countries: the role of soil carbon sequestration. American Journal of Agricultural Economics 85, 1178-1184. Burridge, P., 1980. On the Cliff-Ord test for spatial correlation. Journal of the Royal Statistical Society. Series B (Methodological) 42, 107-108. Capalbo, S.M., Antle, J.M., Mooney, S., Paustian, K., 2004. Sensitivity of carbon sequestration costs to economic and biological uncertainties. Environmental Management 33, S238-S251. Cheng, S., and J. S. Long. 2007. Testing for IIA in the multinomial logit model. Sociological methods and Research 35:583-600. Deng, X., J. Huang, S. Rozelle, and E. Uchida. 2006. Cultivated land conversion and potential agricultural productivity in China. Land Use Policy 23:372-384. Deng, X., J. Huang, S. Rozelle, and E. Uchida. 2008a. Growth, population and industrialization, and urban land expansion of China. Journal of Urban Economics 63:96-115. Fang, J., Chen, A., Peng, C., Zhao, S., Ci, L., 2001. Changes in forest biomass carbon storage in China between 1949 and 1998. Science 292, 2320-2322. Feng, H., Kurkalova, L.A., Kling, C.L., Gassman, P.W., 2006. Environmental conservation in agriculture: land retirement vs. changing practices on working land. Journal of Environmental Economics and Management 52, 600-614. 93 Fotheringham, M.E.C., Brunsdon, C., 1998. Geographically weighted regression: a natural evolution of the expansion method for spatial data analysis. Environment and Planning A 30, 1905-1927. Hartkamp, A. D., K. D. Beurs, A. Stein, and J. W. White. 1999. Interpolation techniques for climate variables. CIMMYT, Mexico, D.F. Huang, Y., Sun, W., 2006. Changes in topsoil organic carbon of croplands in mainland China over the last two decades. Chinese Science Bulletin 51, 1785-1803. Jobbágy, E.G., Jackson, R.B., 2000. The vertical distribution of soil organic carbon and its relation to climate and vegetation. Ecological Applications 10, 423-436. Kelejian, H.H., Prucha, I.R., 1999. A generalized moments estimator for the autoregressive parameter in a spatial model. Internatinoal Economic Review 40, 509-533. Kelejian, H. H., I. R. Prucha. 2001. On the asymptotic distribution of the Moran I test statistic with applications. Journal of Econometrics 104:219-257. Kravchenoko, A., Bullock, D.G., 1999. A comparative study of interpolation methods for mapping soil properties. Agronomy Journal 91, 393-400. Lewis, D. J., and A. J. Plantinga. 2007. Policies for habitat fragmentation: combining econometrics with GIS-based landscape simulations. Land Economics 83:109-127. Li, M., J. Wu, and X. Deng. 2010. Identifying drivers of land-use change in China: A spatial multinomial logit analysis. Working Paper, Oregon State University, OR. Liu, J., M. Liu, D. Zhuang, Z. Zhang, and X. Deng. 2003. Study on spatial pattern of land-use change in China during 1995-2000. Science in China (Series D) 46:373384. Lubowski, R. N., A. J. Plantinga, and R. N. Stavins. 2006. Land-use change and carbon sinks: econometric estimation of the carbon sequestration supply function. Journal of Environmental Economics and Management 51:135-152. Mansfield, C., Phaneuf, D.J., Johnson, R., Yang, J.-C., Beach, R., 2008. Preferences for public lands management under competing uses: the case of Yellowstone National Park. Land Economics 84, 282-305. Nelson, G. C., V. Harris, and S. W. Stone. 2001. Deforestation, land use, and property rights: empirical evidence from Darién, Panama. Land Economics 77:187-205. 94 Nelson, G.C., Hellerstein, D., 1997. Do roads cause deforestation? Using satellite images in econometric analysis of land use. American Journal of Agricultural Economics 79, 80-88. Organization for Economic Co-operation and Development (OECD), 2010. Green Growth Strategy Interim Report: Implementing our Commitment for a Sustainable Future. Meeting of the OECD Council at Ministerial Level, in: Paris. Parks, P.J., hardie, I.W., 1995. Least-cost forest carbon reserves: cost-effective subsidies to convert marginal agricultural land to forests. Land Economics 71, 122-136. Parton, W.J., Scurlock, J.M.O., Ojima, D.S., Gilmanov, T.G., Scholes, R.J., Schimel, D.S., Kirchner, T., Menaut, J.-C., Seastedt, T., Moya, E.G., Kamnalrut, A., J.I., K., 1993. Observations and modeling of biomass and soil organic matter dynamics for the grassland biome worldwide. Global Biogeochemical Cycles 7, 785-809. Pausch, G.R., Kurkalova, L.A., Babcock, B.A., Kling, C.L., 2001. The efficiency of sequestering carbon in agricultural soils. Contemporary Economic Policy 19, 123134. Phaneuf, D.J., Smith, K., Palmquist, R.B., Pope, J.C., 2008. Integrating property value and local recreation models to value ecosystem services in urban watersheds. Land Economics 84, 361-381. Plantinga, A.J., Mauldin, T., Miller, D.J., 1999. An econometric analysis of the costs of sequestering carbon in forests. American Journal of Agricultural Economy 81, 812-824. Polyakov, M., and D. Zhang. 2008. Property tax policy and land-use change. Land Economics 84:396-408. Richards, K.R., Stokes, C., 2004. A review of forest carbon sequestration cost studies: a dozen years of research. Climatic Change 63, 1-48. Schlesinger, W.H., 1977. Carbon balance in terrestrial detritus. Annual Review of Ecology and Systematics 8, 51-81. Small, K.A., Rosen, H.S., 1981. Applied welfare economics with discrete choice models. Econometrica 49, 105-130. State Environmental Protection Administration of China (SEPA), National Bureau of Statistics of China (NBS), 2006. China green national accounting study report 2004. Chinese Academy for Environmental Planning, in: Beijing. 95 Stavins, R.N., 1999. The costs of carbon sequestration: a revealed-preference approach. The American Economic Review 89, 994-1009. Tang, H., Qiu, J., Ranst, E.V., Li, C., 2006. Estimations of soil organic carbon storage in cropland of China based on DNDC model. Geoderma 134, 200-206. Wang, S., Huang, M., Shao, X., Mickler, R.A., Li, K., Ji, J., 2004. Vertical distribution of soil organic carbon in China. Environmental Management 33, S200-S209. Wang, S., Tian, H., Liu, J., Pan, S., 2003. Pattern and change of soil organic carbon storage in China: 1960s-1980s. Tellus 55B, 416-427. Wu, H., Guo, Z., Peng, C., 2003. Distribution and storage of soil organic carbon in China. Global Biogeochemical Cycles 17, 1-11. Yang, Y., Mohammat, A., Feng, J., Zhou, R., Fang, J., 2007. Storage, patterns and environmental controls of soil organic carbon in China. Biogeochemistry 84, 131141. Zhang, F., Li, C., Wang, Z., Wu, H., 2006. Modeling impacts of management alternatives on soil carbon storage of farmland in Northwest China. Biogeosciences 3, 451466. 96 CHAPTER 4 DECOMPOSING THE CHANGE OF CO2 EMISSIONS IN CHINA: A DISTANCE FUNCTION APPROACH MAN LI 97 ABSTRACT This essay examines the sources of change in carbon dioxide (CO2) emissions. It evaluates the relative contributions of the sources to emission abatement using a new empirical approach. The method uses the data envelopment analysis (DEA) technique to decompose emission changes into seven components based on the Shephard output distance function. It allows for cross-sectional analysis under flexible data requirement. The method accounts for factors that increase carbon emissions, as well as decrease them. With the application of decomposing change in China‟s CO2 emissions at the provincial level between the years 1991 and 2006, the study finds that 1) GDP scale effect accounts for the majority of emission increments; 2) the emission index associated with capital is a dominant contributor to emission abatement; and 3) the effects of technical change in production and change in the GDP-composition by sector play positive roles in shrinking emissions. 98 INTRODUCTION One of the most important targets in climate change policy is to reduce carbon dioxide (CO2) emissions. CO2 emissions are going up at a rapid rate. For example, China surpassed the United States to become the largest CO2 emitter of the world in 2006, releasing nearly 1.7 million petagrams of carbon (Pg C) into the atmosphere, which is 2.4 times more than the amount it emitted in 1991 (Marland et al. 2009). Under the 1997 Kyoto Protocol, Annex I countries (i.e., most industrialized countries) committed themselves to cutting their aggregate anthropogenic CO2 equivalent emissions by at least 5% below 1990 levels for the time periods 2008 to 2012. Kyoto‟s Clean Development Mechanism also allows Annex I countries to implement emission-reduction project in developing countries to acquire carbon credits and offset their off-quota of emissions. Consequently, how to reduce CO2 emissions presents great challenges for policy makers of many countries. It is of theoretical and practical significance to understand what factors have driven the change of CO2 emissions. Many of the previous studies used an index decomposition analysis (IDA).1 Examples include Ang and Pandiyan (1997), Han and Chatterjee (1997), Sun (1999), Viguier (1999), Wang et al. (2005), and Wu et al. (2005). A majority of findings point out that gross domestic product (GDP) growth is a main source of CO2 emission increments in most countries, while declining energy intensity contributes to declines in emissions over the time periods 1960 to 2000. The findings of studies on China are consistent with those in the global context (Ang and Pandiyan 1997; Wang et al. 2005; Wu et al. 2005). The major purpose of this essay is to propose an alternative approach to examine the sources of CO2 emissions change and evaluate the relative contributions of the sources to emission abatements. To achieve the objective, we decompose the change into seven components based on the Shephard output distance function under a joint production framework. We apply this method to disaggregate the province-level changes of CO2 emissions in China from 1991 to 2006. Major findings are 1) GDP scale effect accounts for the majority of emission increments; 2) the emission index associated with 99 capital is a dominant contributor to emission abatement; and 3) the effects of technical change in production and GDP-composition change by sector play positive roles in shrinking emissions. Distance function approach is recently applied in the literature for decomposing the change in undesirable production outputs (Pasurka 2006; Zhou and Ang 2008). Those applications differ in the assumption of measuring technical efficiency. For example, Pasurka (2006) assumes an output-based radial measure which scales on good and bad outputs symmetrically; Zhou and Ang (2008) use input distance functions and output distance functions that model include undesirable output to measure changes in efficiency. In this essay we assess productive efficiency through desirable output subvector by holding bad output and inputs constant. This assumption is reasonable because improvement of productive efficiency associated with good outputs is much more common in practice than that associated with bad outputs. Besides, in contrast to Pasurka (2006) and Zhou and Ang (2008) where the effects of all production inputs are treated jointly, our approach possesses a unique feature which allows isolating the impact of every input on emission change. Application of the isolating technique is found in energy studies (Ang et al. 2004; Wang 2007). But to the best of our knowledge, it has not been applied in emission analysis.2 The essay also contributes to the literature in two aspects. First, it uses crosssectional data observed at the provincial level, which overcome the limitation in the existing literature where decomposition typically relies on detailed sector-level data. Second, it investigates the sources of CO2 emission changes in China in the recent sixteen years (1991-2006). Most of previous analyses are conducted in sample periods of pre2000. The remainder of this essay is organized in the following manner. Section 2 describes the methodology. Section 3 applies the approach for decomposing China‟s CO2 emission changes between 1991 and 2006. The final section concludes. 100 METHODOLOGY In the context of joint production settings, the technology at any time period t is described by the set: (4.1) S t K , L, E, Y, C; t : K , L, E can produce Y, C at time t , where K , L, E denote production inputs of capital, labor, and energy, respectively; Y 3 represent all desirable outputs of GDP; C is undesirable outputs of CO2 emissions joint with Y . We group GDP into three major sectors – agriculture, industry, and service, and use elements Y1 , Y2 , and Y3 to represent outputs from those sectors (i.e., Y Y1 , Y2 , Y3 ). CO2 emissions are generated by agriculture, industry, and service that cover all production activities in an economy. The set S t satisfies standard properties of production set such as null-jointness, weak disposability of joint outputs (good and bad), strong disposability of inputs and good outputs, which suffice to define a Shephard output function. In words, the idea of null-jointness means that if some good outputs are produced then some bad outputs must also be produced; weak disposability says that it is costly to reduce the bad outputs because abatement uses resources that otherwise could have been used to maintain or increase desirable outputs; strong disposability implies that 1) good outputs can be disposed of without cost and 2) an increase in inputs cannot congest outputs. Appendix C gives descriptions of the properties through the production set.3 Eq. (4.2a) provides the definition of the Shephard output distance function for the subvector of good outputs Y , which measures how far the observed production of good outputs is from the maximum potential production at time t given inputs, the bad output, and technology. (4.2a) Yt t Dy K t , Lt , E t , Yt , C t ; t inf : K t , Lt , E t , , C S t , Alternatively, we can define a subvector distance function for mixed period analysis, 101 Y Dy K , L , E , Y , C ; t inf : K , L , E , ,C (4.2b) S t , where the superscript represents the time period of observed production activity and t represents the time period of technology, e.g., t 1 . By definition, it is easy to verify that the subvector distance function satisfies three properties: first, Y , C S t ; second, nonincreasing in for Dy K , L , E , Y , C ; t 1 K ,L ,E , the input set K , L , E , C ; third, nondecreasing, homogeneous of degree 1 and convex in good outputs Y . Taking production technology at time period t (i.e., S t ) as a reference, the change of CO2 emissions between time periods t and t 1 is decomposed as in the following equation, (4.3) Dy K t , Lt , E t , Y t , C t ; t C t 1 Y1t 1 Y2t 1 Y3t 1 Ct Y1t Y2t Y3t Dy K t 1 , Lt 1 , E t 1 , Y t 1 , C t 1 ; t C t 1 Y1t 1 Y2t 1 Y3t 1 Dy K t 1 , Lt 1 , E t 1 , Y t 1 , C t 1 ; t Ct Y1t Y2t Y3t Dy K t , Lt , E t , Y t , C t ; t Dy K t , Lt , E t , Y t , C t ; t Y1t 1 Y2t 1 Y3t 1 Y1t Y2t Y3t Dy K t 1 , Lt 1 , E t 1 , Y t 1 , C t 1 ; t 1 Dy K t 1 , Lt 1 , E t 1 , Y t 1 , C t 1 ; t 1 C t 1 Dy K t 1 , Lt 1 , E t 1 , y t 1 , C t 1 ; t Dy K t 1 , Lt 1 , E t 1 , Y t 1 , C t 1 ; t C t Dy K t , Lt , E t , y t , C t ; t : SC TE TC t 1 C t 1 Dy K t 1 , Lt 1 , E t 1 , y t 1 , C t 1 ; t C t Dy K t , Lt , E t , y t , C t ; t , Y3 Y1 Y2 , , where y , representing proportions of output from Y1 Y2 Y3 Y1 Y2 Y3 Y1 Y2 Y3 the three sectors in GDP; superscript t and t 1 denote time periods. The decomposition 102 (4.3) indicates that the emission change between periods t and t 1 is the product of four multipliers. The first one, named as scale effect ( SC ), is the ratio of GDP between the two time periods. The second one is technical efficiency change effect ( TE ), which captures the change in the distance between the observed production and its corresponding potential production between time t and t 1 . The third one is defined as technical change effect ( TC t 1 ), measuring the shift of frontier technology by taking production at time period t 1 as a reference. The technical efficiency change effect and technical change effect are the components of the Malmquist productivity change index and are widely studied in the area of productivity and efficiency analysis (Färe et al. 1994). Although technology scales solely on the good output Y , TE and TC can provide a measure of changes in CO2 emissions associated with technical change and changes in technical efficiency, because emission changes will affect the shape of production frontier and the location of individual production in the production set. When the assumption of constant returns to scale (CRS) is imposed on the reference technology S t , the last multiplier on the right-hand side of Eq. (4.3) can be reorganized such that C t 1 Dy K t 1 , Lt 1 , E t 1 , y t 1 , C t 1 ; t (4.4) C t Dy K t , Lt , E t , y t , C t ; t K t 1 C t 1 Dy where k Dy t 1 , CLt 1 , CEt 1 , y t 1 , CC t 1 ; t t 1 Kt Ct t 1 t t t , CLt , CE t , y t , CC t ; t Dy k t 1 , l t 1 , et 1 , y t 1 ,1; t Dy k t , l t , et , y t ,1; t , K L E , l , and e , denoting capital-carbon ratio, labor-carbon ratio, and C C C energy-carbon ratio, respectively. This rearrangement reveals a desirable property of the Shephard output distance function under CRS assumption, i.e., the function is homogeneous of degree -1 in K , L , E ,C . t t t t As we previously discussed, the assumption of weak disposability of joint outputs GDP and CO2 emissions means that cutting emissions uses resources which otherwise could have been used to maintain or increase good outputs. Hence changes in k, l, and e respectively reflect changes in the 103 allocation of each individual input between good outputs production and CO2 emission abatement. For example, ceteris paribus, any increase in k, l, and e will result in reduction of CO2 emissions. Due to the direct link between energy consumption and CO2 emission, changes of e also capture the effect of fuel switching on CO2 emissions.4 Dy k t 1 , l t 1 , et 1 , y t 1 ,1; t Dy k t , l t , et , y t ,1; t is an Malmquist-type productivity index with the reference technology of S t . It measures a joint effect that results from changes in the ratios of each of three inputs to carbon and GDP-composition change by sector between time periods t and t 1 . We further separate Dy k t 1 , l t 1 , et 1 , y t 1 ,1; t Dy k t , l t , et , y t ,1; t into four parts by isolating the effects of changes in k , l , e , and y between the two time periods. We write the first component in Eq. (4.5) below to illustrate the structure and present the whole decomposition of index Dy k t 1 , l t 1 , et 1 , y t 1 ,1; t Dy k t , l t , et , y t ,1; t in Appendix D. 1 24 (4.5) D k t 1 , l t 1 , et 1 , y t 1 ,1; t 6 D k t 1 , l t , et 1 , y t 1 ,1; t 2 y y t t 1 t 1 t 1 t t t 1 t 1 Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t 2 2 t 1 t 1 t t 1 t 1 t 1 t 1 t D k , l , e , y ,1; t D k , l , e , y ,1; t y y D k t , l t 1 , et , y t 1 ,1; t D k t , l t 1 , et 1 , y t ,1; t y y KC t 2 2 t 1 t t t 1 t 1 t t 1 t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t t t t t 1 t t t 1 t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t 2 6 Dy k t 1 , l t 1 , et , y t ,1; t Dy k t 1 , l t , et , y t ,1; t t t 1 t t t t t t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t . We name KC t as emission index associated with capital. It measures the effect of change in capital-carbon ratio on CO2 emissions. The remaining three components, denoted as LC t , EC t and CM t , can be derived in a similar way. Thus we have 104 (4.6) Dy k t 1 , l t 1 , et 1 , y t 1 ,1; t Dy k t , l t , et , y t ,1; t : KC t LC t EC t CM t . The definitions of LC t and EC t are analogs to that of KC t , respectively measuring the emission indices associated with labor and energy. CM t is defined to measure GDP composition effect. Combining equations (4.3), (4.4) and (4.6) gives that (4.7) C t 1 : SC TE TC t 1 KC t LC t EC t CM t . Ct Notice that decomposition (4.7) uses technology at time t as a reference. Alternatively, change in CO2 emissions can be decomposed with the reference of production technology at time t 1 (i.e., S t 1 ). (4.8) C t 1 : SC TE TC t KC t 1 LC t 1 EC t 1 CM t 1 . Ct To avoid choosing an arbitrary benchmark, we calculate the geometric mean of the right hand side of equations (4.7) and (4.8) and obtain the fundamental decomposition result of this essay as shown in Eq. (4.9). 12 12 C t 1 12 12 SC SC TE TE TC t 1 TC t KC t KC t 1 t C (4.9) LC t LC t 1 12 EC t EC t 1 12 CM t CM t 1 12 : SC TE TC KC LC EC CM . The decomposition result in (4.9) demonstrates that there are seven indices accounting for the change of CO2 emissions between time periods t and t 1 : GDP scale effect, technical efficiency change effect, technical change effect, emission indices associated with capital, labor, and energy, and GDP composition effect. For any one of them, it will contribute to shrinking CO2 emissions if its value is less than one. To calculate the value of the seven indices, this study constructs a production set with an activity analysis model as described in Appendix E. We use data envelopment analysis (DEA) to compute the value of every output distance function in (4.8). DEA is a 105 nonparametric linear programming approach, which enables us to relax functional assumptions on production technology. Based on definition (4.2b) the relevant linear programming problem is specified as in the optimal problem (4.10). D K y , L , E , Y , C ; t s.t. (4.10) t n 1 n N t n 1 n N 1 max z Ynt Y z Cnt C z K nt K N t n 1 n z Ltn L N t n 1 n z Ent E N t n 1 n znt 0, n 1, , N. Compared with the index decomposition method, the approach proposed in this study satisfies all three criteria tests – time-reversal, factor reversal, and zero-value robust – as suggested in a survey by Ang and Zhang (2000). By contrast, out of ten index decomposition methods investigated by the survey, only three passed all tests.5 Meanwhile, the present decomposition provides an alternative approach to quantifying scale effect, composition effect, and technique effect, which are widely used to explain the environmental Kuznets curve (EKC) since Grossman and Krueger (1993).6 In particular, SC and CM in the context of this study are respectively equivalent to scale effect and composition effect applied in EKC literature. The joint effect of TE TC KC LC EC in Eq. (4.9) plays the same role as technique effect. 106 APPLICATION In this section, we decompose the province-level changes of CO2 emissions in China between 1991 and 2006 using the proposed method. Data Data on sixteen years 1991-2006 are collected for the empirical study. Province is the unit of observation and there are 29 observations for each year.7 Variables include CO2 emissions, capital stock, labor input, energy consumption, and GDP by region. Data in value terms are measured at the 1990 real 108 Chinese yuan (CNY). The national CO2 emissions for the time period 1991-2006 are collected from Marland et al. (2009). However, the actual provincial CO2 emissions do not exist because measuring emissions from a large amount of widely dispersed sources would be prohibitively expensive. National Bureau of Statistics of China (NBSC) reports waste gas emissions (WGE) and industrial waste gas emissions (IWGE) at the provincial level for the time period 1991-95. After 1995 only IWGE is reported. Therefore we use provincelevel WGE and IWGE as parameters to proxy for provincial CO2 emissions. The approach follows a recent study by Auffhammer and Carson (2008) on forecasting China‟s CO2 emissions. Appendix F provides the detail on estimation procedure. Column 1 in Table 1 reports the change in CO2 emissions in 29 provinces. It indicates that the emissions increase in all provinces except for Beijng and Heilongjiang. There are no official data available for measuring China‟s capital stock at the provincial level. This essay estimates it with a perpetual inventory approach, which is a typical way of calculating capital stock in the literature. Start with the initial year of 1978 and for any year t 1979, (4.11) , 2006 ,8 capital stock of province n in year t is calculated as Kt ,n 1 Kt 1,n I t ,n 1 t 1978 K1978,n s 1979 1 t t s , I s ,n where K is capital stock, is depreciation rate, I is the real value of investment in fixed assets, and subscript s represents the time period index. According to the formula 107 (4.10), we need to determine the depreciation rate and the initial value of provincial capital stock in 1978. As in Chow (2002), we adopt a depreciation rate of 4% and assume it to be constant across provinces. Following the studies of Hall and Jones (1999) and Henderson et al. (2007), the initial capital stock for province n is estimated as: (4.12) K1978,n I1978,n gn , where g n is the geometric mean of growth rates of real fixed investment over the time period 1978-88 for province n . Time series of investment in fixed assets for the periods 1978-2006 are collected from NBSC (2005; 2006-07). Investment deflators are not available at the provincial level. We use the same province-specific price index as for GDP as in Henderson et al. (2007). Data on labor input and GDP are collected directly from NBSC (2005; 2006-07). Labor input is measured in the number of employees. GDP consists of three sectors: primary, secondary, and tertiary industries. Primary industry is composed of farming, forestry, animal husbandry, and fishery. Secondary industry mainly includes activities of building, mining, manufacture, electricity and gas production. And those not belonging to the former two are classified into tertiary industry, e.g., transportation, trade, finance, education, public service, etc. For convenience, we use Agriculture, Industry, and Service, respectively, representing three sectors. We collect data on total energy consumption by region from NBSC (2005) and NBSC and NDRC (1997-99; 2005-07) and use 104 megagrams standard coal equivalent as the unit. Results and Discussions Before presenting the final results, it is instructive to give an outline of national emission trend in China from 1991 to 2006, which is plotted in Fig.4.1. It shows that CO2 emissions peaked for the first time in 1996 then dropped off and touched the bottom in 1998. After China‟s entry to the World Trade Organization (WTO) in 2001, the emissions went up at a soaring rate of 12% per year. 108 Figure 4.1. National Carbon Dioxide Emissions in China, 1991-2006 Columns 2-7 of Table 4.1 report contributions to CO2 emission change from GDP scale effect, technical efficiency change effect, technical change effect, and the effects of changes in capital-carbon ratio, labor-carbon ratio, energy-carbon ratio, and GDP-composition by sector between 1991 and 2006.9 As shown in column 2, the value of GDP scale effect ( SC ) is greater than one in all provinces in this study. The annual contribution of SC to CO2 emission increments is 11.6%, much higher than the values of other six indices, indicating that SC is a main driver for emission increases. This result is in line with the findings of previous studies (Wang et al. 2005; Wu et al. 2005; Zhou and Ang 2008). Columns 3-4 in Table 4.1 describe the effects of technical efficiency change ( TE ) and technical change ( TC ) on emission change. On average, technical improvement reduces CO2 emissions by 1.63% per year whereas the impact of technical 109 Table 4.1. Change in CO2 Emissions and its Seven Decomposing Indices, 1991-2006 CO2 SC TE TC KC LC EC Region (1) (2) (3) (4) (5) (6) (7) Beijing 0.9327 5.1038 1.0000 0.3509 0.4833 0.8733 0.8603 Tianjin 3.2448 6.1891 1.0000 0.4012 0.7816 1.7809 1.0032 Hebei 3.9790 5.6581 0.9502 0.7950 0.6567 1.5453 1.0004 Shanxi 3.0833 4.8386 0.7303 0.8495 0.7312 1.4066 1.0000 Inner Mongolia 3.1309 6.1853 1.0228 0.7816 0.5769 1.6155 0.9960 Liaoning 1.7890 4.4473 0.9851 0.6069 0.5389 1.3289 1.0000 Jilin 1.1705 4.9314 0.9297 0.7238 0.3950 1.0995 0.9978 Heilongjiang 0.8985 3.8130 0.8926 0.6817 0.4341 0.9550 0.9897 Shanghai 1.8531 5.7876 1.0000 0.3985 0.5782 1.3230 0.9209 Jiangsu 3.0475 7.3023 1.0000 0.6823 0.5599 1.3244 0.9776 Zhejiang 2.7911 7.4071 1.0590 0.7132 0.5518 1.2122 0.8224 Anhui 2.3603 6.0340 0.9803 0.9556 0.4642 1.0890 0.9950 Fujian 3.7017 7.2234 0.9899 0.7523 0.6392 1.3491 0.9586 Jiangxi 2.3234 5.4711 1.0000 1.1411 0.4388 1.1493 0.9819 Shandong 3.0803 6.9385 0.9699 0.7720 0.5723 1.2996 0.9973 Henan 3.2020 5.4488 0.8435 1.2084 0.5801 1.1834 1.0002 Hubei 2.1211 5.1413 1.0870 0.8227 0.4973 1.2243 0.9962 Hunan 1.8173 4.5680 0.9543 1.1318 0.4520 1.0835 0.9925 Guangdong 2.7694 7.1287 1.0000 0.5802 0.6260 1.2571 0.8887 Guangxi 5.5688 5.4680 1.0000 1.2028 0.7741 1.4490 1.0329 Hainan 7.3462 4.6507 1.0000 0.8228 0.9825 2.1505 0.9934 Sichuan 1.8869 4.7069 1.2188 1.0647 0.3765 1.1018 0.9940 Guizhou 1.3191 3.8372 1.1972 1.4104 0.2919 1.0007 0.9920 Yunnan 3.5650 3.9216 1.2739 1.1349 0.6517 1.2359 1.0006 Shaanxi 2.3000 4.4196 0.9910 0.8701 0.5563 1.1415 0.9988 Gansu 1.6829 4.2883 1.2001 0.9900 0.3524 1.0529 0.9983 Qinghai 3.7703 4.1445 1.2969 0.5736 0.9468 1.5437 1.0000 Ningxia 2.9084 4.3732 1.2577 0.7500 0.6435 1.3844 0.9989 Xinjiang 2.8612 3.9363 1.2670 0.6917 0.7445 1.5658 1.0000 Geomean 2.4943 5.1784 1.0293 0.7810 0.5604 1.2769 0.9777 Annual Geomeans 1.0628 1.1159 1.0019 0.9837 0.9621 1.0164 0.9985 # Regions < 1.00 2 0 11 22 29 2 21 # Regions = 1.00 0 0 8 0 0 0 3 # Regions > 1.00 27 29 10 7 0 27 5 CM (8) 1.4343 0.9357 0.9170 0.9986 0.6821 0.9394 0.8140 0.9439 1.1407 0.8438 0.9071 0.8301 0.8325 0.7516 0.7993 0.8397 0.7605 0.7577 0.9574 0.7308 0.9147 0.7492 0.7025 0.7802 0.9516 0.8919 0.8367 0.7923 0.7114 0.8564 0.9897 27 0 2 efficiency change is negligible. The values of TE and TC are distributed unevenly across regions. For the richest metropolises, Beijing and Shanghai, the contribution of 110 TC to abatement is far above the national average. This is most likely a result of highly concentrated top national universities and research institutes in these cities. By contrast, the contribution of TE is trivial because their economies stay on the production frontier for the whole sample periods. Guangdong presents a similar case as a wealthy coastal province where there is the largest amount of special economic zones in the 1990s. With the advantage of location close to Hong Kong and being received preferential treatment from the central government, Guangdong experienced technical progress. However, being on the frontier at the start of the sample periods prevents it from improving in efficiency. However, interior regions tell a different story. In some midland provinces such as Shanxi and Henan, TE plays a positive role in curtailing CO2 emissions as expected. Located adjacent to coastal provinces, those regions benefit from technological catch-up. But it is not true for western provinces including Yunnan, Gansu, Qinghai, Ningxia, and Xinjiang, which are outlying areas in the South/North-west. Geographical disadvantage makes them fall behind the forefront of economic reform. Emission indices associated with capital ( KC ), labor ( LC ), and energy ( EC ) are reported in columns 5-7 of Table 4.1. It is easy to find that KC plays a dominant role in emission abatement over the 16 years. In contrast to KC , LC raises the emissions by 1.64% per year and becomes the second largest source of increments in emissions. The value of EC is almost equal to one, implying that this index has a trifling effect on emission changes. It is in line with the findings of previous study Ang (1999) where carbon factor (i.e., the ratio of CO2 emissions to energy consumption) is proposed to be a less useful indicator in the study of climate change. The last column of Table 4.1 reports GDP composition effect ( CM ) on emissions change. As is shown, CM results in a reduction in emissions of 1.03% annually. In particular, it shrinks more CO2 emissions in the interior provinces, such as Inner Mongolia, Guangxi, Guizhou, and Xinjiang. One reason is that structural transformation from industry to service tends to curtail industrial emissions. It should be noticed that Beijing and Shanghai are the only two regions with a value of CM greater than one. As is known, energy intensity in agricultural sector is much lower than that in 111 industrial and service sectors. The agricultural proportions of GDP are extremely low in these two metropolises, which leads to higher values of CM . Table 4.2. Geometric Means of Annual Changes for Each Consecutive Two-year Period, 1991-2006 Two-year pairs 1991-1992 1992-1993 1993-1994 1994-1995 1995-1996 1996-1997 1997-1998 1998-1999 1999-2000 2000-2001 2001-2002 2002-2003 2003-2004 2004-2005 2005-2006 Annual Geomeans Non-chaining Annual Geomeans CO2 SC TE TC KC LC EC CM (1) 1.0396 1.0568 1.0385 1.1050 1.0719 1.0254 0.7835 1.0441 1.0852 1.1407 1.0917 1.1207 1.1665 1.1245 1.1076 1.0628 (2) 1.1392 1.1512 1.1334 1.1172 1.1157 1.1040 1.0925 1.0844 1.0923 1.0957 1.1059 1.1184 1.1314 1.1267 1.1326 1.1159 (3) 1.0238 1.0179 0.9931 1.0001 0.9900 0.9988 1.0005 1.0112 1.0006 1.0066 1.0060 0.9927 0.9918 0.9929 1.0037 1.0019 (4) 0.9668 0.9655 0.9654 0.9843 0.9982 0.9781 0.9677 1.0239 1.0175 1.0060 0.9985 1.0113 0.9845 1.0200 0.9910 0.9917 (5) 0.9517 0.9389 0.9345 0.9865 0.9509 0.9500 0.8061 0.9495 0.9673 0.9990 0.9800 0.9874 1.0027 0.9782 0.9712 0.9687 (6) 1.0026 1.0150 1.0047 1.0180 1.0118 1.0124 0.9434 0.9935 1.0151 1.0333 1.0128 1.0185 1.0443 1.0297 1.0212 1.0010 (7) 0.9884 0.9979 0.9973 0.9961 1.0105 0.9949 0.9858 1.0023 1.0060 1.0115 0.9944 0.9900 1.0011 0.9957 1.0042 0.9955 (8) 0.9775 0.9821 1.0208 1.0044 1.0001 0.9936 0.9881 0.9835 0.9879 0.9846 0.9957 1.0025 1.0073 0.9826 0.9871 0.9931 1.0628 1.1159 1.0019 0.9837 0.9621 1.0164 0.9985 0.9897 The proposed method also applies to consequential years. Because choice between chaining (based on time series data) and non-chaining (based on the data for the beginning and ending years) analysis will affect the results of five factor TC , KC , LC , EC , and CM ( SC and TE will not be changed if we accumulate the chain-based results by taking their annual geometric means), we alternatively decompose the change of CO2 emissions for each consecutive two-year period between 1991 and 2006.10 Table 4.2 reports the geometric means of 29 observations for every period and the annual geometric means for the whole period from 1991 to 2006. In comparison with the annual geometric means obtained based on the beginning year 1991 and the ending year 2006, we find that the decomposition results are robust to the selection between chaining and non-chaining 112 analysis. The alternative chain-based decomposition slightly changes five factors in magnitude but does not affect the results by much. The non-chaining analysis is less computational intensive because only 36 subvector distance functions are needed to be estimated. By contrast, the chaining analysis requires calculating 540 distance functions ( 540 36 15 periods ). Hence, we recommend adopting non-chaining analysis when applying the proposed approach to long-term decomposition. The sub-period 1997-98 deserves attention. As opposed to the long-run trend, the average provincial emissions fall down by 21.7% from 1997 to 1998. Three factors – KC , LC , and TC – contribute to the sudden drop. In particular, KC (= 0.8061) reduces as many as 19.4% of CO2 emissions in the one-year interval, which is far lower than its annual geometric means 0.9687 (chaining) and 0.9621 (non-chaining); LC (= 0.9434) plays a positive role in shrinking emissions by accounting for 5.66% of the abatements; TC (= 0.9677) drives the decline by 3.3%, lower than the annual average levels 0.9917 (chaining) and 0.9837 (non-chaining). This phenomenon relates to policy oriented industrial transition from energy intensive technology to capital/labor intensive technology and policy oriented changes in resource allocation between increasing good outputs and reducing bad outputs. In 1997 the Chinese regulatory authority passed the Energy Conservation Law, which provides broad guidance for the establishment of energy efficiency policies. Besides, the concurrent industrial restructuring caused the shutdown of thousands of inefficient enterprises including small-scale mines and power plants. The reform shrank energy consumption and hence led to a sudden fall in CO2 emissions because of the direct link between energy consumption and emissions. Meanwhile, capital and labor inputs were increased. The increments supplemented the loss of energy inputs in production on the one hand and were allocated to emission reduction on the other hand. 113 CONCLUDING COMMENTS This essay presents an alternative approach to index decomposition method for disaggregating change of CO2 emissions in the context of joint production settings. Data requirements for the proposed method are much flexible (e.g., it allows using crosssectional data observed at provincial level). Hence, it overcomes the limitation of index decomposition approach which typically relies on detailed sector-level data. Without specifying any functional form for heterogeneous production functions across regions, it can isolate and quantify the effects of changes in GDP scale, production technology, technical efficiency, capital-carbon ratio, labor-carbon ratio, energy-carbon ratio, and GDP-composition by sector. We apply the proposed method to decomposing province-level changes in CO2 emissions of China into seven components for the time period 1991-2006. Caution must be exercised when interpreting the results, which are influenced by the quality of some of the data. Nevertheless, the results provide strong evidence that the GDP scale effect is the main driver behind rising emissions, whereas the emission index associated with capital, technical change effect, and GDP composition effect are joint contributors to emission abatement. Specifically, capital accumulation (i.e., emission index associated with capital) plays a dominant role across the country. Regional analysis indicates an uneven pattern within China: coastal provinces favor technical improvement whereas interior (midland and western) regions enjoy structural transition of GDP; midland areas benefit from technical efficiency catching-up while western provinces fall behind the technology frontier. Our decomposition results offer informative implications for policymakers to reduce CO2 emissions in China. The interior regions of China have great potential of technical efficiency improvement by introducing advanced technology and learning management experiences from coastal provinces. As the key driver of China‟s overall economic growth (Henderson et al. 2007), capital accumulation will still be the most important contributor to abatement across the country in the foreseeable future. But in the long run, China should rely more on technical progress to reduce CO2 emissions. 114 ENDNOTES 1 Ang and Zhang (2000) provide an excellent survey on the methods and application areas of Index decomposition analysis. 2 See Siegel (1945) for a general discussion. 3 Färe et al. (1989) and Färe et al. (2001; 2007) give the details on definitions and discussions for those properties. 4 Our approach can be extended by using more detailed energy data, e.g., electricity and non-electricity or coal, gas, and petroleum products, to capture more information on fuel switching effect and to distinguish it from energy allocation effect. 5 The three index decomposition methods are multiplicative logarithmic mean Disvisia index, additive logarithmic mean Disvisia index, and refined Laspeyres index. 6 Copeland and Tailor (1994) propose formal model-based definitions of the three effects. 7 Tibet is dropped from the dataset due to incomplete observations. Chongqing, which was elevated to a provincial status municipality in 1997, is counted as part of Sichuan province in this study. Beijing, Shanghai, and Tianjin are provincial-level municipalities and hence are treated as provinces. 8 The year 1978 is one of the milestones in Chinese economy since then market reforms started. In addition, the relatively low initial value of capital in 1978 and the high rates of investment ensure that the estimates of the capital stock for our sample period 1991-2006 are not sensitive to the 1978 benchmark value. 9 Results reported in Table 1 are obtained based on a non-chaining analysis, i.e., using data on the beginning year 1991 and the ending year 2006. 10 Thanks for one of the anonymous referees of Ecological Economics pointing out this issue. 115 ACKNOWLEDGEMENTS I am grateful to Rolf Färe and Shawna Grosskopf for their constructive comments on an earlier version of this essay. I gratefully acknowledge two anonymous referees of Ecological Economics whose suggestions have improved the exposition of the essay substantially. 116 REFERENCES Ang, B.W., 1999. Is the energy intensity a less useful indicator than the carbon factor in the study of climate change. Energy Policy 27, 943-946. Ang, B. W., F. L. Liu, and H.-S. Chung. 2004. A generalized Fisher index approach to energy decomposition analysis. Energy Economics 26:757-763. Ang, B.W., Pandiyan, G., 1997. Decomposition of energy-induced CO2 emissions in manufacturing. Energy Econ. 19, 363-374. Ang, B.W., Zhang, F.Q., 2000. A survey of index decomposition analysis in energy and environmental studies. Energy 25, 1149-1176. Auffhammer, M., Carson, R.T., 2008. Forecasting the path of China's CO2 emissions using province-level information. J. Environ. Econ. Manage. 55, 229-247. Chow, G.C., 2002. China's Economic Transformation. Blackwell, Oxford. Färe, R., Grosskopf, S., Lovell, C.A.K., 1989. Multilateral productivity comparisons when some outputs are undesirable: A nonparametric approach. Rev. Econ. Statist. 71, 90-98. Färe, R., Grosskopf, S., Norris, M., Zhang, Z, 1994. Productivity growth, technical progress, and efficiency change in industrialized countries. Amer. Econ. Rev. 84, 66-83. Färe, R., Grosskopf, S., Pasurka Jr., C.A., 2001. Accounting for air pollution emissions in measures of state manufacturing productivity growth. J. Reg. Sci. 41, 381-409. Färe, R., Grosskopf, S., Pasurka Jr., C.A., 2007. Pollution abatement activities and traditional productivity. Ecological Economics 62, 673-682. Grossman, G.M., Krueger, A.B., 1993. Environmental Impacts of a North American Free Trade Agreement, in: Garber, P.M. (Eds), The U.S.-Mexico Free Trade Agreement. MIT Press, Cambridge, MA, pp. Hall, R.E., Jones, C.I., 1999. Why do some countries produce so much more output per worker than others? Quart. J. Econ. 114, 83-116. Han, X., Chatterjee, L., 1997. Impacts of growth and structural change on CO2 emissions of developing countries. World Devel. 25, 395-407. 117 Henderson, D.J., Tochkov, K., Badunenko, O., 2007. A drive up the capital coast? Contributions to post-reform growth across Chinese provinces. J. Macroecon. 29, 569-594. Marland, G., Boden, T., Andres, R.J., 2009. National CO2 emissions from fossil-fuel burning, cement manufacture, and gas flaring: 1751-2006, in: Trends online: A compendium of data on global change. Carbon Dioxide Information Analysis Center, O.R.N.L., U.S. Department of Energy, Oak Ridge Tennessee. National Bureau of Statistics of China, 2006-07. China Statistical Yearbook. China Statistical Press, Beijing. National Bureau of Statistics of China, 2005. Comprehensive Statistical Data and Materials on 55 Years of New China. China Statistical Press, Beijing. National Bureau of Statistics of China and National Development and Reform Commission, 1997-99, 2005-07. China Energy Statistical Yearbook. China Statistical Press, Beijing. Pasurka, C.A., 2006. Decomposing electric power plant emissions within a joint production framework. Energy Econ. 28, 26-43. Siegel, I. H. 1945. The generalized "ideal" index-number formula. Journal of the American Statistical Association 40:520-523. Sun, J.W., 1999. Decomposition of Aggregate CO2 Emissions in the OECD: 1960-1995. The Energy Journal 20, 147-155. Viguier, L., 1999. Emissions of SO2, NOX and CO2 in transition economics: emissions inventories and Divisia index analysis. The Energy Journal 20, 59-87. Wang, C., 2007. Decomposing energy productivity change: A distance function approach. Energy 32, 1326-1333. Wang, C., Jining, C., Zhou, J., 2005. Decomposition of energy-related CO2 emissions in China: 1957-2000. Energy 30, 73-83. Wu, L., Kaneko, S., Matsuoka, S., 2005. Driving forces behind the stagnancy of China's energy-related CO2 emissions from 1996 to 1999: the relative importance of structural change, intensity change and scale change. Energy Policy 33, 319-335. Zhou, P., Ang, B.W., 2008. Decomposition of aggregate CO2 emissions: A productiontheoretical approach. Energy Econ. 30, 1054-1067. 118 APPENDICES 119 APPENDIX C Properties of production set: (1) Null-jointness: K , L, E, Y, C S and C 0 imply Y 0 ; (2) Weak disposability of joint outputs: K , L, E, Y, C S and 0 1 imply Y Y imply K , L, E, Y, C S ; (3) Strong disposability of good outputs: K , L, E, Y, C S and K , L, E, Y, C S ; (4) Strong disposability of inputs: K , L, E, Y, C S . K , L, E, Y, C S and K , L, E K , L, E imply 120 APPENDIX D The whole decomposition of index Dy k t 1 , l t 1 , et 1 , y t 1 ,1; t Dy k t , l t , et , y t ,1; t is (A4.1) 1 24 D k t 1 , l t 1 , et 1 , y t 1 ,1; t 6 D k t 1 , l t , et 1 , y t 1 ,1; t 2 D k t 1 , l t 1 , et , y t 1 ,1; t 2 y y y D k t , l t 1 , et 1 , y t 1 ,1; t D k t , l t , et 1 , y t 1 ,1; t D k t , l t 1 , et , y t 1 ,1; t y y y 2 2 2 Dy k t 1 , l t 1 , et 1 , y t 1 ,1; t Dy k t 1 , l t 1 , et 1 , y t ,1; t Dy k t 1 , l t , et , y t 1 ,1; t Dy k t 1 , l t , et 1 , y t ,1; t t t 1 t 1 t Dy k t , l t , et , y t 1 ,1; t Dy k t , l t , et 1 , y t ,1; t Dy k t , l t , et , y t ,1; t D k , l , e , y ,1; t y 2 6 t 1 t 1 t t t 1 t t t Dy k , l , e , y ,1; t D k , l , e , y ,1; t y t t 1 t t t t t t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t 1 24 D k t 1 , l t 1 , et 1 , y t 1 ,1; t 6 D k t , l t 1 , et 1 , y t 1 ,1; t 2 D k t 1 , l t 1 , et , y t 1 ,1; t 2 y y y D k t 1 , l t , et 1 , y t 1 ,1; t D k t , l t , et 1 , y t 1 ,1; t D k t 1 , l t , et , y t 1 ,1; t y y y 2 2 2 t 1 t 1 t 1 t t t 1 t t 1 t t 1 t 1 t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t Dy k t 1 , l t , et 1 , y t ,1; t Dy k t , l t , et , y t 1 ,1; t Dy k t , l t , et 1, y t ,1; t 2 6 t 1 t 1 t t t t 1 t t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t t 1 t t t t t t t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t 1 24 D k t 1 , l t 1 , et 1 , y t 1 ,1; t 6 D k t , l t 1 , et 1 , y t 1 ,1; t 2 D k t 1 , l t , et 1 , y t 1 ,1; t 2 y y y D k t 1 , l t 1 , et , y t 1 ,1; t D k t , l t 1 , et , y t 1 ,1; t D k t 1 , l t , et , y t 1 ,1; t y y y 2 2 2 t 1 t 1 t 1 t t t t 1 t 1 t t 1 t 1 t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t t 1 t 1 t t t t t t 1 t t 1 t t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t 2 6 t 1 t t 1 t t t t 1 t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t t 1 t t t t t t t D k , l , e , y ,1; t D k , l , e , y ,1; t y y 1 24 D k t 1 , l t 1 , et 1 , y t 1 ,1; t 6 D k t , l t 1 , et 1 , y t 1 ,1; t 2 D k t 1 , l t , et 1 , y t 1 ,1; t 2 y y y D k t 1 , l t 1 , et 1 , y t ,1; t D k t , l t 1 , et 1 , y t ,1; t D k t 1 , l t , et 1 , y t ,1; t y y y 2 2 2 t 1 t 1 t t 1 t t t 1 t 1 t t 1 t t 1 D k , l , e , y ,1; t D k , l , e , y ,1; t D k , l , e , y ,1; t y y y t 1 t 1 t t t t t 1 t t t 1 t t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t 2 6 Dy k t 1 , l t , et , y t 1 ,1; t Dy k t , l t , et , y t 1 ,1; t t 1 t t t t t t t Dy k , l , e , y ,1; t Dy k , l , e , y ,1; t : KC t LC t EC t CM t . 121 APPENDIX E Let n be an index for observation, n 1, , N . At each time period t 1, , T , the production set Kn , Ln , En , Yn , Cn ; t is: S t K , L , E , Y , C ; t : n 1 znt Ynt Y N (A4.2) N z Cnt C t n 1 n z K nt K N t n 1 n z Ltn L N t n 1 n z Ent E N t n 1 n znt 0, n 1, , N, where znt ‟s are intensity variables, denoting the weights assigned to each observation. Superscript t ‟s represent the time period of production technology and superscript ‟s denote the time period of production activity. The activity analysis model demonstrates all technology properties of the Shephard output distance function described in section 2. Nonnegativeness of intensity variables indicates that production technology exhibits CRS. Inequality constraints imply freely disposable properties, whereas joint constraints on good and bad outputs construct weakly disposable property. By imposing the condition N n 1 Cnt 0 , the good and bad outputs are null-jointness. 122 APPENDIX F We conduct a regression with pooled cross-sectional IWGE and WGE 1991-95 (145 observations) and use the result to predict WGE over the time periods 1996-2006. The predicting equation is estimated: (A4.3) WGEn 142.368 1.126 IWGEn n , 72.161 0.019 where n denotes the provincial-level observations. There is convincing evidence that the mean of WGE is statistically associated with IWGE with a two-sided p-value <<0.0001. The adjusted R 2 from this OLS regression is 0.961. To obtain a conversion factor from WGE to CO2 emissions, we conduct an OLS regression on the national-level CO2 emissions with a regressor of national-level WGE. The following equation is estimated: (A4.4) Ct 7.118WGE1991:1997,t 5.241WGE1998:2004,t t , 0.247 0.111 where WGE1991:1997 is aggregate annual WGE for China if t 1991,1997 and 0 otherwise; WGE1998:2004 is aggregate annual WGE for China if t 1998, 2004 and 0 otherwise. We group the sample into two categories as the restructuring of China‟s coal sector in the late 1990s resulted in the closure of thousands of small-scale mines and power plants. There is overwhelming evidence that the mean of CO2 emissions is associated with WGE statistically (two-sided p-value << 0.0001). This model explains 93.5% of variations in national CO2 emissions after adjusting to degrees of freedom. 123 CHAPTER 5 CONCLUSIONS MAN LI 124 The first essay conducted a national-scale analysis to identify the major drivers of land use conversions and to assess the relative importance of these drivers in China. The analysis compiled a unique, national-scale dataset that includes high-quality land use data, and developed two set of multinomial logit models to analyze land-use choice among six major uses for two time intervals of 1988-1995 and 1995-2000. The spatial analysis provides convincing evidence for the existence of spatial dependence. Results indicate that both socioeconomic and geophysical variables affected land-use change in China. GDP growth and agricultural investment had larger impacts on farmland conversion, while population increase and agricultural investment were more influential in grassland loss. The model is used to evaluate the effectiveness of public agricultural investment as a policy tool for farmland protection in traditional agricultural regions, including the Huang-Huai-Hai Plain, Yangtze River Delta, and Sichuan Basin. The results indicate it is more cost-effective for agricultural investment to larger counties with lower farmland coverage (<50%) when the objective is to reduce farmland conversion. The second essay estimated the association between land-use change and SOC density in China and evaluated the welfare effects of the green growth policy that targets resource reservation and environment protection in conjunct with soil carbon sequestration. A statistical SOC density model was developed, which captures spatial autocorrelation and spatial heterogeneity simultaneously and is easily applied to a large region. The spatial analysis provides strong evidence for the existence of spatial nonstationarity in the parameter estimates and spatial autocorrelation in error term. Results indicate that the SOC density is higher in forest and grass lands and lower in unused land in China. GDP growth leads to farmland and grassland loss, deforestation, and idleness, which accelerates soil carbon runoff. Implementation of the green growth policy could generate up to 0.7-1.1 million Mg SOC and 22.2-37.4 million CNY welfare losses annually throughout the simulation period (2001-2050). The marginal welfare loss is approximately ¥15.3/Mg (equivalent to $2.25/Mg) for sequestering about 1 million Mg SOC per year. The results reveal that soil-based carbon sequestration merits consideration. 125 The third essay presented an alternative approach to index decomposition method for disaggregating change of CO2 emissions in the context of joint production settings. Without specifying any functional form for heterogeneous production functions across regions, it can isolate and quantify the effects of changes in GDP scale, production technology, technical efficiency, capital-carbon ratio, labor-carbon ratio, energy-carbon ratio, and GDP-composition by sector. The proposed method was applied to decomposing province-level changes in CO2 emissions of China into seven components for the time period 1991-2006. Results indicate that the GDP scale effect was the main driver behind rising emissions, whereas the emission index associated with capital, technical change effect, and GDP composition effect were joint contributors to emission abatement. Capital accumulation played a dominant role across the country. The results offer informative implications for policymakers to reduce CO2 emissions in China. The interior regions of China have great potential of technical efficiency improvement by introducing advanced technology and learning management experiences from coastal provinces. As the key driver of China‟s overall economic growth, capital accumulation will still be the most important contributor to abatement across the country in the foreseeable future. But in the long run, China should rely more on technical progress to reduce CO2 emissions. 126 BIBLIOGRAPHY Adams, R.M., Adams, D.M., Callaway, J.M., Chang, C.-c., McCarl, B.A., 1993. Sequestering carbon on agricultural land: social cost and impacts on timber markets. Contemporary Policy Issues 11, 76-87. Alig, R., Adams, D.M., McCarl, B.A., Callaway, J.M., Winnett, S., 1997. Assessing effects of mitigation strategies for global climate change with an intertemporal model of the U.S. forest and agriculture sectors. Environmental and Resource Economics 9, 259-274. Ang, B.W., 1999. Is the energy intensity a less useful indicator than the carbon factor in the study of climate change. Energy Policy 27, 943-946. Ang, B. W., F. L. Liu, and H.-S. Chung. 2004. A generalized Fisher index approach to energy decomposition analysis. Energy Economics 26:757-763. Ang, B.W., Pandiyan, G., 1997. Decomposition of energy-induced CO2 emissions in manufacturing. Energy Econ. 19, 363-374. Ang, B.W., Zhang, F.Q., 2000. A survey of index decomposition analysis in energy and environmental studies. Energy 25, 1149-1176. Anselin, L., editor. 2006. Spatial econometrics. Palgrave Macmillan, Basingstoke. Antle, J.M., Capalbo, S.M., Paustian, K., Ali, M.K., 2007. Estimating the economic potential for agricultural soil carbon sequestration in the Central United States using an aggregate econometric-process simulation model. Climatic Change 80, 145-171. Antle, J.M., Diagana, B., 2003. Creating incentives for the adoption of sustainable agricultural practices in developing countries: the role of soil carbon sequestration. American Journal of Agricultural Economics 85, 1178-1184. Auffhammer, M., Carson, R.T., 2008. Forecasting the path of China's CO2 emissions using province-level information. J. Environ. Econ. Manage. 55, 229-247. Burridge, P., 1980. On the Cliff-Ord test for spatial correlation. Journal of the Royal Statistical Society. Series B (Methodological) 42, 107-108. Capalbo, S.M., Antle, J.M., Mooney, S., Paustian, K., 2004. Sensitivity of carbon sequestration costs to economic and biological uncertainties. Environmental Management 33, S238-S251. 127 Carrión-Flores, C., and E. G. Irwin. 2004. Determinants of residential land-use conversion and sprawl at the rural-urban fringe. American Journal of Agricultural Economics 86:889-904. Cheng, S., and J. S. Long. 2007. Testing for IIA in the multinomial logit model. Sociological methods and Research 35:583-600. Chinese Academy of Social Sciences. 2009. “ 中国城镇人口 2008 年末 6.07 亿.” http://www.cpirc.org.cn/news/rkxw_gn_detail.asp?id=10684 (accessed December 1, 2009). Chow, G.C., 2002. China's Economic Transformation. Blackwell, Oxford. Deng, X., J. Huang, S. Rozelle, and E. Uchida. 2006. Cultivated land conversion and potential agricultural productivity in China. Land Use Policy 23:372-384. Deng, X., J. Huang, S. Rozelle, and E. Uchida. 2008a. Growth, population and industrialization, and urban land expansion of China. Journal of Urban Economics 63:96-115. Deng, X., J. Liu, D. Zhuang, J. Zhan, and T. Zhao. 2002. Modeling the relationship of land use change and some geophysical indicators for the interlock area of farming and pasturing in China. Journal of Geographical Sciences 12:397-404. Deng, X., H. Su, and J. Zhan. 2008b. Integration of multiple data sources to simulate the dynamics of land systems. Sensors 8:620-634. Dubin, R.A., 1998. Spatial autocorrelation: A primer. Journal of Housing Economics 7, 304-327. Fang, J., Chen, A., Peng, C., Zhao, S., Ci, L., 2001. Changes in forest biomass carbon storage in China between 1949 and 1998. Science 292, 2320-2322. Färe, R., Grosskopf, S., Lovell, C.A.K., 1989. Multilateral productivity comparisons when some outputs are undesirable: A nonparametric approach. Rev. Econ. Statist. 71, 90-98. Färe, R., Grosskopf, S., Norris, M., Zhang, Z, 1994. Productivity growth, technical progress, and efficiency change in industrialized countries. Amer. Econ. Rev. 84, 66-83. Färe, R., Grosskopf, S., Pasurka Jr., C.A., 2001. Accounting for air pollution emissions in measures of state manufacturing productivity growth. J. Reg. Sci. 41, 381-409. 128 Färe, R., Grosskopf, S., Pasurka Jr., C.A., 2007. Pollution abatement activities and traditional productivity. Ecological Economics 62, 673-682. Feng, H., Kurkalova, L.A., Kling, C.L., Gassman, P.W., 2006. Environmental conservation in agriculture: land retirement vs. changing practices on working land. Journal of Environmental Economics and Management 52, 600-614. Fotheringham, M.E.C., Brunsdon, C., 1998. Geographically weighted regression: a natural evolution of the expansion method for spatial data analysis. Environment and Planning A 30, 1905-1927. Grossman, G.M., Krueger, A.B., 1993. Environmental Impacts of a North American Free Trade Agreement, in: Garber, P.M. (Eds), The U.S.-Mexico Free Trade Agreement. MIT Press, Cambridge, MA, pp. Hall, P., editor. 1966. Von Thünen's Isolated State: An English edition of Der Isolierte Staat by Johann Heinrich Von Thünen. First edition. Pergamon Press. Hall, R.E., Jones, C.I., 1999. Why do some countries produce so much more output per worker than others? Quart. J. Econ. 114, 83-116. Han, X., Chatterjee, L., 1997. Impacts of growth and structural change on CO2 emissions of developing countries. World Devel. 25, 395-407. Hartkamp, A. D., K. D. Beurs, A. Stein, and J. W. White. 1999. Interpolation techniques for climate variables. CIMMYT, Mexico, D.F. Henderson, D.J., Tochkov, K., Badunenko, O., 2007. A drive up the capital coast? Contributions to post-reform growth across Chinese provinces. J. Macroecon. 29, 569-594. Huang, Y., Sun, W., 2006. Changes in topsoil organic carbon of croplands in mainland China over the last two decades. Chinese Science Bulletin 51, 1785-1803. Jobbágy, E.G., Jackson, R.B., 2000. The vertical distribution of soil organic carbon and its relation to climate and vegetation. Ecological Applications 10, 423-436. Kelejian, H.H., Prucha, I.R., 1999. A generalized moments estimator for the autoregressive parameter in a spatial model. Internatinoal Economic Review 40, 509-533. Kelejian, H. H., and I. R. Prucha. 2001. On the asymptotic distribution of the Moran I test statistic with applications. Journal of Econometrics 104:219-257. 129 Klier, T., and D. P. McMillen. 2008. Clustering of auto supplier plants in the United states: generalized method of moments spatial logit for large samples. Journal of Business and Economic Statistics 26:460-471. Kravchenoko, A., Bullock, D.G., 1999. A comparative study of interpolation methods for mapping soil properties. Agronomy Journal 91, 393-400. Lewis, D. J., and A. J. Plantinga. 2007. Policies for habitat fragmentation: combining econometrics with GIS-based landscape simulations. Land Economics 83:109-127. Li, M., J. Wu, and X. Deng. 2010. Identifying drivers of land-use change in China: A spatial multinomial logit analysis. Working Paper, Oregon State University, OR. Liu, J., M. Liu, D. Zhuang, Z. Zhang, and X. Deng. 2003. Study on spatial pattern of land-use change in China during 1995-2000. Science in China (Series D) 46:373384. Long, H., G. K. Heilig, X. Li, and M. Zhang. 2007. Socio-economic development and land-use change: analysis of rural housing land transition in the transect of Yangtse River, China. Land Use Policy 24:141-153. Long, H., X. Wu, W. Wang, and G. Dong. 2008. Analysis of urban-rural land-use change during 1995-2006 and its policy dimensional driving forces in Chongqing, China. Sensors 8:681-699. Lubowski, R. N., A. J. Plantinga, and R. N. Stavins. 2006. Land-use change and carbon sinks: econometric estimation of the carbon sequestration supply function. Journal of Environmental Economics and Management 51:135-152. Mansfield, C., Phaneuf, D.J., Johnson, R., Yang, J.-C., Beach, R., 2008. Preferences for public lands management under competing uses: the case of Yellowstone National Park. Land Economics 84, 282-305. Marland, G., Boden, T., Andres, R.J., 2009. National CO2 emissions from fossil-fuel burning, cement manufacture, and gas flaring: 1751-2006, in: Trends online: A compendium of data on global change. Carbon Dioxide Information Analysis Center, O.R.N.L., U.S. Department of Energy, Oak Ridge Tennessee. National Bureau of Statistics of China, 2006-07. China Statistical Yearbook. China Statistical Press, Beijing. National Bureau of Statistics of China, 2005. Comprehensive Statistical Data and Materials on 55 Years of New China. China Statistical Press, Beijing. 130 National Bureau of Statistics of China and National Development and Reform Commission, 1997-99, 2005-07. China Energy Statistical Yearbook. China Statistical Press, Beijing. Nelson, G. C., V. Harris, and S. W. Stone. 2001. Deforestation, land use, and property rights: empirical evidence from Darién, Panama. Land Economics 77:187-205. Nelson, G.C., Hellerstein, D., 1997. Do roads cause deforestation? Using satellite images in econometric analysis of land use. American Journal of Agricultural Economics 79, 80-88. Organization for Economic Co-operation and Development (OECD), 2010. Green Growth Strategy Interim Report: Implementing our Commitment for a Sustainable Future. Meeting of the OECD Council at Ministerial Level, in: Paris. Ostwald, M., and D. Chen. 2006. Land-use change: impacts of climate variations and policies among small-scale farmers in the Loess Plateau, China. Land Use Policy 23:361-371. Parks, P.J., hardie, I.W., 1995. Least-cost forest carbon reserves: cost-effective subsidies to convert marginal agricultural land to forests. Land Economics 71, 122-136. Parton, W.J., Scurlock, J.M.O., Ojima, D.S., Gilmanov, T.G., Scholes, R.J., Schimel, D.S., Kirchner, T., Menaut, J.-C., Seastedt, T., Moya, E.G., Kamnalrut, A., J.I., K., 1993. Observations and modeling of biomass and soil organic matter dynamics for the grassland biome worldwide. Global Biogeochemical Cycles 7, 785-809. Pasurka, C.A., 2006. Decomposing electric power plant emissions within a joint production framework. Energy Econ. 28, 26-43. Pausch, G.R., Kurkalova, L.A., Babcock, B.A., Kling, C.L., 2001. The efficiency of sequestering carbon in agricultural soils. Contemporary Economic Policy 19, 123134. Phaneuf, D.J., Smith, K., Palmquist, R.B., Pope, J.C., 2008. Integrating property value and local recreation models to value ecosystem services in urban watersheds. Land Economics 84, 361-381. Pinkse, J., and M. E. Slade. 1998. Contracting in space: an application of spatial statistics to discrete-choice models. Journal of Econometrics 85:125-154. Plantinga, A.J., Mauldin, T., Miller, D.J., 1999. An econometric analysis of the costs of sequestering carbon in forests. American Journal of Agricultural Economy 81, 812-824. 131 Polyakov, M., and D. Zhang. 2008. Property tax policy and land-use change. Land Economics 84:396-408. Richards, K.R., Stokes, C., 2004. A review of forest carbon sequestration cost studies: a dozen years of research. Climatic Change 63, 1-48. Schlesinger, W.H., 1977. Carbon balance in terrestrial detritus. Annual Review of Ecology and Systematics 8, 51-81. Seto, K. C., and R. K. Kaufmann. 2003. Modeling the drivers of urban land use change in the Pearl River Delta, China: intergating remote sensing with socioeconomic data. Land Economics 79:106-121. Siegel, I. H. 1945. The generalized "ideal" index-number formula. Journal of the American Statistical Association 40:520-523. Small, K.A., Rosen, H.S., 1981. Applied welfare economics with discrete choice models. Econometrica 49, 105-130. State Environmental Protection Administration of China (SEPA), National Bureau of Statistics of China (NBS), 2006. China green national accounting study report 2004. Chinese Academy for Environmental Planning, in: Beijing. Stavins, R.N., 1999. The costs of carbon sequestration: a revealed-preference approach. The American Economic Review 89, 994-1009. Sun, J.W., 1999. Decomposition of Aggregate CO2 Emissions in the OECD: 1960-1995. The Energy Journal 20, 147-155. Tang, H., Qiu, J., Ranst, E.V., Li, C., 2006. Estimations of soil organic carbon storage in cropland of China based on DNDC model. Geoderma 134, 200-206. Viguier, L., 1999. Emissions of SO2, NOX and CO2 in transition economics: emissions inventories and Divisia index analysis. The Energy Journal 20, 59-87. Wang, C., 2007. Decomposing energy productivity change: A distance function approach. Energy 32, 1326-1333. Wang, C., Jining, C., Zhou, J., 2005. Decomposition of energy-related CO2 emissions in China: 1957-2000. Energy 30, 73-83. Wang, S., Huang, M., Shao, X., Mickler, R.A., Li, K., Ji, J., 2004. Vertical distribution of soil organic carbon in China. Environmental Management 33, S200-S209. 132 Wang, S., Tian, H., Liu, J., Pan, S., 2003. Pattern and change of soil organic carbon storage in China: 1960s-1980s. Tellus 55B, 416-427. Wu, H., Guo, Z., Peng, C., 2003. Distribution and storage of soil organic carbon in China. Global Biogeochemical Cycles 17, 1-11. Wu, L., Kaneko, S., Matsuoka, S., 2005. Driving forces behind the stagnancy of China's energy-related CO2 emissions from 1996 to 1999: the relative importance of structural change, intensity change and scale change. Energy Policy 33, 319-335. Wu, J., R. M. Adams, C. L. Kling, and K. Tanaka. 2004. From microlevel decisions to landscape changes: an assessment of agricultural conservation policies. American Journal of Agricultural Economics 86:26-41. Wu, J., and S.-H. Cho. 2007. The effect of local land use regulations on urban development in the Western United States. Reginal Science and Urban Economics 37:69-86. Yang, Y., Mohammat, A., Feng, J., Zhou, R., Fang, J., 2007. Storage, patterns and environmental controls of soil organic carbon in China. Biogeochemistry 84, 131141. Zhang, F., Li, C., Wang, Z., Wu, H., 2006. Modeling impacts of management alternatives on soil carbon storage of farmland in Northwest China. Biogeosciences 3, 451466. Zhang, Y., and S. Zhao. 2003. Analysis of land use change in Horqin Sandy land and its surrounding area during the past 15 years (In Chinese). Journal of Natural Resources 18:174-181. Zhou, P., Ang, B.W., 2008. Decomposition of aggregate CO2 emissions: A productiontheoretical approach. Energy Econ. 30, 1054-1067.