UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Essays on China's international trade Li, Bingjing 2016

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


24-ubc_2016_september_li_bingjing.pdf [ 6.21MB ]
JSON: 24-1.0307177.json
JSON-LD: 24-1.0307177-ld.json
RDF/XML (Pretty): 24-1.0307177-rdf.xml
RDF/JSON: 24-1.0307177-rdf.json
Turtle: 24-1.0307177-turtle.txt
N-Triples: 24-1.0307177-rdf-ntriples.txt
Original Record: 24-1.0307177-source.json
Full Text

Full Text

Essays on China’s International TradebyBingjing LiB.Soc.Sc., The Chinese University of Hong Kong, 2009M.Phil., The Chinese University of Hong Kong, 2011A THESIS SUBMITTED IN PARTIAL FULFILLMENT OFTHE REQUIREMENTS FOR THE DEGREE OFDOCTOR OF PHILOSOPHYinThe Faculty of Graduate and Postdoctoral Studies(Economics)THE UNIVERSITY OF BRITISH COLUMBIA(Vancouver)July 2016c© Bingjing Li 2016AbstractThis dissertation studies the impact of international trade on China along different dimen-sions.The first chapter investigates how the decline in trade barriers and the resulting exportexpansion affected human capital accumulation in China. Following a theoretically consis-tent approach, I exploit variations in regional exposures to high- and low-skill export demandshocks, which stems from the diverse initial industry composition across prefectures, and dif-ferential skill intensities across industries. I find that high-skill export shocks raise both highschool and college enrollments, while low-skill export shocks depress both. The amplifieddifferences in skill abundance across prefectures reinforce the initial industry specialization.These findings suggest a mutually reinforcing relationship between regional comparative ad-vantage and skill formation.The second chapter examines the impacts of export expansion on air pollution and healthoutcomes in China. To disentangle trade-induced scale and composition effects from tech-nique effect, I construct two export shocks at the prefecture level: (i) PollutionExportShockrepresents the pollution content of export expansion measured in pounds of pollutants perworker; (ii) ExportShock measures the export exposure in dollars per worker. The two mea-sures differ because prefectures specialize in different products: while two prefectures mayexperience the same shock in dollar terms, the one specializing in dirty industries has a largerPollutionExportShock. I find that a higher PollutionExportShock increases both pollution andmortality. A higher ExportShock tends to reduce pollution and mortality, but the effect is notalways statistically significant. I also provide evidence that export expansion affects mortalitythrough the channel of air pollution.The third chapter provides evidence that over-export of grains aggravated China’s GreatFamine. I exploit cross-county variation in grain export exposure which stems from theirdifferences in suitability of cultivating different crops that experienced variable export shocksin early famine years. I find that a county’s suitability in high-export-exposure crops is posi-tively associated with its famine severity. However, the correlation is statistically insignificantfor low-export-exposure crops. Moreover, for high-export-exposure crops, the correlation be-tween suitability and famine severity declines with distance to railroad. These patterns areconsistent with the possibility that excessive grain exports severely reduced food availabilityfor domestic consumption.iiPrefaceChapter 2 Trade, Pollution and Mortality in China of this dissertation is a joint work withProfessor Matilde Bombardini. I was involved throughout each stage of the research: collect-ing and preparing data, designing empirical method, carrying out estimation, organizing andpresenting results, and writing several subsections of the manuscript.iiiTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi1 Export Expansion, Skill Acquisition and Industry Specialization: Evidencefrom China . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Related Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41.3 A Model with Endogenous Skill Supply . . . . . . . . . . . . . . . . . . . . . 51.3.1 Education Decision . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61.3.2 Production and Trade . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.3.3 Preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.3.4 Trade Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81.3.5 Steady State Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . 81.4 Export Demand Shocks to Regional Markets . . . . . . . . . . . . . . . . . . 91.4.1 Reduced Form Relation . . . . . . . . . . . . . . . . . . . . . . . . . . 91.4.2 Export Demand Shocks: From National to Local . . . . . . . . . . . . 101.5 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.5.1 Local Economies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.5.2 Population Censuses . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.5.3 Trade and Tariff Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 121.5.4 Regional Export Demand Shocks . . . . . . . . . . . . . . . . . . . . 121.5.5 Industry Output and Other Socioeconomic Data . . . . . . . . . . . . 131.6 Effects of Export Demand Shocks on School Enrollment . . . . . . . . . . . . 131.6.1 Baseline Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141.6.2 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151.7 Effect of Change in Skill Supply on Industry Specialization . . . . . . . . . . 181.8 Counterfactual Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201.8.1 Counterfactual Proportional Changes . . . . . . . . . . . . . . . . . . 201.8.2 Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21iv1.8.3 Eliminating External Trade . . . . . . . . . . . . . . . . . . . . . . . . 221.8.4 Further Trade Liberalization . . . . . . . . . . . . . . . . . . . . . . . 231.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242 Trade, Pollution and Mortality in China . . . . . . . . . . . . . . . . . . . . 352.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352.1.1 Relation to the Literature . . . . . . . . . . . . . . . . . . . . . . . . 372.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392.2.1 Local Economies and Employment Data . . . . . . . . . . . . . . . . 392.2.2 Export and Tariff Data . . . . . . . . . . . . . . . . . . . . . . . . . . 402.2.3 Pollution Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402.2.4 Mortality Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412.2.5 Other Demographic, Socioeconomic and Wind Data . . . . . . . . . . 422.2.6 Quality Assessment of the Chinese Data Pollution and Mortality . . . 422.3 Preliminary Event Study: the 2002 US Steel Safeguard Measures . . . . . . . 432.4 Theoretical framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442.4.1 Changes in Transport Costs: Deriving Export Demand Shocks . . . . 462.5 Empirical Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462.5.1 Pollution Export Shocks and Export Shocks . . . . . . . . . . . . . . 472.5.2 Specification 1: Total Effect of Export Shocks on Mortality . . . . . . 482.5.3 Identification Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . 492.5.4 Specification 2: Pollution Concentration Channel . . . . . . . . . . . 512.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512.6.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512.6.2 Results for Specification 1: Total Effect of Export Shocks on Mortality 522.6.3 Results for Specification 2: Pollution Concentration Channel . . . . . 532.6.4 Effects of Pollution on Infant Mortality by Cause of Death . . . . . . 542.7 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552.7.1 Robustness: Neighboring Shocks and Wind Direction . . . . . . . . . 552.7.2 Robustness: Future shocks . . . . . . . . . . . . . . . . . . . . . . . . 552.7.3 Robustness: Mortality by Gender . . . . . . . . . . . . . . . . . . . . 562.7.4 Robustness: Mortality of Young Children Aged 1-4 . . . . . . . . . . 562.7.5 Energy Production . . . . . . . . . . . . . . . . . . . . . . . . . . . . 562.7.6 Imports . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 572.7.7 Other Controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 572.7.8 Alternative Measures of External Demand Shocks . . . . . . . . . . . 582.7.9 Change in Infant Mortality Rate and Export Expansion by IndustryGroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582.8 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593 Grain Exports and China’s Great Famine . . . . . . . . . . . . . . . . . . . 753.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 753.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763.2.1 Rural Institutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763.2.2 The 1959-1961 Great Famine . . . . . . . . . . . . . . . . . . . . . . . 773.2.3 The Role of International Trade . . . . . . . . . . . . . . . . . . . . . 78v3.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793.3.1 Population Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793.3.2 Historical Trade Data on Agricultural Products . . . . . . . . . . . . 803.3.3 Crop Productivity Data . . . . . . . . . . . . . . . . . . . . . . . . . . 813.3.4 Data on Agricultural Production and Export Exposure . . . . . . . . 823.3.5 Distance to Railroad Data . . . . . . . . . . . . . . . . . . . . . . . . 833.3.6 Historical Weather Data . . . . . . . . . . . . . . . . . . . . . . . . . 833.4 Empirical Evidence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 843.4.1 Famine Severity and Crop Suitability . . . . . . . . . . . . . . . . . . 843.4.2 Famine Severity, Crop Suitability and Distance to Railroad . . . . . . 863.4.3 Famine Severity and Crop Specialization . . . . . . . . . . . . . . . . 873.4.4 Famine Severity and Export Shock . . . . . . . . . . . . . . . . . . . 883.4.5 Robustness: Pre-Famine, Famine and Post-Famine Years . . . . . . . 893.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104AppendicesA Appendix for Chapter 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111A.1 Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111A.1.1 Share of Educated Workers . . . . . . . . . . . . . . . . . . . . . . . . 111A.1.2 Expected Productivity of Educated and Uneducated Workers . . . . 111A.1.3 Prices and Trade Flows . . . . . . . . . . . . . . . . . . . . . . . . . . 112A.2 Reduced Form Relation between Export Shocks and School Enrollment . . . 113A.2.1 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113A.2.2 Export Demand Shocks: From National to Local . . . . . . . . . . . . 115A.3 Data Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116A.3.1 Administration Division and Industrial Classifications . . . . . . . . . 116A.3.2 Trade Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117A.3.3 Employment and Output Data at 2-digit CSIC codes . . . . . . . . . 117A.4 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120A.4.1 Educational Provision and Export Shocks . . . . . . . . . . . . . . . . 120A.4.2 Migration Pattern and Export Shocks . . . . . . . . . . . . . . . . . . 120A.4.3 Effects on Export Demand Shocks on School Enrollment of Locals . . 121A.4.4 Heterogeneous Reponses to Export Shocks of Different DemographicGroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122A.4.5 Dropping One Province/Two Provinces at a Time . . . . . . . . . . . 122A.4.6 Children under Compulsory Schooling . . . . . . . . . . . . . . . . . . 122A.4.7 Other Market Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . 123A.4.8 Disaggregated Export Demand Shocks . . . . . . . . . . . . . . . . . 123A.4.9 Change in Skill Supply and Industry Specialization: Different Measures 124A.5 Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132viB Appendix for Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134B.1 Data Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134B.1.1 Administration Division: Consistent Prefectures . . . . . . . . . . . . 134B.1.2 Industrial Classifications . . . . . . . . . . . . . . . . . . . . . . . . . 134B.1.3 Prefecture Level Data on Wind Direction . . . . . . . . . . . . . . . 134B.1.4 Employment Weighted and Wind Direction Weighted Neighboring Ex-port Pollution Shocks . . . . . . . . . . . . . . . . . . . . . . . . . . . 135B.1.5 Input-Output Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . 136B.1.6 Data Quality of Air Pollution: Comparison of Official Data and USEmbassy Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136B.1.7 Data Quality of IMR: Comparison of Cohort Size across Censuses . . 137B.2 Additional Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137B.2.1 Robustness Checks: Effect of Export Shocks on Changes in PM2.5Concentration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137C Appendix for Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147C.1 Additional Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . 147C.2 Data Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151C.2.1 Yearly County Population Size . . . . . . . . . . . . . . . . . . . . . . 151C.2.2 Historical Trade Data from Different Sources . . . . . . . . . . . . . . 151viiList of Tables1.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291.2 Changes in School Enrollment and Export Shocks . . . . . . . . . . . . . . . 301.3 Changes in School Enrollment and Export Shocks: Controlling the ExportShocks of Other Prefectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311.4 Changes in School Enrollment, Export Shocks and Import Shocks . . . . . . . 321.5 Changes in Skill Supply and Change in Industry Specialization . . . . . . . . 331.6 Counterfactual: Eliminating External Trade . . . . . . . . . . . . . . . . . . . 341.7 Counterfactual: Further Trade Liberalization . . . . . . . . . . . . . . . . . . 342.1 Effect of 2001-03 Steel Safeguards on API . . . . . . . . . . . . . . . . . . . . 642.2 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 652.3 Change in Infant Mortality Rate and Shocks: 2SLS . . . . . . . . . . . . . . . 662.4 Changes in Pollutant Concentration and Shocks: 2SLS . . . . . . . . . . . . . 672.5 Changes in Infant Mortality Rate and Changes in Pollutant Concentration:2SLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 682.6 Elasticity of IMR to pollutant concentration in other studies . . . . . . . . . 682.7 Changes in Infant Mortality Rate and Shocks by Causes of Death: SO2, 2SLS 692.8 Change in Infant Mortality Rate and Shocks: Neighboring Shocks,2SLS . . . 702.9 Change in Infant Mortality Rate and Future Shocks . . . . . . . . . . . . . . 712.10 Change in Infant Mortality Rate and Shocks by Gender: 2SLS . . . . . . . . 712.11 Change in Mortality Rate of Children aged 1-4 and Shocks: 2SLS . . . . . . . 722.12 Change in Infant Mortality Rate and Shocks: SO2 Robustness . . . . . . . . 732.13 Change in Infant Mortality Rate and Export Expansions by Industry Groups 743.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 993.2 Famine Severity, Crop Suitability and Distance to Railroad . . . . . . . . . . 1003.3 Famine Severity and Production Sepcialization . . . . . . . . . . . . . . . . . 1013.4 Famine Severity and Export Exposure . . . . . . . . . . . . . . . . . . . . . . 1013.5 Robustness: The Effect of Export Exposure on Population Size: Pre-Famine,Famine and Post-Famine Years . . . . . . . . . . . . . . . . . . . . . . . . . . 1023.6 Output, Procurement, Export and Export Price by Crop . . . . . . . . . . . . 103A.1 Skill Intensities of 2-digit Manufacturing Industries . . . . . . . . . . . . . . . 119A.2 Educational Provision and Export Export Exposure . . . . . . . . . . . . . . 125A.3 Migration and Export Shocks . . . . . . . . . . . . . . . . . . . . . . . . . . . 125A.4 Changes in Share of Immigrants, Change in School Enrollment and ExportShocks: Different Samples by Migration Status . . . . . . . . . . . . . . . . . 126A.5 Changes in School Enrollment and Export Shocks: by Samples . . . . . . . . 127viiiA.6 Dropping One Province/Two Provinces at a Time . . . . . . . . . . . . . . . 128A.7 Change of School Enrollment and Export Shocks: Ages under CompulsorySchooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128A.8 Changes of Market Employment, Home Production, Unemployment Rate andExport Shocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129A.9 Changes in School Enrollment, Export Shocks and Import Shocks: Disaggre-gated Education Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130A.10 Changes in Skill Supply and Change in Industry Specialization: Different Mea-sures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131A.11 Calibration and Data Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . 133A.12 Calibrated αk and βk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133B.1 Correlation of pollution data of US embassy and MEP (AQI) . . . . . . . . . 141B.2 Concentration of PM2.5 (PM10) from different sources (mg/m3, 2013) . . . . 141B.3 Changes in Infant Mortality Rate and Shocks by Causes of Death: TSP, 2SLS 142B.4 Changes in Infant Mortality Rate and Shocks by Causes of Death: NO2, 2SLS 143B.5 Change in Infant Mortality Rate and Shocks: TSP, Robustness . . . . . . . . 144B.6 Change in Infant Mortality Rate and Shocks: NO2, Robustness . . . . . . . . 145B.7 Change in PM2.5 Concentration and Shocks: Robustness . . . . . . . . . . . 146C.1 Famine Severity and Weather Conditions . . . . . . . . . . . . . . . . . . . . 149C.2 First Stage Results of IV Regressions . . . . . . . . . . . . . . . . . . . . . . . 150ixList of Figures1.1 Spatial Distribution of Export Demand Shocks: 2000-05 . . . . . . . . . . . . 251.2 School Enrollment by Ages over Years . . . . . . . . . . . . . . . . . . . . . . 261.3 School Enrollment across Prefectures of Different Periods . . . . . . . . . . . 261.4 Change in Educational Attainment and Export Shocks: Different Age Groups 271.5 Divergence in Education Attainment . . . . . . . . . . . . . . . . . . . . . . . 282.1 Mechanism relating Export Shocks to Mortality . . . . . . . . . . . . . . . . . 622.2 The relationship between Log(Exports) and Log(ExportTariff) . . . . . . . . 622.3 Distribution of Export Pollution Shocks over Decades, SO2 . . . . . . . . . . 633.1 Export and Import (1955-1961) . . . . . . . . . . . . . . . . . . . . . . . . . . 913.2 Composition of Grain Exports (1955-1961) . . . . . . . . . . . . . . . . . . . 923.3 Average Exports over 1958-1960 versus Average Exports over 1955-1957 . . . 923.4 Spatial Distribution of Potential Yields by Crops . . . . . . . . . . . . . . . . 933.5 Railway Network in China over Time . . . . . . . . . . . . . . . . . . . . . . . 943.6 Relative Pirce over 1955-1961 . . . . . . . . . . . . . . . . . . . . . . . . . . . 953.7 Kernel Densities of Precipitation and Temperature in Spring and Summer:1959-1961 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 963.8 The Correlation between Crop Suitabilities and Birth-Cohort Size over Time 973.9 The Correlation between Crop Suitabilities and Birth-Cohort Size over Timeby Distance Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98A.1 Regions in China . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132B.1 Timeline of Steel Safeguard . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139B.2 Steel Export and Production Share over 2000-04 . . . . . . . . . . . . . . . . 139B.3 Predicted and Actual Population Size of Age 10 in 2010 . . . . . . . . . . . . 140B.4 Wind Direction Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140C.1 Survivor Cohort Size — Mean and Coefficient of Variation across Counties . . 147C.2 Population Size by Year — Mean and Coefficient of Variation across Counties 147C.3 China’s Historical Railway Network (CIA, 1967) . . . . . . . . . . . . . . . . 148C.4 Trade Flows in 1967 from Different Sources (Top 15 Trading Partners) . . . . 151xAcknowledgementsThe past five years of pursuing my academic career at UBC have been challenging and yetrewarding. I would like to take this opportunity and express my sincere appreciation to allthose standing by me during my Ph.D. study. Without their support and encouragement,the dissertation presented herein would not have been possibleFirst and foremost, I would like to express the deepest gratitude to my supervisor ProfessorMatilde Bombardini for her invaluable guidance and support throughout my Ph.D. program.Her open-mindedness and her enthusiasm to knowledge sharing and teaching help me exploreand discover my research interest. Her sharp sense in economics inspired me to view thingsfrom different perspectives. Her continuous encouragement keeps my chin up whenever I hita wall in research or life. I feel so blessed to have had worked with Matilde, who is not onlya great academic advisor but also a wonderful friend.My sincere thanks also go to other members of my thesis committee. Professor DavidGreen is an extremely helpful mentor. His motivation, insights and constructive advice areessential for the completion of this work. With his guidance and encouragement, I am able tobridge international trade and labor economics in my job market paper. I am also indebtedto Professor Tomasz Swiecki. His help is indispensible to each stage of my dissertation, fromforming ideas to final revision. His thoughtful and detailed comments greatly improve thequality of my work.I am also grateful to other faculty, staffs and my beloved friends at UBC for their support.I would like to thank Vanessa Alviarez, Keith Head, Hiro Kasahara and all other participantsin the UBC Trade Study Group for their feedback and fruitful discussions on my research.I also owe my gratitude to Professor Brian Copeland for his comments and insightful sug-gestions on the second chapter of my dissertation. I am also grateful to Siwan Anderson,Giovanni Gallipoli, Amartya Lahiri, Wei Li, Henry Siu, Francesco Trebbi and Yaniv Yedid-Levi for their suggestions and advises at various stages of my Ph.D. study. I also thankMaureen Chin for all the administrative support throughout the program, especially duringthe job market season.Finally, I owe a tremendous amount of gratitude to my dearest parents and husband. Nowords can express how grateful I am for their unconditional love and endless support.xiChapter 1Export Expansion, Skill Acquisitionand Industry Specialization:Evidence from China1.1 IntroductionThe impact of “Made in China” on labor markets has been studied extensively in recent years.Pierce and Schott (2012), Ebenstein et al. (2012), Autor, Dorn and Hanson (2013), Autor etal. (2014), Utar (2014) and Balsvik, Jensen and Salvanes (2015) find that the manufacturingsector in developed countries saw deteriorating employment conditions because of importcompetition from China. However, little is known about the flip side of the coin, i.e., theconsequences of China’s export expansion on its own labor market.1 Drawing on a rich dataset on sub-national economies, the present paper examines the differential impact of exportexpansion on local markets within China between 1990 and 2005. The paper has two focuses.First, it analyzes the effects of export demand shocks on human capital accumulation ordecumulation in local economies, by exploiting cross-market variation in export exposurestemming from initial differences in industrial specialization. Second, it investigates thedynamics of regional industrial specialization resulting from the trade-induced shifts in skillsupply.Understanding the linkages among trade, regional comparative advantage and humancapital accumulation is important. In recent decades, export-oriented policies have beenassociated with industrialization and economic growth in many developing countries. Nev-ertheless, there are few studies inquiring the long term effects of these export-led growthstrategies. In particular, as is suggested by Findlay and Kierzkowski (1983), trade may ex-acerbate economic differences across countries through its effect on educational attainment,which is thought to be a key ingredient for sustained economic development.China provides an interesting setting in which to study the relations. Over the periodspanned by the main data (1990-2005), China experienced two waves of trade liberalization.The first took place in 1992, when Deng Xiaoping made his famous southern tour to Guang-dong province and reasserted the economic reforming and opening agenda. Favorable tradepolicies were adopted from that time and greatly boosted exports. Exports quadrupled inthe 1990s, from 62 billion USD in 1990 to 248 billion USD in 2000. Accession to the WTOin 2001 further accelerated export expansion. Total exports tripled from 2000 to 2005, from1To the best of my knowledge, Han, Liu and Zhang (2012) is the only paper studying the differential effectsof trade liberalization on different regions within China. Employing the geographic distance to the coast as aproxy for export exposure, they find that China’s accession to the WTO is significantly associated with risingwage inequality and skill premium in high-exposure regions.1248 billion to 759 billion USD. In the meantime, there was a substantial reduction in exportbarriers. The effective export tariff decreased from 8.4% in 1990 to 3.5% in 2005. In addition,different industries experienced various declines in export tariffs, with the standard devia-tion being 7.9%. Several features of China’s economy make the analysis and identification inthis study feasible. The country’s large geographic diversity is associated with considerablevariation in skill endowment and industry composition across regions. Moreover, the house-hold registration system greatly restricts the inter-region migration (Tombe and Zhu, 2014).Because of such labor market frictions, China’s micro regions are effectively like local labormarkets, which allows me to conduct the analysis at the sub-national level.I extend a two-factor, multiple-sector and multiple-region Eaton and Kortum (2002)model by endogenizing skill formation. In the framework, each agent is endowed with two-dimensional productivities, which represent the aptitude at being an educated or uneducatedworker. Based on labor market conditions, agents sort themselves into schools. From themodel, I derive a simple reduced form equation linking changes in school enrollment to ex-ogenous export demand shocks that differentially affect the education decision. The causaleffect of export expansion on skill acquisition is carefully studied using an empirical approachguided by the model. Specifically, during a given period, regions experience different low-skilland high-skill export demand shocks, which stem from the diverse initial industry composi-tion across regions and the various skill intensities of different industries. The shocks alterthe skill premium in local markets and hence the incentives to acquire education.To construct the regional low- and high-skill export shocks in a theoretically consistentway, I proceed as follows. First, to isolate the external export demand shocks from otherfactors that may also be associated with export growth, such as productivity and labor supply,I only employ the component of exports that is predicted by the change in tariffs faced byChinese exporters over time in different sectors. Second, the predicted export expansion isallocated to various regions according to initial sectoral employment shares. Finally, theapportioned exports are attributed to high skilled (low skilled) labor according to sectoralskill intensities and divided by the total amount of high skilled (low skilled) workers. Thehigh-skill (low-skill) export shock can therefore be interpreted as export exposure in dollarsper skilled (unskilled) worker.Next, school enrollments of young people aged 16 to 22 across China’s prefectures arerelated to the regional export shocks at different skill levels during the period 1990 to 2005.A 1000 USD low-skill export shock is estimated to reduce high-school enrollment rate by 3.1percentage points and college enrollment rate by 2.4 percentage points. Conversely, a 1000USD high-skill export shock raises the high-school enrollment rate by 0.8 percentage pointand college enrollment rate by 1 percentage point.I address a series of issues that may contaminate these results. In particular, one mayconcern that low-skill (high-skill) export shock may lower (increase) the average educationalattainment of immigrants and increase (lower) the average educational attainment of out-migrants. To examine the potential confounding effect introduced by selective migration, Istudy how export shocks affect the educational attainment of young workers aged between23-35. I show that export shocks have insignificant effect on educational attainment of theseolder age groups, whose education decisions were made before the shocks occurred.To address the concern that an individual’s education decision could be affected by thelabor market conditions in other prefectures, I include variables of non-local export shocks,2i.e., shocks in neighboring prefectures and inverse distance weighted export shocks elsewherein China. I find that the local labor market condition affects educational choice independentof cross-border spillovers. I also delve on the importance of imports on educational choiceand show that low-skill import shock raises both high school and college enrollments, whilehigh-skill import shock depresses both. The finding that import shocks have opposite effectsto export shocks is reassuring.To study whether the enlarged differences in skill abundance reinforce regional industryspecialization, I link the change in industry employment share from 2000 to 2010 to thechanges in school enrollment rate between 1990 and 2000, and instrument the latter with theexport demand shocks. I find that an increase in college enrollment reduces the employmentshare of low-skill industries and raise the employment share of high-skill industries in thesubsequent decade. The 2SLS estimates suggest that a 10 percentage point rise in collegeenrollment rate during the 1990s increases the employment share of a high-skill industryby 0.87 percentage point, while it lowers the employment share of a low-skill industry by0.78 percentage point in the 2000s. The finding is aligned with the cross-country evidenceof Romalis (2004) that regions saw their production further shift towards sectors that moreintensively employ the fast accumulating factors.These empirical findings suggest a mutually reinforcing relationship between regional com-parative advantage and skill formation. Due to nationwide trade liberalization, regions thatinitially specialized in low-skill industries received relatively larger low-skill export shocks,which deskilled the local labor force. The change in skill supply strengthened their initialcomparative advantage. As a result, these regions became more specialized in the low-skillsector in the subsequent period, making them more prone to low-skill export shocks in thefuture. The converse will be the case for a region initially specialized in high-skill industries.To assess the general equilibrium effects of trade liberalization on human capital accumu-lation in China, I calibrate the model to bilateral trade flows and China’s micro-level data andevaluate the implications of past and future trade liberalization for China. The counterfac-tual analysis finds that trade liberalization in past decades lowered educational attainmentfor most regions in China. In addition, further globalization will lead to further regionaldivergence in educational attainment and industry specialization. A 30 percent reductionin external trade cost is found to reduce school enrollment in the North Coast region (themost negatively affected region) by 8 percents and increase school enrollment in the NorthMunicipality region (the most positively affected region) slightly by 1 percent.The remainder of the paper is organized as follows. Section 1.2 provides a literaturereview. Section 1.3 lays out a trade model with endogenous skill formation. Section 1.4constructs regional export demand shocks following the model. Section 1.5 describes thedata set and summary statistics. Section 1.6 examines the effects of export expansions onschool enrollment. Section 1.7 investigates the evolution of regional industry specializationresulting from trade-induced shifts in skill supply. Section 1.8 calibrates the model and studiesthe quantitative implications of further trade liberalization for China. Section 1.9 concludesthe paper.31.2 Related LiteratureThis paper relates to the literature on the dynamic Heckscher-Ohlin (HO) Model, whichembeds endogenous factor formation into the classic HO framework. Seminal works includeStiglitz (1970), Findlay and Kierzkowski (1983) and Borsook (1987). The model predictsthat trade increases the return to the abundant factor in a country, which induces furtheraccumulation of this factor. Moreover, the amplified differences in factor abundance willin return strengthen a country’s initial comparative advantage and industry specialization.Therefore, a country with abundant unskilled labor will be further deskilled and become morespecialized in labor-intensive sectors after trade liberalization.However, another strand of literature shows that trade can raise the skill premium andhence encourage skill acquisition even in the context of developing countries (Burstein andVogel, 2012; Harris and Robertson, 2013; Danziger, 2014). According to Burstein and Vogel(2012), a reduction in trade costs induces within-sector reallocation of production towardmore productive and skill-intensive firms, which reinforces (counteracts) the HO forces inskill-abundant (skill-scarce) countries. Under certain parameters, they show that trade liber-alization increases skill premium even in developing countries. Harris and Robertson (2013)build a model endogenizing capital and skill formation. In their setting, capital and skilled la-bor are complementary in production and trade liberalization increases the return to capital.Calibrated to data from China and India, the model predicts that the skill premium increasesin response to the tariff removal, which generates substantial accumulation of human capitalin both countries. My findings of a negative impact of low-skill export shock on skill acquisi-tion should not be taken as a rejection of the models emphasizing within-sector reallocationand capital accumulation mechanisms. Instead, the findings in this paper suggest that thebetween-sector Stolper-Samuelson effect appears to be strong.2Relatively few empirical studies attempt to estimate the effects of trade on human capitalformation. Edmonds and Pavcnik (2005) investigate the reduction of child labor associatedwith the rising price of rice when Vietnam removed its sanctions on rice exports. Edmonds,Pavcnik and Topalova (2010) examine the adverse effects of import competition on children’sschooling following India’s tariff reform in the 1990s. Both studies focus on young children3and find dominating income effects, i.e. trade policies alter household income which in turnaffects child labor supply.My paper shares more similarities with the works of Blanchard and Olney (2014) andAtkin (2015), which show that the skill intensity of exports matters. In particular, usingcross-country panel data and a gravity based instrumental variable (IV) technique, Blan-chard and Olney (2014) show that export expansion in the agricultural and low-skill manu-facturing sectors reduces average years of schooling whereas export expansion in the high-skillmanufacturing sector increases it. Employing Mexican micro-level data, Atkin (2015) findsthat the arrival of less-skilled export manufacturing jobs increases school dropouts at age 16.2Models emphasizing within-sector reallocation and capital accumulation have received mixed empiricalsupport in the context of China. Using Chinese manufacturing firm data, Ma, Tang and Zhang (2014) findthat firms become less capital intensive after exporting. To rationalize the findings, they propose a model ofheterogeneous firms producing multiple products with different capital intensities. The model predicts thatin a labor abundant country like China, exporting firms’ capital intensities decline due to product churning,i.e., exporting firms allocate more resources to produce labor-intensive products.3Edmonds and Pavcnik (2005) study the children aged 6–15 and Edmonds, Pavcnik and Topalova (2010)study the children aged 10–14.4This paper attempts to build on these earlier contributions by bringing in several additionalelements. First, I develop a theoretically consistent approach to study the effect of exportexpansions on skill acquisition. Guided by the model, I construct regional high-skill andlow-skill export demand shocks, and study their differential effects on schooling. Second, Idraw on a rich sub-national data set and the within-country approach necessarily controlsfor several confounders that are not accounted for by country-level panel studies. Althoughthe context of the study is a developing country, the vast geographic diversity of China al-lows me to examine the effects of export shocks on regions that are initially skill abundant.Third, I take a further step and investigate the feedback effect of trade-induced human capitalaccumulation or decumulation on regional industry specialization.This paper also fits in the rapidly growing literature that employs the variation in re-gional initial differences in industry composition to study the differential effects of trade onlocal economies within a country, including Topalova (2010), Hakobyan and McLaren (2010),Autor, Dorn and Hanson (2013), Kovak (2013), Dix-Carneiro and Kovak (2015), and others.However, unlike these studies, which focus on import shocks, I am more interested in theexport demand shocks generated by the rest of the world.1.3 A Model with Endogenous Skill SupplyThe framework has three building blocks. The first is a standard Roy model of educationchoice, which builds on the work of Lagakos and Waugh (2013), Hsieh et al. (2013) andBurstein, Morales and Vogel (2015). Second, I augment the multi-region and multi-sectorRicardian model in Costinot, Donaldson and Komunjer (2012) with multiple factors of pro-duction. Due to the variation in regional comparative advantage and sectoral skill intensities,the skill content of trade shocks differs across regions. Third, I embed the Roy model in ageneral equilibrium framework and examine the effects of reduction in trade barriers on skillpremium and educational attainment.Specifically, the model features multiple regions (i = 1, ..., N) and multiple sectors (k =1, ...,K) with different levels of skill intensity. Production in each sector requires inputs fromboth skilled and unskilled workers. The total supply of skilled and unskilled labor in region iare Hi and Li. I assume that there is no migration among regions but workers are perfectlymobile across sectors.4 Also, labor is inelasitically supplied.A unit mass of workers are born at each unit of time, and each of them lives for a periodof length T . At birth, each worker is endowed with a vector of “individual productivities”,denoted by {zh, zl}, which represent the efficiency of being a skilled worker and an unskilledworker, respectively. By assumption, zh and zn are randomly drawn from the multivariate4I abstract away from internal migration for the transparency of the model. The potential confoundingeffects introduced by migration are examined in the empirical part of the paper. The assumption on migrationacross regions is the same as Autor, Dorn and Hanson (2013). Alternatively, as is shown in Galle, Rodr´ıguez-Clare and Yi (2015), one can allow perfect migration within a country and assume that heterogeneous workersin some regions are more closely attached to some sectors. The latter framework also shows that a national-level trade liberalization has different effects across regions within a country and suggests a Bartik-style indexof region-level trade shock for empirical analysis.5Fre´chet distribution:5F (zh, zl) = exp[−(z− κ1−νh + z− κ1−νl )1−ν ].The parameter κ > 1 determines the dispersion of efficiency units, with a higher value of κcorresponding to smaller dispersion. The parameter ν ∈ [0, 1) governs the correlation betweenzh and zl. A higher value of ν increases this correlation, and ν = 0 corresponds to the casewhere zh and zl are independent. Every worker decides about whether to enter the labormarket right away as an unskilled worker with zl efficiency units of labor, or spend a periodof time ϕ receiving an education and become a skilled worker equipped with zh efficiencyunits of labor.1.3.1 Education DecisionThe lifetime income of a skilled worker is∫ Tϕ wi,hzhe−rtdt = wi,hr (e−rϕ − e−rT )zh, and thelifetime income of an unskilled worker is∫ T0 wi,lzle−rtdt = wi,lr (1 − e−rT )zl, where wi,h andwi,l denote the wage per efficiency unit of skilled labor and unskilled labor respectively. Inequilibrium, the marginal workers are indifferent between obtaining an education or not, andhence the school enrollment rate in region i is determined bypii,h = Pr(zhwi,h(e−rϕ − e−rT ) ≥ zlwi,l(1− e−rT )) = 1µ− κ1−νi + 1, (1.1)where µi =e−rϕ−e−rT1−e−rTwi,hwi,lsummarizes the regional return to schooling, which is determinedby the local skill premiumwi,hwi,l. Equation (1.1) shows that the school enrollment rate increaseswith the return to schooling. Furthermore, a larger value of κ, which implies higher densityof workers at the margin, raises the responsiveness of schooling to µi. A higher value of ν,which lessens the role of comparative advantage, also increases this responsiveness. In thesteady state, pii,h also has a interpretation of share of educated workers in region i. Then,the share of uneducated workers is pii,l = 1− pii,h.As is shown in Appendix A.1.2, the expected productivities of educated and uneducatedworkers in the steady state areE(zh|zh > zl/µi) = γpi−(1−ν)/κi,h and E(zl|zl > zhµi) = γpi−(1−ν)/κi,l ,where γ = Γ(κ−1κ ) is the gamma function evaluated at (κ−1)/κ. Then, the effective suppliesof skilled labor and unskilled labor in the steady state areHi = (T − ϕ)γpi1−(1−ν)/κi,h and Li = Tγpi1−(1−ν)/κi,l . (1.2)According to Equation (1.2), the total skilled labor supply increases with the share of educatedworkers at a diminishing rate. Intuitively, higher education return induces workers who arerelatively unproductive as educated workers to nonetheless select into school, which lowers theaverage productivity of skilled workers. For a similar reason, total unskilled labor increaseswith the share of uneducated workers at a decreasing rate.5We can interpret the vector of “individual productivities” as a random start of talents across two occu-pations, h and l. An individual needs to receive education for a period of length ϕ to launch a career inoccupation h. This is a simplified version of the case in Hsieh et al. (2013). A similar setting can also befound in Lagakos and Waugh (2013) and Burstein, Morales and Vogel (2015).61.3.2 Production and TradeThe production side of the model extends Costinot, Donaldson and Komunjer (2012) byincluding multiple factors of production. Each good k may come in an infinite numberof varieties indexed by ω ∈ Ω ≡ {1, ...,+∞}. Production of the ωth variety of good krequires inputs of both skilled labor (hi,k(ω)) and unskilled labor (li,k(ω)) with the followingtechnologyyi,k(ω) = ψi,k(ω)hi,k(ω)αk li,k(ω)1−αk ,where ψi,k(ω) is the total factor productivity (TFP) of producing the ωth variety of good kin region i, and αk is the income share of skilled labor, with α1 < α2 < ... < αK . I assumeψi,k(ω) is a random variable drawn independently for each triplet (i, k, ω) from the Fre´chetdistribution:Ψi,k(ψ) = e−(ψ/ψi,k)−ε with ψi,k > 0 and ε > 1.Note that the parameter of fundamental productivity in sector k, ψi,k, varies across regions,and the parameter ε captures intra-industry heterogeneity. The variable cost of producingthe ωth variety of good k in region i, vi,k(ω), is then expressed byvi,k(ω) =1ψi,k(ω)(wi,hαk)αk( wi,l1− αk)1−αk .Firms in region i and sector k face an iceberg cost, τij,k, to sell to market j, i.e., they mustship τij,k units of output for one unit to arrive in region j. It is assumed that τij,k ≥ 1 ∀ j 6= iand τii,k = 1. Therefore, the marginal cost of selling each unit in market j is vi,k(ω)τij,k.Markets are assumed to be perfectly competitive. Hence, for each variety of good k,region i can sell in region j at price pij,k(ω) = vi,k(ω)τij,k. Consumers in region j opt tosource individual varieties from the lowest cost location, which implies the prevailing priceof ωth variety of good k, pj,k(ω), satisfiespj,k(ω) = mini{vi,k(ω)τij,k}.1.3.3 PreferencesThe representative consumer’s utility in region j is defined over the goods from sector k =1, ...,K asCj =K∏k=1Cβkj,k,where βk > 0 is the exogenous preference parameter satisfying∑Kk=1 βk = 1 and Cj,k is thetotal consumption of the composite good k in region j,Cj,k =( N∑i=1∑ω∈Ωij,kcj,k(ω)(σ−1)/σ)σ/(σ−1),where σ > 1 + ε is the elasticity of substitution between different varieties.6 Ωij,k ≡ {ω ∈Ω|pij,k(ω) = mini′{pi′j,k(ω)}} is the set of varieties of good k imported by region j from6The restriction is a technical assumption that guarantees the existence of a well-defined CES price index.7region i. cj,k(ω) is region j’s consumption of ωth variety of good k. The consumption isdetermined by cj,k(ω) = (pj,k(ω)Pj,k)1−σβkEj , where Ej is the total expenditure of region j, andPj,k = (∑Ni∑Ωij,kpj,k(ω)1−σ)1/(1−σ) denotes the price index of sector k in region j.1.3.4 Trade FlowsTrade is assumed to be balanced. Thus, Ej = Yj where Yj denotes total income. As is shownin Appendix A.1.3, the value of exports of good k from region i to j, Xij,k, is determined byXij,k = λij,kβkYj ,where λij,k denotes the fraction of region j’s expenditure on good k allocated to the goodsproduced in region i and satisfies the following equation:λij,k =(vi,kτij,k)−ε∑i′(vi′,kτi′j,k)−ε ,where vi,k =1ψi,k(wi,hαk)αk( wi,l1−αk)1−αk is the average variable cost of producing good k in regioni.1.3.5 Steady State EquilibriumDenote Yi,k as sector k’s value of output in region i. National income is determined bythe sum of sectoral value outputs, i.e.∑k Yi,k = Yi. In the steady state, the N(K + 5)endogenous variables {wi,h, wi,l, Hi, Li, {Yi,k}k=1,...,K , Yi}i=1,...,N state are determined by thefollowing labor market clearing conditions:wi,hHi =K∑k=1αkYi,k, (1.3)wi,lLi =K∑k=1(1− αk)Yi,k, (1.4)the goods market clearing condition:Yi,k =N∑j=1λij,kβj,kYj , (1.5)and the factor supply equations:Hi = (T − ϕ)γ(1( 1−e−rT(e−rϕ−e−rT )wi,lwi,h)κ1−ν + 1)1−(1−ν)/κ, (1.6)Li = Tγ(1− 1( 1−e−rT(e−rϕ−e−rT ),wi,lwi,h)κ1−ν + 1)1−(1−ν)/κ. (1.7)81.4 Export Demand Shocks to Regional MarketsIn this section, I derive a theoretical link between export demand shocks from the rest ofthe world (ROW) and education decisions in China. This simple framework underlies theempirical measures of regional export shocks and the identification strategy.1.4.1 Reduced Form RelationThe exogenous shocks in this model come from changes in iceberg costs, {τˆij,k}, and changesin productivity, {ψˆi,k}, where i is a region in China and j is a region in the ROW, andhats over variables denote log changes (xˆ ≡ d lnx). To derive the reduced form relation,each region in China is treated as a small open economy, and hence the exogenous shocks toregion i have no effect on the income level of other regions. I only consider the instantaneousresponse of school enrollment to the exogenous shocks so that Hˆi = Lˆi = 0.7 As exact changesin iceberg cost, τˆij,k, are not observable from data, for the purpose of empirical analysis, theiceberg cost of exporting good k from region i in China to region j in the ROW is assumed totake the form τij,k = τj,kτ˜ij,k. Here τj,k captures the costs such as tariff, exchange rate, andinstitutions, which all exporters in China incur, and τ˜ij,k represents the idiosyncratic coststhat apply to region i, such as the local transportation infrastructure.As is shown in Appendix A.2, the impact of external demand shocks induced by {τˆj,k}on region i’s school enrollment rate is as follows:pˆii,h = −c1∑khi,k∑j∈ROWγij,k(1−λij,k)τˆj,k+c2∑kli,k∑j∈ROWγij,k(1−λij,k)τˆj,k+ν({ψˆi,k, ˆ˜τij,k}),(1.8)where ci,1, ci,2 > 0 are general equilibrium scaling factors; hi,k ≡ Hi,k/Hi and li,k ≡ Li,k/Lidenote, respectively, sector k’s employment share in skilled and unskilled labor of region i;The export demand shock to sector k is the sum of τˆj,k’s weighted by γij,k(1− λij,k), whereγij,k ≡ Xij,k/Yi,k denotes the revenue share from market j for sector k in region i. On theone hand, as γij,k measures how important market j is as an outlet of good k from regioni, a higher weight is assigned to τˆj,k if γij,k is larger. On the other hand, λij,k capturesthe market share of region i’s good k in market j, so a lower weight is assigned to τˆj,kif λij,k is larger. Intuitively, a large λij,k indicates that region i has high productivity orlow iceberg cost in sector k relative to all other regions. Hence, a small change in τj,k haslittle effect on export demand. ν({ψˆi,k, ˆ˜τij,k}) is the residual term subsuming the effect ofproductivity shocks {ψˆi,k} and idiosyncratic iceberg cost shocks {ˆ˜τij,k}. The sectoral shocksare then aggregated to the regional level, using employment share as weights. Equation(1.8) implies that school enrollment increases more when the declines in iceberg cost aremore pronounced in the sectors employing larger share of skilled labor. I consider the terms∑k hi,k∑j 6=i γij,k(1− λij,k)τˆj,k and∑k li,k∑j 6=i γij,k(1− λij,k)τˆj,k as high-skill and low-skill7Note that conditions (1.3)-(1.5) must be satisfied in any equilibrium, while conditions (1.6) and (1.7) onlyhold in the steady state. In deriving Equation (1.8), I consider the response of skill premium and schoolenrollment to exogenous shocks in the short run by disturbing Equations (1.3)-(1.5), while keeping the laborsupply Hi and Li constant. In the model, time is continuous and workers are continuously distributed, andhence the mass of a cohort born at an instant of time is zero. Therefore, a change in education decision of aspecific cohort does not affect total factor supplies, i.e., Hˆi = 0 and Lˆi = 0.9export demand shocks. Industry-level supply shocks deriving from productivity changes areabsorbed in the residual term.This weighted-average structure resembles the empirical approach in the literature on thelocal effects of trade (Topalova, 2010; Edmonds, Pavcnik and Topalova, 2010; Autor, Dornand Hanson, 2013; Kovak, 2013). However, trade shocks differ by skill levels, reflecting thedifferent skill content embodied in different industries. This approach shares similarities withDix-Carneiro and Kovak (2015).1.4.2 Export Demand Shocks: From National to LocalIn order to employ Equation (1.8) for empirical analysis, I make the following assumptions:(1) the general equilibrium scaling factor (ci,1 and ci,2) are the same across China’s regions(i.e. ci,1 = c1 and ci,2 = c2); (2) λij,k ≈ 0, that is each region in China has small market sharein region j in the ROW; and (3)Xi ROW,kXCH ROW,k≈ Ei,kEk , that is the share of region i in China’stotal exports of good k is approximated by the region’s share of national employment inthat industry. With these restrictions in place, the change in school enrollment in region ibecomespˆii,h ≈ c˜1∑kHi,kEi,kEi,kEk∆XkHi− c˜2∑kLi,kEi,kEi,kEk∆XkLi+ ν({ψˆi,k, ˆ˜τij,k}), (1.9)where c˜1, c˜2 > 0; Ei,k/Ek denotes the share of region i in China’s employment of sectork; Hi,k/Ei,k and Li,k/Ei,k denote the employment share of skilled and unskilled workersin sector k of region i, respectively; ∆Xk ∝∑j∈ROW XCHj,kτˆj,k is the change in nationalexports induced by {τˆj,k}. The approximation is detailed in Appendix A.2.Following Equation (1.9), the main measures of high-skill and low-skill export demandshocks are constructed as∆ExportLSit =∑kLik0Eik0Eik0Ek0∆XktLi0and ∆ExportHSit =∑kHik0Eik0Eik0Ek0∆XktHi0. (1.10)Note that the measures of export shocks are constructed as the interaction of the industryfactor intensity, the regional initial industry composition, and export changes at the sectorallevel. Specifically, to build the high-skill demand shock, the national export shock of sectork, ∆Xk, is apportioned to region i according to its share of national industry employment inthe base period, Eik0/Ek0. Then, this regional export expansion is attributed to high skilledlabor according to the sector’s skill intensity, Hik0/Eik0, and normalized by the amount ofhigh skilled labor Hi0. By construction, ∆ExportHSit represents export exposure in dollar perskilled labor. Similarly, ∆ExportLSit represents export exposure in dollar per unskilled labor.Exports of industries with different skill intensities expand by varying levels, and atvarying times, inducing different export exposures of various types of workers across regionsin China. As skill intensity reflects inherent technological requirements of an industry, it issimilar across regions. Therefore, the differences in regional export shocks mainly stem fromthe differences in local industry composition in the base period. In addition, differences inskill intensity across sectors provide the identification of parameters c1 and c2.101.5 Data1.5.1 Local EconomiesIn the empirical analysis, a local economy is a prefecture, an administrative division in Chinaranking between province and county. Prefectures are matched across census years accordingto the 2005 administration division of China, so that the data have a geographic paneldimension. There are 340 prefectures, with median land area of 13,152 km2 and medianpopulation of 3.2 million in year 2000.Under China’s household registration system, most migrant workers have restricted accessto public health, education and social services, which in effect imposes significant barriers oninter-regional labor mobility.8 According to the 2000 census, less than 4.5% of the populationaged between 16 and 59 changed their prefecture of residence during the previous five years.This number increased slightly to 4.8% in 2005.9 In contrast, the five-year migration rateacross states was around 12.5% for the US in 2000 (Kaplan and Schulhofer-Wohl, 2013) andthe five-year migration rate across districts in India (a similar administrative division toprefecture) was around 13.5% in 2007 (Marden, 2015).1.5.2 Population CensusesI use data from a 1% subsample of the 1990 and 2000 China Population Censuses, and a20% subsample of the 2005 China 1% Population Sampling Survey (mini census).10 Thecensus data contain information such as region of residence, migration, school enrollment,educational attainment, demographic characteristics, employment status, occupation, andindustry.Data on Education. As of 1986, Chinese law mandated nine years of compulsory schooling(six years of primary education and three years of junior secondary education). It alsorequires all children to attend school by the age of 7.11 For most of the analysis, the sampleis restricted to young people aged at least 16, who should have already finished compulsoryschooling. Specifically, I separately examine school enrollment of people of high school age(16 to 18 years), and college age (19 to 22 years). School enrollment rates at the prefecturelevel are constructed for each age group.Figure 1.2 shows national school enrollment rates by age over the census years. Theimprovement of school enrollment is remarkable. From 1990 to 2005, enrollment rates in-creased by 30.5 and 12.7 percentage points for high school and college respectively. Theseimprovements are more or less evenly split between the periods 1990 to 2000 and 2000 to2005. It is also worth noting that, although it is improving over time, the enforcement ofcompulsory schooling is imperfect, especially for junior secondary education. Figure 1.3 plotsprefecture school enrollment rates against lag period’s enrollment rates for both high school8Tombe and Zhu (2014) estimate the migration cost across province amounted to 1.5 times annual incomein the 1995-2000 period, and declined slightly to 1.3 times annual income between 2000 to 2005.9The figure computed is from 20% subsample of the 2005 China 1% Population Sampling Survey.10The 2005 China 1% Population Sampling Survey is like a medium-term small-scale census. It surveys 1%of the population and the questionnaire is very similar to the regular censuses. In my sample, the number ofobservations in 2005 is around 20% of those in 1990 and 2000.11Provinces had different effective dates for implementing the compulsory education law.11and college. Most of the prefectures lie above the 45◦ line, suggesting the improvement inschool enrollment was a nationwide phenomenon. More importantly, there is large variationin the increase in school enrollment conditional on the same initial level.Data on Industry Employment. To construct export shocks as defined in (1.10), I usethe employment data from the 1990 census. The census records region of residence andindustry of employment at 3-digit Chinese Standard Industrial Classification (CSIC) codes(1984 version). To study the industry reallocation in Section 1.7, I also collect prefectureemployment data at 2-digit CSIC level for censuses 1990, 2000 and 2010, which are assembledand published by the provincial statistics bureau.1.5.3 Trade and Tariff DataFrom the UN Comtrade Database, I obtain data on China’s export and import values at4-digit International Standard Industrial Classification (ISIC) level for years 1992, 2000 and2005.12 Data on China’s export tariffs imposed by destination countries at 4-digit ISIC levelare collected from the TRAINS Database. The tariff faced by Chinese exporters in a 4-digitISIC industry k during year t is computed according toTariffXkt =∑cExportChina,c,k,t−1ExportChina,k,t−1Tariffckt,where Tariffckt denotes the tariff imposed by country c on goods of industry k during periodt. The tariffs are weighted by the country’s share in China’s total exports of good k in thelag period and then aggregated to the industry level. These weights are constructed usingthe trade flow data from three years earlier. Data on import tariffs imposed by China on4-digit ISIC industries are collected from the TRAINS Database. To match the trade andtariff data with sectoral employment data from the 1990 census, I concord them into 3-digitCSIC codes. All export and import data are inflated to 2005 US dollar using the ConsumerPrice Index from China Statistical Yearbooks.1.5.4 Regional Export Demand ShocksTo build the empirical counterpart of the regional export demand shocks as is defined in(1.10), I first run the regressionlnExportkt = β ln(TariffXkt ) + γk + φt + µkt, (1.11)where TariffXkt is the weighted export tariff of industry k in year t, and γk and φt are industryand year fixed effects. Then, the local export demand shocks are constructed according to∆ExportLSit =∑kLik0Eik0Eik0Ek0∆ ̂ExportktLi0and ∆ExportHSit =∑kHik0Eik0Eik0Ek0∆ ̂ExportktHi0.where ̂Exportk,t = exp(βˆ ln(TariffXkt ) + γˆk + φˆt) is the exponent of the fitted value fromregression (1.11). High skilled labor H is considered as the set of workers with college121992 is the first year when the ISIC export data is available for China from UN Comtrade.12education or above, and low skilled labor L is considered as the set of worker with highschool education or below.Note that the exogeneity of the conventional Bartik-style instrument relies on the as-sumptions that other time-varying, region specific determinants of the outcome variable areuncorrelated with (1) a region’s initial industry composition, and (2) industry shocks at thenational level. The latter requirement could be violated if an industry clusters in a specificregion and the region also specializes in that industry. The above strategy potentially ad-dresses this concern, as shocks at the national level are induced by external demand shocksfrom the ROW due to changes in export tariffs, which are deemed to be exogenous.Panel B in Table 1.1 shows the descriptive statistics of export demand shocks, ∆ExportLSand ∆ExportHS , by time period. The mean low- and high-skill shocks from 1990 to 2000 are$183 and $311, respectively. Export growth accelerated between 2000 and 2005, with meanlow-skill and high-skill shocks being $542 and $916. Because of the substantial geographic dif-ferences in industry composition, the variation in export exposure is large across prefectures.The difference in low-skill export shock between prefectures at the 25th and 75th percentilewas $171 during the first period, and $503 in the second period. For high-skill export expan-sion, the differences amounted to $419 and $1194, respectively. Panels A and B of Figure 1.1show the distribution of low-skill and high-skill export demand shocks across prefectures inChina during period 2000 to 2005. Prefectures are outlined in gray and provinces are outlinedin black. Notice that export shocks are larger in the east of the country. The differences ofthe high-skill and low-skill shocks, i.e., ∆ExportHS −∆ExportLS , are displayed in Panel C.The differential export shocks, ranging from -$263 to $1077, reflect industry specializationacross prefectures. The differences across prefectures are stark, even within a province.1.5.5 Industry Output and Other Socioeconomic DataData on industry output come from Chinese Industrial Annual Survey. Other socioeconomicvariables at the prefecture level, including GDP per capita, fiscal expenditure on education,sex ratio, and share of population with urban Hukou, come from various provincial statisticalyearbooks and population censuses. The distance to the nearest port for each prefecture iscalculated using information from the World Port Index. More details about the data sourcescan be found in Appendix B.1.1.6 Effects of Export Demand Shocks on School EnrollmentThis section examines the effects of export demand shocks on high school and college en-rollments following the framework outlined in Section 1.4.2 and using the export demandshocks constructed in Section 1.5.4. The estimation strategy identifies the effects of exportexpansion induced by the reduction in export tariff on the education decision. The identi-fication relies on the assumption that the changes in export tariff are uncorrelated with theprefecture-specific shocks to industry productivity and trade costs.131.6.1 Baseline ResultsI evaluate the effects of export shocks on school enrollment by estimating the followingequation:∆Enrollit = βLS∆ExportLSit + βHS∆ExportHSit + φpt + εit, (1.12)where ∆Enrollit is the change in school enrollment in the prefecture i between t − 1 and t.The regression stacks the first differences for the two periods, 1990 to 2000 and 2000 to 2005,and includes province×year dummies (φpt). By introducing φpt, I flexibly account for theprovince-specific shocks in different periods, and hence the identification comes from within-province variation in export exposure. In all regressions, standard errors are clustered at theprovince level to account for the potential serial correlation over time and across prefectureswithin a province.The results of the baseline regression (1.12) are presented in Column (1) of Table 1.2.Panel A and Panel B report the results for high school enrollment (age group 16–18) andcollege enrollment (age group 19–22), respectively. Export demand shocks have a statisticallysignificant effect on school enrollment. Specifically, a $1000 increase in export per unskilledworker reduces high school enrollment rate by 5.8 percentage points and decreases collegeenrollment rate by 2.1 percentage points. Conversely, a $1000 increase in export per skilledworker raises enrollment rates by 0.8 percentage point for high school and 1.2 percentagepoint for college. Column (2) incorporates a set of concurrent socioeconomic shocks thatmight independently affect education choices. It includes change in log GDP per capita,change in log fiscal expenditure, change in average age, change in the proportion of boys, andchange in the proportion of population with urban Hukou. These controls leave the mainresults unaffected. Column (3) additionally includes the start-of-the-period school enrollmentrate and log GDP per capita to account for the prefecture-specific trends that may correlatewith the initial conditions of education and economic development. Again, the regressionresults remain stable.A potential threat to the identification strategy is that the unobserved time-varyingprefecture-specific determinants of school enrollment may correlate with a prefecture’s initialindustry structure. For example, if the prefectures initially specializing in high-skill industriesreceive a larger supply shock in education provision than the prefectures initially specializingin low-skill industries, the results will be confounded. To address this concern, Column (4)introduces two additional control variables constructed as followsPLIi0 =∑k∈TrLik0Eik0Eik0Ei0and PHIi0 =∑k∈TrHik0Eik0Eik0Ei0.where PLIi0 and PHIi0 stand for prefecture i’s low-skill and high-skill intensity in the baseperiod, respectively. These controls serve two purposes. First, the sum of PLIi0 and PHIi0equal to the employment share of the tradable sector of prefecture i in the base period.Hence their inclusion isolates the variation of export demand shocks stemming from thewithin-tradable sector industry composition from the variation arising from the importanceof tradable sector for local employment. Second, given the same size of tradable sector, alarger PHIi0 indicates that a prefecture is more specialized in high-skill industries. Therefore,the inclusion of PLIi0 and PHIi0 captures the initial differences in industry specialization.It is found that the estimates are insensitive to these controls.14Instead of explicitly controlling for the initial industry composition, Column (5) augmentsthe regression model with prefecture dummies, which effectively accounts for the prefecture-specific linear time trend of school enrollment. As a result, the coefficients are identifiedthrough the time variation of export shocks within a prefecture. The results remain similarto those in Column (4), which alleviates the concern that the estimates are confounded bythe different secular trends across prefectures that are associated with the initial industryspecialization.Column (6) replaces the change in school enrollment (a flow variable) with the change ineducational attainment (a stock variable) as dependent variable. The dependent variable inPanel A is change in the share of population aged 16 to 18 with some high school education orabove, and the dependent variable in Panel B is change in the share of population aged 19 to22 with some college education or above. The estimates remain qualitatively similar. Quanti-tatively, however, the effects of low-skill export demand shock on educational attainment aresmaller in magnitude than the effects on school enrollment. This finding provides suggestiveevidence that a low-skill export shock not only discourages young people from proceedingonto high school/college, but also increases the dropout rate of those in high school/college.To gauge the magnitude of the estimated effects of export demand shocks, we can considerthe differential changes in school enrollments associated with the interquartile ranges of low-skill and high-skill export shocks between 1990 and 2005 (which were, respectively, $681 perlow skill worker and $1600 per high skill worker). The point estimates in Column (4) implythat the high school and college enrollment rates in the prefectures at the 25th percentileof ∆ExportLS increased, respectively, by 2.3 and 2.1 percentage points more than those inthe prefectures at the 75th percentile. Similarly, the high school and college enrollment ratesin prefectures at the 25th percentile of ∆ExportHS increased, respectively, by 0.8 and 1.1percentage points less than those in the prefectures at the 75th percentile.1.6.2 RobustnessExport Expansions in Other Prefectures. The identification strategy relies on theassumption that only the local labor market condition is relevant to the education decision.Nevertheless, the locations of education and employment could be separate in the sense thatan individual attains more or less education locally in response to export shocks elsewhere.Two approaches are adopted to investigate the effects of non-local export shocks:(a) I construct the weighted average of export shocks of the neighboring prefectures accordingto:∆Exports,Nit =∑r∈Neighboriθir∆Exportsrt,where Neighbori denotes the set of prefectures sharing a border with prefecture i andθir =Er0∑r′∈Neighbori Er′0is the employment share of prefecture r among the neighboring pre-fectures of i.(b) Controlling for neighboring prefectures’ export shocks may not be sufficient, as an indi-vidual could respond to the expansion of a distant export manufacturing hub. Alternatively,I employ the inverse spatial distance among prefectures, to weight and aggregate the shocks15of all other prefectures as follows∆Exports,Nit =∑r 6=i1dir∆Exportsrt,where dir denotes the distance between prefectures i and r.13Table 1.3 presents the regression results incorporating the controls ∆Exports,Nit . The leftand right panels give the results for approaches (a) and (b), respectively. Regardless of whichway the non-local export shock is constructed, the estimated effects of local export shocks areinsensitive and resemble those in Table 1.2. These findings suggest that the local labor marketcondition affects educational choice independently of cross-border spillovers. In addition, low-skill non-local export shocks are found to have an adverse effect on local school enrollmentin most specifications. When constructed following approach (b), high-skill non-local exportshocks are estimated to have positive effects on local school enrollments.Import Shocks. An important element missing in the analysis thus far is the importanceof imports from the ROW into China. In Table 1.4, I add import shocks ∆ImportLSit and∆ImportHSit to the regression analysis. These shocks are constructed in a similar way to theexport shocks discussed in Section 1.5.4, with data on exports and export tariffs replaced byimports and import tariffs. As is predicted by the model, a low-skill import shock increasesthe skill premium and hence raises the school enrollment rate, and the opposite is the casefor a high-skill import shock. As shown in Columns (1) and (4), a $1000 low-skill importshock enhances high school and college enrollment rates by 1.1 and 1.7 percentage pointsrespectively. A $1000 high-skill import shock depresses high school and college enrollmentrates by 0.2 and 0.2 percentage point respectively. Columns (2) and (5) augment the spec-ification with the controls PLI and PHI. The estimates remain statistically significant atconventional levels for both age groups, with the exception of high school enrollment in thecase of low-skill import shock.An issue with measuring exposure to import competition is that the imports to China notonly include final goods purchased by domestic consumers, but also intermediate inputs andcapital goods purchased by firms. The latter may substitute or complement the skilled or un-skilled workers, shifting the skill demand and altering education incentives. If this is the case,the estimated impacts of import shocks may pick up effects other than import competition.To isolate the potential confounding effects introduced by imports of intermediate and capi-tal goods, Column (3) and (6) include the controls ∆ImportLS,CIit and ∆ImportHS,CIit , whichmeasure the exposure to imports of intermediate and capital goods per low skilled workerand per high skilled worker.14 Conditional on the import shocks of intermediate and capital13I normalize the distance so that min(dir)=1. As a result, the distance measure does not carry any unit.14The intermediate inputs and capital goods are defined according to UNCTAD SoP product groups.∆ImportLS,CIit and ∆ImportHS,CIit are constructed according to∆ImportLS,CIit =∑kLik0Eik0Eik0Ek0∆ ̂ImportCIktLi0and ∆ImportHS,CIit =∑kHik0Eik0Eik0Ek0∆ ̂ImportCIktHi0,where ̂ImportCIkt denotes the change in imports of capital and intermediate goods belonging to industry k,which is predicted by the changes in import tariffs. More details can be found in the Appendix B.1.16goods, the low-skill and high-skill import shocks are found to have opposite and statisticallysignificant effects on the school enrollments, as predicted by import competition.Although the measures of import shocks are potentially subject to endogeneity issues15,the opposite effects of the export and import shocks detected in Table 1.4 are reassuring.Moreover, compared to Table 1.2, the estimated coefficients of export demand shocks changelittle. This finding alleviates the concern that the measures ∆Exportsit may in part pick upimport shocks, if the sectors experiencing larger export shocks are also the ones experiencinglarger import shocks.Migration. The results could be confounded by selective migration. It is plausible that low-skill export expansions lower the average education of immigrants and increase the averageeducation of out-migrants. The converse could be true for the case of high-skill exportexpansion. If it is the case, export shocks could change the average education level of thelabor force by altering the composition of migrants. In Appendix A.4.2, I investigate theeffects of the regional export shocks on the migration flows of workers with different educationlevels, and find suggestive evidence for the selective migration. Nevertheless, I show that thisdoes not seem to quantitatively affect the effects emphasized by my model.To evaluate the potential bias introduced, I estimate Equation (1.12) separately for eachage group between 18 and 35 with the dependent variable replaced by the change in edu-cational attainment. The idea is that if export demand shocks indeed affect the educationchoices through altering the contemporary labor market condition, the effects should be largeramong the school age young people than in the older groups, as the education decision ofthe latter groups was made prior to the shocks. However, if instead the estimates of βsare driven by selective migration, it is expected that low- and high-skill shocks should havesimilar effects on all age groups, if migration costs are similar for the young workers. Theestimates and the 95% confidence intervals are plotted in Figure 1.4. The upper panel showsthe effect of low- and high-skill export shocks on the share of the population with some highschool education, and the lower panel shows the results for college education attainment.Consistent with the findings in Table 1.2, export shocks have statistically significant effecton the educational attainment of school age young people. However, the effects are greatlydampened from age 23 onwards. This finding suggests that estimates of βs are unlikely to beseverely biased by selective migration.Inflows of migrants could also dilute the educational effect of export expansion on localyoung people if, for example, the expansion of low-skill exports attracts low-skilled immi-grants, offsetting the negative effect of the shocks on local skill premium. It is also possiblethat the local labor market is segmented in such a way that the low-skill export manufactur-ing sector employs only migrant workers. In both scenarios, low-skill export expansion mighthave only a small effect on local education decisions. Appendix A.4.3 explores this possibilityby estimating Equation (1.12) separately for the sample of non-migrants.16 It shows thattheir education decisions are responsive to both low- and high-skill export shocks. Moreover,the estimates of βs are statistically similar to those of the baseline sample, which contains15Unlike export tariffs which are imposed by other countries, import tariffs are set by Chinese governmentand may be part of a well-planned development strategy. In such a scenario, changes in import tariffs acrossindustries could systematically correlate with industry productivity shocks.16This robustness check is restricted to the period 2000-2005, as the information on between-prefecturemigration is not available in the 1990 census.17both locals and immigrants.Another concern related to migration is that students could go to university outside theirhome prefecture in response to trade induced change in local education return. (This problemis less a concern for high school.) It could lead to measurement error in the college enroll-ment rate that correlates with export shocks. In particular, the baseline measure of collegeenrollment rate understates the actual level for prefectures that receive a large high-skill ex-port shock and constrained by education capacity. As a result, the estimate of ∆ExportHSitwill bias towards zero. To alleviate the concern, I re-define college enrollment share to beCollegeEnrollit =EnrollNMit +EnrollEMitPopNMit +EnrollEMit. Here, PopNMit is the population of non-migrants agedbetween 19-22 in prefecture i and year t, EnrollNMit is the amount of students enrolled inschool, EnrollEMit is the amount of students who emigrates from prefecture i within past fiveyears and are enrolled in school elsewhere. Column (7) of Appendix Table A.4 reports the re-gression result using this alternative outcome measure. Again, the estimates are statisticallysimilar to those of the baseline result.Other Robustness Checks. In Appendix A.4, I demonstrate the robustness of the basicresults to different concerns and many additional specifications. Appendix A.4.1 addressesthe concern that change in education provision could confound the effect of export shockson education demand. In particular, I show that change in education fiscal expenditure(or number teachers) are weakly correlated with a prefecture’s initial skill level. A low-skillexport shock tends to increase local fiscal expenditure on education, suggesting that localgovernments, especially of prefectures that have larger exposure to low-skill export shock,could alter education supply to offset the export-induced shock on demand side. Importantly,in the case that the low-skill export exposure is correlated with local educational provision,estimates in Table 1.2 are still likely to provide a lower bound (in magnitude) of the effect oflow-skill export shock on education demand.Appendices A.4.4 and A.4.5 confirm that the main findings are not driven by a particulardemographic group or region. Appendix A.4.6 shows that the school enrollment of youngchildren subject to compulsory schooling was barely affected by export expansion during thesample periods. Appendix A.4.7 examines the effects of export shocks on the youth’s labormarket outcomes and finds a mirror pattern for the market employment. Lastly, AppendixA.4.8 disaggregates the measures of export exposure into low-, medium- and high-skill shocks,and obtains consistent results.1.7 Effect of Change in Skill Supply on IndustrySpecializationIn the dynamic context of Findlay and Kierzkowski (1983), trade enlarges the differences infactor abundance, which reinforces the initial comparative advantage and industry special-ization. This section investigates the effect of trade-induced human capital accumulation ordecumulation on evolution of industry specialization. Specifically, I estimate the following18equation∆ShareMikt =∑K∈{LS,MS,HS}γK1(k ∈ K)∆HighSch.Enrollit−1+∑K∈{LS,MS,HS}δK1(k ∈ K)∆College.Enrollit−1 + X′ik,tζ + ψpk + υikt(1.13)where ∆ShareMik,t is the change in employment share of industry k in the manufacturing sectorof prefecture i during the period 2000 to 2010, ∆HighSch.Enrollit−1 and ∆College.Enrollit−1are the changes in high school and college enrollments in prefecture i during the period 1990to 2000, and 1(k ∈ K) is a dummy variable equal to 1 if industry k belongs to industry groupK, where K can be low-skill, medium-skill or high-skill.17 The changes in school enrollmentare allowed to have differential effects on industries belonging to different industry groups,which are captured by the coefficients γK and δK . The vector Xik,t contains industry k’sstart-of-the-period employment share in prefecture i and its changes in employment sharein the period 1990 to 2000. Province×industry fixed effects ψpk are included to captureunobserved provincial industry policies.Changes in school enrollment in the previous decade could be endogenous. For exam-ple, local government could change educational provision to adapt to industrial planning.To address this potential endogeneity problem, in some specifications, I instrument the in-teraction terms 1(k ∈ K)∆HighSch.Enrollit−1 and 1(k ∈ K)∆College.Enrollit−1 with1(k ∈ K)∆ExportLSit−1 and 1(k ∈ K)∆ExportHSit−1. The IV regressions thus estimate theimpact of export-induced changes in skill supply on the change in industry specialization inthe later decade. The identification relies on the assumption that unobserved shocks affect-ing the evolution of industry specialization in the 2000s are uncorrelated with local industrycomposition in 1990 and export tariff changes during the 1990s.The regression results are presented in Table 1.5. The OLS and 2SLS estimates areshown in Columns (1) and (2) respectively. Both find a significant effect of changes in collegeenrollment between 1990 and 2000 on changes in industry employment in the period 2000 to2010. The 2SLS estimates imply that a 10 percentage point increase in college enrollmentin the 1990s reduces the employment share of an industry belonging to low-skill group by0.87 percentage point, and raises the employment share of an industry belonging to high-skillgroup by 0.78 percentage point over the period 2000 to 2010. These IV estimates are largerin magnitude than OLS estimates in Column (1). Columns (3) and (4) repeat the regressionsin Columns (1) and (2), but replace the change in employment share with the change inoutput share in the manufacturing sector as dependent variable. The OLS estimates areinsignificantly different from zero. However, the 2SLS results are close to those in Column(2), albeit less precisely estimated.Thus far, the regression analysis has been restricted to industry composition withinthe manufacturing sector. To check the sensitivity of the results to the inclusion of non-manufacturing sector, in Columns (5) and (6), the dependent variable ∆Shareikt is re-definedas the change in manufacturing industry k’s share of total employment in prefecture i over theperiod 2000 to 2010. As is shown in Column (6), the estimates from the 2SLS regression areconsistent with baseline results. As a robustness exercise, I substitute the high school/college17The grouping of the manufacturing industries is discussed in Appendix B.1.19enrollment rates with the share of population with some high school/college education inEquation (1.13) as independent variable, and repeat the exercise in Table 1.5. The detailsand the results are discussed in Appendix A.4.9. The basic results are robust when the flowmeasures of skill supply are replaced by the stock measures.1.8 Counterfactual AnalysisThis section uses the model proposed in Section 1.3 to address the following counterfactualquestions: How did the trade liberalization in the past decades affect the aggregate welfare,income distribution and human capital accumulation across regions in China? What are theimplications of further trade liberalization on regional divergence in educational attainmentand industry specialization? Due to the lack of data on bilateral trade flows among prefecturesand between prefectures and the ROW, the quantitative analysis below defines regions andindustries at more aggregated level.1.8.1 Counterfactual Proportional ChangesLet xˆ ≡ x′/x denote the proportional change in any variable x between the initial andcounterfactual equilibria. The following system of equations solves for changes in wages(wˆi,h,wˆi,l), income (Yˆi), sectoral output (Yˆj,k), trade flows (λˆij,k) and labor supply (Hˆi, Lˆi):Yˆi,kYi,k =∑jλ′ij,kβkYˆjYj (1.14)λ′ij,k =λij,k(wˆαki,hwˆ1−αki,l τˆij,k)−ε∑i′ λi′j,k(wˆαki′,hwˆ1−αki′,l τˆi′j,k)−ε (1.15)wˆi,h = Hˆ−1i∑khi,kYˆi,k (1.16)wˆi,l = Lˆ−1i∑kli,kYˆi,k (1.17)Hˆi =( (wˆi,h/wˆi,l) κ1−νpii,h(wˆi,h/wˆi,l)κ1−ν + (1− pii,h))1−(1−ν)/κ(1.18)Lˆi =( 1pii,h(wˆi,h/wˆi,l)κ1−ν + (1− pii,h))1−(1−ν)/κ(1.19)∑kYˆi,kYi,k = YˆiYi (1.20)The above system is derived from equilibrium conditions 1.3-1.7 and maps the exogenouschanges in iceberg cost {τˆij,k} to outcomes {wˆi,h, wˆi,l, Yˆi, Yˆi,k, λ′ij,k, Hˆi, Lˆi}, given parameters{αk, βk, ε, κ, ν} and initial values {Yi,k, Yi, λij,k, hi,k, li,k, pii,h} from the data.The solution to the above system can also be used to capture changes in other relevantobjects, namely, school enrollment and aggregate welfare. It is straightforward to show thatpˆii,h =(wˆi,h/wˆi,l)κ1−νpii,h(wˆi,h/wˆi,l)κ1−ν + (1− pii,h)20Cˆi =∏k(λˆii,k)−βk/ε(wˆαki,hwˆ1−αki,l )−βk1.8.2 CalibrationIn this section, I calibrate the model to match fundamental moments in the data. The equi-librium system in Equations (1.14)-(1.20) takes parameters {αk, βk, ε, κ, ν} and certain initialvalues {Yi,k, Yi, λij,k, hi,k, li,k, pii,h} as given. The parameters include the skill intensities ofdifferent sectors (αk), the preference weight on goods from different sectors (βk), and theparameters governing the firm productivity distribution (ε) and worker productivity distri-bution (κ and ν). The initial values include a region’s sectoral outputs (Yi,k), total income(Yi), trade share (λij,k), sectoral employment shares (hi,k and li,k), and school enrollmentrate (pii,h). I briefly discuss the calibration here and provide a summary in Table A.11. Moredetails can be found in Appendix A.5.Parameters Observable from Data. To calibrate the trade share λij,k, I employ China’sregional input-output (IO) data for the year 2007. The IO table provides data on the bilateraltrade flows of 14 industries18 between 8 regions in China and each region’s trade with theROW. A map of the regions is shown in Appendix A.5. With the trade flow data, I thencalculate the expenditure share λij,k = Xij,k/∑i′ Xi′j,k, where Xij,k denotes the spending byregion j on good k from region i. The IO table also provides the data on sectoral output,Yi,k, and total income, Yi, for regions in China. The corresponding data from the ROWare obtained from the World Input-Output Database (WIOD) 2007, with the industriesaggregated to the same level as China’s regional IO data.The data on sectoral employment share for skilled and unskilled labor (hi,k and li,k) areobtained from the 2005 mini census for regions in China, and from WIOD Socio EconomicAccounts (WIOD SEA) 2007 for the ROW. For school enrollment rates, I employ data fromthe 2005 mini census for regions in China, and the 2005 data from Barro and Lee (2010)for the ROW. pii,h is calibrated as the share of population aged 20 to 24 with some collegeeducation. I assume that preferences and industry skill intensities are homogeneous acrossregions. αk is calibrated as the income share of workers with some college education inindustry k using data from the 2005 mini census. To calibrate βk, I use China’s regional IOtable and calculate the expenditure share of each sector k for China as a whole. The valuesof calibrated αk and βk are shown in Table A.12.Trade Elasticity. Simonovska and Waugh (2014) propose an unbiased estimator for ε usinga simulated method of moments. For the following quantitative analysis, I use their preferredestimate of ε = 4.14.Calibration of Worker Productivity Distribution. Following the strategy adopted byHsieh et al. (2013), I calibrate the parameter κ using within-group wage variation. As aworker’s wage in the model equals the value of her marginal product, variation in individualproductivity maps into variation in wages across workers. Let wzi be the wage of individual18The original table contains 17 industries. I aggregate all the non-tradable industries into one sector. Asa result, the modified table has 13 tradable industries and one non-tradable industry.21z in region i. It is straightforward to show that in region i the wages of educated workers(wzi|Educ) and uneducated workers (wzi|Uneduc) follow the Fre´chet distributions:wzi|Educ ∼ Fre´chet(wκi,hpi−(1−ν)i,h , κ) and wzi|Uneduc ∼ Fre´chet(wκi,l(1− pii,h)−(1−ν), κ)Hence, the coefficient of variation of wages within a region×education group in our modelsatisfiesCV 2 =V arianceMean2=Γ(1− 2/κ)(Γ(1− 1/κ))2 − 1 (1.21)where Γ(·) denotes the gamma function. Hence, κ is implicitly determined by the within-group coefficient of variation of wage.Income data come from the 2005 mini census. I restrict the sample to full-time salariedworkers aged between 23 and 55 with monthly income of at least 300 RMB.19 To estimate κ,I take residuals from the following cross-sectional regressionlnwzg = ρg + X′zξ + εzgwhere lnwzg is the log wage of individual z belonging to group g, ρg denotes region×educationgroup dummies, and Xz is a vector of individual controls including dummies for age, gender,occupation, industry, marital status, urban residence status, workplace ownership type20,and working hours in the previous week. Then, I calculate the mean and variance of theexponent of these wage residuals, and solve κ numerically according to Equation (1.21). Thepoint estimate of κ is 3.41.The parameter ν, which governs the correlation of zh and zl, cannot be directly cali-brated.21 Hence, following Hsieh et al. (2013), I use ν = 0.1 as the baseline value, and showthat the main findings are not sensitive to the alternative values.1.8.3 Eliminating External TradeTo investigate the effects of external trade liberalization in past decades across regions, I settrade costs between China’s region and the ROW to infinity.22 The counterfactual equilibriumis simulated using Equations (1.14)-(1.20). Table 1.6 displays the changes in educationalattainment and gains from external trade. Most regions in China are found to have higher skillpremium and hence higher educational attainment in the counterfactual where the externaltrade is shut down. The only exception is the North Municipalities, where school enrollmentis higher relative to other regions and the ROW, which makes it more prone to receiving19Full-time salaried workers are defined as those who are not self-employed and worked at least 30 hoursin the previous week. In 2005, depending on a region’s economic development, the minimum monthly wageranged from 235 to 690 RMB across China.20The workplace ownership type indicates whether an individual is working in a state-owned, collectivelyowned or privately owned enterprises.21The parameter ν captures the absolute advantage of an individual. Its calibration requires separateobservations of the wages of educated and uneducated workers. However, from the available data, it isdifficult to infer the counterfactual wage of an educated (uneducated) worker being an uneducated (educated)worker.22i.e. τˆij,k =+∞ if i 6=ROW & j =ROW;+∞ if i =ROW & j 6=ROW;1 otherwise.22high-skill shocks when external trade is liberalized. The last column presents the gains fromexternal trade, which are equal to the minus of the percentage changes in welfare. It showsthat the gains from external trade are unevenly distributed across regions. Not surprisingly,coastal regions gain more from external trade. The most open region, the South Coast, enjoysan increase in welfare of 7.3% whereas the inland Central region only experiences a slightincrease of 1.6%.1.8.4 Further Trade LiberalizationAs is established in Sections 1.6 and 1.7, surging trade in past decades exacerbated the differ-ences in skill abundance across regions in China, which in turn reinforced regional comparativeadvantage and industry specialization. These findings suggest that ongoing globalization hasimportant ramifications in terms of regional divergence in China. The general equilibrium ef-fects of further trade liberalization are analyzed in Table 1.7. Specifically, I lower the icebergcosts between China’s region and the ROW by 30% and evaluate the effects on school enroll-ment and welfare. Columns (2) and (3) show the results of a uniform reduction in trade costsfor all industries. Excepting the North Municipalities, most regions in China will experiencea decline in skill premium and consequently educational attainment in the counterfactual. Inaddition, the gains from further trade liberalization are substantial. The coastal regions willexperience increases in welfare of more than 15%. Even in the inland areas, welfare will riseby more than 5%.China is expected to engage in more trade with other developing countries in the future,which will bring in positive high-skill demand shocks. To investigate the associated implica-tions, I repeat the exercises in Columns (2) and (3), but only lower the external trade costs forthe high-skill manufacturing industries by 30%.23 Under this scenario, Central Coast, SouthCoast, and North Municipalities will see rises in skill premium and educational attainment,while the other regions will experience declines in skill premium and decumulation in humancapital. A reduction of trade barriers for the high-skill industries will also generate sizableincreases in welfare, ranging for 2.1% for North Coast to 9.6% for South Coast. Columns (6)and (7) repeat the exercises in Columns (2) and (3), but only lower the external trade costsfor the low-skill manufacturing industries by 30%.24 All regions in China will be deskilled bysuch a trade reform. Additionally, the increases in welfare are moderate, as shown in Column(7).To better assess the effects of trade on regional divergence in educational attainment,I conduct the counterfactual analysis in Column (2) repeatedly, by gradually reducing theiceberg cost by 0 to 50%. The results are displayed in upper left panel of Figure 1.5, whichplots the relation between proportional changes in school enrollment (pˆi) and the changes iniceberg cost (τˆ) across regions. Except in the North Municipalities, school enrollment declinesas the iceberg cost is reduced. Moreover, educational attainment diverges across regions astrade liberalization deepens. The upper right panel corresponds to scenarios in which tradeliberalization is restricted to high-skill industries, and the bottom left panel corresponds toscenarios in which trade liberalization is restricted to low-skill industries. The pattern of23The high-skill industries include Chemicals and Chemical Products, Machinery, Transportation Equip-ment, and Electrical and Optical Equipment.24The low-skill industries include Textiles, Apparels, Footwear and Leather Products, Wood and Productsof Wood and Cord, Non-Metallic Products, and Manufacturing, n.e.c..23divergence is observed in both cases.1.9 ConclusionBy analyzing local labor markets that are subject to differential export demand shocks ac-cording to initial patterns of industry specialization, this paper examines the effects of exportexpansion on skill acquisition in China over the period 1990 to 2005. Export expansion isfound to have differential effects on school enrollment, depending on a prefecture’s initialindustry composition. Prefectures initially specializing in high-skill industries experiencedrelatively faster human capital accumulation relative to prefectures initially specializing inlow-skill industries. These findings suggest that the between-sector Stolper-Samuelson effectappears to be strong in the context of China. Moreover, trade-induced enlarged differencesin skill abundance across prefectures reinforced initial industry specialization patterns.The findings suggest that international trade could amplify the differences in skill abun-dance across countries due to the mutually reinforcing relationship between comparative ad-vantage and skill formation. Although the benefits of international trade are often stressed,developing countries specializing in low-skill sector could experience a decline in educationattainment when integrating into the world market. The present paper ignores externalitiesof education. However, if positive externalities of education are strong, the decumulation ofhuman capital could undermine long-term gains from trade in a developing country. Theseconsiderations warrant future research on optimal policies balancing the trade-off betweenpromoting educational attainment and increasing export growth in low-skill sectors in thecontext of developing countries.24Figure 1.1: Spatial Distribution of Export Demand Shocks: 2000-05Panel A: Export Shock: Low Skill Panel B: Export Shock: High SkillPanel C: Differential Export Shocks25Figure 1.2: School Enrollment by Ages over YearsCompulsory SchoolingHigh SchoolCollege0. Rate7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22Age1990 2000 2005Figure 1.3: School Enrollment across Prefectures of Different Periods0.,20000 .2 .4 .6 .8 1Enrollment,1990High School Enrollment0.2.4.6Enrollment,20000 .2 .4 .6Enrollment,1990College Enrollment0.,20050 .2 .4 .6 .8 1Enrollment,2000High School Enrollment0.2.4.6Enrollment,20050 .2 .4 .6Enrollment,2000College Enrollment26Figure 1.4: Change in Educational Attainment and Export Shocks:Different Age Groups−.04−.020.0218 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35AgeHigh School, Low Skill−.010.01.0218 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35AgeHigh School, High Skill−.04−.020.0218 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35AgeCollege, Low Skill−.010.01.0218 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35AgeCollege, High SkillNote: The upper panel shows the effects low/high-skill export shocks on high school education attainment. Thelower panel shows the effects low/high skilled export shocks on college education attainment. All regressionsare weighted by the start of the period prefecture’s share of the age group population. All regressions controlfor the start-of-the-period share of population with some HS/college education and change in the log GDPper capita. Standard errors are clustered at province level.27Figure 1.5: Divergence in Education Attainment0.τˆˆAll sectors0.τˆˆHigh−skill sectors0.τˆˆLow−skill sectors  NortheastNorth MunicipalitiesNorth CoastCentral CoastSouth CoastCentralSouthwestNorthwest28Table 1.1: Summary Statisticsmean std 10th 25th 50th 75th 90th(1) (2) (3) (4) (5) (6) (7)Panel A: Change in School Enrollment: ∆Enroll16-18, 90-00 0.124 0.113 -0.024 0.056 0.129 0.204 0.26019-22, 90-00 0.043 0.090 -0.026 0.007 0.036 0.080 0.12416-18, 00-05 0.156 0.098 0.026 0.089 0.164 0.223 0.28119-22, 00-05 0.065 0.081 -0.017 0.016 0.057 0.111 0.16616-18, 90-05 0.281 0.149 0.077 0.175 0.293 0.387 0.47219-22, 90-05 0.108 0.111 0.010 0.042 0.098 0.159 0.250Panel B: Export Shocks (1000USD): ∆ExportLS, 90-00 0.183 0.223 0.016 0.049 0.101 0.220 0.479HS, 90-00 0.311 0.453 0.000 0.028 0.154 0.447 0.746LS, 00-05 0.542 0.664 0.046 0.147 0.306 0.650 1.382HS, 00-05 0.916 1.373 0.001 0.084 0.452 1.278 2.234LS, 90-05 0.724 0.886 0.062 0.195 0.406 0.876 1.856HS, 90-05 1.226 1.824 0.001 0.113 0.608 1.713 2.998Notes: Number of prefecture is 340.29Table 1.2: Changes in School Enrollment and Export Shocks∆Enroll ∆Enroll ∆Enroll ∆Enroll ∆Enroll ∆Highsch./∆College(1) (2) (3) (4) (5) (6)Panel A: Age 16-18∆ExportLS -0.058*** -0.051*** -0.024** -0.034*** -0.058** -0.019*(0.014) (0.008) (0.009) (0.011) (0.028) (0.009)∆ExportHS 0.008*** 0.007*** 0.006*** 0.005** 0.009** 0.009***(0.002) (0.002) (0.002) (0.002) (0.003) (0.002)Province×Year Y Y Y Y Y YControls Y Y Y Y YInitial Conditions Y Y Y YPHI and PLI Y YPrefecture YN 680 673 673 673 673 673R2 0.399 0.644 0.681 0.693 0.924 0.857Panel B: Age 19-22∆ExportLS -0.021*** -0.021*** -0.019*** -0.031*** -0.039** -0.023***(0.003) (0.005) (0.004) (0.005) (0.015) (0.002)∆ExportHS 0.012*** 0.010*** 0.010*** 0.007*** 0.010** 0.004**(0.002) (0.002) (0.002) (0.002) (0.005) (0.002)Province×Year Y Y Y Y Y YControls Y Y Y Y YInitial Conditions Y Y Y YPHI and PLI Y YPrefecture YN 680 673 673 673 673 673R2 0.145 0.597 0.695 0.761 0.920 0.770Notes: All regressions are weighted by the start-of-the-period prefecture’s age group population. Controlsinclude change in average age, change in sex ratio, change in share of Han ethnic group, change in share ofpopulation with urban Hukou, change in log fiscal expenditure on education, and change in log GDP per capita.Initial conditions include the start of period school enrollment rate and log GDP per capita. Standard errorsare clustered at the province level. *** p<0.01, ** p<0.05, * p<0.130Table 1.3: Changes in School Enrollment and Export Shocks: Controlling the Export Shocks of Other Prefectures(a) Neighbor Prefectures (b) All Other PrefecturesAge 16-18 Age 19-22 Age 16-18 Age 19-22(1) (2) (3) (4) (5) (6) (7) (8)∆ExportLS -0.034*** -0.057** -0.032*** -0.038*** -0.031*** -0.054* -0.029*** -0.036**(0.010) (0.025) (0.004) (0.013) (0.010) (0.028) (0.005) (0.016)∆ExportHS 0.006*** 0.009** 0.007*** 0.011** 0.008*** 0.012*** 0.008*** 0.011**(0.002) (0.003) (0.002) (0.005) (0.002) (0.004) (0.002) (0.005)∆ExportLS,N -0.004 -0.004 -0.005*** -0.007* -0.053** -0.062 -0.031*** -0.023(0.003) (0.005) (0.001) (0.004) (0.023) (0.043) (0.010) (0.026)∆ExportHS,N 0.001 0.000 -0.000 -0.001 0.037** 0.040* 0.017*** 0.008(0.001) (0.001) (0.001) (0.002) (0.014) (0.023) (0.006) (0.013)Province×Year Y Y Y Y Y Y Y YControls Y Y Y Y Y Y Y YInitial Conditions Y Y Y Y Y Y Y YPHI and PLI Y Y Y YPrefecture Y Y Y YN 673 673 673 673 673 673 673 673R2 0.695 0.924 0.767 0.923 0.703 0.926 0.764 0.920Notes: All regressions are weighted by the start-of-the-period prefecture’s age group population. Controls include change in average age,change in sex ratio, change in share of Han ethnic group, change in share of population with urban Hukou, change in log fiscal expenditure oneducation, and change in log GDP per capita. Initial conditions include the start of period school enrollment rate and log GDP per capita.Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, * p<0.131Table 1.4: Changes in School Enrollment, Export Shocks and Import ShocksAge 16-18 Age 19-22(1) (2) (3) (4) (5) (6)∆ExportLS -0.027*** -0.036*** -0.023*** -0.023*** -0.033*** -0.033***(0.009) (0.012) (0.007) (0.004) (0.005) (0.007)∆ExportHS 0.007*** 0.006** 0.006** 0.011*** 0.008*** 0.009***(0.002) (0.002) (0.002) (0.002) (0.002) (0.002)∆ImportLS 0.011* 0.007 0.015** 0.017* 0.008* 0.008*(0.006) (0.007) (0.006) (0.009) (0.004) (0.005)∆ImportHS -0.002*** -0.001** -0.002*** -0.002** -0.001** -0.001***(0.001) (0.001) (0.000) (0.001) (0.000) (0.000)∆ImportLS,CI -0.054*** -0.016(0.012) (0.016)∆ImportHS,CI 0.001 -0.001(0.003) (0.003)Province×Year Y Y Y Y Y YControls Y Y Y Y Y YInitial Conditions Y Y Y Y Y YPHI and PLI Y Y Y YN 673 673 673 673 673 673R2 0.682 0.693 0.698 0.697 0.762 0.762Notes: All regressions are weighted by the start-of-the-period prefecture’s age group population. Controlsinclude change in average age, change in sex ratio, change in share of Han ethnic group, change in share ofpopulation with urban Hukou, change in log fiscal expenditure on education, and change in log GDP per capita.Initial conditions include the start of period school enrollment rate and log GDP per capita. Standard errorsare clustered at the province level. *** p<0.01, ** p<0.05, * p<0.132Table 1.5: Changes in Skill Supply and Change in Industry Specialization∆ShareM ∆ShareEmployment Output EmploymentOLS 2SLS OLS 2SLS OLS 2SLS(1) (2) (3) (4) (5) (6)LS ×∆HighSch.Enrollt−1 0.025*** 0.132*** -0.003 0.153 0.012*** 0.080***(0.007) (0.050) (0.012) (0.098) (0.003) (0.029)MS ×∆HighSch.Enrollt−1 -0.010 -0.020 -0.008 -0.078 0.001 0.017*(0.007) (0.030) (0.010) (0.100) (0.001) (0.010)HS ×∆HighSch.Enrollt−1 -0.015** -0.113** 0.013 -0.069 -0.003 -0.025**(0.006) (0.046) (0.011) (0.092) (0.002) (0.012)LS ×∆College.Enrollt−1 -0.022*** -0.087*** -0.006 -0.088* -0.009*** -0.063***(0.008) (0.034) (0.008) (0.050) (0.002) (0.019)MS ×∆College.Enrollt−1 0.008 0.009 0.007 0.010 -0.002* -0.020***(0.006) (0.018) (0.009) (0.053) (0.001) (0.007)HS ×∆College.Enrollt−1 0.013** 0.078*** 0.001 0.079* -0.000 0.015*(0.006) (0.030) (0.011) (0.042) (0.002) (0.009)∆ShareEmpMt−1 Y Y Y Y∆ShareEmpt−1 Y YProvince×Ind Y Y Y Y Y YN 8,721 8,721 9,180 9,180 8,721 8,721Notes: All regressions control for the start of the period industry employment/output share. Standard errors are clusteredat the prefecture level. *** p<0.01, ** p<0.05, * p<0.133Table 1.6: Counterfactual: Eliminating External Trade(1) (2) (3) (4)Region piihwˆi,hwˆi,lpˆiih %GNortheast 0.167 1.008 1.027 2.778North Municipalities 0.452 0.993 0.986 6.353North Coast 0.118 1.013 1.044 1.805Central Coast 0.290 1.022 1.060 6.713South Coast 0.136 1.007 1.024 7.300Central 0.116 1.006 1.019 1.602Southwest 0.088 1.005 1.016 2.792Northwest 0.117 1.006 1.022 1.722Notes: Gains from trade %G = (1− Cˆ)× 100.Table 1.7: Counterfactual: Further Trade LiberalizationAll Ind. High skill ind. Low skill ind.wˆi,hwˆi,lpˆiih %Cwˆi,hwˆi,lpˆiih %Cwˆi,hwˆi,lpˆiih %C(1) (2) (3) (4) (5) (6) (7) (8) (9)Northeast 0.981 0.940 8.347 0.999 0.997 4.275 0.991 0.972 1.724North Municipalities 1.003 1.006 16.109 1.008 1.018 8.255 0.993 0.986 2.471North Coast 0.976 0.920 5.562 0.994 0.981 2.148 0.988 0.962 1.285Central Coast 0.974 0.930 15.212 1.013 1.034 6.996 0.972 0.927 3.382South Coast 0.989 0.965 15.933 1.012 1.038 9.614 0.984 0.948 1.774Central 0.989 0.964 5.003 0.996 0.986 3.283 0.996 0.987 0.765Southwest 0.989 0.961 8.206 0.998 0.992 4.559 0.998 0.992 1.467Northwest 0.989 0.964 5.541 0.997 0.991 3.521 0.995 0.982 1.00934Chapter 2Trade, Pollution and Mortality inChina2.1 IntroductionAmong the many extraordinary dimensions of China’s economic growth in the last 3 decadesis the contemporaneous spectacular boom in export performance: the annual export growthrate was 14% during the 1990s and 21% during the 2000s. At the same time, this rapideconomic growth has been accompanied by concerns that many of the benefits deriving fromhigher incomes may be attenuated by the similarly rapid deterioration in the environmentand increase in pollution. According to Ebenstein et al. (2015) many of the gains in healthoutcomes have been slowed down by a simultaneous rise in the concentration of pollutants. Inthis paper we study how the export boom in China between 1990 and 2010 affected pollutionand and infant mortality across different prefectures. We exploit variation across prefecturesin the initial pattern of comparative advantage to build local export shocks. This is similarto the approach by Topalova (2010), Kovak (2013) and Autor, Dorn and Hanson (2013) tostudy the effects of import competition on employment, with the important difference thathere we are interested in the export demand shock generated by the rest of the world. Wetherefore build an instrument that captures the part of export increase that is predicted bythe change in tariffs faced by Chinese exporters over time in different sectors.Why are we interested in this specific component of output growth? One may wonderwhy we set out to measure the effect of an expansion deriving from foreign demand. After all,doesn’t an extra unit of output generate the same extra pollution regardless of whether it isgoing to be exported or consumed domestically? The answer is generally no, unless they arecaused by the same type of underlying shock. That would be the case if, for example, therewas a decline in internal transport costs that increased domestic demand for a certain good,everything else equal. This would be similar to a foreign demand shock caused by a dropin tariff on that good. In general, though, all kinds of unobserved shocks, like productivityimprovements, may affect domestic output and exports, but they most likely would affectemissions differently. We therefore need an instrument that can isolate demand shocks, inthis case foreign, from other unobserved sources of output growth. To further motivateour focus on foreign demand shocks faced by China, let us restate that the period between1990-2010 is one of extraordinary integration of China in the world economy and is thereforea quantitatively important driver of output growth whose environmental consequences webelieve are worth studying.Our analysis relies on variation in export growth at the local level. Therefore the tariff-predicted export expansion at the national level is apportioned to the various prefecturesaccording to initial sectoral employment shares. We construct two export shocks at the35prefecture level: (i) PollutionExportShock represents the pollution content of export ex-pansion and is measured in pounds of pollutant per worker; (ii) ExportShock measures thedollars per worker associated with export expansion. The two measures differ because pre-fectures specialize in different products and while two prefectures may experience the sameexport shock in dollar terms, the one specializing in a polluting sector, like steel, experi-ences a larger PollutionExportShock. Borrowing the language of Grossman and Krueger(1995) and Copeland and Taylor (2003), ExportShock captures both a scale and an incomeeffect (or technique effect): prefectures with larger initial employment in industries that ex-perience large export shocks will see their income and production increase. The variablePollutionExportShock captures the interaction of export expansion and pollution intensity:prefectures with larger initial employment in industries that both experience large exportshocks and have high emission intensity are expected to become more polluted.We find that a one standard deviation increase in PollutionExportShock increases infantmortality by an additional 2.3 infant deaths per one thousand live births. The magnitude ofthis effect has to be gauged in the context of the evolution of infant mortality over this period.Between 1990 and 2010 infant mortality rate in China went from 36 per thousand to 5 deathsper thousand live births, but this decline hides substantial heterogeneity. Between 2000 and2010 for example, the 75th percentile prefecture experienced a decline of 23.7 deaths, while the25th percentile prefecture saw a drop of only 8.7. What our results imply is that increasedpollution due to specialization and exports explains about 17% of this differential declineacross different areas in China. Importantly, using the same data as Chen et al. (2013), wecan show that such effects are concentrated on mortality due to cardio-respiratory conditions,which are supposedly the most sensitive to air pollution. On the contrary, mortality due toother causes does not respond to any of the export shocks.While we find a strong and consistent impact of PollutionExportShock on mortality, theeffect of ExportShock is less robust. We find that ExportShock tends to decrease mortality,but the effect is statistically significant only during the decade 2000-2010 (during whichexport expansion was an order of magnitude bigger than during the 1990’s).Why do PollutionExportShock and ExportShock affect mortality? The next questionwe tackle is the quantification of the channels through which these two shocks influencehealth outcomes. The most intuitive channel through which PollutionExportShock affectsmortality is pollutant concentration, while ExportShock may affect mortality through differ-ent channels. On the one hand, an increase in income due to export expansion may increasethe demand for clean environment and the consumption of healthcare services which wouldin turn improve health outcomes. On the other hand, it may also increase the consumptionof environmentally unfriendly goods like cars, which would in turn raise pollution. Our iden-tification relies on the assumption that conditional on ExportShock, PollutionExportShockaffects mortality only through the channel of air pollution.We show that a positive PollutionExportShock increases the concentration of SO2, whileExportShock tends to reduce it. In the decade 2000-2010 a one standard deviation increasein PollutionExportShock increases SO2 concentration by 6.3 µg/m3 while a one standarddeviation increase in ExportShock decreases SO2 concentration by 1.9 µg/m3, but this lat-ter effect is not significant. These changes represent respectively 22.5% and −6.7% of theinterquartile range, i.e. the difference between the 25th and 75th percentile prefectures interms of SO2 change during 2000-2010.36Finally, we close the loop by showing how pollution affects infant mortality, a link whichhas been studied before, but for which we offer a different identification strategy. We find theelasticity of infant mortality to SO2 to be 0.9, which is quantitatively similar to the estimateby Tanaka (2015) of 0.82 for China, albeit during a different time period. The elasticity ofIMR to PM2.5 is 2.1, which is not directly comparable to the estimate of 1.73 we have forChina by Chen et al. (2013) because the pollutant in that case is total suspended particles.A potential explanation of this larger effect is that PM2.5 is considered much more fatal dueto the smaller diameter of the particles.We are careful in addressing a series of issues that may affect confidence in these results. Inparticular one may be concerned that official sources for data on pollution may systematicallymisreport pollutant concentrations in order to hide imperfect compliance with environmentalregulation from the public. In this regard we check the correlation of the official daily pollutionlevels with the levels reported by the American Embassy and Consulates in 5 Chinese cities.We show that the correlation is above 94 percent for most of the series even though thelevels reported by the American Embassy are generally higher than the ones reported byChinese official sources. Another issue that we delve on is the quantitative importance oftrade policy shocks for the overall structure of production and level of pollution. We take aspecific episode, the steel safeguard tariffs imposed by the US in 2002-2003 to show that forprefectures with heavy steel production pollution decreases relative to control prefectures in2001 and increases back up in 2003.We go through a number of robustness checks to rule out pre-existing trends, and weinclude other variables like export shocks in neighboring and upwind prefectures, importshocks, local energy production and take into account input-output linkages that transmitforeign demand shocks to upstream industries. We also analyze the results by gender and byage, finding a relatively homogeneous effect across different groups.Given the battery of results that we present, what can we say about whether the exportboom in China was good or not for the environment, and ultimately for the health of (inparticular young) individuals? Given our strategy, the extent to which we can answer thisquestion is limited to the cross-section of prefectures we analyze. We can only say whether aprefecture that was more exposed to export expansion than another benefited or not, relativeto that prefecture. If all prefectures in China benefited from trade expansion by a commonamount, then this is a factor that we obviously cannot capture. That said, our results docapture all benefits that should accrue differentially to different prefectures proportionallyto their export exposure, and this is measured by our variable ExportShock. What we findis that, on average, the effect of ExportShock is smaller and less significant than the effectof PollutionExportShock and therefore this implies that the export boom in China broughtabout an increase in infant mortality in the average prefecture, with important differencesacross prefectures.2.1.1 Relation to the LiteratureOur study contributes to two main strands of the literature, the one related to trade and pol-lution and the one studying the effect of pollution on mortality. The first generally addressesthe question of whether international trade affects pollution through a variety of channels.Employing the language introduced by Grossman and Krueger (1995), Copeland and Taylor(2003) and Copeland and Taylor (2004), increased international trade can: i) lead to a more37intense scale of production which increases pollution (scale effect) ; ii) induce specialization,which could reduce or increase pollution depending on whether a country specializes in cleanor dirty industries (composition effect); and iii) generate an increase in income which wouldraise the demand for better environmental quality (technique effect). Antweiler, Copelandand Taylor (2001) find that emissions across several world cities depend positively on thescale of economic activity and the capital abundance of the country and depend negativelyon income. Their main finding in relation to the trade-environment link is that, as a countryis more open to trade, on average emissions are lower. Their cautiously optimistic conclusionis that trade may be good for the environment. By their own recognition, the issue of iden-tification and unobserved heterogeneity is not fully solved in their paper. Differently fromthe panel approach of Antweiler, Copeland and Taylor (2001), Frankel and Rose (2005) limittheir analysis to a cross-section of countries and employ a geography-based IV approach.They identify that, controlling for income, increased trade leads to lower emissions. Ourcontribution is to take a step further in the direction of identifying the causal effect of tradeon environmental quality and health. Our within-country approach necessarily controls forseveral unobserved variables that are not accounted for by country-level panel studies, but wealso adopt several techniques to deal with other potential sources of endogeneity. The costof our approach, relative to country-level analysis, is that we necessarily ignore national-levelgeneral equilibrium effects.Our results also highlight the importance of industry composition for explaining differ-ences in pollution across Chinese prefectures. Although the approach and the question issomewhat different, it still relates to the importance of the composition effect studied forexample by Levinson (2009), who finds it to be very small in the US. With a state-of-the-artmicro-founded trade model, a recent paper by Shapiro and Walker (2015) confirms Levin-son’s finding that composition effects play very little role in explaining the dramatic declinein emissions in the US in the 1990s and 2000s. Essentially the fact that trade with developingcountries has allowed the US to shift towards the production of cleaner products like serviceshas played a smaller role than the overall decline in emission intensity at the product level.We contribute to a second literature that studies the impact of pollution on mortality, inparticular of infants. The reason why infant mortality is often chosen as a relevant outcomeis not only that young children are a particularly vulnerable members of society and thatper se may be of particular interest, but also because their health outcomes are more closelyrelated to immediate environmental conditions, while adults’ health may be the consequenceof factors accumulated over the course of many years. These studies are conducted bothin developed countries like Chay and Greenstone (2003a), Chay and Greenstone (2003b),Currie and Neidell (2005) and Currie, Neidell and Schmieder (2009), but also in developingcountries, like Greenstone and Hanna (2014), Arceo, Hanna and Oliva (2016) and Tanaka(2015). It is important to consider the main differences between a study like ours and theone for example by Currie et al. (2015). Currie et al.’s study is based on much richer data:individual birth records with a host of socio-economic characteristics of the mother and weeklymeasures for pollution. The high quality of the data is their primary line of attack on the issueof unobserved heterogeneity potentially generating a spurious correlation between mortalityand pollution. We believe that the disadvantages generated by access to more aggregate datain our paper are at least partially compensated by the advantage of an instrumental variableapproach to account for unobserved heterogeneity.38In terms of specific studies on trade and pollution in China, we are only aware of a fewpapers. An earlier paper by Dean (2002) considers the link between openness and water pol-lution across Chinese provinces, but it essentially exploits national level measures of opennessand therefore estimates the relationship using pure time variation whereas our entire strategyrelies on exploiting differential shocks within China. The most closely related paper to oursis by de Sousa, Hering and Poncet (2015) who exploit city-level variation in exports and findthat increased processing trade in China leads to lower pollution. In the energy and environ-mental science literature some studies, like Lin et al. (2016) and Yan and Yang (2010) haveaddressed the global impact of China’s trade on various pollutants, but they do not identifythe effect on China itself and its air quality.The rest of the paper proceeds as follows. Section 2.2 describes the various data sources,while section 2.2.6 probes the quality of specific variables, like air quality and mortality.Section 2.3 shows, through a simple event study, that even a single trade policy measureimposed by the US affected air quality in China’s steel-manufacturing areas. In section2.5 we construct our export shock measures and present our identification strategy in twoparts: i) we first show the reduced form effect of PollutionExportShock and ExportShockon mortality; ii) we then show that export shocks affect mortality through pollution. Section2.6 discusses our main results and 2.7 reports a number of robustness checks. We concludein section Data2.2.1 Local Economies and Employment DataThe unit of analysis is a prefecture in China, which is an administrative division rankingbetween province and county. Prefectures are matched across census years according to the2005 administration division of China, so that the data has a geographic panel dimension.There are 340 prefectures, with median land area of 13,152 km2 and median population of3.2 million in year 2000. The size of the prefecture makes it a relevant unit of analysis.Moreover, migration rate among prefectures is low, due to the household registration system.According to the 2000 Census, there are less than 4.5% of population aged between 16 and59 who changed the prefecture of residence during the past five years. This inter-prefecturemigration rate increased slightly to 4.8% in 2005.25 In contrast, the five-year migration rateacross states is around 12.5% for US in 2000 (Kaplan and Schulhofer-Wohl, 2013) and thefive-year migration rate across districts in India (which is a similar administrative division asprefecture in China) is around 13.5% in 2007 (Marden, 2015).The information on industry employment structure by prefecture is from the 1% sub-samples of the 1990 and 2000 China Population Censuses. The census data contain relevantinformation regarding the prefecture of residence and the industry of employment at 3-digitChinese Standard Industrial Classification (CSIC) codes.2625The figure is from 20% subsample of 2005 China 1% Population Sampling Survey.261990 Census employs CSIC 1984 version and 2000 Census employs the CSIC 1994 version. We reconcilethe two versions and create a consistent 3-digit CSIC code. There are 148 industries in the manufacturingsector.392.2.2 Export and Tariff DataFrom the UN Comtrade Database, we obtain the data on China’s export and import valuesat 4-digit International Standard Industrial Classification (ISIC) Rev.3 codes for years 1992,2000 and 2010.27 Data on export tariffs faced by Chinese exporters by destination countriesand import tariffs imposed by Chinese government on 4-digit ISIC Rev.3 industries are fromthe TRAINS Database. We map the trade and tariff data to the 3-digit CSIC sectoralemployment data from the population censuses, using the concordances between ISIC andCSIC.2.2.3 Pollution DataIndustry Pollution IntensityWe construct the pollution intensity for each 3-digit CSIC industry, employing the data fromthe World Bank’s Industrial Pollution Projection System (IPPS) and China’s environmentyearbooks published by Ministry of Environmental Protection (MEP). The IPPS is a list ofemission intensities, i.e., emission per value output, of a wide variety of pollutants by 4-digitStandard Industrial Classification (SIC) codes. These data were assembled by the WorldBank using the 1987 data from the US EPA emissions database and manufacturing census.28We aggregate the data to 3-digit CSIC level and consider the pollutants sulfur dioxide (SO2),total suspended particles (TSP ) and nitrogen dioxide (NO2) in the analysis.To address the concern that China’s industry pollution intensity may be uniformly higherthan that of the US, we use the data on 2-digit sector pollution intensity from MEP to adjustthe level. In particular, the pollution intensity for each pollutant p, of 3-digit CSIC industryk, is imputed following the steps: (i) using industry output as weight,29 aggregate the 3-digitIPPS data γp,IPPSk to 2-digit CSIC γp,IPPSK , where K is a 2-digit CSIC sector; (ii) for each3-digit industry, calculate the ratio of its pollution intensity to the corresponding 2-digitsector pollution intensity, i.e., rp,IPPk = γp,IPPSk /γp,IPPSK ; (iii) impute pollution intensity for3-digit CSIC according to γp,MEPk = rp,IPPk × γp,MEPK . Therefore, while the level of industrypollution intensity is adjusted, the within sector heterogeneity retains the feature of the IPPSdata. To account for the changing industry pollution intensity over time, we use the 1996and 2006 data from MEP and construct measures for each decade t. The pollution intensityγp,MEPkt is employed to build the pollution export shock as is discussed in Section 2.5.Data on Pollution ConcentrationInformation on annual average concentrations of SO2 is collected for the years 1992, 2000 and2010. The data is obtained from China’s environment yearbooks, which report the data on airpollution for 77, 100, and 300 cities/prefectures for years 1992, 2000 and 2010, respectively.30We supplement this main dataset with the information gathered from the provincial/city271992 is the first year when the export data is available for China.28To our best knowledge, there is no analogous data at such disaggregated level for China.29The data on output by industry is from Chinese Industrial Annual Survey.30Only SO2 is continuously published in China’s environmental yearbooks over the sample period. Theconcentration of TSP was reported in 1990 and 2000, however, it was replaced by PM10 in 2010.40statistical yearbooks, government reports and bulletins. Restricting to the prefectures withat least two readings, an unbalanced panel data is complied which covers 203 prefectures.To address the concern that the pollution readings may be subject to manipulations bylocal governments, we employ the satellite information on PM2.5 provided by NASA.31 TheNASA dataset contains information on three-year running mean of PM2.5 concentrationfor grids of 0.1 degree by 0.1 degree since 1998. Adjacent grid points are approximately 10kilometers apart. For the purpose of our analysis, we employ the data of years 2000 and2010 and construct the decadal change in PM2.5 at prefecture level. Specifically, for eachcounty-year observation, we calculate the average PM2.5 concentration using the data of thegrid points that lie within the county. Then the county-level data is aggregated to prefecturelevel, weighted by the county population.Daily Data on Air Pollution in ChinaThe daily data on air pollution index (API), an overall measure of ambient air quality, isfrom Ministry of Environmental Protection of China. The dataset records the API of majorcities in China starting from 5th June 2000. The number of cities covered increases from 42in the early sample period to 84 by the end of our sample period. Over 2001 to 2005, theaverage API declined from 83 to 73.322.2.4 Mortality DataOur empirical analysis mainly focuses on infant to circumvent the problem of unknown life-time exposure to pollution and other unobserved factors that jeopardize health. Because ofthe low migration rates of pregnant women and infants, the location of the exposure is mostlikely be the location of residence. Moreover, the environmental hazards are more likely totake immediate effects on infants due to their vulnerability. Also, infant deaths present alarge loss of life expectancy. From our data, the mortality rate in the first year of life is higherthan that in the next thirty years combined.The infant mortality rates (IMR) are constructed from the China Population Census foryears 1982, 1990, 2000 and 2010. Each census recorded the number of births and deathswithin a household during the last 12 months before the census was taken. The total numberof deaths of the age 0 is collected for every county, and then aggregated the to prefecture level.The total number of births by prefecture is derived in the same way. The infant mortalityrate is defined as the number of deaths of age 0 per 1000 of live births. In addition to IMR,we assemble the data on mortality rate of young children aged 1-4 at prefecture level for years1990, 2000 and 2010.33We supplement the census mortality data with the vital statistics obtained from the 1992and 2000 China’s Disease Surveillance Points (DSP) system. The DSP collects birth anddeath registration for 145 nationally representative sites, covering approximately 1% of the31We use the Global Annual PM2.5 Grids from MODIS, MISR and SeaWiFS Aerosol Optical Depth (AOD),v1 (1998-2012) dataset from NASA’s Socioeconomic Data and Applications Center (SEDAC). The data onPM2.5 are derived from Aerosol Optical Depth satellite retrievals, using the GEOS-Chem chemical transportmodel, which accounts for the time-varying local relationships between AOD and PM2.5.32For the balanced panel of 42 cities, the average API declined from 86 to 73.33The mortality rate of young children aged 1-4 is defined asDeaths1−4Deaths1−4+Poplution1−4 × 1000.41national population. The data recorded whether or not an infant died within a calendar year,and if he did, the cause of death, using the International Classification of Disease 9th Revision(ICD-9) codes. We match each surveillance site to the prefecture where it is located,34 andclassify the causes of death into several categories: cardio-respiratory illness, infant-specificcauses (including congenital anomalies and perinatal conditions), digestive illness, infectiousillness, malnutrition, and external causes (including accidents and violence).352.2.5 Other Demographic, Socioeconomic and Wind DataWe collected other demographic and socioeconomic variables at prefecture level, includingGDP per capita, provision of medical care, sex ratio at newborns, share of population withdifferent education attainments, share of agricultural employment, and population densityby prefecture from various provincial statistical yearbooks and population censuses. Thedistance to the nearest port for each prefecture is calculated using the information fromWorld Port Index. In addition, for the period of 2000 to 2009, we obtain information onoutput by 3-digit CSIC industry, fossil fuel energy production, and production shares ofstate-owned enterprises (SOE) and foreign firms at prefecture level, from Chinese IndustrialAnnual Survey.36 The data on wind direction for each prefecture is from NOAA IntegratedSurface Global Hourly Data. Appendix B.1 provides more detailed discussions on the datasources and construction of variables.2.2.6 Quality Assessment of the Chinese Data Pollution and MortalityIn this section we address the concern that official reports from the Chinese government maynot be fully reliable due to the desire to under-report pollution and mortality. With regards topollution, in order to assess the severity of underreporting, we have to consider the incentivesof officials at various levels of government in the period considered between 1990 and 2010.As reported by Chen et al. (2013), although the data on pollution were collected starting inthe late 1970’s, they were not published until 1998, so it is unlikely that fear of public uproarwould be a concern for local officials. More importantly, in a number of studies Jia (2014) andJia, Kudamatsu and Seim (2015) report that officials most likely perceived local economicgrowth to be the criterion for promotion, rather than environmental quality. In fact Jia (2014)shows how increased pollution is a byproduct of the quest for higher economic growth byaspiring politicians. Moreover, our identification strategy compares the changes in pollutantconcentration of prefectures with different initial industrial specialization. Therefore, ourresults will be contaminated only if the pollution data were systematically manipulated forprefectures with different initial industry composition. Despite all these considerations, onemight still be concerned that our pollution measures are very noisy, so in the Appendix we34The surveillance sites are primarily at the county level.35Cardio-respiratory illness includes all causes under ICD-9 codes 25-28, 31 and 32; infant specific illnessincludes the causes under codes 44 and 45; digestive illness includes the causes under codes 33 and 44;infectious illness includes the causes under 1-7; external reason include all the causes under codes from 47-53and E47-E56; and the malnutrition include the causes under code 19. In 1992, the shares of infant deaths dueto cardio-respiratory, infant-specific, digestive, infectious, external causes and malnutrition are 28.5%, 46.6%,1.9%, 6.2%, 5.8% and 0.3%, respectively.36The data set includes all the state owned firms and non-state firms with revenue above 5 million RMB(approximately 800 thousand US dollar).42corroborate our data by showing that the official Chinese daily data on air quality has acorrelation of at least 0.94 with the US Consulate or Embassy data, depending on the city.In the appendix we also show the results of an exercise aimed at detecting over- or under-reporting of infant mortality. In essence we compare the number of 10-year-old children in aprefecture in a given census year with the expected number of 10-year-old children based onthe reported mortality and birth figures from the last population census (a decade earlier).We find a correlation of 0.98 between these two measures, which of course cannot perfectlycoincide due to unaccounted-for migration.2.3 Preliminary Event Study: the 2002 US Steel SafeguardMeasuresBefore building comprehensive measures of export shocks that cover the entire manufacturingsector, it is worthwhile to pause and consider whether export shocks generated by foreigndemand are realistically large enough to affect local pollution measures. First, we shouldconsider that exports are a large share of China’s output. The average annual share of exportsto GDP was 27% in the 1990’s and 17% in the 2000’s, so it is plausible that changes in foreigndemand may substantially alter the pattern of production. Nevertheless we investigate theissue through a specific event, i.e., the US imposition of safeguard tariffs on imports ofsteel products in March 2002 (and its removal in December 2003). On 28th June 2000, theUS Trade Representative (USTR) requested the US International Trade Commission (ITC)commence a Section 201 investigation on whether steel imports of 612 different 10-digit HSproduct categories were causing injury to the domestic industry. The USITC investigationcovered imports with a combined value of some $17 billion, more than half of total USimports of steel in 2001. On 22nd October 2001, the ITC announced its findings that 85%of the imported products subject to investigation had caused injury to the domestic steelindustry, and in December 2001, the ITC announced its non-binding recommendation forsafeguard tariffs and quotas. On 20th March 2000, President George W. Bush announced theapplication of safeguard tariffs and quotas on 272 different 10-digit HS product categories,which were significantly higher than that recommended by the USITC Commissioners.37(Read, 2005; Bown, 2013) To retaliate against the US safeguard measures and mitigate theassociated diversionary effects, the EU imposed temporary safeguard measures on its steelimports on 28 March 2002 and introduced the definite safeguard measures on 29 September2002. According to the World Bank Temporary Trade Barriers Database, 225 different 10-digit HS product categories were subject to investigation, among which 53 were imposed bysafeguard tariffs. China also initiated its own safeguard measures on 20th May 2002. On4th December 2003, the US lifted all the safeguard tariffs and the EU and China removedtheir measures in the same month. Figure B.1 summarizes the timeline of the 2002 steelsafeguards, with different colors indicating different stages including investigation, provisionalmeasures, final measures and scheduled liberalization. Figure B.2 shows that, relative to otherindustries, steel export dropped in 2001 and 2002, and re-bounced to the pre-safeguard levelin 2004. The export contraction may contribute to the decline in output during 2002.37185 products received a 30% tariff, 60 received a 15% tariff, 15 received a 13% tariff and 7 received a 8%tariff in the first year. In 19th March 2003, the tariffs for each of the categories stepped down to 24%, 12%,10% and 7%.43As documented by Read (2005), it is generally understood that tariffs were imposed forreasons related to domestic political consideration and are unlikely to be related in theirtiming and magnitude to events happening in China. This event is useful for our studybecause it pertains an industry whose activity is highly polluting, like steel. According to theWorld Bank IPPS data on emission intensity employed by Levinson (2009), SIC industry 331and 332 are in the top 10% industries by emissions of both SO2 and particulate matter.38 Weexploit this event to detect whether a temporary protection measure in the US that raisedimport tariffs on several steel products affected air quality relatively more in prefectures thatproduce more steel in China. We are interested in the differential level of air quality insteel-producing regions relative to other regions before and after the steel safeguards. Thespecification we employ is the following:APIit = αi + βry + γrm + ShareSteeli ×NoSGt + εit ,where APIit is the Air Pollution Index in prefecture i on day t, ShareSteeli is equal to theshare of employment in steel sectors39 and the dummy NoSG is equal to 1 in the periodbefore the investigation and after the revocation of the safeguards (the time window of thesteel safeguard corresponds to the days between 28/06/2001 and 04/12/2003).40 Finally, αiis a prefecture dummy, while βry and γrm are, respectively, region-year and region-monthdummies.41 Standard errors are clustered at the prefecture level. The different specificationsin Table 2.1 employ different samples of months and interact the variable ShareSteeli withdifferent policy time dummies: AfterSGt is equal to 1 in the period after the termination ofthe safeguard policy and BeforeSGt is defined similarly. Our findings indicate that beforeand after the period during which the policy was in place, the API was higher in steel-producing cities. The effect is not very large, but strongly significant. Column (8) of Table2.1 indicates that the API (which averages 78.7) decreased by 2.53 × 0.8 = 2 points duringthe policy months in prefectures that had the average share of steel employment relative toa prefecture that had no employment in steel.2.4 Theoretical frameworkIn this section we present a simple Ricardian model of trade and pollutant emissions thatrationalizes our empirical specification. The setup is a standard Eaton-Kortum style model(see Eaton and Kortum, 2002) with multiple sectors as in Costinot, Donaldson and Komunjer(2012) and fixed emission intensities by sector. Consider a world economy that features38Steel mills closures have been used in Pope III (1989), Ransom and Pope III (1995) and Pope III (1996)to detect the effect of particulate matter concentration on health outcomes.39Employments under the code 32 and 34 of the CSIC 2002 classification, which pertain to steel and steel-related products. Employment data is from the 2000 population census.40We find that exports from China to the US start declining during the investigation phase. This “investiga-tion effect”, originally analyzed theoretically and empirically by Staiger and Wolak (1994), was also detectedin the same context by Bown (2013).41There are 8 regions: Northeast (Heilongjiang, Jilin and Liaoning), North Municipalities (Beijing andTianjin), North Coast (Hebei and Shandong), Central Coast (Shanghai, Jiangsu and Zhejiang), South Coast(Guangdong, Fujian and Hainan), Central(Henan, Shanxi, Anhui, Jiangxi, Hubei and Hunan), Southwest(Guangxi, Chongqing, Sichuan, Guizhou, Yunnan and Tibet) and Northwest (Inner Mongolia, Shanxi, Gansu,Qinghai, Ningxia and Xinjiang).44multiple prefectures in China, indexed by i = 1, ..., NC , and multiple regions in the rest ofthe world (henceforth ROW), indexed by i = NC + 1, ..., N , and K sectors, k = 1, ...,K.Each sector features multiple varieties, indexed by ω. Preferences are described by a Cobb-Douglas upper-tier utility function (with consumption shares βk) and a lower-tier CES utilityfunction. Each sector is characterized by an emission intensity γk which is equal to the ratioof emissions divided by the value of output and is assumed to be fixed.42 There is only onefactor of production, labor, and the production function for variety ω of good k in region itakes the following linear form:Qik(ω) = zik(ω)Lik(ω) ,where Lik(ω) denotes the labor employed in region i to produce variety ω of good k. The asso-ciated labor productivity is represented by zik(ω), and it is drawn from a Fre´chet distributionFik (·), that is:Fik(z) = exp[−(z/zik)−θ] for all z ≥ 0 .We assume that there is a large non-manufacturing sector that also employs only labor andthat determines the wage wi. Trade between regions is costly and τijk denotes the iceberg costof shipping good k from region i to j. We maintain the standard assumption that τijk ≥ 1 ifj 6= i and τiik = 1. Markets are assumed to be perfectly competitive, and each region importsfrom the lowest cost supplier. The producer price for each variety ω is given by pik(ω) =wi/zik(ω). The value output of sector k in region i is Yik =∑ω pik(ω)Qik(ω) = wiLik.Following Eaton and Kortum (2002), the value of exports of good k from prefecture i inChina to region j in the ROW is determined by:Xijk = λijkβkYj ,where λijk denotes the share of expenditure on good k in region j that is allocated to theproducts from prefecture i. This share λijk depends on production and transportation costsaccording to the following expression:λijk =(wiτijk/zik)−θ∑Ni′=1(wi′τi′jk/zi′k)−θ .We can calculate the size of each sector in each region, as approximated by the employmentin the sector Lik, as follows:wiLik =N∑j=1Xijk =N∑j=1λijkβkYj (2.1)Finally, total emissions is simply given by:Pi =K∑k=1γkYik .42We assume fixed emission intensities not only because of simplicity, but mostly because we do not haveaccess to micro-data at the prefecture level that would allow us to test predictions regarding the effect of tradeon production techniques. The assumption of fixed emission intensities is also made in Shapiro (2015).452.4.1 Changes in Transport Costs: Deriving Export Demand ShocksThe exogenous shocks in this model come from changes in iceberg costs {τˆiRk}, where i isa prefecture in China and we denote by R the set of all other regions in the rest of theworld. Hats over variables denotes log changes (xˆ ≡ d lnx). We assume that all regions inChina face the same export cost and that internal trade costs remain unchanged, that isτˆiRk = τˆi′Rk = τˆRk and τˆii′k = 0. Total differentiation of equation (2.1) giveswidLik =XiRkXRkdXRk , (2.2)where dXRk is the change in exports of good k from China to the ROW due to a change intransport costs.43 Then, the total change in emissions is given by:dPi = γcXiRcXRcdXRc + γdXiRdXRddXRd . (2.3)We have access to measures of pollution concentration at the region i, Ci, where air pollutantlevels are measured per cubic meter. Since the amount of air is roughly proportional to theunderlying area of land, our measure of pollutant concentration Ci will be proportional toPiTi, where Ti denotes the land area of region i. We can therefore rewrite equation (2.3) as:dCi ∝∑kγkXiRkXRkdXRkTi︸ ︷︷ ︸PollutionExportShocki. (2.4)Equation (2.3) sheds light on how external demand shocks at the national level lead to dif-ferential environmental impact across prefectures in China. In particular, we show thata prefecture receives a larger export pollution shock if it specializes in dirty industriesthat experienced larger declines in trade costs. The reduced form pollution export shockPollutionExportShocki captures both scale and composition effects as highlighted in Antweiler,Copeland and Taylor (2001). Moreover, the weighted average structure of export pollutionshock resembles the empirical approach in the literature on the local effects of trade (Autor,Dorn and Hanson, 2013; Topalova, 2010; Kovak, 2013). However, it reflects the pollutioncontent embodied in the trade cost induced export growth.2.5 Empirical SpecificationIn this section we lay out our empirical methodology and explain our identification strategy.We start by describing how we construct our key explanatory variables, PollutionExportShock(expansion in exports measured in units of pollutant) and ExportShock (expansion in ex-ports measured in dollars). We then introduce our instrument for export expansion, which43More specifically equation (2.2) is derived as follows:widLik = −θ(1− λRk)XiRkXRkXRkdτRkτRk=XiRkXRkdXRk ,where dXRk = −θ(1− λRk)XRk dτRkτRk and λRk is China’s total market share of good k in the ROW.46consists of tariffs faced by Chinese exporters in different sectors. Finally we introduce our twooutcome variables: mortality and pollution concentration. We illustrate schematically thecausality links that we will explore in Figure 2.1. “Export” tariffs, i.e., tariffs that Chinese ex-porters face, affect the extent of export expansion, measured by PollutionExportShock andExportShock, which ultimately affect mortality, potentially through pollution concentrationas further discussed below.2.5.1 Pollution Export Shocks and Export ShocksIn this section we build the empirical measures that captures exports shocks and that willassess the relationship between export expansion and pollution changes in individual pre-fectures in China. We expect increased exports to affect pollution through two potentialchannels, which we capture with two types of export shocks.i) Increased foreign demand induces an increase in total manufacturing production asderived in the model (the scale effect). Increased exports may also increase local wages andprofits, which, through an income effect may increase the demand for clean air, thus reducingpollution. Although this technique effect is ignored by our model in section 2.4, it will bepotentially contained in the effect of ExportShockit.ii) The amount of additional pollutants generated as will depend on whether expan-sion is concentrated in dirty or clean industries. These scale and composition effects arecaptured in our theoretical framework in section 2.4 and will be measured by the variablePollutionExportShockpit.We focus on channel ii) first. Because we will analyze the outcome in terms of differentpollutants, we denote by γpkt the pollution intensity for pollutant p, i.e. the total amount ofemissions in sector k, P pkt, divided by the value of output Ykt so that γpkt =P pktYkt. If we hadaccess to the value of exports by individual prefecture i then we could find the impact ofexport expansion on local pollution changes using the following equation:∆Cpit =∑kγpkt∆XiRktTi, (2.5)where ∆Cpit measures the change in concentration of pollutant p in prefecture i between yeart − 1 and year t and ∆XiRkt is the analogous change in export value from prefecture i insector k. Prefecture level exports could in principle be calculated from firm-level customsdata, but such data are not available for the earlier time period in our sample. Moreover, weargue that, even if such data were available, they may be more likely to be affected by localproductivity shocks, which we are attempting to isolate from our analysis. In line with themodel prediction in equation 2.4, we rewrite the link between pollution concentration andlocal export expansion as follows:∆Cpit =∑kγpktXiRk,t−1XRk,t−1∆XRktTi. (2.6)where ∆XRkt denotes China’s export growth in industry k, Lik,t−1 and Lk,t−1 are respectivelyprefecture i’s employment in industry k and China’s total employment in industry k in theprevious period t − 1. We use employment share Lik,t−1Lk,t−1 to proxy for a prefecture’s export47share in industry k,XiRk,t−1XRk,t−1 because export data at the prefecture level are not available forthe earlier time period (1990) of our sample.44Thus far, we posit that local pollution concentration is proportional to emissions nor-malized by land area. Empirically, however, this approach faces the problem that someprefectures include vast desert areas where pollutant presence is very low, and populationis almost absent. Ideally we would like to know the size of the populated area, but in theabsence of such information we use the size of the population (as proxied by the size ofthe workforce in i) to replace land area. Therefore PollutionExportShockpit, our empiricalmeasure of export-induced pollution in prefecture i, is constructed as follows:PollutionExportShockpit =∑kγpktLik,t−1Li,t−1∆XRktLk,t−1, (2.7)and it measures the pounds of pollutant p associated with export expansion measured ona per worker basis. The normalization by local employment that we discussed above servesthe additional purpose of making our PollutionExportShockpit measure easily comparable toour second measure of export shock, which we define simply as ExportShockit. This secondmeasure, which addresses channel ii), i.e. the impact of income on environmental outcomes,is constructed as follows:ExportShockit =∑kLik,t−1Li,t−1∆XRktLk,t−1, (2.8)and it measures the dollar value of export expansion in prefecture i, also on a per workerbasis. This is the equivalent of the change in value of imports per worker at the commutingzone level in Autor, Dorn and Hanson (2013). The variation across prefectures of our twomeasures, PollutionExportShockpit and ExportShockit stems from initial differences in localindustry employment structure, a feature common to the Bartik approach (see Bartik, 1991).We analyze more in detail the properties of these shocks in the context of our discussion ofidentification, which we cover in the section Specification 1: Total Effect of Export Shocks on MortalityIn this section we describe our approach to identifying the causal impact of a decline intransport costs on pollution and mortality across prefectures in China. Our first specificationis the following:∆IMRit = α1 + α2PollutionExportShockpit + α3ExportShockit + εit (2.9)where ∆IMRit is the change in infant mortality date in prefecture i between year t− 1 andt, while εit is an error term that captures other unobserved factors. In the following sectionwe address issues related to endogeneity. In order to focus attention, we provide concreteexamples of demand and supply shocks that could threaten identification and describe themethodologies we employ to tackle them.44We also employ the more theoretically correct export shares for the second decade (2000-2010) in one ofour robustness checks.482.5.3 Identification StrategyOur basic specification (2.9) relates infant mortality to our two export shocks, ignoring otherpotential socio-economic determinants that could be important drivers of mortality. Wetherefore include several control variables that capture education, provision of health ser-vices, ethnic composition and income. Even after the inclusion of such variables, we arestill concerned that the error term εit may be affected demand and supply factors that arecorrelated with our export shock measures.Bartik ApproachThe first type of shocks we may be concerned about is local productivity or factor supplychanges that may affect local output and exports and affect pollution concentration at thesame time. Both measures PollutionExportShockpit and ExportShockit, through a Bartikapproach, tackle this issue by not employing export expansion at the local level, but ratherusing a weighted average of national export expansion. As usual this approach relies on theassumption that other time-varying, region-specific determinants of the outcome variable areuncorrelated with:(a) region’s initial industry composition, and(b) industry shocks at the national level.We will address issue (a) in three ways. The first is to control for pre-existing trends ininfant mortality, so that we can account for the possibility that a region initially specializedin polluting industries may be on a different trajectory in terms of overall health outcomes.The second way we address issue (a) is to check that we cannot predict current infant mortalitychanges using future export shocks, thereby again ensuring that the two are not driven bya common unobserved factor. Our third approach to dealing with (a) is to control for thefollowing variable, PollEmploymentpit, which measures the level of pollution implied by theinitial employment structure in prefecture i:PollEmploymentpit =∑kγpktLik,t−1Li,t−1. (2.10)Essentially we are concerned that regions initially specialized in dirty industries may justhave initially more lax regulation and therefore be prone to relax such regulations even more.We may then mistake such effect as the consequence of export expansion. Controlling forPollEmploymentpit makes sure that we are comparing two prefectures with the same initialaverage level of specialization in dirty industry, which likely summarizes their attribute to-wards regulation, among other factors. Consider two prefectures specializing, respectively, insteel and cement and assume both sectors have very similar pollution intensities. As a result,the two prefectures have similar value of PollEmploymentp, indicating they have similarinitial pollution level. Nevertheless, they may experience different PollutionExportShockp,if for example, steel receives a larger external demand shock.4545Notice that this is an average implied pollution level and it does not control for the entire composition ofemployment. If we were to control for the the entire vector of employment shares there would be of course novariation in the variables of interest PollutionExportShockpit and ExportShockit.49Export Tariffs and Export ShocksOne of the main concerns regarding the Bartik approach is that an industry clusters in aspecific region and the region also highly specializes in it. In this case, requirement (b) will beviolated because the national shock will essentially coincide with the local shock. And this iswhen the Bartik “apportionment” will fail to generate an exogenous local shock starting fromnational shocks. In this case we need to introduce variation that can isolate national exportchanges due to changes purely in foreign demand. Otherwise both PollutionExportShockpitand ExportShockit may capture other types of local (in this case also national) supply shocks.We choose to employ variation in average tariff rates faced by China when exporting to therest of the world, which we denote by ExportTariff . More specifically ExportTariffkt insector k and time t is defined as a weighted average of the tariff τkjt imposed by country j,where the weights depend on time t− 1 Chinese exports to country j, Xkj,t−1:ExportTariffkt =∑jXkj,t−1Xk,t−1τkjt .We believe changes in these tariffs to be mainly determined by political considerations inother countries and therefore to be rather exogenous to China’s internal shocks. Neverthelesswe need to check that changes in ExportTariff are indeed uncorrelated with various shockswithin China. In particular we verified that changes in ExportTariffkt are uncorrelatedwith: (i) changes in domestic demand across different sectors; (ii) changes in value addedper worker (as a proxy for productivity growth) across sectors; (iii) emission intensities (i.e.cleaner industries are not being liberalized at a different pace from dirty ones).We posit that the growth in total exports can be explained by a decrease in the level oftariffs faced by exporters, so we adopt the following specification:lnXkt = ηk + φt + γ ln(1 + ExportTariffkt) + εkt , (2.11)where ηk and φt are sector and time fixed effects. We report the results of this regressionin Figure 2.2. The estimated coefficient implies that a 1% increase in the tariff faced byexporters decreases exports by 7.8%. Our estimate is within the range of gravity equationestimates of the effect of bilateral trade frictions as in (Head and Mayer, 2014), althoughtoward the higher end. We obtain the fitted value of the logarithm of exports in equation(2.11), then take the exponential of such predicted value to obtain X̂kt:X̂kt = exp(ηˆk + φˆt + γ̂ ln(1 + ExportTariffkt)) . (2.12)We employ predicted exports from (2.12) in changes, i.e., ∆X̂kt, to construct instrumentsfor our export shocks of interest. Note that ∆X̂kt is the empirical counterpart of dXRk asdiscussed in section 2.4.We estimate equation (2.9) using instrumental variables that are constructed using pre-dicted exports derived in the previous section, equation (2.12). The two instrumental vari-ables ̂PollutionExportShockpit and ̂ExportShockit are constructed as follows:̂PollutionExportShockpit = ∑kγpktLik,t−1Li,t−1∆X̂ktLk,t−1,50̂ExportShockit = ∑kLik,t−1Li,t−1∆X̂ktLk,t−1.Importantly, this IV approach also addresses another problem that may bias our estimates,i.e. omitted variables. Imagine for example that export expansion and domestic demandwere positively correlated. Since we do not have data on changes in domestic demand forthe earlier period, our estimates of α2 may be upward biased because it would contain theeffect of export expansion and domestic demand. Because, as we said earlier, ExportTariffis uncorrelated with domestic demand changes, we have a valid instrument to tackle thisproblem of omitted variable bias.2.5.4 Specification 2: Pollution Concentration ChannelOur second specification attempts to identify the specific channels through which exportshocks affect mortality. In particular we posit that PollutionExportShockpit affects mortalityonly through its effect on pollution concentration while ExportShockit may affect mortalitythrough its potential negative effect on pollution or through its general impact on income,which may increase demand for healthcare and in general affect living conditions of children.These considerations are represented in the diagram of Figure 2.1 and are reflected in ourchoice of specification, which is composed of two equations. The first is the mortality equation,which is similar to (2.9):∆IMRit = δ1 + δ2∆PollConcpit + δ3ExportShockit + νit , (2.13)where ∆PollConcpit is change in pollutant p concentration in prefecture i between year t− 1and year t. We again use an IV approach with instrumental variables ̂PollutionExportShockpitand ̂ExportShockit to disentangle the effect on mortality of increases in pollution caused byexport expansion and income effects of export booms. Let us reiterate that the exclusionrestriction here is that PollutionExportShockpit does not independently affect mortality oncepollution concentration is accounted for.The second equation is the pollution concentration equation and it relates export shocksto PollConcpit:∆PollConcpit = ρ1 + ρ2PollutionExportShockpit + ρ3ExportShockit + µit (2.14)with the same instruments ̂PollutionExportShockpit and ̂ExportShockit employed to identifythe causal effects of different export shocks on pollution concentration in a given prefecture.2.6 Results2.6.1 Summary StatisticsBefore delving into the results we briefly describe the data summarized in Table 2.2. We focuson the two outcome variables of interest, infant mortality rate (IMR) and pollution concen-tration, and on the two shocks of interest, PollutionExportShockpit and ExportShockit. InPanel A we see that IMR has declined dramatically over the period 1982-2010 from an av-erage of 36 deaths per thousand live births to just above 5 per thousand. Moreover, there is51substantial heterogeneity in infant mortality both in levels and in changes over time. Morespecifically the 1982 IMR was 14 in the prefecture at the 10th percentile and 67 at the90th percentile. In 2010 a similar disparity persists: at 10th percentile IMR is 1.4, while atthe 90th it is almost 11, so we may conclude that in relative terms heterogeneity in infantmortality across provinces has increased. This is a pattern we can detect by looking at thepercentiles of decade changes in IMR. Between 1990 and 2000 for example, although on aver-age all preceture saw a decline in IMR, the prefectures at the 90th percentile saw an increaseby 9 per thousand. We seek to explain part of this pattern through export shocks that havedifferentially hit different prefectures.Panel B shows that different Chinese prefectures are exposed to very different sulfur diox-ide and particulate matter concentrations46. While the average prefecture in 2000 featureda concentration of SO2 about 43 micrograms per cubic meter, this measure went from 12µg/m3 at the 10th percentile to 92 µg/m3 at the 90th percentile. To put these numbersinto perspective 20 µg/m3 is the 24-hour average recommended by the World Health Orga-nization,47 which implies that 75% of Chinese cities did not comply with the recommendedthreshold in 2000. The data on changes in SO2 concentration over time show even moreheterogeneity. Although the average prefecture saw a decline of 5 µg/m3, the standard devi-ation of the change was 33 and more than half the cities saw a deterioration in sulfur dioxideconcentration during the 2000s.Panel C reports the variable PollutionExportShockpit as change in pounds of pollutantembodied in exports per worker in a given prefecture. Although it is not easy to gauge themagnitude of this shock, it is easy to verify that it varied substantially, since for all pollutants,SO2, TSP and NO2 the standard deviation of the shock is most of the time higher than themean. The two maps in Figure 2.3 show that the variation was not clustered in certainprovinces, and that even within provinces different prefectures experienced different levels ofPollutionExportShockit.Panel D reports the variable ExportShockit as change in exports in 1000 dollars perworker. Notice first that the export in the 2000s was one order of magnitude larger thanthe shock in the 1990’s. During the 1990’s the average prefecture saw an increase in exportsper worker of 151 dollars, while in the 2000s that figure was 1,440 dollars. In both periodsthe standard deviation is larger than the mean, with substantial heterogeneity displayed byexport shocks (in the 2000’s the 10th percentile prefecture saw an increase of only 220 dollars,while the one at the 90th percentile experienced a surge of 3,100 dollars per person).2.6.2 Results for Specification 1: Total Effect of Export Shocks onMortalityIn this section we report the results of estimating the effect of our two shocks of interestPollutionExportShock and ExportShock on infant mortality as shown in equation (2.9).The results appear in Table 2.3. The three panels differ by the pollutant employed to con-struct the PollutionExportShock variable. Panels A, B and C report, respectively, the effectsof the export content of sulfur dioxide, total suspended particles and nitrogen dioxide. Allcolumns present instrumental variables regressions as detailed in section 2.5.2. In column46The concentration of PM2.5 here comes from NASA.47The data are obtained from “Air quality guidelines: global update 2005: particulate matter, ozone,nitrogen dioxide, and sulfur dioxide” published by World Health Organization.52(1) we find a positive and significant (at the 1% level) effect of a pollution export shock oninfant mortality. Columns (1) through (7) employ different controls and fixed effects. Inparticular column (1) controls for region×year dummies to account for omitted factors thatcould be contemporaneously evolving in different regions and decades in China and that couldbe correlated with our export shock variables. Column (2) controls for the initial value ofthe following variables: log GDP per capita, overall mortality rate, agriculture employmentshare and population density. Column (3) controls for contemporaneous changes in the fol-lowing variables: log GDP per capita, share of boys, share of population with middle schooleducation, share of population with high school education or above, number of hospital bedsper capita, agricultural employment share, and distance to the nearest port. In column (4)we add a 2nd degree polynomial in the lag change in IMR, which flexibly control for theprefecture-specific pre-determined trends in IMR. The regression result is unaffected, a factwhich alleviates the concern that our estimates are confounded if the different secular trendsacross prefectures which are associated with the initial industrial composition. In column (6),in order to address the same concern, we control for the average initial pollutant emissionsimplied by the start-of-the-period employment structure, i.e. PollEmploymentpit as describedby equation (2.10). Column (7) replaces Region×Year dummies with Province×Year dum-mies, thereby reducing the amount of variation in the export shocks and the coefficient ofinterest. The last three specifications all find a negative, but not always significant effect ofthe variable ExportShockit, which captures the effect of export expansion on local incomeand therefore health outcomes. In addition, all regressions pass the weak instrument test.For example, the Kleibergen-Paap rk Wald F statistic in column (7) is 28.3, well above theStock-Yogo critical value for 10% maximal IV size.We now comment on the magnitude of these effects. One extra pound per worker of SO2raises IMR by 0.317-0.537 extra deaths per thousand live births depending on the specifi-cation. Because of its ability to better account for local changes in unobservable variables,our preferred specification is in column (7). Because the magnitude of this shock varies bydecades, it is worth explaining the resulting effects separately. A one standard deviationincrease in PollutionExportShockSO2 in the 1990’s brings about 1.1 extra deaths per thou-sand (5.9% of a standard deviation in IMR change over the same period). The equivalentnumber for the 2000’s is 2.5 extra infant deaths per thousand live births (17.5% of a standarddeviation in IMR change over the same period). Using the statistically significant estimatein column (7) of Panel C to measure the effect of ExpostShock on mortality, we find thata 1990’s standard deviation increase in export per capita causes 0.17 fewer deaths per thou-sand live births, while the equivalent effect for a 2000’s standard deviation is 1.78 fewerdeath per thousand live births. We will evaluate the robustness of these results to additionalconsiderations in section Results for Specification 2: Pollution Concentration ChannelSo far we explored the “reduced form” effect of export shocks on mortality, but we haveignored the channels through which exports shocks operate. In this section we explore theeffect that export shocks have on mortality through pollution concentration as summarizedby equations (2.13) and (2.14) and the diagram in Figure 2.1. Both equations are estimatedby instrumental variables employing the instruments described in section 2.5.4. Table 2.4reports the effects of the two types of export shocks on the air concentration of SO2 (Panel53A) and PM2.5 (Panel B). We find that PollutionExportShock has a positive and signifi-cant effect on pollution concentration for both pollutants, as expected through compositionand scale effects, while ExportShock has a negative effect on pollution, as implied by thetechnique effect, but the effect is often not statistically significant. The set of controls andfixed effects are analogous to Table 2.3. According to column (3) of Panel A, a one 2000’sstandard deviation increase in PollutionExportShock causes SO2 concentration to rise by anadditional 6.3 µg/m3, while a one 2000’s standard deviation increase in ExportShock causesconcentration to fall by 1.9 µg/m3 (not statistically significant).In Table 2.5 we explore the effect of pollution concentration on infant mortality, stillallowing for export shocks to have a separate effect on IMR through the effect of exportshocks on income. Notice that in this table we report, for comparison with other studies, theOLS estimates of the relationship between mortality and pollution concentration. Columns(1), (3) and (5) show that such correlation is not significantly different from zero, a result thatis easily explained by the fact that a rise in pollution concentration can be due to increasedeconomic activities and the associated increase in income can reduce mortality. Nevertheless,we find a significant and positive effect of the change in pollution concentration on mortalityonce we adopt an IV approach. This is true for both pollutants SO2 and PM2.5. Becausethe link between pollution concentration and infant mortality has been estimated by otherstudies, to make our results comparable, we express them in terms of elasticities. We findthat the elasticity of IMR to PM2.5 to be 2.1, while the elasticity of IMR to SO2 is 0.9.Table 2.6 reports estimates from other studies to facilitate comparison.Our estimate of the elasticity of IMR to SO2 concentration is quite similar to the oneestimated for China by Tanaka (2015) which is 0.82. There is no direct comparison for theelasticity or IMR to PM2.5 and our estimate of 2.1 is higher than estimates based on TSPand PM10. We believe this higher elasticity is justified by the higher risk of damage causedby fine particulate matter which is capable of penetrating more deeply in the lungs. Thestronger effect of fine particulate matter is documented in (Pope III et al., 2002).2.6.4 Effects of Pollution on Infant Mortality by Cause of DeathIn this section we provide additional evidence to corroborate the finding that the increasedmortality we detect is indeed due to pollution. We employ a source of data that has beenpreviously explored in Chen et al. (2013) to measure the effect of increased pollution due tocoal-fueled heating on mortality in China. Because we only have data for 117 prefectures forthe first decade, we cannot include the same rich set of fixed effects as we did in Table 2.3.We still include region fixed effects and control for pollution implied by initial industrial com-position PollEmployment. In Table 2.7 we report the estimates for PollutionExportShockusing SO2 as pollutant, but the appendix reports the equivalent tables for TSP and NO2.We find that only mortality due to cardio-respiratory causes is sensitive to the pollutioncontent of export. Infant mortality classified as related to digestive, infectious, external andinfant-specific causes do not appear to be sensitive to export shocks. These results reassureus that we are finding the effects where it is reasonable to expect them.542.7 RobustnessIn this section we return to the baseline “reduced form” results of Table 2.3 and introducea number of control variables that account for several potential confounding factors. Eachcolumn of Table 2.12 is discussed in the following sections and while this table employsPollExportShockSO2 , appendix Tables B.5 and B.6 perform the same robustness checks forother sources of air pollution. The analogous robustness tests for the link of export shocksand pollution concentration are discussed in Appendix B. Robustness: Neighboring Shocks and Wind DirectionIn Table 2.8 we consider the impact of export shocks experienced by neighboring prefectures.In principle, there are at least two channels through which export shocks in neighboringprefectures can affect local pollution and health outcomes. The first one is through interme-diate goods. If a neighboring prefecture of prefecture i experiences an increase in exports, itwill increase its demand for intermediate inputs, some of which may come from prefecturei and could in turn generate extra production and therefore pollution and adverse healthconsequences in prefecture i. The other possibility is that pollution generated in nearbyprefectures may simply be transported by wind to other areas. In order to capture these ef-fects we construct two variables, which are explained in detail in the appendix. The variablePollutionExportShockp,Nit is an average of the pollution export shocks experienced by prefec-tures that share a border with prefecture i. In order to measure the effect of wind-transportedpollution, variable WindPollutionExportShockp,Nit gives a larger weight to neighboring pre-fectures that are in the upwind position relative to prefecture i. If the intermediate channelis the only one active, then WindPollutionExportShockp,Nit should not affect mortality in ionce we control for PollutionExportShockp,Ni,t . The regression results are presented in Table2.8. As shown in columns (1) and (2), regardless of the weight, a positive neighboring exportpollution shock is estimated to have significantly positive effect on IMR. More importantly,the estimates of the coefficients for local shocks remain similar to the baseline regression withthe inclusion of non-local shocks. This finding suggests that the local pollution affect IMRindependently of cross-border spillovers. The regression in column (3) includes both mea-sures of neighboring shock. We find that WindPollutionExportShockk,N has significantlypositive effect on IMR, while the estimated coefficient of PollutionExportShockk,N becomesinsignificant. We take this finding as suggestive evidence that the neighboring pollution shockaffects IMR by bringing in wind-borne pollutants. Columns (4) and (5) repeat the regres-sions in Columns (2) and (3), with WindPollutionExportShockk,N constructed using theinformation from the two nearest weather stations instead of the nearest. The results remainrobust.2.7.2 Robustness: Future shocksAs we have already mentioned, one of the drawbacks associated with the Bartik approach isthat the initial industrial composition may be correlated with other unobserved characteris-tics that also affect infant mortality. More concretely, we may be concerned that prefecturesinitially highly specialized in polluting sectors may have already been experiencing relativelyhigher increases in child mortality because, for example, they had a more lax enforcement55of environmental regulation. Aside from pre-trends controls which we have introduced pre-viously, here we perform a falsification exercise where we regress the current change in IMRon future pollution export shocks. Table 2.9 finds no correlation between IMR and futureshocks and moreover the sign is reversed relative to our main regression in Table 2.3. Thisfinding suggests that prefectures hit by larger export shocks were not already experiencingrelatively higher mortality rates.2.7.3 Robustness: Mortality by GenderTable 2.10 investigates the effects of export expansion on IMR by gender, using the speci-fication of column (5) in Table 2.3. The results are qualitatively similar for both genders,which is consistent with the priori that air pollution harms infant’s health indiscriminately.Quantitatively, however, it is found that the effects on girls are larger in magnitude thanboys. One possible explanation is that in the context of China, due to the traditional pref-erence for boys, parents could be more likely to take measures to minimize a newborn son’sexposure to pollution or to seek medical treatment for his illness. This echoes the findingsin Jayachandran (2009) that the air pollution caused by wildfires in Indonesia had largeradverse effect on the mortality of newborn girls than boys.2.7.4 Robustness: Mortality of Young Children Aged 1-4Table 2.11 examines the effect of export expansion on the mortality in early childhood. Wefollow the specification (2.9), but replace the dependent variable with change in mortalityrate (MR) of children aged 1-4. Column (1) shows that one extra pound per worker of SO2increases MR of the age 1-4 by 0.037 per thousand. Moreover, it is estimated that 1000USD export expansion reduces MR of age 1-4 by 0.11 per thousand. The coefficients of bothPollutionExportShockSO2 and ExportShock are significant at 1% level. Due to the lack ofdata on mortality rate of children aged 1-4 from the 1982 census, we are not able to controlfor the pre-trends of MR for the full sample. Column (2) includes the quadratic terms ofchange in IMR in the previous decade, which in effect account for the common secular trendsof IMR and MR in the early childhood. The coefficients of interest change little.48 Columns(3)-(4) and (5)-(6) present the consistent results for TSP and NO2, respectively.2.7.5 Energy ProductionThe analysis so far has employed, in constructing our PollutionExportShock, data that onlyaccounts for the direct emissions generated in the production process, but it does not includeemissions due to the generation of electricity needed for production. The reason why onlydirect emissions are usually included in the intensity measures is that one would need toknow the source of the electricity and that depends on the region where firms are located,regardless of the industry. For our purpose, if electricity generation is not accounted forin our pollution export shock, we may be under-estimating the increase in pollution due toexport expansion. At the same time, electricity generation may happen in other provincesand sufficiently far from where production takes place, so that its effect will not be felt in the48We verify that the results remain robust if we restrict the sample to years 2000 and 2010 and control forthe quadratic terms of change in MR for age 1-4 in the previous decade.56prefecture where the export-induced demand for power is arising.49 In order to address theseconcerns, we introduce a control at the prefecture level which is the amount (also measuredin dollar value of output per worker) of electricity generated by fossil fuel. The magnitudeand significance of the PollutionExportShock is not affected, but we find that expansion inenergy production is a significant predictor of increases in infant mortality.2.7.6 ImportsWe have so far disregarded imports in accounting for the link between trade, pollution andmortality. There are three reasons for this asymmetry of treatment. First, China’s tradesurplus has grown considerably over 1990-2010 from 5.5 billions to 336 billions, implying thegap between exports and imports has widened.50 The second reason is that, in principle,increase import inflows may have two distinct effects on pollution. If imports replace lo-cal production, then increased imports would likely reduce pollution and mortality, but ifimports are concentrated in intermediate inputs, then a surge in imports may spur furtherlocal production and cause pollution. The third reason is that, while we are reasonably con-fident about the exogeneity of changes in tariffs faced by Chinese exporters, import tariffsestablished by China are unlikely to be uncorrelated with other industry factors that causepollution. Despite all these considerations, we present a specification where we introduce aPollutionImportShock, constructed analogously to our PollutionExportShock by just re-placing export value with import value and employing ̂PollutionImportShock constructedusing import tariffs as an instrument.51 Column (2) of table 2.12 finds an insignificant effectof PollutionImportShock on mortality.2.7.7 Other ControlsIn column (3) we control for a prefecture-level variable, introduced by Li (2016), which mea-sures the extent to which export expansion increases demand for high skill workers. Theconcern is that what PollutionExportShock is capturing expansion in low-skill industries(rather than dirty ones) and that this may be correlated with enforcement of environmentalregulations, if regions with more educated workers demand higher environmental standards.Although this scenario seems to be confirmed in the data, our coefficient of interest is notaffected. In column (4) we introduce two variables that capture the role played by SOE’s(state-owned enterprises) as studied by Dean and Lovely (2010). We confirm their findingsin that the share of SOE’s is positively correlated with increases in mortality, while the shareof foreign firms is not. While we cannot exclude other channels that link these variables, ourfindings are consistent with the logic that SOE’s are subject to less stringent controls andtherefore may be disproportionately responsible for increased pollution. In column (5) weintroduce a variable that captures differences in environmental regulation due to the TwoControl Zones (TCZ), as described by Tanaka (2015). Our finding of a positive relation-49As an example, in 2008 in Guangdong around 40% of the electricity was imported from outside theprovince. (China Energy Statistical Yearbook, 2009)50Autor, Dorn and Hanson (2013) also focus on imports from China to the US due to the large and growingtrade imbalance between the two countries.51The construction of ̂PollutionImportShock is analogous to that of ̂PollutionExportShock with exporttariffs replace by import tariffs.57ship between mortality and the TCZ dummy is consistent with the view that those stricterenvironmental regulations were introduced in more polluted cities.2.7.8 Alternative Measures of External Demand ShocksOne may concern that external demand shocks may induce production expansion of inter-mediate goods, and as a result extra emissions of pollutant. In particular, our measure willunderstate the pollution shocks in the prefectures specializing in dirty intermediate goods. Toalleviate this concern, we use the information from China’s input-output tables and constructalternative pollution export shock and income export shock as follows:52PollutionExportShockp,IOit =∑kγpktLik,t−1Li,t−1∆Y ktLk,t−1,ExportShockIOit =∑kLik,t−1Li,t−1∆Y ktLk,t−1.∆Y kt is the component of industry k of the vector (I − C)−1∆Xt, where I is an identitymatrix, and C is the matrix of input-output coefficients and ∆Xt is the vector of indus-try export expansion during the period t. We also construct the corresponding instrumentŝPollutionExportShockp,IOit and ̂ExportShockp,IOit by replacing ∆Xt with tariff-predicted ex-port growth ∆X̂t. Aligned with our baseline results, column (6) shows that pollution exportshocks has significantly positive effect on IMR, while the estimated coefficient of income ex-port shock is statistically indifferent from zero. In addition, we find that a standard deviationincrease in PollutionExportShockIOit of SO2 increases IMR by 2.8 per thousand of live births,which magnitude is similar to our baseline findings.53Finally, for the period 2001-2010 we have data on output, instead of just exports, so we re-place exports with total production and create a PollutionOutputShock and anOutputShock,but still adopt the same IV strategy described in section 2.5.4. The result is completely inline with the ones in Table Change in Infant Mortality Rate and Export Expansion byIndustry GroupsAs is discussed in Section 3.3, we employ the information of pollution intensity from theUS data to impute China’s 3-digit industry pollution intensity. A potential concern is thatan industry’s emission intensity varies across countries and over time, due to different andchanging abatement technologies. In this section, we use the information of industry pollutionranking and examine the effects of export expansion of different industry groups on IMR.The concern of measurement error in industry pollution intensity could be alleviated for tworeasons. First, the relative pollution intensities across industries are likely to be inherentin the dissimilarities of production processes.54 In other words, the ranking of pollution52We use 1997 input-output table to construct export shocks over 1992-2000 and 2007 input-output table toconstruct export shock over 2000-2010. Results remain similar if we use 1997 input-output table to constructexport shocks for both decades. More details can be found in the Appendix B.1.53The standard deviation of PollutionExportShockSO2,IOit is 38.2 over the period 1992 to 2010.54For example, the production of steel is dirtier than electronic products in both 90’s and 00’s, and in bothChina and US.58intensities is likely to preserve over space and time. Second, the grouping of industrieseffectively reduces the measurement errors introduced by a handful of industries.Specifically, the CSIC industries are ranked according to the pollution intensity of SO2,and the ones belonging to the bottom, middle and top tertiles are classified into Clean,Medium and Dirty group, respectively. The measures of local economy’s export expansion ofdifferent pollution intensity groups are constructed according toExportShockKit =∑k∈KLik,t−1Li,t−1∆XktLk,t−1,where K denotes the sector which industry k belongs to, and K ∈ {Clean,Medium,Dirty}.By construction, ExportShockKit captures the exposure in dollar per worker to export expan-sion in sector K. We instrument ExportShockKit with the IV constructed as follows:̂ExportShockkit = ∑k∈KLik,t−1Li,t−1∆X̂ktLk,t−1,where ∆X̂kt is the tariff-predicted export growth of industry k over the period t as derived insection 2.5.3. Then, we investigate the effects of export shocks of different pollution intensitygroups on IMR by estimating the following equation∆IMRit = κ1 + κ2ExportShockCit + κ3ExportShockMit + κ4ExportShockDit + νit .Columns (1)-(4) in Table 2.13 report the estimated coefficients of overall, clean, mediumand dirty export expansion on IMR, respectively. We only detect a significant effect of dirtyexport expansion on IMR. It is estimated that a 1000 USD ExportShockD increases IMR by11.3 per thousand births. Column (5) include all clean, medium and dirty export expansionsas independent variables in one regression. The estimated coefficient of ExportShockD re-mains similar to column (4). Moreover, we find a significant effect of clean export expansionon IMR, with a 1000 USD ExportShockC reducing IMR by 0.8 per thousand births. Thesecounteracting effects illustrate that the effect of export on pollution depends on whetherexpansion is concentrated in dirty or clean sectors.To account for the prefecture-specific pre-trend that may be correlated with the initialindustry structure, in column (6), we additionally control for a prefecture’s start-of-the-period employment share of clean, medium and dirty industry groups, i.e., EmpShareKit−1 =∑k∈KLik,t−1Li,t−1 . As a result, the variation of export exposures stems from the differences inindustrial mix within dirty, medium and dirty sectors. Despite the identification is madethrough different data variation, the estimated coefficient of ExportShockD remains similarto that in column (4). However, the estimated coefficient for ExportShockC becomes in-significant. In conclusion, consistently with our baseline results in Table 2.3, we find that alarger pollution content of export increases IMR.2.8 Concluding RemarksIn the 20 years between 1990 and 2010 China experienced a tremendous increase in itsexposure to international trade. China’s exports went from 62 billion USD in 1990 to 1.559billion USD in 2010, while its trade surplus went from 8.7 billion in 1990 to 336 billion USDin 2010. Even as a share of GDP, trade surplus has been increasing over time: it was 2%in the 1990’s and almost 6% in the 2000’s. In this paper we ask whether this export boomgenerated additional pollution that affected health outcomes during the same period. Weare particularly interested in isolating the effect of increased demand from abroad due to thereduction in trade barriers because export expansion can be due to a number of factors, amongwhich productivity changes. Why are we disregarding productivity changes? This is becausethose shocks would also increase exports, but it would be hardly appropriate to attribute theirenvironmental and health consequences as due to trade. The thought experiment we havein mind is to hold technology constant and consider only export expansion due to increaseddemand from countries that now impose lower tariffs on goods coming from China. Ouridentification strategy relies on two components: i) we instrument export expansion withtariff reductions faced by Chinese exporters; ii) we exploit initial differences in the industrialcomposition of different prefectures to costruct local export shocks.We find that the pollution content of export affects mortality. Take two prefectures withinthe province Jiangsu, north of Shanghai: Zhenjiang experiences an PollutionExportShocklarger than Taizhou by about a standard deviation. Our estimates imply that Zhenjiang seesits infant mortality increase relative to Taizhou by 1.1 and 2.5 extra deaths per thousand livebirths respectively in the 1990’s and 2000’s because of export expansion. This is equivalent toabout 15% of the variation in decline in IMR across prefectures. In reality during the 2000’sZhenjiang’s IMR declined by less than Taizhou’s IMR55. We find that a plausible reason whymortality increases is that PollutionExportShock affects concentration of air pollutants: ourestimates again imply that Zhenjiang sees an increase in SO2 concentration by 6.3 µg/m3(about 20% of the standard deviation of SO2 change over the period) relative to Taizhou.Although the dollar-value export shock has the opposite sign and tends to decrease infantmortality and pollution, we find that effect to be less often significant.We have purposely emphasized that our identification strategy based on differences acrossprefectures in export exposure is meant to identify only the relative differences in mortalitycaused by pollution induced by trade. Although we still believe that it is the most cautiousinterpretation, we offer now a quantitative assessment of the effects we have analyzed, butnow applied to the entire country and the whole time period. For the period 1990-2000,with an average number of births of 55,723 per prefecture per year, our calculation implythat PollutionExportShock caused a total of 66,064 extra deaths.56 For the period 2000-2010, when the export shocks in dollar terms are ten times those in the 1990’s, the samecalculation yields an additional 179,409 infant deaths. Although the effect of ExportShockis not always significant, we still employ it to quantify its beneficial effect on mortality. Theeffect of ExportShock reduced the number of deaths by 3,075 in 1990-2000 and by 21,632in 2000-2010. The net effect expressed in percentage was an increase in mortality by 0.032and 0.113 percentage points. As a reference, Currie and Neidell (2005) find that the declinein carbon monoxide in California over the 1990’s caused a reduction in infant mortality of55The infant mortality rate actually declined by 6.8 in Zhenjiang and by 8.8 in Taizhou56This number is calculated by multiplying the average PollExportShock of 2.2 by the coefficient 0.317and the number of births 55723. The number is then divided by 1000 because mortality is expressed relativeto 1000 live births. The resulting number is multiplied by 10 years and the number of prefectures (340) andfinally divided by 2 because the export shock is calculated over a 10 years period and we assume that theincrease happened in equal increments over the 10 years.600.02 percentage points. So our effect is similar to theirs for the 1990’s, but about 5 timeslarger in the 2000’s, which is reasonable since exports grow at an extraordinary pace duringthat period. Our effect is much smaller (it is about one tenth) relative to the one found byJayachandran (2009) of an increase in mortality by one percentage point due to Indonesianwildfires.The main take-away from this paper is that China might have greatly benefited fromincreased access to world markets, but some regions bore a disproportionate cost of this rapidexport expansion in terms of higher infant mortality due to their more rapidly deterioratingenvironmental quality. This paper represents a first attempt to quantify the extent of theextra burden born by those regions that because of comparative advantage specialize in highlypolluting products.61Figure 2.1: Mechanism relating Export Shocks to Mortality“Export”	  tariffs	  Pollu1on	  Export	  Shock	  Export	  Shock	  Pollu1on	  Concentra1on	  Mortality	  Figure 2.2: The relationship between Log(Exports) and Log(ExportTariff)Slope: −7.815           (1.288)−2−10123ln(Exportkt)−.1 −.05 0 .05 .1ln(1+ExportTariffkt)Note: Both axes report residuals of the variable regressed on time and sector fixed effects62Figure 2.3: Distribution of Export Pollution Shocks over Decades, SO263Table 2.1: Effect of 2001-03 Steel Safeguards on APIDep. Var: API (1) (2) (3) (4) (6) (7)Before× Share 2.518** 2.592**(1.059) (1.048)After × Share 1.628*** 1.540***(0.513) (0.556)NoSG× Share 2.553*** 2.534***(0.465) (0.498)Prefecture Y Y Y Y Y YYear×Region Y Y Y Y Y YMonth×Region Y Y Y Y Y YTime Window 07/00 – 07/00– 06/02– 03/02– 07/00– 07/00–09/02 12/02 05/05 08/05 05/05 08/05N 36,915 41,238 64,903 76,953 96,084 103,810R2 0.417 0.407 0.363 0.351 0.372 0.376Note: Standard errors are clustered at the prefecture level. *** p<0.01, ** p<0.05, * p<0.164Table 2.2: Summary Statisticsmean std 10th 25th 50th 75th 90thPanel A: IMR (number of deaths per 1000 births)IMR, 1982 36.105 24.659 14.648 20.037 28.729 45.096 66.790IMR, 1990 31.428 23.915 11.420 16.239 24.198 39.084 61.088IMR, 2000 23.642 17.994 7.720 11.182 18.834 29.627 46.127IMR, 2010 5.132 5.972 1.374 2.067 3.362 5.708 10.944∆IMR, 90-00 -7.786 18.709 -23.745 -13.759 -6.358 1.561 9.562∆IMR, 00-10 -18.510 14.455 -37.307 -23.704 -14.696 -8.748 -4.756Panel B: Changes in Pollution concentrations(µg/m3)SO2, 1992 86.354 76.636 20 39 64 104 173SO2, 2000 43.445 38.757 12 19 31 55 92∆SO2, 90-00 -41.459 55.883 -87 -55 -31 -9 0∆SO2, 00-10 -5.624 33.235 -45 -14 1 14 23PM2.5, 2000 34.156 19.662 10.857 18.327 31.603 48.369 61.311∆PM2.5, 00-10 12.073 9.887 0.576 4.174 11.917 18.481 24.929Panel C: Export Pollution Shocks (pounds per worker)SO2, 90-00 2.199 3.471 0.074 0.360 1.219 2.681 5.408TSP , 90-00 2.133 2.915 0.091 0.374 1.229 2.624 5.164NO2, 90-00 0.609 0.813 0.022 0.114 0.348 0.831 1.637SO2, 00-10 8.109 7.996 1.582 2.984 5.494 9.997 18.540TSP , 00-10 9.821 9.216 2.241 3.678 6.794 12.136 21.983NO2, 00-10 2.768 2.788 0.517 0.984 1.889 3.457 6.699Panel D: Export Shocks (1000 dollars per worker)ExShock, 90-00 0.151 0.220 0.004 0.019 0.072 0.178 0.423ExShock, 00-10 1.440 2.223 0.217 0.399 0.785 1.532 3.10065Table 2.3: Change in Infant Mortality Rate and Shocks: 2SLSDep. Var: ∆IMR (1) (2) (3) (4) (5) (6) (7)Panel A: SO2PollExShockSO2 0.449*** 0.422*** 0.475*** 0.427*** 0.537*** 0.506*** 0.317***(0.058) (0.141) (0.130) (0.099) (0.125) (0.136) (0.121)ExShock -0.573 -0.500 -0.215(0.485) (0.508) (0.470)Panel B: TSPPollExShockTSP 0.424*** 0.393*** 0.450*** 0.399*** 0.461*** 0.415*** 0.255***(0.057) (0.121) (0.123) (0.097) (0.117) (0.108) (0.093)ExShock -0.427 -0.330 -0.148(0.496) (0.503) (0.449)Panel C: NO2PollExShockNO2 1.442*** 1.588*** 1.823*** 1.550*** 2.318*** 2.118*** 1.383***(0.187) (0.435) (0.431) (0.335) (0.507) (0.521) (0.367)ExShock -1.245** -1.069* -0.648(0.535) (0.564) (0.465)Region×Year Y Y Y Y Y YInitial Conditions Y Y Y Y Y YContemporary Shocks Y Y Y Y Y∆IMRt−1 & ∆IMR2t−1 Y Y Y Y∑k γpktLikt−1Lit−1Y YProvince×Year YN 680 673 673 673 673 673 673Notes: All regressions are weighted by population of age 0. Initial conditions include start of period GDP percapita, overall mortality rate, agriculture employment share and population density. Contemporaneous shocksinclude change in log GDP per capita, change in share of boys, change in share of population with middle schooleducation, change in share of population with high school education or above, change in number of hospital bedsper capita, change in agricultural employment share, a dummy of provincial capital city and distance to the nearestport. Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, * p<0.166Table 2.4: Changes in Pollutant Concentration and Shocks: 2SLSDep. Var: ∆Pollutant Concentration (1) (2) (3)Panel A: ∆SO2PollExShockSO2 0.747*** 1.050*** 0.785**(0.249) (0.254) (0.353)ExShock -1.425* -0.856(0.805) (0.961)N 268 268 268Panel B: ∆PM2.5PollExShockTSP 0.203*** 0.239*** 0.181***(0.060) (0.074) (0.066)ExShock -0.721 -0.622(0.522) (0.523)N 340 340 340Region(×Year) Y Y YInitial Conditions Y Y YContemporaneous Shocks Y Y Y∆IMRt−1 & ∆IMR2t−1 Y Y Y∑k γpktLikt−1Lit−1 YNotes: All regressions are weighted by population of age 0. Initial conditions include start of period GDPper capita, overall mortality rate, agriculture employment share and population density. Contemporaneousshocks include change in log GDP per capita, change in share of boys, change in share of population withmiddle school education, change in share of population with high school education or above, change in numberof hospital beds per capita, change in agricultural employment share, a dummy of provincial capital city anddistance to the nearest port. Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, *p<0.167Table 2.5: Changes in Infant Mortality Rate and Changes in Pollutant Concentration:2SLSOLS 2SLS OLS 2SLS OLS 2SLSDep. Var: ∆IMR (1) (2) (3) (4) (5) (6)Panel A: SO2 Concentration and IMR∆SO2 -0.016 0.299** -0.016 0.298** -0.016 0.464*(0.033) (0.150) (0.033) (0.150) (0.034) (0.278)ExShock -0.071 0.052 -0.071 -0.084(0.395) (0.431) (0.388) (0.483)N 268 268 268 268 268 268Panel B: PM2.5 Concentration and IMR∆PM2.5 0.073 1.317* 0.049 0.925* 0.044 1.454*(0.077) (0.769) (0.081) (0.489) (0.083) (0.752)ExShock -1.552*** -0.777 -1.525*** -0.612(0.451) (0.526) (0.432) (0.726)N 340 340 340 340 340 340Region(×Year) Y Y Y Y Y YInitial Conditions Y Y Y Y Y YContemporaneous Shocks Y Y Y Y Y Y∆IMRt−1 & ∆IMR2t−1 Y Y Y Y Y Y∑k γpktLikt−1Lit−1 Y YNotes: All regressions are weighted by population of age 0. Initial conditions include start of period GDPper capita, overall mortality rate, agriculture employment share and population density. Contemporaneousshocks include change in log GDP per capita, change in share of boys, change in share of population withmiddle school education, change in share of population with high school education or above, change innumber of hospital beds per capita, change in agricultural employment share, a dummy of provincial capitalcity and distance to the nearest port. Standard errors are clustered at the province level. *** p<0.01, **p<0.05, * p<0.1Table 2.6: Elasticity of IMR to pollutant concentration in other studiesCountry SO2 TSP PM10Arceo et al. (2015) Mexico 0.42Chay and Greenstone (2003a and b) US 0.28-0.63Chen et al.(2013) China 1.73Tanaka (2010) China 0.82 0.9568Table 2.7: Changes in Infant Mortality Rate and Shocks by Causes of Death:SO2, 2SLSCardio- Infant DigestiveRespiratory SpecificDep. Var: ∆IMRC (1) (2) (3) (4) (5) (6)PollExShockSO2 0.276*** 0.459** -0.318 -0.079 0.014 0.013(0.088) (0.185) (0.428) (0.758) (0.021) (0.060)∆Export -2.571 -4.101 0.291(1.799) (5.687) (0.345)Region Y Y Y Y Y Y∑k γpktLikt−1Lit−1 Y Y YN 117 117 117 117 117 117Infectious External MalnutritionCauses(7) (8) (9) (10) (11) (12)PollExShockSO2 0.043 -0.010 -0.026 0.136 -0.013 -0.019(0.035) (0.083) (0.072) (0.206) (0.017) (0.035)ExShock -0.444 -0.164 0.003(0.526) (1.461) (0.174)Region Y Y Y Y Y Y∑k γpktLikt−1Lit−1 Y Y YN 117 117 117 117 117 117Notes: All regressions are weighted by total birth. All regressions start of period overall mortalityrate, change in log GDP per capita, change in hospital beds per capita, and change in agriculturalemployment share. Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, *p<0.169Table 2.8: Change in Infant Mortality Rate and Shocks:Neighboring Shocks,2SLSNearest stations Nearest two stationsDep. Var: ∆IMR (1) (2) (3) (4) (5)Panel A: SO2PollExShockSO2 0.448*** 0.399*** 0.382*** 0.410*** 0.399***(0.129) (0.123) (0.121) (0.123) (0.122)ExShock -0.782 -0.779 -0.662 -0.777 -0.705(0.562) (0.546) (0.551) (0.550) (0.553)PollExShockSO2,N 0.312** -0.393 -0.282(0.145) (0.304) (0.303)Wind PollExShockSO2,N 0.436*** 0.808*** 0.410*** 0.680**(0.144) (0.282) (0.142) (0.277)Panel B: TSPPollExShockTSP 0.383*** 0.355*** 0.344*** 0.358*** 0.350***(0.106) (0.102) (0.102) (0.103) (0.103)ExShock -0.635 -0.632 -0.531 -0.628 -0.561(0.543) (0.517) (0.516) (0.518) (0.522)PollExShockTSP,N 0.338** -0.370 -0.271(0.146) (0.300) (0.298)Wind PollExShockTSP,N 0.438*** 0.774*** 0.420*** 0.668**(0.143) (0.294) (0.141) (0.285)Panel C: NO2PollExShockNO2 1.923*** 1.741*** 1.670*** 1.778*** 1.736***(0.552) (0.529) (0.525) (0.529) (0.531)ExShock -1.323* -1.254* -1.125* -1.263* -1.190*(0.692) (0.651) (0.666) (0.657) (0.672)PollExShockNO2,N 0.854** -0.884 -0.578(0.427) (0.770) (0.783)Wind PollExShockNO2,N 1.157*** 1.995*** 1.094** 1.653**(0.436) (0.735) (0.446) (0.780)Region×Year Y Y Y Y YInitial Conditions Y Y Y Y YContemporary Shocks Y Y Y Y Y∆IMRt−1 & ∆IMR2t−1 Y Y Y Y Y∑k γpktLikt−1Lit−1Y Y Y Y YN 673 673 673 673 673Notes: All regressions are weighted by population of age 0. Initial conditions include start ofperiod GDP per capita, overall mortality rate, agriculture employment share and population density.Contemporaneous shocks include change in log GDP per capita, change in share of boys, change inshare of population with middle school education, change in share of population with high schooleducation or above, change in number of hospital beds per capita, change in agricultural employmentshare, a dummy of provincial capital city and distance to the nearest port. Standard errors areclustered at the province level. *** p<0.01, ** p<0.05, * p<0.170Table 2.9: Change in Infant Mortality Rate and Future ShocksSO2 TSP NO2Dep. Var: ∆IMR (1) (2) (3) (4) (5) (6)PollExShockpt+1 -0.291 -0.138 -0.240 -0.011 -0.899 0.151(0.216) (0.207) (0.245) (0.249) (0.989) (1.206)ExShockt+1 -2.033 -2.581 -3.130(2.337) (2.489) (2.433)∑k γpkt+1Likt+1Lit+10.050 0.042 -0.053 -0.070 -0.076 -0.129(0.181) (0.182) (0.060) (0.060) (0.555) (0.564)N 333 333 333 333 333 333Notes: All regressions are weighted by population of age 0. All regressions in Panel A include allthe controls in the specification (6) of Table 2. Panel B controls for start of period overall mortalityrate. Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, * p<0.1Table 2.10: Change in Infant Mortality Rate and Shocks by Gender: 2SLSBoys GirlsSO2 TSP NO2 SO2 TSP NO2Dep. Var: ∆IMR (1) (2) (3) (4) (5) (6)PollExShockp 0.440*** 0.343*** 1.859*** 0.588*** 0.490*** 2.401***(0.131) (0.099) (0.476) (0.164) (0.137) (0.610)ExShock -0.453 -0.294 -0.964* -0.576 -0.374 -1.194**(0.535) (0.538) (0.585) (0.508) (0.463) (0.554)Region×Year Y Y Y Y Y YInitial Conditions Y Y Y Y Y YContemporary Shocks Y Y Y Y Y Y∆IMRt−1 & ∆IMR2t−1 Y Y Y Y Y Y∑k γpktLikt−1Lit−1 Y Y Y Y Y YN 673 673 673 673 673 673Notes: All regressions are weighted by population of age 0. Initial conditions include start of period GDP percapita, overall mortality rate, agriculture employment share and population density. Contemporaneous shocksinclude change in log GDP per capita, change in share of boys, change in share of population with middle schooleducation, change in share of population with high school education or above, change in number of hospital bedsper capita, change in agricultural employment share, a dummy of provincial capital city and distance to the nearestport. Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, * p<0.171Table 2.11: Change in Mortality Rate of Children aged 1-4 and Shocks: 2SLSSO2 TSP NO2Dep. Var: ∆MR1−4 (1) (2) (3) (4) (5) (6)PollExShockp 0.037*** 0.036*** 0.028*** 0.027*** 0.132*** 0.129***(0.010) (0.010) (0.007) (0.006) (0.033) (0.033)ExShock -0.110*** -0.110*** -0.096*** -0.095*** -0.137*** -0.135***(0.032) (0.031) (0.032) (0.031) (0.035) (0.034)Region×Year Y Y Y Y Y YInitial Conditions Y Y Y Y Y YContemporary Shocks Y Y Y Y Y Y∑k γpktLikt−1Lit−1Y Y Y Y Y Y∆IMRt−1 & ∆IMR2t−1 Y Y YN 673 673 673 673 673 673Notes: All regressions are weighted by population of age 1-4. Initial conditions include start of period GDP percapita, overall mortality rate, agriculture employment share and population density. Contemporaneous shocksinclude change in log GDP per capita, change in share of boys, change in share of population with middle schooleducation, change in share of population with high school education or above, change in number of hospital bedsper capita, change in agricultural employment share, a dummy of provincial capital city and distance to the nearestport. Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, * p<0.172Table 2.12: Change in Infant Mortality Rate and Shocks: SO2 RobustnessEnergy Import High-skill Share of TCZ IO Adjusted OutputProduction Shocks Shock Ownership ShocksDep. Var: ∆IMR (1) (2) (3) (4) (5) (6) (7)PollExShockSO2 0.440*** 0.798*** 0.458*** 0.409*** 0.476*** 0.072**(0.109) (0.230) (0.126) (0.119) (0.137) (0.028)ExShock -1.779*** -0.864* 2.493** -1.772*** -0.502 0.154(0.417) (0.501) (1.140) (0.396) (0.527) (0.346)∆EnergyProd 0.278**(0.130)PollImShockSO2 -0.235**(0.108)HighSkillShock -7.297***(2.801)Share SOE 7.093**(3.431)Share Foreign 0.200(3.439)TCZ 2.691***(0.760)PollOutputShockSO2 0.189***(0.066)OutputShock -0.258(0.473)Region×Year Y Y Y Y Y Y YInitial Conditions Y Y Y Y Y Y YContemporary Shocks Y Y Y Y Y Y Y∆IMRt−1 & ∆IMR2t−1 Y Y Y Y Y Y Y∑k γpktLikt−1Lit−1Y Y Y Y Y Y YN 340 673 673 340 673 673 340Notes: All regressions are weighted by population of age 0. Initial conditions include start of period GDP per capita, overallmortality rate, agriculture employment share and population density. Contemporaneous shocks include change in log GDPper capita, change in share of boys, change in share of population with middle school education, change in share of populationwith high school education or above, change in number of hospital beds per capita, change in agricultural employment share,a dummy of provincial capital city and distance to the nearest port. Standard errors are clustered at the province level. ***p<0.01, ** p<0.05, * p<0.173Table 2.13: Change in Infant Mortality Rate and Export Expansions by Industry GroupsDep. Var: ∆IMR (1) (2) (3) (4) (5) (6)ExShock 0.373(0.534)ExShockC -0.095 -0.908** -0.480(0.551) (0.443) (0.587)ExShockM 0.372 1.636 3.296(1.668) (1.549) (2.415)ExShockD 11.178*** 11.514*** 9.655***(2.417) (2.553) (3.037)Region×Year Y Y Y Y Y YInitial Conditions Y Y Y Y Y YContemporaneous Shocks Y Y Y Y Y Y∆IMRt & ∆IMR2t Y Y Y Y Y YLag Emp. Share YN 673 673 673 673 673 673Notes: All regressions are weighted by population of age 0. Initial conditions include start of period GDPper capita, overall mortality rate, agriculture employment share and population density. Contemporaneousshocks include change in log GDP per capita, change in share of boys, change in share of population withmiddle school education, change in share of population with high school education or above, change in numberof hospital beds per capita, change in agricultural employment share, a dummy of provincial capital city anddistance to the nearest port. Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, *p<0.174Chapter 3Grain Exports and China’s GreatFamine3.1 IntroductionDuring 1959 to 1961, a famine raged across China and resulted in 16.5 to 30 million excessdeaths and 30 to 33 million lost or postponed births. In terms of population loss, thisfamine is the worst demographic disaster in human history. The existing literature studiesdifferent facets of the Great Famine and arrives at the consensus that it is a consequenceof multiple interrelated institutional failures during the Great Leap Forward (GLF) period.In this paper, we show that despite dramatic output falls during the early famine years, thecentral government increased grain export to exchange for capital goods and to acceleratethe repayment of debt to the Soviet. We collect international trade data by grain crops andfind that during famine years, certain grain crops like rice and soybean experienced muchlarger export growths than other crops like wheat and maize. The surge in exports of rice andsoybean relative to wheat is aligned with the changes in international relative prices duringthe period. Exploiting cross-county variation in grain export exposure, which stems fromdifferences in the suitability of cultivating different crops and distance to railway, we provideempirical evidence that the excessive grain exports reduced food available for consumption inrural areas and as a result aggravated famine severity. In particular, we find that counties thathave higher suitability in cultivating high-export-exposure crops experienced more dire faminesituation. However, a county’s suitability in low-export-exposure crops was uncorrelated withits famine severity. We show that this pattern is hard to explain with pre-existing trends thatare associated with regional comparative advantage in cultivating different crops. Moreover,we find that for high-export-exposure crops, the correlation between productivity and famineseverity declines with a county’s distance to railroad.Our study complements the literature on the causes of China’s Great Famine in severalways. First, it is the first to provide empirical evidence at micro-level to show that over-export of grains is associated with the spatial pattern of famine severity. Second, unlike theprevious studies that mainly rely on province-level panel data,57 we compile a unique dataseton famine severity, crop specialization, distance to railway, export exposure at the countylevel. We conduct the analysis using within-province variation in order to better accountfor unobserved heterogeneity across provinces that may confound the over-export channelemphasized by the paper.The paper is closely related to studies which show that a fall in agricultural output andover-procurement of grains from rural sector were the key contributors to the Great Famine.57The exception is Meng et al. (2015), who also employ a county-level analysis to complement their province-level main results.75Lin and Yang (2000) find that both food availability decline, i.e., decline in grain output,and entitlement failure, i.e., over-procurement of grains due to urban bias policies, contributeto the famine severity. Meng et al. (2015) document a surprising positive relationship be-tween regional agricultural productivity and mortality. They argue that the more productiveregions experienced larger absolute production drops during famine years, and hence theywere subject to over-procurement of grains under progressive and inflexible government pro-curement policies. We argue that over-procurement of grains during famine years were inpart caused by the surge in international exports. Unlike grain procurement for domesticdistributional purposes, grain exports undoubtedly lowered food availability for domesticconsumption. We find that only productivity of crops that had large export exposure, i.e.,rice and soybean, was positively associated with famine severity, while productivity of othercrops, i.e., wheat, was uncorrelated with famine severity. This pattern cannot be solely ex-plained by over-procurement of grain crops as a whole. In particular, we document thatcrops with small export exposure (wheat) actually experienced larger growth in procurementrelatively to crops with large export exposure (rice and soybean).The remainder of this paper is organized as follows. Section 3.2 introduces the back-ground of the Great Famine and the role of international trade. Section 3.3 describes thedataset. Section 3.4 provides empirical evidence that over export of grains during famineyears aggravated famine severity. Section 3.5 concludes the paper.3.2 BackgroundIn this section, we briefly discuss the background of China’s 1959-1961 famine, including ruralinstitutions, the basic facts of this demographic crisis and the role of international trade.3.2.1 Rural InstitutionsThe Communist Party of China (CCP) started collectivization in 1952, in the hope of trans-forming Chinese agriculture from fragmented household farming into large-scale mechanizedproduction. The initial phase of collectivization (1952-1957) was cautious and smoothed.The production unit was in the form of elementary or advanced cooperative, and usuallyconsisted of 20 to 200 households. Peasants joined the various forms of cooperatives on avoluntary basis and retained the right of withdrawal. Production was planned and organizedat the level of cooperative and a household’s income depended on their inputs of land, cap-ital goods and labor. In the period 1952-1957, agricultural output grew continuously at anaverage annual rate of 4.6%. (Lin, 1990; Li and Yang, 2005)In 1958, the CCP launched the Great Leap Forward movement and adopted radicalheavy-industry oriented policies. To achieve the lofty goals set by the GLF, more resourceshad to be extracted from the vast rural sector which consisted of approximately 80% of thepopulation at the time. Being impatient of the lukewarm growth in agricultural output,the central planners decided to take an aggressive approach and further amalgamate ruralcollectives into massive communes. By the end of the year, 24,000 communes had been setup, with an average size of 5000 households and 10,000 acres. Compulsory participation incommunes became an official policy, and private property rights of lands and capital goodswere abolished. Harvest and storage of agricultural goods were conducted at the commune76level and private markets for trading foods were virtually eliminated. Peasants no longerreceived pecuniary rewards for their effort input but instead free foods were supplied incommunal kitchens. The communal movement, nevertheless, was followed by the collapse inagricultural outputs. The grain output plunged by 15% in 1959 and reached only about 70%of the 1958 level in 1960 and 1961. (Lin, 1990; Lin and Yang, 2000; Li and Yang, 2005; Menget al., 2015)Aside from production, the distribution and consumption of grains were also intensivelycontrolled by the central government. Under an in-kind agricultural tax system, the centralplanner set a target of grain procurement to meet the needs of planned urban consumption,industrial inputs, reserve requirement and international trade. After harvests, local govern-ments collected grains to fulfill their quota obligations, and peasants kept grains retainedafter the procurement. This system was progressive and rigid in the sense that local manda-tory quotas were set prior to a agricultural season according to the region’s past grain output,and could not be adjusted to the actual quantity harvested. To fund the GLF campaign, thegovernment raised the procurement of grains from 46 million tons in 1957 to 52 million tonsin 1958, and the total procurement reached 64 million tons in 1959 when the grain outputslumped. (Lin and Yang, 2000; Meng, et al., 2015)3.2.2 The 1959-1961 Great FamineThe Great Famine over 1959-1961 resulted in 16.5 to 30 million excess deaths and 30 to 33million lost or postponed births.58 According the official statistics, the national death ratejumped from 11.98 per thousand in 1958 to 25.43 per thousand in 1960 when the faminewas most severe. In the meanwhile, birth rate dropped from 29.22 to 20.86 per thousand.Although the famine is a nationwide calamity, there existed considerable differences in famineexposures across regions. For example, while Jiangsu province had an rise in death rate from9.4 to 18.4 per thousand from 1958 to 1960, its neighbor Anhui province experienced adramatic increase in death rate from 12.3 to 68.6 per thousand. Moreover, the famine waslargely restricted to the rural sector for two reasons. First, the central government gavehigh priority to urban grain supplies, and hence urban food rations were seldom below thesubsistence level. Second, stringent controls over rural-urban migration and even rural-ruralmigration prohibited starving people from fleeing famine stricken regions. (Lin and Yang,2000; Meng et al., 2015)The existing literature on China’s Great Famine revolves around finding the primarycause of this nationwide calamity. A first strand of the literature raised the factors thatexplained the sudden decline in food-availability. They include factors related to the plungein agricultural output, e.g., a succession of natural disasters (Yao, 1999), forced commualiza-tion and removal of exit right (Lin, 1990), diversion of resources from agricultural to heavyindustry due to the GLF (Li and Yang, 2005), and also factors causing waste of food, e.g.,consumption inefficiency in communal kitchens (Chang and Wen, 1998; Yang and Su, 1998).A second series of papers focused on the factors resulting in entitlement failures, which includeover-procurement of grains from rural sector because of urban-biased food policy (Lin andYang, 2000), and the rigid and progressive procurement policy that caused over-procurementof grains from regions that suffered larger negative productivity shock (Meng et al., 2015).58The estimates of excess deaths and lost/postponed births come from several studies that carefully examinethe demographic data, including Coale (1981) Ashton et al. (1984) and Yao (1999) among others.77The literature also point out the macro implications of the surge in net grain export in theperiod 1958-1960 (Ashton et al., 1984; Johnson, 1998).For a massive and widespread famine like the one in China during 1959-1961, there couldbe a complicated set of factors that interacted and reinforced each other and culminated ina demographic catastrophe. The famine ended in 1962 together with the modification ofpolicies and institutions along multiple dimensions. Extreme policies from the GLF wereabandoned. The central government substantially increased grain imports, and transferreda large amount of grains to the rural sector. Rural institutions were altered and eventuallyresembled those in the pre-GLF years: the role of communes was diminished and productionwas managed by the elementary or advanced cooperatives; a compensation scheme for effortwas restored and communal kitchens were abolished; grain procurement rates were reducedand rural trade fairs were reopened. Nevertheless, grain output in 1962 remained 18.2% lowerthan the level in 1957, and the pre-famine grain production level was not regained until 1966(Lin, 1990; Meng et al., 2015).3.2.3 The Role of International TradeChina in the 1950s pursued development policies that heavily biased towards industrializa-tion. As a result, the central government harshly squeezed the agricultural sector to expediteindustrial development and subsidize urbanites (Lin and Yang, 2000). The exports of agricul-tural goods and grains comprised around 40% and 15% of the total exports before the famine.Hence, to some extent, the country’s capacity in obtaining foreign exchange to facilitate in-dustrialization relied on exporting agricultural goods and especially grains. Moreover, sincescarce foreign exchange was reserved mainly for importing industrial equipment, China barelyimport grains until 1961. The zealous industrial policies during the GLF further distortedthe balance across sectors. In 1959, despite the plunge in output, the total grain exportssurged to 4.1 million tons, and the large net grain exports continued through out 1960.The rapid deterioration of relationship with the USSR since 1959 partially contributedto the rise of grain exports in famine years while the leadership knew that some of its peoplewere starving (Riskin, 1998; Yao, 1999). The sino-soviet political tension escalated in June1960 when the USSR withdrew its economic advisers and specialists working in China. TheCCP Politburo immediately decided to accelerate the repayment of Soviet loan from 16 yearsto 5 years. The accumulated debt owing to the USSR amounted to around 1.5 billion RMB,which was approximately 14 times the trade surplus in 1958. To meet the repayment timeline,a “trade group” was set up to restrict imports and oversee the collection of commodities forexport (Garver, 2016).59Net grain exports over the period of 1958 to 1960 totaled around 9.6 million tons. Meng etal. (2015) estimate that one kilogram of grains contains 3,587 calories and daily average caloricneed is 804 to 1,871.60 Combined with these estimates, the net grain exports during 1958 to59This “trade group” was led by high-ranking officials including the premier Zhou Enlai, the vice premiersLi Fuchun and Li Xiannian (See the CPC Central Committee emergency notification of the campaign forcommodity procurement and export. http://cpc.people.com.cn/GB/64184/64186/66667/4493401.html). Theleadership were aware of the food deficiency and hardship of increasing grain exports, but Mao claimed “TheYan’an period was hard too, but eating pepper didnt kill anybody. Our situation now is much better thanthen. We must tighten our belts and struggle to pay off the debt within five years.” (Garver, 2016).60As is detailed in Meng et al. (2015), the daily caloric need is calculated based on caloric requirementsby age and sex recommended by the United States Department of Agriculture (USDA) and the demographic781960 translates into energy that would have covered the caloric needs of 16.7 to 38.9 millionpeople for three years. These estimates are commensurate with the total population lossduring the Great Famine. In 1961, pressured by food scarcity, China substantially increasedgrain imports, resulting in a net grain imports of 4.4 million tons which provided 23.2 to 53.9million person-year caloric needs. As is also pointed out by Ashton et al. (1984), Johnson(1998) and Riskin (1998), massive excess death could have been avoided had the governmentacted swiftly to stop exports and start large-scale imports of grains.3.3 DataOur dataset is compiled from multiple sources. In this section, we discuss the data sourcesand the construction of different measures.3.3.1 Population DataThe analysis of population growth at the county level uses the 1953 and 1964 China Popu-lation Censuses data from the Michigan China Data Center.61 Over 1953 to 1964, China’spopulation grew by 19%, from 0.58 to 0.69 billion. Following Meng et al. (2015), we restrictour main sample to 20 provinces with majority Han population. (These provinces are Anhui,Beijing, Fujian, Guangdong, Hebei, Heilongjiang, Henan, Hubei, Hunan, Jiangsu, Jiangxi,Jilin, Liaoning, Shaanxi, Shandong, Shanghai, Shanxi, Sichuan, Tianjin and Zhejiang.) Thesample comprises more than 85% of China’s population in 1953.62 We further drop urbancounties and counties which experienced changes in boundary over 1953 to 1964.63 As aresult, our main sample is composed of 1164 counties. Hereafter, we refer to this data asthe 1953-1964 sample. As is shown in Table 3.1, over the period 1953-1964, the populationsize increased by 0.18 log points. Moreover, there is substantial heterogeneity in populationgrowth across counties over time, with the interquartile range of ∆ lnPop equaling to 0.18.We also use information on birth-cohort size of survivors from a 1% sample of the 1990China Population Census. The data allow us to study the population size of birth-cohortsthat had different famine exposures. Again, we restrict the sample to the provinces as inthe 1953-1964 sample and have 1295 rural counties in the 1990 sample.64 Figure C.1 in thestructure in China from the 1953 Population Census. The authors show that for China as a whole in the1950s, 1871 calories were needed per person-day on average for heavy labor and normal child development.Also, on average, an individual only needed 804 per day to stay alive.61The county-level data from 1953 census only contains information on total number households and pop-ulation, with no further demographic breakdowns. Therefore, we can only restrict our analysis to the overallpopulation growth.62Meng et al. (2015) exclude Sichuan province due to the lack of data. As we don’t have similar issue,the counties in Sichuan province are included in our main sample. There were 29 province-level divisions inChina during in the 1950s. (The present day provinces Hainan and Chongqing used to belong to provincesGuangdong and Sichuan, respectively.) We exclude five province-level autonomous regions, Tibet, Xinjiang,Inner Mongolia, Guangxi and Ningxia, where people faced different economic and political policies due tohistorical and political reasons. We also exclude Qinghai, Guansu, Yunnan and Guizhou, where large ethnicminority population lived in autonomous communities.63There are 159 urban counties and 157 counties that experienced boundary changes.64The sample size is different from the 1953-1964 sample because the administrative divisions are differentin different census years. In addition, for the birth-cohort size analysis, we don’t need worry the countieswhich boundary evolved over time.79Appendix presents the mean and the coefficient of variation (cv) of survivor population sizeacross different birth cohorts. Similar to Meng et al. (2015), the size of cohorts born in thefamine years falls below that of the other cohorts. More importantly, the variance in birth-cohort size across counties surges in the famine years and reverts back to the pre-famine levelby year 1963. This suggests that there exists considerable variation in the famine severityacross counties.To perform robustness checks, we also collect yearly data on population spanning theperiod of 1953 to 1964 at the county level from various sources. The majority of data isfrom numerous volumes of Local Gazetteers, which is periodically published by Chinese localgovernments and contains relevant information concerning history, economy, administration,development, and so on.65 We employ the secondary data collected and complied by Cao(2005).66 We supplement this data with information collected from various statistical bookspublished by the provincial Statistics Bureau.67 Similar to the 1953-1964 sample, we restrictthe sample to rural counties with stable administrative boundary between 1954 to 1963. Thefinal dataset is a balanced panel of 856 counties for years 1953, 1957, 1961 and 1964. FigureC.2 shows the mean and the cv of population size across years. The growth in populationsignificantly deviates from the trend during over 1957 to 1961. In addition, the cross-countyvariance of population size increases during 1957 to 1961 and declines over 1961 to 1964.These patterns are largely consistent with the findings in Figure C.13.3.2 Historical Trade Data on Agricultural ProductsWe collected data on trade flows by agricultural products from the China Customs StatisticsYearbooks (1955-1961). The recently declassified data provide rich information on product-level bilateral trade flows between China and other countries. More details about this dataare discussed in Appendix C.2.2. The upper panel of Figure 3.1 shows the trade flows betweenChina and the rest of the world. Both export and import increased up to 1959 and Chinaalways maintained a trade surplus. The lower panel presents China’s exports and imports ofgrain products.68 Export of grain products comprised 12.1%-17.6% of the total export overthe period of 1955 to 1960. China barely imported grain products until 1961. Moreover,grain exports climbed to its historical height during the onset of famine. The net export ofgrain products increased from 0.64 billion RMB (1.92 milliion tons) in 1957 to 0.91 billionRMB (2.62 million tons) in 1958, to 1.32 billion RMB (4.05 million tons) in 1959, and to0.84 billion RMB (2.77 million tons) in 1960. In 1961, China switched from being a netexporter to being a net importer of grains with net imports amounting to 0.62 billion RMB(4.4 million tons). Over the period 1962 to 1966, China remained as a grain net importer ofgrain products (Lin and Yang, 2000).The aggregate data cannot reveal the composition of types of grain crops that were65Local Gazetteers is often regarded as the authoritative encyclopedia of a locality in China (typicallyprovince, city, or county). See Xue (2010) for detailed descriptions. This archival data has been used in recentstudies. For example, Chen, Li and Meng (2013) collect data on the year in which ultrasound machines wereintroduced in different counties. Almond, Li and Zhang (2013) collect data on the timing of land reform andgrain outputs across counties.66The dataset provides an unbalanced panel of 713 counties belonging to the provinces in our sample.67See Appendix for detailed description of data sources.68The grain products include soybean, rice, wheat, maize, millet, sorghum, barley, buckwheat, beans andflour.80actually exported and the changes in composition over time. Figure 3.2 shows that soybeanand rice are the two most important export goods, which together made up 81% to 95% oftotal grain exports across years. More importantly, different crops had different exposuresto export shock. Figure 3.3 reports the growth of average export in 1958-1961 over that in1955-1957 by crops. The exports of rice and soybean expanded respectively by 6.4 and 2.03million tons, while the exports of wheat, maize and other grain products increased slightlyby 0.99, 0.72 and 0.06 million tones respectively. In the empirical analysis, we study thecross-county differences in export exposures to rice, soybean and wheat, which experiencedthe largest surge in export during the famine period.The increase in rice and soybean exports relative to wheat exports is also aligned withthe changes in relative prices over the period. As is shown in Panel A of Figure 3.6, theexport price of rice was higher than that of wheat through out the period 1955-1960. Therelative price surged in 1958 and remained higher than the pre-1958 level in 1959 and 1960.We also find a similar evolution of relative price of soybean to wheat. Panel B displays theexport price of rice from Thailand relative to that of wheat from the US over years. This timeseries resembles that in Panel A, which ensures that the evolution of relative price was not afeature unique to exports from China, but rather driven by international demand and supplyforces. Lastly, Panel C shows that the domestic price of rice and soybean increases relativeto that of wheat over time in the US. All these findings suggest that the central governmentchose to expand exports of crops with increasing relative price in need of meeting the loftyindustrialization targets and repaying external debts.3.3.3 Crop Productivity DataThe data on productivity of cultivating different crops are from the Food and AgricultureOrganization (FAO)’s Global Agro-Ecological Zones (GAEZ) V3.0 database, which provideshigh resolution information on potential yields of different crops under various technologies at5×5 arc-minute grid level (approximately 9.25 km×9.25 km at the equator).69 The potentialyields are estimated using agronomic models and based on climate conditions, soil type,elevation and topography. Unlike directly observed yields, the potential yields at a givenlocation are a function of local biophysical conditions, and hence they are plausibly exogenousto other economic activities. We construct the potential yields of rice, soybean and wheatat the county level, by computing the average potential yields of grids that fall within thecounty boundary.70Figure 3.4 illustrates the distribution of potential yields of rice, soybean and wheat acrosscounties in China. The differences in the spatial distributions across crops are stark. On theone hand, the regions most suitable for rice cultivation are in the areas of the Yangtze Plainand the south coast. On the other hand, the regions most productive in producing soybeanand wheat are concentrated in the northeast region and the middle and lower reaches of the69This dataset has been widely used in recent studies, including Nunn and Qian (2011), Meng et al. (2015),and Bustos et al. (2016) among others.70We use data on potential yields under low-level input technology, i.e., production is based on rain-fedirritation, low-level of mechanization, utilization of fertilizer and chemicals for pest and disease control. Weconsider that the low-level input technology better describes the technologies used by Chinese farmers in the1950s and early 1960s. The potential yield of rice is taken as the maximum value of the potential yields ofwetland rice and dryland rice.81Yellow River.71 One may worry that counties clustered in certain area with high suitabilityof cultivating a certain crop may be systematically different along unobserved characteristicsfrom other counties. To alleviate the concern, we include province fixed effects into theregression analysis and hence we examine the effect of different crop’s productivity on famineseverity using the within-province variation.For the purpose of the analysis, we standardize the potential yields of different crops, andthey are denoted by ψci , where c ∈ {rice, soybean, wheat}.723.3.4 Data on Agricultural Production and Export ExposureOur empirical analysis also requires data on a region’s share in national production of dif-ferent crops. To obtain these variables, we use the recently declassified data from CountyStatistics on Cultivated Area and Output of Different Crops (1957), which is published by theChinese Ministry of Agriculture.73 These data reflect the agriculture production across Chi-nese counties before the GLF and was made available to the public only recently. Thereforewe consider that it is less likely to have been misreported by the famine-era government.To measure a region’s specialization in different crops, we construct the variables Sci bynormalizing the region’s output shareY ciY c (where Yc =∑i Yic denotes the total output of cropc in China) by its population share PopiPop . We consider a county was more specialized in cropc if its output share is larger than its population share, and hence was more exposed to anexport shock of crop c. In Table 3.1, we see that the mean of the normalized output shareis around 0.9 across different crops, which suggest that on average, a county’s productionis more or less proportional to its population size. In addition, counties exhibit notabledifferences in crop specialization. For example, the standard deviation of normalized outputshare equals 0.98 for rice.Using the data of regional output share, we also construct a Bartik-style index to measurea county’s exposure to the surge in exports of grain crops during famine years, that is,∆Exporti =∑c∈{r,s,w}YicYc∆ExportcPopi,where ∆Exportc = Avg.Exportc,58−60−Avg.Exportc,55−57 is the difference between crop c’saverage export during 1958-1960 and its average export during 1955-57. Specifically, to buildcounty i’s export exposure ∆Exporti, the national export growth in crop c is apportionedto county i according to its share in national output of that crop YicYc in 1957. Then, thisimputed regional export growth is normalized by the local population Popi. By construction,∆Exporti captures the per capita exports.7471The correlations are -0.23 between rice and wheat, -0.27 between rice and soybean and 0.71 between wheatand soybean.72To be precise, the standardized variable has mean zero and unit variance.73To the best of our knowledge, this statistical book is the only available source that provides data onagricultural production at county level by crop before the GLF.74According to FAO (2003), the caloric contents per gram of rice, soybean and wheat are similar (4.12-4.16for rice, 4.07 for soybean and 3.78-4.12 for wheat). Because of the similar caloric content of different crops,we consider the measure ∆Exporti well captures the caloric loss due to export shocks.823.3.5 Distance to Railroad DataThe map of historical railroad network in China is obtained from the Communist ChinaMap Folio (1967)m published by the US Central Intelligence Agency (CIA). Although it waspublished in 1967, to the best of our knowledge, this map provides information on railwaysystem closest to that in the famine era. As discussed in CIA (1967), despite the steadyprogress in railway expansion before 1958, the construction of new lines was suspended orabolished during the chaotic GLF period. Between 1962 and 1967, constructions focused onthe southwest provinces of Guizhou and Yunan, which are excluded from our main sample.We digitize the scanned map as displayed in Figure C.3 using ArcGIS software. Panel A inFigure 3.5 shows the railway network in 1967. Every province except Tibet was connectedto the railway system by at least one main line. Nevertheless, the rail network density in1967 was much lower compared to that in 1990 (as is shown in Panel B).75 We calculate thedistance to railroad for each county as the shortest distance between its centroid and the railnetwork.3.3.6 Historical Weather DataThe historical weather data are taken from Terrestrial Air Temperature and Precipitation:Monthly and Annual Time Series (1950-1996), Version 1.01, which provides monthly av-erages of temperature and precipitation at 0.5 × 0.5 degree grid level (approximately 56km×56 km at the equator).76 The grid-level estimates are interpolated from an average of20 weather stations, with correction for elevation. The grid data are mapped to counties.Specifically, for each county-year-month observation, we calculate the average temperatureand precipitation using the data of grids that overlap with the county territory. Then, for eachcounty-year observation, we construct variables of average temperature and precipitation inspring (February, March and April) and in summer (May, June and July).The upper panels in Figure 3.7 plot the distribution of spring and summer precipitationin famine years (1959-61) and the distributions of their historical means (1950-58). For bothseasons the distributions of year 1959 are more skewed to the right than those of the historicalmean, indicating that most counties experienced more precipitation than regular years. Thisis in line with the reported widespread floods in 1959.77 In contrast, the distribution of theyear 1960 have a larger density at small values, a pattern which is more evident in summerand is consistent with the documented extensive drought in 1960.78 In addition, year 1961had more precipitation in spring and fewer precipitation in summer than regular years. Thelower panels of Figure 3.7 shows the corresponding distribution for temperature. The famineyears have warmer springs than regular years. In summer, the distribution in year 1959 andyear 1960 resemble that of the historical mean, while the distribution of year 1961 has larger75The map on railway system for 1990 is from NASA Socioeconomic Data and Applications Center(SEDAC)’s Fundamental GIS Digital Chart of China (1993) database.76This dataset has been used in several recent studies, including Dell et al. (2012) and Meng et al. (2015).77The Yellow River flooded the east China in July 1959. According to the International Disaster Center,it is the third deadest natural disaster in China’s history. (See http://www.emdat.be/database) There wereother regions suffered from massive flooding as well. For example, floods inundated 810 thousand hectares inGuangdong in June (Ashton et al., 1984).78As is discussed in Ashton et al. (1984), the hardest regions were Hebei, Shandong, Shanxi, and HenanProvinces, where droughts lasted for 6-7 months.83densities at large values of temperature.3.4 Empirical EvidenceIn this section, we provide empirical evidence that grain export exposure is associated withfamine severity across counties in China. We investigate this correlation with three ap-proaches. Specifically, we relate a county’s famine severity to its i) suitability in producingdifferent crops ψci ; ii) actual specialization in different crops Sci ; and iii) overall grain exportshock ∆Exporti. Furthermore, we corroborate our results by using different proxies of famineseverity and bringing in extra variation in export exposure introduced by a county’s distanceto railroad.3.4.1 Famine Severity and Crop SuitabilityIn this subsection, we investigate the correlation between the county population growth andsuitability of cultivating different crops by estimating the following regression:∆ ln(Pop)i = αrψri + αsψsi + αwψwi + γp + ui , (3.1)where ∆ ln(Pop)i is the change in log population between 1953 and 1964 in county i; ψri , ψsiand ψwi denote the standardized productivity of producing rice, soybean and wheat respec-tively in county i; and γp is the province fixed effect. The difference in population growth rateis used to proxy for the famine severity. Province fixed effect in specification (3.1) capturesthe unobserved province-specific factors that affect population growth, such as incentives ofprovincial leaders to adopt radical policies during the GLF (Lin and Yang, 1998; Yang andSu, 1998; Chen and Kung, 2011). As a result, identification comes from within-provincevariation in productivity of cultivating different crops. In all regressions, standard errors areclustered at the prefecture level to account for the potential spatial correlation across countieswithin a prefecture.Conditional on productivity of other crops, a higher productivity in cultivating crop chas offsetting effects on the change in population during famine years. On the one hand, ahigher productivity in crop c provides more food to a county, which tends to alleviate thefamine severity. On the other hand, a county could have a larger export exposure in crop c ifit has a higher productivity in c relative to other crops, which aggravates local food shortage.We expect that the positive effect is more or less homogeneous across crops as they providesimilar caloric content. However, the magnitude of the adverse effect tends to be differentacross crops and depends on a crop’s export growth at the national level. As a result, thenet effect of productivity in crop c also depends on its national export growth.Column (1) in Table 3.2 reports the result of the baseline regression (3.1). The estimateof coefficient αr implies that a standard deviation increase in rice productivity is associ-ated with a 0.017 lower log population growth (around 10% of a standard deviation in logpopulation growth in the same period). We also find that one standard deviation increasein soybean productivity brings about a 0.039 lower log population growth (around 22% ofa standard deviation in log population growth in the same period). In contrast, the pro-ductivity in cultivating wheat is positively associated with log population growth, with theestimated coefficient αw equaling 0.025. All these estimates are statistically significant at84conventional levels. In column (2), we control for famine-year weather conditions, whichinclude spring temperature and its squared term, summer temperature and its squared term,spring precipitation and its squared term, and summer precipitation and its squared term ofyears 1959, 1960 and 1961. The results remain similar to the baseline specification, exceptthat the estimate of coefficient αw becomes insignificant. The findings that rice and soybeanproductivies have overall negative effects while wheat productivity has a weak positive effecton population growth are consistent with the fact that rice and soybean experienced muchlarger export expansion than wheat.The estimated effects of famine-year weather conditions on population growth are reportedin Table C.1 in Appendix. We find that the amount of rainfall in summer 1959 has asignificant effect on log population growth. Combining the estimated coefficients of 1959summer precipitation and its squared term indicates that the rainfall starts exerting negativemarginal effect on population growth when it exceeds 26cm (1.5 standard deviation above theaverage 1959 summer precipitation). This finding suggests that natural disasters like floodingcould play a role at the start of the famine. Nevertheless, except a mild significantly positiveeffect of 1960 summer temperature, we don’t find any significant effect of the remainingweather conditions.A drawback of using the inter-census population growth as a proxy for famine severity isthat it also incorporates population changes in non-famine years. It is possible that a higheragricultural productivity is positively (negatively) associated with population growth beforeand after the famine. If it is the case, our estimates in Table 3.2 tend to under-estimate(over-estimate) the effects of different crop productivities on famine severity. Nevertheless,the finding that productivities of crops with high export exposure (i.e., rice and soybean)have larger adverse effects on population growth than that of crops with low export exposure(i.e., wheat) is evidence suggesting over export during the early famine years could aggravatethe famine severity. That being said, we adopt two approaches to evaluate the robustnessof these results. First, we formally test pre-existing trends using a subsample of counties forwhich we have data on population for each year of 1953, 1957, 1961 and 1964. The resultsare described in section 3.4.5. Second, we use the data from the 1990 census and examine theeffects of different crops’ suitability on population size of survivors of famine birth-cohortsversus that of non-famine birth-cohorts. Specifically, we estimate the following regression:ln(Pop)il =1965∑l=1950δr`I`l ψri +1965∑l=1950δs`I`l ψsi +1965∑l=1950δw` I`l ψwi +X′ilθ + γpl + φi + uil , (3.2)where ln(Pop)il is the log population of cohort l in county i as is observed from the 1990census79; I`l is a dummy variable that equals 1 if the cohort l is born in year `; Xil contains aset controls, including variables of spring and summer temperature in county i and year l andtheir squared terms, and spring and summer precipitation in county i and year l and theirsquared terms; and γpl and φi are province×cohort and county fixed effects, respectively.The standard errors are clustered at the prefecture level. Because ψci s are time invariant andequation 3.2 includes county fixed effects, the estimates of δc`s must be interpreted relativeto a baseline year, which we take to be 1950, i.e., we normalize δc1950 to be 0. Therefore, the79For this analysis, we restrict the sample to cohorts born between 1950 to 1965. The pre-1949 periodis excluded because of civil wars before People’s Republic of China was founded. The post-1966 period isexcluded because of upheavals associated with the 1966-1976 Cultural Revolution.85coefficients δc`s reflect the changing relationship between population size and crop suitabilityover birth-cohorts.If due to export, a higher productivity in crop c is associated with loss of populationduring the famine, we expect the adverse effect to be largest among cohorts born over theperiod 1959 to 1961. This is because the effect of famine severity on mortality was largestamong young children (Ashton, et al., 1984), and fertility rate could decline more in countieswith more severe food shortage.The upper-left panel of Figure 3.8 plots the point estimates and 95% confidence intervalsof the coefficient δr` . The correlation between rice productivity and birth-cohort size is statis-tically similar to zero for most years during the period 1950-1957. The correlation dips belowzero from 1958 to 1960, but starts reverting back to pre-famine level from 1961. The upperright panel displays the correlation between soybean productivity and birth-cohort size overtime. The estimate of δs` remains stable and insignificant different from zero during 1950 to1958, but starts declining from 1959 and reaches the trough in 1961. These findings suggestthat a higher rice productivity or a higher soybean productivity is associated with a shortfallin population size of famine birth-cohorts. However, this pattern is absent for wheat. As isshown in the lower-left panel, the correlation between wheat productivity and cohort size isstatistically indifferent from zero for the famine birth-cohorts. As is discussed earlier, thedifference in population size between famine birth-cohorts and non-famine birth-cohorts canbe taken as an alternative proxy for the famine severity. Hence, the findings in Figure 3.8agree with those in Table Famine Severity, Crop Suitability and Distance to RailroadRail transport was the dominant mode of freight transport in the famine era. Accordingto National Bureau of Statistics China, rail transport comprised more than 75% of the to-tal freight transport during 1958 to 1961.80 Moreover, because of the weight of grains, theirtransport largely relied on railroad.81 Therefore, conditional on the same crop productivities,counties closer to railroad have larger exposure to grain export due to their lower transporta-tion cost. If grain export indeed aggravate famine severity, we expect that the magnitude ofthe adverse effect of crop suitability diminishes with distance to railway. Furthermore, thisdiminishing relationship should be more pronounced for the high-export-exposure crops. Totest this hypothesis, we classify counties into ShortDistance or LongDistance group, based onwhether their distance to railroad is below or above the median distance. Then, we estimatethe following regression:∆ ln(Pop)i =∑dαrdIdi ψri +∑dαsdIdi ψsi +∑dαwd Idi ψwi + χd + γp + ui , (3.3)where ∆ ln(Pop)i and γp are the same as those defined for equation (3.1); Idi is a indicatorvariable that equals 1 if county i belongs to distance group d; and χd is the distance group fixedeffects. The coefficient αcd reflects the distance-specific relationship between crop suitabilityand famine severity.80Historical data on freight transport by mode are obtained from the 60 Years of New China StatisticalBook published by National Bureau of Statistics China.81According to Donaldson (2016), for agricultural goods, freight rate of railroad transport is considerablylower than that of other transport mode in the colonial India context.86Column (3) in Table 3.2 reports the results of regression (3.3). We find that one standarddeviation higher rice suitability decreases population growth by 0.033 log points in Short-Distance counties and by 0.015 log point in LongDistance counties. The difference betweenαrShort and αrLong is statistically significant at 5% level. We also find that soybean suitabilityhas significantly negative effect on population growth in ShortDistance counties, but smalland insignificant effect in LongDistance counties. In contrast, wheat suitability has insignifi-cant effect on population growths for both distance groups. In column (4), we group countiesinto tertiles according to their distance to railway, and hence have three distance groups, i.e.,ShortDistance, MidDistance and LongDistance. Consistent with the results in column (3), forrice and soybean, the adverse effect of crop suitability monotonically decline with distance.For wheat, however, the estimated coefficients are always insignificant, regardless of distancegroup.In Figure 3.9, we estimate equation (3.2) separately for the ShortDistance and LongDis-tance county groups. The upper panels show the estimated coefficients for the high-export-exposure crops, i.e., rice and soybean. Regardless of distance to railroad, the coefficients offamine birth-cohorts drop below those of non-famine birth-cohorts. More importantly, thedips are more pronounced for ShortDistance group, which implies that conditional on thesame crop productivities, famine was less severe in the county further away from railroads.In addition, we don’t find any negative effect of wheat suitability on the size of famine birth-cohorts. Altogether, the findings from the birth-cohort size analysis are consistent with thosein Table Famine Severity and Crop SpecializationThus far, our analysis relies on biophysics-based measures of crop suitability to proxy acounty’s export exposures. One may question to what extent these measures capture theactual grain production pattern and grain export exposure across counties in China. Toalleviate this concern, we use the actual pre-famine agricultural production data and examinethe correlation of a county’s population growth and its specialization in different crops byestimating the following regression:∆ ln(Pop)i = βrSri + βsSsi + βwSwi + γp + ui , (3.4)where ∆ ln(Pop)i and γp are the same as those defined for equation (3.1). Sci is countyi’s normalized output share of crop c as defined in section 3.3.4, which reflects productionspecialization of crop c across counties before the GLF. When Sci > 1, it indicates that countyi produces crop c proportionally more than its population. These “surplus” outputs are mostlikely to be subject to procurement as national exports of crop c expand. Hence, we considerthat a county with a higher Sci has a larger export exposure in crop c. In the famine years,the over-procurement to cater for export needs could aggravate local food shortage.Columns (1) in Table 3.3 shows the regression results. Consistent with the results inTable 3.2, we find that a higher specialization in rice and soybean is associated with a lowerlog population growth over the period 1953 to 1964. Moreover, the correlation between theextent of specialization in wheat and log population growth is insignificantly different fromzero. Column (3) augments the regression model with controls of weather conditions infamine years. These controls leave the baseline results unaffected.87A caveat of using output share in 1957 as a proxy for regional export exposure is thatin addition to regional differences in fundamental productivity, it may incorporate other un-observed county-specific characteristics and transitory shocks. One may worry that thoseunobserved characteristics and shocks could independently affect the famine severity throughchannels other than export induced over-procurement. To address this concern, we instru-ment the normalized output share of different crops Sci with their productivity ψci . The firststage regression results are reported in Table C.2. We find that a higher productivity ofcultivating crop c raises its own output share, and the effect is statistically and economicallysignificant. For example, as is shown in column (1), one standard deviation higher potentialyield of rice is estimated to increase normalized output share of rice by 0.38. In addition, wefind negative cross-crop productivity effect, that is, a higher productivity of other grain cropstends to depress a crop’s output share conditional on its own productivity. Columns (2) and(4) in Table 3.3 present the IV estimates of equation (3.4). Regardless of the specification, theIV estimates of βr and βs are always negative and significant, while the IV estimates of βware statistically insignificant. Qualitatively, these findings align with with the OLS regressionresults. Quantitatively, however, we find that the IV estimates are larger in magnitude thanthe OLS estimates.3.4.4 Famine Severity and Export ShockIn this subsection, we relate famine severity to regional grain export shock in famine years,which variation stems from cross-county differences in crop specialization and various nationalexport growth across crops. Specifically, we estimate the following regression:∆ ln(Pop)i = κ∆Exporti + γp + ui , (3.5)where ∆ ln(Pop)i and γp are the same as those defined for equation (3.1). ∆Exporti is aBartik-style index as defined in section 3.3.4, which measures the per capita decrease in foodavailability due to grain export growths in the first three years of the GLF (1958-1960). Acounty has a larger export exposure, if relatively to the rest of the country, it specializes morein crops that had larger export growth at the national level.Column (1) in Table 3.4 presents the regression results. One kilogram export shock isestimated to reduce log population growth by 0.064 log points. To gauge the magnitude of theestimated effect of export shock, we can consider the differential changes in log populationassociated with a standard deviation difference in ∆Export (which is 0.745 kilogram percapita per famine year). The point estimate column (1) imply that a standard deviation inexport shock is associated with a 0.048 lower log population growth (around 27% of a standarddeviation in log population growth during 1953-1964). Column (3) includes the weathercontrols and the estimate of δ remains statistically similar to that of the baseline specification.In columns (2) and (4), we instrument the export shock ∆Export with suitability of differentcrops ψci . The IV estimates are significantly negative, but they are always larger in magnitudethan their OLS counterparts. The first stage results are reported in columns (7) and (8) inTable C.2. We find that a higher rice or a higher soybean productivity increases a county’sgrain export exposure, while a higher wheat productivity has insignificant effect on it.883.4.5 Robustness: Pre-Famine, Famine and Post-Famine YearsIn this section, we examine the time-varying effect of crop suitability (crop specialization andgrain export shock) on population, using a cross-county panel data that covers years 1953,1957, 1961 and 1964. This exercise addresses the following concern: if counties that are moresuitable (specialized) in cultivating rice and soybean are on a different trajectory in termsof population growth, our results would just pick up a pre-determined trend instead of aeffect of grain export exposure on famine severity. Specifically, we separate the effect of cropsuitability (crop specialization and grain export shock) on population size in the pre-famine(1957), famine (1961) and post-famine (1964) years. If our baseline results are contaminatedby the county-specific secular trends, we would expect crop suitability (specialization) to becorrelated with population size in the non-famine periods.First, we consider the following equation that relates the crop suitability to county pop-ulation across years:ln(Pop)it =∑tηrτIτt ψri +∑tηsτIτt ψsi +∑tηwτ Iτt ψwi + γpt + φi + uit , (3.6)where ln(Pop)it is the log population of county i in year t ∈ {1953, 1957, 1961, 1964}; Iτt isa dummy variable that equals 1 if year t equals to τ ; and γpt and φi are province×year andcounty fixed effects, respectively. The standard errors are clustered at the prefecture×yearlevel. The coefficients ηcτ captures the time-varying correlation between population size andcrop suitability. As regression (3.6) includes county dummies, the absolute level of the effectof crop suitability ψci on population size cannot be identified. For this reason, the effect in1953 is normalized to zero, i.e., ηc1953 = 0. If crop suitability only affects population growthduring the famine years, we would expect estimated coefficients of ηcs to be 0 in 1957 andnon-zero in 1961, i.e., ηc1957 = 0 and ηc1961 6= 0. We also expect the coefficients to be stabilizedafter the famine, i.e., ηc1961 ≈ ηc1964.The estimates of equation (3.6) are reported in column (1) of Table 3.5. Regardless ofcrop, we do not detect any significant effect of crop suitability on population growth over 1953to 1957, i.e., the estimated coefficients ηc1957 are insignificantly different from zero. In con-trast, we find significantly negative correlation between suitability of high-export-exposurecrops (rice and soybean) and population size in 1961. This finding suggests that in thefamine period, the population in rice- or soybean-suitable counties begins to decline relativeto counties that are not suitable. Moreover, the adverse effect of rice (and soybean) produc-tivity on population persists till 1964, with the estimated coefficient of ηr1961 (ηs1961) beinginsignificantly different from that of ηr1964 (ηs1964). Column (2) include the controls of weatherconditions and the estimates remain stable.Columns (3) re-estimates equation (3.6), but replace suitability measures ψci s with cropspecialization measures Sci . The findings are largely consistent with those in columns (1) and(2). Notice that in column (3), we detect a mild negative effect of soybean specializationon population size in 1957. Nevertheless, this effect is not robust when we include weathercontrols in column (4). Moreover, the estimate of αc1961 is almost three times larger thanthat of αc1957. In addition, we find that wheat specialization is positively associated withpopulation growth in the pre- and post-famine years. Lastly, we link the county populationover time to shock in grain export during the famine period, by estimating the following89regression:ln(Pop)it =∑tρτIτt ∆Exporti + γpt + φi + uit , (3.7)where ln(Pop)it, Iτt , γpt and φi are the same as defined in specification (3.6). ∆Exporti is thesame as defined in section 3.3.4 which captures the shock in grain export in the period 1958to 1960. The coefficients ρτ s reveal the correlation between population size and ∆Exporti ineach year, and ρ1953 is normalized to be 0. As the export shock is specific to the period 1958to 1960, we would expect that it only exerts negative effect on population growth in famineperiod, but insignificant effect in the pre- and post-famine years. That is, ρ1957 ≈ 0, ρ1961 < 0and ρ1961 ≈ ρ1964. Column (5) report the estimates of equation (3.7). Indeed, we find thatpopulation in large-export-shock counties decreases relative to the low-export-shock countiesover 1957 to 1961. Such differential changes are absent in the pre-famine (1953-1957) andpost-famine (1961-1964) eras.The placebo tests in columns (1) to (6) provide supporting evidence for the followingclaims: i) our baseline results are unlikely to be severely biased by the county-specific pre-existing trend in population growth that are correlated with crop suitability or specialization;and ii) the inter-census (1953 to 1964) population growth is arguably a close proxy for severityof the 1959-1961 famine.3.5 Concluding RemarksBy analyzing counties that are subject to heterogeneous grain export exposure due to theirdifferent suitability and specialization in cultivating different crops, this paper examines theeffect of over-export on the severity of China’s 1959-1961 famine. Our main findings are:productivity in rice and soybean (crops that had large export surge during famine years) arepositively correlated with famine severity; however, productivity of wheat (a crop that thathad small export shock during famine years) are uncorrelated (sometimes negatively corre-lated) with famine severity. Moreover, for the high-export-exposure crops, the correlationbetween productivity and famine severity decline with distance to railroad. These resultsare robust when we use counties’ specialization in different crops to proxy for grain exportexposure.Can this data pattern be explained by the overall over-procurement of grain crops? Toanswer this question, we resort to the historical data on procurement and export by crops, asreported in Table 3.6. We find that the procurement of wheat grew faster than those of riceand soybean over the famine years. For example, over 1957-1959, the procurement of wheatincreased by 38.3%, while the procurement of rice and wheat increased by 22.1% and 21.3%,respectively. Nevertheless, only 8.7% of the increment in wheat procurement was to cater forthe growth in export, while the proportion was 26.2% for rice and was 52.8% for soybean.If the spatial pattern of famine severity is mainly driven by grain procurement for domesticdistributional purpose, we would expect that counties that have higher productivity in wheatto be more severely hit by famine relative to counties that have higher productivity in riceand soybean. Our data, nevertheless, show an opposite pattern.90Figure 3.1: Export and Import (1955-1961)46.5553.32 52.6463.3275.1163.1445.9020406080100 million RMB1955 1956 1957 1958 1959 1960 1961Total Export36.3242.08 43.1652.8165.37 62.4344.53020406080100 million RMB1955 1956 1957 1958 1959 1960 1961Total Import7.598.636.389.3513.218.522.35051015100 million RMB1955 1956 1957 1958 1959 1960 1961Export of Grain Products.26 0 0 .29 0 .048.46051015100 million RMB1955 1956 1957 1958 1959 1960 1961Import of Grain Products91Figure 3.2: Composition of Grain Exports (1955-1961)01234million tons1955 1956 1957 1958 1959 1960 1961Wheat Soybean Rice Maize OthersFigure 3.3: Average Exports over 1958-1960 versus Average Exports over 1955-1957.992. Export 1958−60 − Avg. Export 1955−57 (million tons)Wheat Soybean Rice Maize Others92Figure 3.4: Spatial Distribution of Potential Yields by CropsPanel A: Rice Panel B: SoybeanPanel C: Wheat93Figure 3.5: Railway Network in China over TimePanel A: 1960Panel B: 199094Figure 3.6: Relative Pirce over 1955-1961.811.21.41.6Price Soybean/Price Wheat11. Rice/Price Wheat1955 1956 1957 1958 1959 1960 1961YearPanel A: China Export Price1. Rice/Price Wheat1956 1957 1958 1959 1960 1961YearPanel B: International Price1. Soybean/Price Wheat2. Rice/Price Wheat1955 1956 1957 1958 1959 1960 1961YearPanel C: US Domestic PriceRelative Price Rice Relative Price SoybeanNote: Panel A shows the export price of rice and soybean to that of wheat. The data on export price is from China CustomsStatistics Yearbooks (1955-1961) by Customs of China. Panel A shows the export price of rice from Thailand (5% broken) to thatof wheat from the US (No.2 hard winter wheat). The data is obtained from Palacpac (1977). Panel C shows the domestic priceof rice and soybean relative to that of wheat in the US and the data are collected from Statistical Abstract of the United States(1960,1963) by the US Census Bureau.95Figure 3.7: Kernel Densities of Precipitation and Temperature in Spring and Summer: 1959-19610.05.1.150 5 10 15 20 25Precipitation, cmSpring0. 10 20 30 40Precipitation, cmSummer0.05.1.15−10 −5 0 5 10 15 20 25Temperature, Celsius degreeSpring0. 15 20 25 30Temperature, Celsius degreeSummer1959 19601961 Avg. 1950−5896Figure 3.8: The Correlation between Crop Suitabilities and Birth-Cohort Size over Time−.2−.15−.1−.050.05.11950 1955 1960 1965CohortRice−.2−.15−.1−.050.05.11950 1955 1960 1965CohortSoybean−.2−.15−.1−.050.05.11950 1955 1960 1965CohortWheatCoef. 95% CI97Figure 3.9: The Correlation between Crop Suitabilities and Birth-Cohort Size over Time by Distance Groups−.1−.0501950 1955 1960 1965CohortRice−.1−.050.051950 1955 1960 1965CohortSoybean−.1−.050.051950 1955 1960 1965CohortWheatNear Far98Table 3.1: Summary Statisticsmean std 75th-2525∆ lnPop 0.178 0.176 0.178Normalized Output Share: ScRice 0.910 0.981 1.754Soybean 0.919 1.352 0.899Wheat 0.922 1.062 1.303∆Export (tons per capita) 1.159 0.745 1.212Distance to Railroad (kilometer) 67.576 64.683 74.716Spring Temperature (C◦)1959 9.087 5.177 5.9781960 9.210 5.591 5.6061961 8.827 5.295 4.934Summer Temperature (C◦)1959 22.677 2.903 3.8731960 22.518 2.899 3.8491961 23.472 2.993 4.194Spring Precipitation (cm)1959 7.407 6.782 12.1501960 5.611 4.961 8.5451961 7.073 6.993 11.695Summer Precipitation (cm)1959 14.641 7.768 10.7251960 13.887 5.957 9.9281961 12.964 5.538 8.045Notes: The summary statistics is based on the 1953-1964 sample.99Table 3.2: Famine Severity, Crop Suitability and Distance to RailroadDep. Var:∆ ln(Pop)i,53−64 (1) (2) (3) (4)ψr -0.017*** -0.022***(0.006) (0.008)ψs -0.039** -0.032**(0.015) (0.016)ψw 0.025* 0.017(0.013) (0.011)ψr × ShortDistance -0.033*** -0.039***(0.010) (0.010)ψr ×MidDistance -0.024**(0.009)ψr × LongDistance -0.015* -0.019*(0.008) (0.010)ψs × ShortDistance -0.043** -0.053***(0.017) (0.019)ψs ×MidDistance -0.035**(0.016)ψs × LongDistance -0.011 0.011(0.022) (0.033)ψw × ShortDistance 0.016 0.023(0.013) (0.015)ψw ×MidDistance 0.008(0.014)ψw × LongDistance 0.009 0.011(0.017) (0.026)Province Y Y Y YWeather Controls Y Y YN 1,164 1,164 1,164 1,164R2 0.355 0.408 0.413 0.419Notes: Weather controls include spring temperature and its squared term, summer temperature and its squaredterm, spring precipitation and its squared term, and summer precipitation and its squared term for famine years1959, 1960 and 1961. Robust standard errors are clustered at the prefecture level. *** p<0.01, ** p<0.05, *p<0.1100Table 3.3: Famine Severity and Production SepcializationDep. Var: (1) (2) (3) (4)∆ ln(Pop)i,53−64 OLS IV OLS IVSr -0.040*** -0.067*** -0.046*** -0.081***(0.008) (0.022) (0.010) (0.029)Ss -0.045*** -0.099** -0.040*** -0.060*(0.009) (0.044) (0.007) (0.035)Sw -0.001 0.032 0.002 0.028(0.008) (0.026) (0.008) (0.030)Province Y Y Y YWeather Controls Y YN 1,164 1,164 1,164 1,164R2 0.397 0.308 0.443 0.420Notes: Weather controls include spring temperature and its squared term, summer temperature and itssquared term, spring precipitation and its squared term, and summer precipitation and its squared termfor famine years 1959, 1960 and 1961. Robust standard errors are clustered at the prefecture level. ***p<0.01, ** p<0.05, * p<0.1Table 3.4: Famine Severity and Export ExposureDep. Var: (1) (2) (3) (4)∆ ln(Pop)i,53−64 OLS IV OLS IV∆Export -0.064*** -0.083*** -0.069*** -0.120***(0.012) (0.023) (0.012) (0.037)Province Y Y Y YWeather Controls Y YN 1,164 1,164 1,164 1,164R2 0.375 0.371 0.430 0.412Notes: Weather controls include spring temperature and its squared term, summer temperature and itssquared term, spring precipitation and its squared term, and summer precipitation and its squared termfor famine years 1959, 1960 and 1961. Robust standard errors are clustered at the prefecture level. ***p<0.01, ** p<0.05, * p<0.1101Table 3.5: Robustness: The Effect of Export Exposure on Population Size: Pre-Famine, Famine and Post-Famine YearsDep. Var:ln(Pop)it(1) (2) (3) (4) (5) (6)ψr × 1957 -0.007 -0.009 Sr × 1957 0.001 0.001 ∆Export 0.000 -0.002(0.005) (0.006) (0.012) (0.013) ×1957 (0.013) (0.013)ψr × 1961 -0.015*** -0.019*** Sr × 1961 -0.042*** -0.046*** ∆Export -0.046*** -0.051***(0.006) (0.006) (0.012) (0.011) ×1961 (0.012) (0.011)ψr × 1964 -0.020*** -0.024*** Sr × 1964 -0.037*** -0.037*** ∆Export -0.044*** -0.043***(0.005) (0.006) (0.012) (0.012) ×1964 (0.012) (0.013)ψs × 1957 -0.030 -0.027 Ss × 1957 -0.023* -0.024(0.026) (0.027) (0.014) (0.015)ψs × 1961 -0.045* -0.043* Ss × 1961 -0.062*** -0.061***(0.026) (0.026) (0.012) (0.012)ψs × 1964 -0.032 -0.035 Ss × 1964 -0.076*** -0.071***(0.027) (0.026) (0.013) (0.014)ψw × 1957 0.025 0.024 Sw × 1957 0.035* 0.040*(0.029) (0.029) (0.020) (0.021)ψw × 1961 0.045 0.042 Sw × 1961 0.025 0.022(0.028) (0.027) (0.020) (0.019)ψw × 1964 0.044 0.047* Sw × 1964 0.037* 0.039**(0.029) (0.027) (0.021) (0.019)County Y Y Y Y Y YProvince×Year Y Y Y Y Y YWeather Controls Y Y YN 3,432 3,432 3,432 3,432 3,432 3,432R2 0.988 0.988 0.988 0.988 0.988 0.988Notes: Weather controls include spring temperature and its squared term, summer temperature and its squared term, spring precipitation and itssquared term, and summer precipitation and its squared term. Robust standard errors are clustered at the prefecture×year level. *** p<0.01, **p<0.05, * p<0.1102Table 3.6: Output, Procurement, Export and Export Price by CropYear Output (10000 tons) Procurement (10000 tons)Rice Soybean Wheat Rice Soybean Wheat1955 7803 912 2297 1999.6 485.7 802.01956 8248 1024 2480 1909.7 493.2 753.51957 8678 1005 2364 2092.6 468.6 719.01958 8085 867 2259 2200.4 463.4 863.51959 6937 876 2218 2554.0 568.3 994.91960 5973 639 2217 1845.1 302.3 825.71961 5364 621 1425 1453.0 253.4 356.4Year Export (10000 tons) Export Price (RMB/kg)Rice Soybean Wheat Rice Soybean Wheat1955 71.1 111.5 6 0.42 0.34 0.301956 107.6 117.3 4.8 0.39 0.32 0.361957 52.6 115.9 0.1 0.37 0.33 0.381958 137.4 124.6 2.6 0.36 0.32 0.21959 173.4 168.5 24.0 0.38 0.31 0.251960 112.6 112.5 16.7 0.36 0.3 0.271961 43.7 41.3 0.1 0.28 0.24 0.29Notes: Output data by crop are collected from Statistics of agriculture in the fifty years of China,(1949-1998) by National Statistics Bureau; procurement data by crop are collected from HistoricalAccounts on China Food Production by the Chinese Ministry of Commerce; data on export and exportprice are obtained from China Customs Statistics Yearbooks (1955-1961) by Customs of China.103BibliographyAlmond, Douglas, Hongbin Li, and Shuang Zhang. 2013. “Land Reform and Sex Selection inChina.” NBER Working Paper.Antweiler, Werner, Brian R. Copeland, and M. Scott Taylor. 2001. “Is Free Trade Good forthe Environment?” The America Economic Review 94(4): 877-908.Arceo, Eva, Rema Hanna, and Paulina Oliva. 2016. “Does the Effect of Pollution on InfantMortality Differ between Developing and Developed Countries? Evidence from MexicoCity.” The Economic Journal 126(591): 257-280.Ashton, Basil, Kenneth Hill, Alan Piazza and Robin Zeitz. 1984. “Famine in China, 1958-61.”Population and Development Review 10(4): 613-645.Atkin, David. 2015. “Endogenous Skill Acquisition and Export Manufacturing in Mexico.”Working Paper.Autor, David H., David Dorn, and Gordon H. Hanson. 2013. “The China Syndrome: LocalLabor Market Effects of Import Competition in the United States.” The America EconomicReview 103(6): 2121-2168.Autor, David H., David Dorn, Gordon H. Hanson, and Jae Song. 2015. “Trade Adjustment:Worker-level Evidence.” The Quarterly Journal of Economics 129(4): 1799-1860.Bartik, Timothy J.. 1991. Who Benefits from State and Local Economic Development Poli-cies? Kalamazoo, Mich. : W.E. Upjohn Institute for Employment Research.Balsvik, Ragnhild, Sissel Jensen, and Kjell G. Salvanes. 2015. “Made in China, Sold inNorway: Local Labor Market Effects of an Import Shock.” Journal of Public Economics127: 137-144.Berger, Daniel, William Easterly, Nathan Nunn, and Shanker Satyanath . 2013. “CommercialImperialism? Political Influence and Trade During the Cold War.” The American EconomicReview 103(2): 863-896.Blanchard, Emily, and William W. Olney. 2014. “Globalization and Human Capital Invest-ment: How Export Composition Drives Educational Attainment.” Working Paper.Borsook, Ian. 1987. “Earnings, Ability and International Trade.” Journal of InternationalEconomics 22(3): 281-295.Bown, Chad P.. 2013. “How Different are Safeguards from Antidumping? Evidence from USTrade Policies toward Steel.” Review of Industrial Organization 42(4): 449-481.104Brandt, Loren, Johannes V. Biesebroeck, and Yifan Zhang. 2012. “Creative Accounting orCreative Destruction? Firm-level Productivity Growth in Chinese Manufacturing.” Journalof Development Economics 97(2): 335-351.Burstein, Ariel, and Jonathan Vogel. 2012. “International Trade, Technology and the SkillPremium.” Working Paper.Burstein, Ariel, Eduardo Morales, and Jonathan Vogel. 2015. “Accounting for Changes inBetween-Group Inequality.” Working Paper.Bustos, Paula, Bruno Caprettini and Jacopo Ponticelli. 2016. “Agricultural Productivityand Structural Transformation. Evidence from Brazil.” The American Economic ReviewForthcoming.Cao, Shuji. 2005. “The Great Famine.” Hong Kong: Time International Publishing Ltd..Chang, Gene Hsin, and Guanzhong James Wen. 1998. “Food Availability versus ConsumptionEfficiency: Causes of the Chinese Famine.” China Economic Review 9(2): 157-166.Chay, Kenneth Y., Michael Greenstone. 2003a. ”Air Quality, Infant Mortality, and the CleanAir Act of 1970.” NBER Working Paper No. 10053.Chay, Kenneth Y., Michael Greenstone. 2003b. ”The Impact of Air Pollution on Infant Mor-tality: Evidence from Geographic Variation in Pollution Shocks Induced by a Recession.”Quarterly Journal of Economics 118(3): 1121-67.Chen, Yuyu, Avraham Ebenstein, Michael Greenstone, and Hongbin Li. 2013. ”Evidence onthe Impact of Sustained Exposure to Air Pollution on Life Expectancy from China’s HuaiRiver Policy.” Proceedings of the National Academy of Sciences 110(6): 12936-41.Chen, Shuo, and James Kai-sing Kung. 2011. “The Tragedy of the Nomenklatura: CareerIncentives and Political Radicalism during China’s Great Leap Famine.” American PoliticalScience Review 105(1): 27-45.Chen, Yuyu, Hongbin Li and Lingsheng Meng. 2013. “Prenatal Sex Selection and MissingGirls in China.” The Journal of Human Resources 48(1): 36-70.CIA, The U.S. Central Intelligence Agency. 1967. “Communist China Map Folio.”Coale, Ansley J.. 1981. “Population Trends, Population Policy, and Population Studies inChina.” Population and Development Review 46(1): 1-34.Copeland, Brian R., and M. Scott Taylor. 2003. Trade and the Environment: Theory andEvidence. Princeton: Princeton University Press.Copeland, Brian R., and M. Scott Taylor. 2004. “Trade, Growth and the Environment.”Journal of Economic Literature 42(1): 7-71.Costinot, Arnaud, Dave Donaldson, and Ivana Komunjer. 2012. “What Goods Do CountriesTrade? A Quantitative Exploration of Ricardo’s Ideas.” Review of Economic Studies 79(2):581-608.105Currie, Janet, Lucas Davis, Michael Greenstone, and Reed Walker. 2015. ”EnvironmentalHealth Risks and Housing Values: Evidence from 1,600 Toxic Plant Openings and Clos-ings.” The American Economic Review 105(2): 678-709.Currie, Janet, and Matthew Neidell. 2005. ”Air Pollution and Infant Health: What Can WeLearn from California’s Recent Experience?” The Quarterly Journal of Economics 120(3):1003-1030.Currie, Janet, Matthew Neidell, and Johannes F. Schmieder. 2009. ”Air Pollution and InfantHealth: Lessons from New Jersey” Journal of Heath Economics 28(3): 688-703.Danziger, Eliav. 2014. “Skill Acquisition and the Dynamics of Trade-Induced Inequality.”Working Paper.de Sousa, Jose´, Laura Hering and Sandra Poncet. 2015. “Has Trade Openness ReducedPollution in China?” CEPPI Working Paper.Dean, Judith M.. 2002. “Does Trade Liberalization Harm the Environment? A New Test.”Canadian Journal of Economics 35(4): 819-842.Dean, Judith M., and Mary E. Lovely. 2010. “Trade Growth, Production Fragmentation, andChina’s Environment.” China’s Growing Role in World Trade. Ed. R. Feenstra and S. Wei.Chicago: NBER and University of Chicago Press.Dean, Judith M., Mary E. Lovely and Hua Wang. 2009. “Are Foreign Investors Attractedto Weak Environmental Regulations? Evaluating the Evidence from China.” Journal ofDevelopment Economics 90(1): 1-13.Dell, Melissa, Benjamin F. Jones, and Benjamin A. Olken. 2012. “Temperature Shocks andEconomic Growth: Evidence from the Last Half Century.” American Economic Journal:Macroeconomics 4(3): 66-95.Dix-Carneiro, Rafael, and Brian K. Kovak. 2015. “Trade Liberalization and the Skill Pre-mium: A Local Labor Markets Approach.” American Economic Review: Papers and Pro-ceedings 105(5): 551-557.Donaldson, Dave. 2016. “Railroad of the Raj: Estimating the Impact of TransportationInfrastructure.” The American Economic Review Forthcoming.Eaton, Jonathan, and Samuel Kortum. 2002. “Technology, Geography, and Trade.” Econo-metrica 70(5): 1741-1779.Ebenstein, Avraham, Margaret McMillan, Yaohui Zhao, and Chanchuan Zhang. 2012. “Un-derstanding the Role of China in the “Decline” of US Manufacturing.” Working Paper.Ebenstein, Avraham, Maoyong Fan, Michael Greenstone, Guojun He, Peng Yin and MaigengZhou. 2015. “Growth, Pollution, and Life Expectancy: China from 1991-2012.” AmericanEconomic Review: Papers and Proceedings 105(5): 226-231.Edmonds, Eric V., and Nina Pavcnik. 2005. “The Effect of Trade Liberalization on ChildLabor.” Journal of International Economics 65(2): 401-419.106Edmonds, Eric V., Nina Pavcnik, and Petia Topalova. 2010. “Trade Adjustment and HumanCapital Investment: Evidence from India Tariff Reform.” American Economic Journal:Applied Economics 2(4): 42-75.FAO, Food and Agriculture Organization of the United Nations. 2003. “Food Energy–Methods of Analysis and Conversion Factors.” FAO Food and Nutrition Paper 77.Findlay, Ronald, and Henryk Kierzkowski. 1983. “International Trade and Human Capital:A Simple General Equilibrium Model.” Journal of Political Economy 91(6): 957-978.Frankel, Jeffrey A., and Andrew K. Rose. 2005. “Is Trade Good or Bad for the Environment?Sorting Out the Causality.” Review of Economics and Statistics 87(1): 85-91.Garver, John W.. 2016. “China’s Quest: The History of the Foreign Relations of the People’sRepublic of China.” New York: Oxford University Press.Galle, Simon, Andre´s Rodr´ıguez-Clare and Moises Yi. 2015. “Slicing the Pie: Quantifyingthe Aggregate and Distributional Effects of Trade.” Working Paper.Greenstone, Michael, and Rema Hanna. 2014. “Environmental Regulation, Air and WaterPollution, and Infant Mortality in India.” The America Economic Review 104(10): 3038-3072.Grossman, Gene M., and Alan B. Krueger. 1995. “Economic Growth and the Environment.”The Quarterly Journal of Economics 110(2): 353-377.Hakobyan, Shushanik, and John McLaren. 2010. “Looking for Local Labor-Market Effects ofNAFTA.” NBER Working Paper 16535.Han, Jun, Runjuan Liu and Junsen Zhang. 2012. “Globalization and wage inequality: Evi-dence from urban China” Journal of International Economics 87(2): 288-297.Harris, Richard G., and Peter E. Robertson. 2013. “Trade, Wages and Skill Accumulation inthe Emerging Giants” Journal of International Economics 89(2): 407-421.Head, Keith, and Thierry Mayer. 2014. “Gravity Equations: Workhorse, Toolkit, Cookbook.”Handbook of International Economics Vol.4, ed. Gita Gopinath, Elhanan Helpman andKenneth Rogoff. New York: Elsevier.Head, Kieth, and Thierry Mayer and John Ries. 2010. “The Erosion of Colonial Trade Linkageafter Independence.” Journal of International Economics 81(1): 1-14.Hsieh, Chang-Tai, Erik Hurst, Charles I. Jones and Peter J. Klenow. 2013. “The Allocationof Talents and U.S Economic Growth.” Working PaperJayachandran, Seema. 2009. “Air Quality and Early-Life Mortality Evidence from Indonesia’sWildfires.” The Journal of Human Resources 44(4): 916-954.Jia, Ruixue. 2014. ”Pollution for Promotion.” Working Paper.107Jia, Ruixue, Masasyuki Kudamatsu and David Seim. 2015. ”Political Selection in China:The Complementary Roles of Connections and Performance.” Journal of the EuropeanEconomic Association 13(4): 631-668.Johnson, D. Gale. 1998. “China’s Great Famine: Introductory Remarks.” China EconomicReview 9(2): 103-109.Kaplan, Greg, and Sam Schulhofer-Wohl. 2013. “Understanding the Long-Run Decline inInterstate Migration.” Working Paper, Federal Reserve Bank of Minneapolis.Kovak, Brian K.. 2013. “Regional Effects of Trade Reform: What is the Correct Measure ofLiberalization.” The America Economic Review 103(5): 1960-1976.Kung, James Kai-sing, and Justin Yifu Lin. 2003. “The Causes of China’s Great Leap Famine,1959-1961.” Economic Development and Cultural Change 52(1): 51-73.Lagakos, David, and Michael E. Waugh. 2013. “Selection, Agriculture, and Cross-CountryProductivity Differences.” The America Economic Review 103(2): 948-980.Levinson, Arik. 2009. ”Technology, International Trade, and Pollution from US Manufactur-ing.” The American Economic Review 99(5): 2177-92.Li, Bingjing. 2016. ”Export Expansion, Skill Acquisition and Industry Specialization: Evi-dence from China.” Working Paper, University of British Columbia.Li, Wei, and Dennis Tao Yang. 2005. “The Great Leap Forward: Anatomy of a CentralPlanning Disaster.” Journal of Political Economy 113(4): 840-877.Lin, Justin Yifu. 1990. “Collectivization in China’s Agricultural Crisis in 1959-1961.” Journalof Political Economics 98(6): 1228-1252.Lin, Jintai, Da Pan, Steven J. Davis, Qiang Zhang, Kebin He, Can Wang, David G. Streets,Donald J. Wuebbles, and Dabo Guan. 2014. ”China’s International Trade and Air Pollutionin the United States.” Proceedings of the National Academy of Sciences 111(5): 1736-1741.Lin, Justin Yifu, and Dennis Tao Yang. 1998. “On the Causes of China’s Agricultural Crisisand the Great Leap Famine.” China Economic Review 9(2): 125-140.Lin, Justin Yifu, and Dennis Tao Yang. 2000. “Food Availability, Entitlements and the Chi-nese Famine of 1959-61.” Economic Journal 110(460): 840-877.Ma, Yue, Heiwai Tang and Yifan Zhang. 2014. “Factor Intensity, Product Switching and Pro-ductivity: Evidence from Chinese Exporters.” Journal of International Economics 92(2):349-362.Marden, Samuel. 2015. “The Agriculture Roots of Industrial Development: ’Forward Link-ages’ in Reform Era China.” Working Paper.Matsuura, Kenji, and Cort Willmott. 2007. Terrestrial Air Temperature and Precipitation:Monthly and Annual Time Series: (1950-1996) Version 1.01. University of Delaware.http://climate.geog.udel.edu/ climate/.108Meng, Xin, Nancy Qian, and Pierre Yared. 2015. “The Institutional Causes of China’s GreatFamine, 1959-1961.” Review of Economic Studies 82(4): 1568-1611.Nunn, Nathan, and Nancy Qian. 2011. “The Impact of Potatoes on Old World Populationand Urbanization.” Quarterly Journal of Economics 126(2): 593-650.Palacpac, Adelita C.. 1977. “World Rice Statistics.” The International Rice Research Insti-tute, Department of Agricultural Economics.Pierce, Justin R., and Peter K. Schott. 2012. “The Surprising Swift Decline of U.S. Manu-facturing Employment.” NBER Working Paper 18655.Pope III, C. Arden. 1989. “Respiratory Disease Associated with Community Air Pollutionand a Steel Mill, Utah Valley.” American Journal of Public Health 75(5): 623-628.Pope III, C. Arden. 1996. “Adverse Health Effects of Air Pollutants in a Nonsmoking Popu-lation.” Toxicology 111(1): 149-155.Pope III, C. Arden, Richard T. Burnett, Michael J. Thun, Eugenia E. Calle, Daniel Krewski,Kazuhiko Ito, and George D. Thurston. 2002. “Lung Cancer, Cardiopulmonary Mortality,and Long-term Exposure to Fine Particulate Air Pollution.” JAMA 287(9): 1132-1141.Ransom, Michael R., and C. Arden Pope III. 2005. “External Health Costs of a Steel Mill.”Contemporary Economic Policy 13(2): 86-97.Read, Robert. 2005. “The Political Economy of Trade Protection: the Determinants andWelfare Impact of the 2002 US Emergency Steel Safeguard Measures.” The World Economy28(8): 1119-1137.Riskin, Carl. 1998. “Seven Questions about the Chinese Famine of 1959-61.” China EconomicReview 9(2): 111-124.Romalis, John. 2004. “Factor Proportions and the Structure of Commodity Trade.” TheAmerica Economic Review 94(1): 67-97.Shapiro, Joseph S.. 2015. ”Trade, CO2, and the Environment.” Working Paper, Yale Univer-sity.Shapiro, Joseph S., and Reed Walker. 2015. ”Why is Pollution from U.S. Manufacturing De-clining? The Roles of Trade, Regulation, Productivity, and Preferences.” NBER WorkingPaper.Staiger, Robert W., and Frank A. Wolak. 1994. ”Measuring Industry-Specific Protection: An-tidumping in the United States.” Brookings Papers on Economic Activity: Microeconomics1994: 51-118.Stiglitz, Joseph E.. 1970. “Factor Price Equalization in a Dynamic Economy.” Journal ofPolitical Economy 78(3): 456-488.Simonovska Ina, Michael E. Waugh. 2014. “The Elasticity of Trade: Estimates and Evidence.”Journal of International Economics 92(1): 34-50.109Tanaka, Shinsuke, 2015. “Environmental Regulations on Air Pollution in China and TheirImpact on Infant Mortality.” Journal of Health Economics 42: 90-103.Topalova, Petia. 2010. “Factor Immobility and Regional Impacts of Trade Liberalization:Evidence on Poverty from India.” American Economic Journal: Applied Economics 2(4):1-41.Tombe, Trevor, and Xiaodong Zhu. 2014. “Trade, Migration and Regional Income Differences:Evidence from China.” Working PaperUnited States Census Bureau. 1960 and 1963. “Statistical Abstract of the United States.”Utar, Hale. 2014. “When the Floodgates Open: ”Northern” Firms’ Responses to Removal ofTrade Quotas on Chinese Goods.” American Economic Journal: Applied Economics 6(4):226-250.Xue, Susan. 2010. “New Local Gazetteers from China.” Collection Building 29(3): 110-118.Yan, Yunfeng, and Laike Yang. 2010. “China’s Foreign Trade and Climate Change: A CaseStudy of CO2 Emissions.” Energy Policy 38(1): 350-356.Yang, Dali Li., and Fubing Su. 1998. “The Politics of Famine and Reform in Rural China.”China Economic Review 9(2): 111-124.Yao, Shujie. 1999. “A Note on the Causal Factors of China’s Famine in 1959-1961.” TheJournal of Political Economy 107(6): 1365-1369.110Appendix AAppendix for Chapter 1A.1 ProofsA.1.1 Share of Educated WorkersNote that the share of educated workers is determined by pih = Pr(zh ≥ zl/µ). LetG(zh, zl) = ∂F (zh, zl)/∂zl = exp[−(z− κ1−νh + z− κ1−νl )1−ν ]κ(z− κ1−νh + z− κ1−νl )−νz− κ1−ν−1lthen,pii =∫ ∞0(1−G(zl/µ, zl))dzl = 1/(µ−κ1−ν + 1).A.1.2 Expected Productivity of Educated and Uneducated WorkersFirst, I show that after education decisions, the ex-post productivity of educated workers fol-lows the distribution Fre´chet(1+µ−κ,κ), and the ex-post productivity of uneducated workersfollows the distribution Fre´chet(1 + µκ,κ). To see this, the distribution of productivity ofeducated workers isPr(zh ≤ z|Educated) = Pr(zh ≤ z|zh ≥ zl/µ) = Pr(zh ≤ z, zh ≥ zl/µ)Pr(zh ≥ zl/µ) .Note thatPr(zh ≤ z, zh ≥ zl/µ) =∫ µz0(G(z, zl)−G(zl/µ, zl))dzl = 11 + µ−κ1−νexp[−(1 +µ− κ1−ν )1−νzκ].As Pr(zh ≥ zl/µ) = 11+µ− κ1−ν, we havePr(zh ≤ z|Educated) = e−(1+µ−κ1−ν )1−νz−κ .Similarly, it can be shown thatPr(zl ≤ z|Uneducated) = e−(1+µκ1−ν )1−νz−κ .Therefore,E(zh|Educated) = γpi−(1−ν)/κh and E(zl|Uneducated) = γ(1− pih)−(1−ν)/κ.where γ = Γ(1− 1/κ). The supplies of skilled and unskilled labor are then determined byH = (T − ϕ)pihE(zh|Educated) = (T − ϕ)γpi1−(1−ν)/κhandL = T (1− pih)E(zl|Uneducated) = Tγ(1− pih)1−(1−ν)/κ.111A.1.3 Prices and Trade FlowsPrices. For good k, the prices which region i presents to region j, pij,k(ω), follow thedistribution Gij,k(p) = 1 − exp[−(vi,kτij,k)−εpε]. As in Eaton and Kortum (2002), pricedistributions have three useful properties:(a) For each good k, the probability that region i provides a variety at the lowest price inregion j, λij,k satisfiesλij,k = Prob(pij,k(ω) = mini′{pi′j,k(ω)}) =∫ ∞0∏i′ 6=i[1−Gi′j,k(p)]dGij,k(p) = (vi,kτij,k)−ε∑i′(vi′,kτi′j,k)−ε .(b) The price distribution of good k prevailing in region j isGj,k(p) = 1−N∏i=1(1−Gij,k(p)) = 1− exp[−∑i′(vi′,kτi′j,k)−εpε].Therefore, the price index of good k in region j isPj,k = (∑i(vi,kτij,k)−ε)−1εΓ(ε+ 1− σkσk)11−σk .(c) The price of good k that region j actually buys from region i also has the distribu-tion Gj,k(p). To see this, conditional on the lowest price supplier being i, the pricedistribution of pj,k(ω) satisfiesProb(pj,k(ω) ≤ p|pij,k(ω) = mini′{pi′j,k(ω)}) = 1λij,k∫ p0∏i′ 6=i[1−Gi′j,k(q)]dGij,k(q) = Gj,k(p).Trade Flows. The fraction of region j’s expenditure on good k allocated to the goodsproduced in region i satisfiesXij,kXj,k=∑ω∈Ω[pj,k(ω)1(pij,k(ω) = mini′{pi′j,k(ω)})]1−σk∑ω∈Ω pj,k(ω)1−σk= λij,kE[pj,k(ω)1−σk |pij,k(ω) = mini′{pi′j,k(ω)}]E[pj,k(ω)1−σk ]= λij,kwhere the second equality employs the strong law of large numbers for independent andidentically distributed random variables and the continuous mapping theorem, and the lastequality uses the properties (b) and (c) of price distributions. Therefore, λij,k denotes notonly the share of varieties of good k consumed by region j that originate from region i, butalso the expenditure of region j on good k from region i.112A.2 Reduced Form Relation between Export Shocks andSchool EnrollmentThis appendix shows how the export shocks affect school enrollment in prefecture i in theshort run, by disturbing the equilibrium system (1.3)-(1.5) with the exogenous changes iniceberg cost τˆij,k (τˆii,k = 0) and productivity ψˆi,k. In this section xˆ ≡ ∆ ln(x) = ∆x/xdenotes the percentage change in any variable x between the initial and new equilibrium.Each prefecture is treated as a small open economy, and by assumption, the export shocksto prefecture i have no effect on the wages and income levels of other regions. In addition, Ionly consider the short-term effect such that Hˆi = Lˆi = 0.A.2.1 LinearizationTaking log-linearization of Equations (1.3) and (1.4) obtainswˆih =∑khi,kYˆi,k (A.1)wˆil =∑kli,kYˆi,k (A.2)where hi,k and li,k denote the share of skilled labor and unskilled labor allocated to sector kin region i, respectively. Taking log-linearization of equation (1.5) obtainsYˆi,k = −ε∑jγij,k(1− λij,k)(αkwˆi,h + (1− αk)wˆi,l − ψˆi,k + τˆij,k) + γii,k(θiwˆi,h + (1− θi)wˆi,l)= − (δi,kεαk − γii,kθi)︸ ︷︷ ︸ai,k1wˆi,h − (δi,kε(1− αk)− γii,k(1− θi))︸ ︷︷ ︸ai,k2wˆi,l + δi,kεψˆi,k − ετˆi,k(A.3)for each sector k. Here γij,k denotes prefecture i’s revenue share in sector k that is frommarket j, and θi =wi,hHiwi,hHi+wi,lLiis the income share of skilled workers in prefecture i. Tosimplify the notations, I define δi,k =∑j γij,k(1− λij,k) and τˆi,k =∑j 6=i γij,k(1− λij,k)τˆij,k.The system of linear Equations (A.1)-(A.3) can be written in the matrix form:1 0 . . . 0 ai,11 ai,120 1... ai,21 ai,22.... . . 0......0 . . . 0 1 ai,K1 ai,K2−hi,1 −hi,2 . . . −hi,K 1 0−li,1 −li,2 . . . −li,K 0 1Yˆi,1Yˆi,2...Yˆi,Kwˆi,hwˆi,l=δi,1εψˆi,1 − ετˆi,1δi,2εψˆi,2 − ετˆi,2...δi,Kεψˆi,K − ετˆi,K00where ai,k1 = δi,kεαk − γii,kθi and ai,k2 = δi,kε(1 − αk) − γii,k(1 − θi). To solve for theendogenous variables wˆih and wˆil, the system is rewritten into more compact matrix notation[IK×K AK×2Φ2×K I2×2] [YˆK×1wˆ2×1]=[SˆK×102×1].113Cramer’s rule and the rule for the determinant of a partitioned matrix are used to solve forthe changes in skilled and unskilled wages:wˆih =det(Xh −ΦSh)det(I−ΦA) and wˆil =det(Xl −ΦSl)det(I−ΦA)where Xh =[0 00 1]Xl =[1 00 0]Sh =δi,1εψˆi,1 − ετˆi,1 ai,12δi,2εψˆi,2 − ετˆi,2 ai,22......δi,Kεψˆi,K − ετˆi,K ai,K2 Sl =ai,11 δi,1εψˆi,1 − ετˆi,1ai,21 δi,2εψˆi,2 − ετˆi,2......ai,K1 δi,Kεψˆi,K − ετˆi,K .It can be shown thatwˆih − wˆil = −ε∑k li,k(δi,kε− γii,k + 1)det(I−ΦA)∑khi,k(−δi,kψˆi,k + τˆi,k)+ε∑k hi,k(δi,kε− γii,k + 1)det(I−ΦA)∑kli,k(−δi,kψˆi,k + τˆi,k). (A.4)The following proof shows that det(I−ΦA) > 0. To simplify the notations, the subscript iis dropped.Proof: Note that det(I−ΦA) = ∑Kk=1 hk(ak1+1)∑Kk=1 lk(ak2+1)−∑Kk=1 hkak2∑Kk=1 lkak1.I use proof by induction to show thatZ(m) =m∑k=1hk(ak1 + 1)m∑k=1lk(ak2 + 1)−m∑k=1hkak2m∑k=1lkak1 > 0, ∀ m ≥ 1.The following properties are employed: (1) as α1 > α2 > ... > αK ,h1l1> h2l2 > ... >hKlK; (2)ak1 + ak2 + 1 = δkε− γkθh + 1 > 0 ∀ k.First, it is straightforward to show that Z(1) = h1l1(a11 + a12 + 1) > 0. Second, supposeZ(m− 1) > 0. It needs to prove that Z(m) > 0. Note thatZ(m) = [hm(am1 + 1) +m−1∑k=1hk(ak1 + 1)][lm(am2 + 1) +m−1∑k=1lk(ak2 + 1)]− [hmam2 +m−1∑k=1hkak2][lmam1 +m−1∑k=1lkak1]> hmlm(am1 + am2 + 1) +m−1∑k=1lkhm(am1ak2 − am2ak1 + am1 + ak2 + 1)+m−1∑k=1lmhk(am2ak1 − am1ak2 + am2 + ak1 + 1)>m−1∑k=1lkhm(am1 + am2 + ak1 + ak2 + 2) > 0114where the first inequality comes from the assumption that Z(m − 1) > 0. The secondinequality uses the facts that lkhm < lmhk and am2ak1−am1ak2+am2+ak1+1 > 0 ∀ k < m.82Therefore, det(I−ΦA) > 0.Then, the change in skill premium can be related to the exogenous changes in iceberg costsτˆij,k in the formwˆi,h− wˆi,l = −bi,1∑khi,k∑j 6=iγij,k(1−λij,k)τˆij,k + bi,2∑kli,k∑j 6=iγij,k(1−λij,k)τˆij,k +ν({ψˆi,k})(A.5)where bi,1 =ε∑k li,k(δi,kε−γii,k+1)det(I−ΦA) > 0, bi,2 =ε∑k hi,k(δi,kε−γii,k+1)det(I−ΦA) > 0 and ν({ψˆi,k}) is theresidual term capturing the effect of productivity shocks {ψˆi,k}k=1,...,K .Log-linearizing equation (1.1) obtainspˆii,h =κ1− ν (1− pii,h)(wˆi,h − wˆi,l) (A.6)Substituting equation (A.5) into (A.6) obtainspˆii,h = −ci,1∑khi,k∑j 6=iγij,k(1−λij,k)τˆij,k+ci,2∑kli,k∑j 6=iγij,k(1−λij,k)τˆij,k+ν˜({ψˆi,k}) (A.7)where ci,1 =κ1−ν (1− pii,h)bi,1 > 0 and ci,2 = κ1−ν (1− pii,h)bi,2 > 0.A.2.2 Export Demand Shocks: From National to LocalThis section derives a reduced form relationship mapping sectoral demand shocks from theROW to changes in school enrollment of local economies in China. In particular, the icebergcost of exporting good k from region i in China to region j in the ROW is assumed to takethe form τij,k = τj,kτ˜ij,k, where τj,k captures the common costs incurred by all the exportersin China, and τ˜ij,k captures the idiosyncratic costs applying to prefecture i. Also, I assumeτˆij,k = 0 if j ∈ China, that is the iceberg costs between China’s prefectures are constant.Then Equation (A.7) can be rewritten aspˆii,h = −ci,1∑khi,k∑j∈ROWγij,k(1−λij,k)τˆj,k+ci,2∑kli,k∑j∈ROWγij,k(1−λij,k)τˆj,k+ν˜(ψˆi,k, ˆ˜τij,k)(A.8)I assume λij,k ≈ 0, i.e., each prefecture in China has a small market share in region j of theROW. Note that by definition, hi,k = Hi,k/Hi, li,k = Li,k/Li and γij,k = Xij,k/Yi,k, thenEquation (A.8) can be rewritten aspˆii,h ≈ −ci,1∑kHi,kHiYi,k∑j∈ROWρij,kXCHj,kτˆj,k+ci,2∑kLi,kLiYi,k∑j∈ROWρij,kXCHj,kτˆj,k+ν˜(ψˆi,k, ˆ˜τij,k)where XCHj,k is China’s total exports of good k to region j in the ROW, and ρij,k =Xij,k/XCHj,k denotes prefecture i’s share in China’s exports of good k to region j in the82Note that am2ak1−am1ak2+am2+ak1+1 = σmσkε2(αk−αm)+σmεγk(αm−θ)+σkεγm(θ−αk)+σmε(1−αm) + σkεαk − γmθ− γk(1− θ) + 1 > σmσkε2(αk −αm) + σmεγk(1− θ) + σkεγmθ− γmθ− γk(1− θ) + 1 > 0.115ROW. Due to lack of data on regional exports and outputs by sectors, ρij,k is approximatedby prefecture i’s share of China’s employment in sector k, Ei,k/Ek,83 and Yi,k is approximatedby prefecture i’s employment in sector k, Ei,k. Thenpˆii,h ≈ c˜i,1∑kHi,kEi,kEi,kEk∆XkHi− c˜i,2∑kLi,kEi,kEi,kEk∆XkLi+ ν˜(ψˆi,k, ˆ˜τij,k) (A.9)where c˜i,1, c˜i,2 > 0 and ∆Xk =∑j∈ROW ∆XCHj,k ∝∑j∈ROW XCHj,kτˆj,k denotes the changein national exports induced by {τˆj,k}. Based on Equation (A.9), high-skill and low-skillexport shocks are defined as∆ExportHSit =∑kHik0Eik0Eik0Ek0∆XktHi0(A.10)∆ExportLSit =∑kLik0Eik0Eik0Ek0∆XktLi0(A.11)From Equation (A.10), the national export expansion of sector k, ∆Xk, is apportioned toprefecture i according to its share of national industry employment Eik0/Ek0 in the baseperiod. The regional export expansion is attributed to high-skilled workers according tothe sector’s skill intensity, Hik0Eik0 , and normalized by the amount of high skilled labor Hi0.Therefore, ∆ExportHSit can be interpreted as high skill demand shock induced by exportdemand shock. Similarly, ∆ExportLSit can be interpreted as low skill demand shock inducedby export demand shock.A.3 Data AppendixA.3.1 Administration Division and Industrial ClassificationsConsistent PrefecturesEach prefecture is assigned a four-digit code in the censuses. The codes can change over years,usually because urbanization changes rural prefectures (“Diqu”) to urban prefectures (“Shi”),which does not necessarily mean re-demarcation. The changing boundaries of prefecturesthreatens the consistency of the defined local economies over time. To address the problem,I map counties in 1990, 2000 and 2010 to prefectures where they belong to in 2005. Bythis construction, I have consistent 340 prefectures over years. The municipalities of Beijing,Chongqing, Shanghai and Tianjin are treated as prefectures in this paper.Industrial ClassificationsData in this paper comes from multiple sources that adopt different industrial classifications.I map the data on employment, output, trade flows, tariffs, and so on to consistent 3-digitCSIC (1984) codes as follows: (1) ISIC data is converted to CSIC, using the concordancebuilt by Dean and Lovely (2010), which cross-matches the 4-digit CSIC (2002) codes andISIC Rev.3 codes. (2) Data at 4-digit CSIC (2002) is converted to 3-digit CSIC (1984), usingthe concordance built by the author.83A similar simplification is made in Autor, Dorn and Hanson (2013).116A.3.2 Trade DataExport Tariff and Import at 3-digit CSICData on China’s export tariffs imposed by destination countries for 4-digit ISIC Rev.3 indus-tries are collected from the TRAINS Database. The tariff faced by the Chinese exporters ina 4-digit ISIC industry k during year t is computed according toTariffXkt =∑cExportChina,c,k,t−1ExportChina,k,t−1Tariffcktwhere TariffXckt denotes the tariff imposed by country c on goods of industry k during periodt. The tariffs are weighted by the country’s share in China’s total exports of good k in thelag period, i.e.ExportChina,c,k,t−1ExportChina,k,t−1 , and then aggregated to the industry level. These weightsare constructed using the trade flow data from three years previously. The export tariffs for3-digit CSIC industries are calculated as the weighted average of the corresponding 4-digitISIC tariffs.Data on import tariffs imposed by China on 4-digit ISIC Rev.3 industries are collectedfrom the TRAINS database. Import tariffs on 3-digit CSIC industries are computed as theweighted average of the associated 4-digit ISIC tariffs.Imports of Intermediate and Capital GoodsFrom the TRAINS database, I extract the data on imports and import tariffs of 6-digitHarmonized System (HS) products classified as Intermediate and capital goods by UNC-TAD Stages of Processing (SoP). I conduct the following steps to construct the measures∆Imports,CIkt employed in the main text: (1) Using the concordance provided by UN WITS,the data are mapped to 4-digit ISIC Rev.3 industries, and then aggregated to the 3-digitCSIC level; (2) the fitted value of imports of intermediate and capital goods, ̂ImportCIkt , isobtained from the regressionln ImportCIkt = β ln(TariffM,CIit ) + γk + φt + µkt ;(3) The intermediate and capital goods import shocks are constructed as∆ImportLS,CIit =∑kLik0Eik0Eik0Ek0∆ ̂ImportCIktLi0and ∆ImportHS,CIit =∑kHik0Eik0Eik0Ek0∆ ̂ImportCIktHi0.A.3.3 Employment and Output Data at 2-digit CSIC codesThis section describes the sources of data used in section 1.7. The 2-digit CSIC manufac-turing industry employment data and the local total employment data for each county arecollected from the population censuses. The 1990 employment data, which covers China’sentire population, is from the University of Michigan’s China Data Center. Data for year2000 and for 2010 is from various books of Population Census Data Assembly published bythe provincial statistics bureaux,84 which cover 10% of the population. These county-leveldata are then aggregated the to prefecture level using the concordance described in AppendixA.3.1.84The data for province Shangdong is not available for 2010.117There are 27 manufacturing industries, which are consistently defined across census years.For the purpose of analysis, I classify these industries into three groups according to the skill-intensity. To be specific, the industries are ranked by the share of college educated workersin their employment in 1990. Industries belonging to the bottom 33% , middle 33%, and top33% are considered as low-, medium- and high-skill groups, respectively. Table A.1 lists theindustries by skill group.The data on industry output is from the Chinese Industrial Annual Survey 2000 and2009. This data contains detailed micro-data for all state-owned firms and non-state firmswith revenues above 5 million RMB (approximately US$800,000). These firms account foraround 90% of the value of manufacturing outputs. Brandt, Biesebroeck and Zhang (2012)provide a detailed description of the data and show that the data can be aggregated almostperfectly to reflect the data reported in the Chinese Statistical Yearbooks.118Table A.1: Skill Intensities of 2-digit Manufacturing IndustriesLow-skill Industries Medium-skill IndustriesFood Processing and Production (13-14) Beverage (15)Garments, Footwear and Related Products (18) Tobacco (16)Leather, Furs, Down and Related Products (19) Textiles (17)Furniture Manufacturing (21) Timber Processing and Related Products (20)Cultural, Educational and Sporting Goods (24) Pulp, Paper, Paper Products (22)Plastic Products (30) Printing and Publishing (23)Nonmetal Mineral Products (31) Petroleum Processing and Coking (25)Metal Products (34) Raw Chemical Materials and Chemical Products (26)Manufacturing, n.e.c. (42) Rubber Products (29)High-skill IndustriesMedical and Pharmaceutical Products (27)Chemical Fiber (28)Smelting and Pressing of Ferrous Metals (32)Smelting and Pressing of Nonferrous Metals (33)Machinery (35-36)Transport Equipment (37)Electric Equipment and Machinery (39)Electronic and Telecommunications Equipment (40)Instruments, Meters, Cultural and Office Machinery (41)Notes: 2-digit CSIC codes in the parentheses.119A.4 RobustnessA.4.1 Educational Provision and Export ShocksOne may worry that the baseline results reported in Table 1.2 is driven by the shocks on thesupply side of education rather than change in demand due to external export shocks. Thisconcern is compounded against the backdrop of large expansion in higher education in Chinasince 1999. Before formally address this concern, let’s re-emphasize that the identification ofthe effect of export expansion on school enrollment comes from regional variation in industrycomposition. Therefore, a supply side shock will only bias the estimates of ∆ExportLS and∆ExportHS if it is correlated with a region’s initial industry structure, e.g., a prefecture witha higher high-skill export exposure due to its initial industry specialization also received alarger increase in educational provision.In this section, I investigate the correlation between a prefecture’s change in educationalprovision and its export exposure. Column (1) of Table regresses change in log of educationfiscal expenditure on PHI and PHI as defined in section 1.6.1. The correlations betweenchange in educational provision and prefecture’s initial skill level are insignificant. Column(2) reports the effect of ∆Exports on education fiscal expenditure. The estimated coefficientof ∆ExportLS implies that a prefecture that experiences a larger low-skill export shocktends to see a larger increase in fiscal expenditure on education. The estimate is significantat 10% level. The estimate of ∆ExportLS is positive, although insignificant. The findingsuggests that local governments, especially of prefectures that have larger exposure to low-skill export shock, could alter education supply to offset the export-induced shock on demandside. Columns (3) and (4) replace the dependent variable by change in log of number of middleschool teachers. (Due to data limitation, the sample is restricted to the second period.) Noneof the estimates are statistically significant.In summary, we find weak evidence that shock in education provision is correlated withexport exposures. This echos the finding that the estimates in Table 1.2 are insensitiveto controls of change in education fiscal expenditure and regional initial skill level. Moreimportantly, if local expansion in education is associated with export shocks, the estimatesin Table 1.2 are still likely provide a lower bound (in magnitude) of the effect of low-skillexport shock on education demand.A.4.2 Migration Pattern and Export ShocksThis section examines the effects of regional export shocks on the migration flows usingdata from the 2000 and 2005 population censuses on the prefecture of residence five yearspreviously. The sample is restricted to individuals of prime working age (those aged 23 to55). The immigration and emigration rates of prefecture i are defined asIMRsit = IMsit/(IMsit + Popsit−1) and EMRsit = EMsit/Popsit−1where IM sit and EMsit denote, respectively, the total immigrants and emigrants of prefecturei with education level s during the past five years, and Popsit−1 is the total population ofeducation level s in prefecture i five years ago.Columns (1) of Table A.3 relates the immigration pattern of unskilled workers to export120shocks by estimating the following regression:∆IMRLSit =∑s∈{LS,HS}θs∆Exportsit + φp + εit (A.12)where ∆IMRLSit is the change in immigration rate of workers with high school education orlower between 2000 and 2005, and φp denotes the province dummy. Low-skill export shocksare found to attract low-skilled immigrants, whereas high-skill export shocks deter them. Theestimates imply that a $1000 low-skill export expansion increases ∆IMRLSit by 2.3 percentagepoints, and a $1000 low-skill export expansion decreases ∆IMRLSit by 0.2 percentage points.Column (2) re-estimates Equation (A.12) but changes the dependent variable to ∆IMRHSit ,i.e., the change in the immigration rate of workers with college education or above between2000 and 2005. The estimated coefficients θk are close to zero and statistically insignificant,suggesting that export shocks have little effect on the immigration flow of skilled workers.Columns (3) and (4) show the effects of export shocks on the emigration rates of workerswith different educational attainment. Low-skill export shocks reduce the out-migration ofworkers with high school education or lower, and induce more emigration of workers withcollege education or above. However, the effects of high-skill export shocks are found to beinsignificant regardless of the type of worker.Column (5) replaces the dependent variable with ∆IMLSitEMLSit(i.e., the change in the ratioof inflows to outflows of low-skilled workers) to study the effects on net flows of low-skilledworkers. Consistent with the finding in Columns (1) and (3), low-skill export shock inducesmore low-skilled immigrants relative to emigrants, and high-skill export shock lowers theratio. Column (6) shows that export shocks have no significant effect on the net migrationflow of skilled workers. Conclusively, low-skill export expansion lowers the average educationof immigrants and increases the average education of emigrants. The converse is true inthe case of high-skill export expansion. Export shocks are found to alter the composition ofmigrant workers mainly through their effect on the migration pattern of low-skilled workers.A.4.3 Effects on Export Demand Shocks on School Enrollment of LocalsLarge inflows of immigrants could reduce the responsiveness of non-migrant education tolocal export shocks. To investigate this possibility, I look at data from the 2000 and 2005population censuses which record the prefecture of residence 5 years previously. Columns(1) of Table A.4 examines the effects of export shocks on immigrant share of the high schoolage population (∆ IMPop). The estimated coefficients of ∆Exports are small and insignificantlydifferent from zero. Columns (2) and (3) present the effects of export shocks on the highschool enrollment rate for the baseline sample (immigrants and locals) and non-migrantsample, respectively. If the negative (positive) educational impact of low-skill (high-skill)export shock is purely explained by the inflows of low skilled (high skilled) young workers,the estimate of βLS (βHS) is expected to diminish significantly when the sample is restrictedto non-migrants. However, the estimated coefficients of the two samples are found to bestatistically similar. Given the evidence from Columns (1)–(3), export shocks are unlikely tochange the high school enrollment rate by altering the composition of school age populationalone.Columns (4)–(6) report analyses analogous to Columns (1)–(3), but focus on the popu-lation aged 19 to 22. Again, export shocks are found to have little effect on the inflow of121college age immigrants. The effect of low-skill (high-skill) export shock is still estimated to besignificantly negative (positive) when the sample is restricted to non-migrants, albeit smallerin magnitude than that of the baseline sample.A.4.4 Heterogeneous Reponses to Export Shocks of DifferentDemographic GroupsIn Table A.5, I investigate the heterogeneous responses of different demographic groups toexport shocks, using the specification of Column (4) in Table 1.2. Columns (1) and (2) showthe effects of export shocks on school enrollment by gender. The estimates are statisticallysimilar for the two genders and remain statistically significant at conventional levels for alloutcomes, with the exception of boys’ high school enrollment in the case high-skill exportshocks, where the estimate is marginally insignificant, with a p-value of 0.12.Columns (3) and (4) present the regression results for the samples of young people withurban and rural Hukou, respectively. Regardless of the outcome variables and samples,the estimates retain the expected signs and significance. In addition, the estimated effectsof export shocks on high school enrollment are quantitatively similar across samples. Incontrast, export shocks are found to have larger effects on college enrollment in magnitude inthe urban sample than in the rural sample. In particular, the estimate of the high-skill exportshock is statistically larger for young people with urban Hukou, which may be explained byurban residences proximity to colleges.A.4.5 Dropping One Province/Two Provinces at a TimeTo test whether the findings are driven by a particular geographic region, I estimate thespecification of Column (4) in Table 1.2, and drop one province or two provinces at a time.Panel A of Table A.6 shows the results for dropping one province at a time. For the sampleof young people aged 16 to 18, the estimates of βLS range from -0.050 to -0.026, and theestimates of βHS range from 0.005 to 0.007. For the sample of young people aged 19 to 22,the estimates of βLS range from -0.034 to -0.030, and the estimates of βHS range from 0.006to 0.008. All these estimates are significant at the conventional level. Panel B presents theresults for dropping two provinces at a time. The estimates are not sensitive to the exclusionof any combination of two provinces.A.4.6 Children under Compulsory SchoolingIn principle, if the law of compulsory schooling is perfectly enforced, then primary and juniorsecondary education are not relevant margins for education choices in this study. However, asis discussed in Section 1.5, although it is improving over time, the enforcement of compulsoryschooling in China is imperfect, especially for children aged 13 to 15. Table A.7 presents theregression results when the sample is restricted to the children aged between 7 to 15. Columns(1) to (3) show that export shocks have no effect on the school enrollment of primary schoolage children. Columns (4) to (6) find small but significant educational impacts of high-skillexport shock for the age group 13 to 15. The estimates imply that a $1000 high-skill exportshock raises middle school enrollment by 0.2 to 0.4 percentage point. Column (5) also detectsa significantly negative effect of low-skill export shock on middle school enrollment, with a122$1000 exposure lowering it by 0.6 percentage point. It is found in the unreported resultsthat the significant effects of export shocks on school enrollment for the 13–15 age group aremainly driven by fifteen-year-olds. As is shown in Figure 1.2, the enforcement of compulsoryschooling is weakest for this marginal age group.A.4.7 Other Market OutcomesTable A.8 re-estimates the specification in Column (4) of Table 1.2, with the dependentvariable replaced by different variables related to market outcomes. Columns (1) and (4)report the effect of export expansion on market employment. The estimates resemble thosein Table 1.2, but with flipped signs. As is shown in Columns (2) and (5), exports shocks areestimated to have no effect on home production. This result is in contrast to the findings ofEdmonds and Pavcnik (2005) and Edmonds, Pavcnik and Topalova (2010) that trade affectschildren’s engagement in domestic work through the income effect. In addition, Columns(3) and (6) show that low-skill export shocks decrease the unemployment rate of youngpeople aged 16 to 22, but the effects of high-skill shocks are statistically insignificant. This isconsistent with the priori that low-skill export expansion provides more jobs for young peoplewho drop out of school, whereas the job opportunities generated by high-skill shocks are notrelevant to them.A.4.8 Disaggregated Export Demand ShocksIn this section, I construct the export shock Exports at three levels of skill intensity, wheres ∈ {LS,MS,HS}, following a strategy similar to that described in Section 1.5.4. Specifically,∆ExportLSit =∑kLik0Li0∆ ̂ExportktEk0, ∆ExportMSit =∑kMik0Mi0∆ ̂ExportktEk0,and ∆ExportHSit =∑kHik0Hi0∆ ̂ExportktEk0.where the low-skilled workers (L) are those with middle school education or lower, themedium-skilled workers (M) are those with some high school education, and the high skilledworkers (H) are those with some college education.Table A.9 presents the regression results using modified export shocks. Columns (1) and(3) show that low-skill shock has an adverse effect on high school and college enrollment,whereas the effect of high-skill shock is estimated to be significantly positive. It is worthnoting that ∆ExportMSit has two offsetting effects on high school and college enrollments.One the one hand, positive medium-skill shocks increase the demand for high school educatedworkers. On the other hand, positive medium-skill shocks discourage young people frompursuing college education. As secondary education is a prerequisite for college, the firstchannel encourages high school and college enrollments, but the second channel tends todepress both. As is shown in Columns (1) and (3), the effects of medium-skill shock arefound to be insignificant. Columns (2) and (4) augment the model with the controls PLIand PHI and the estimates change little. Lastly, Columns (3) and (6) control the prefecturedummies. The effects of ∆ExportLSit are less precisely estimated. Specifically, the effect of123low-skill shock on high school enrollment becomes marginal insignificant, with a p-value of0.14. In addition, a significantly negative effect of medium-skill shock is detected for the16-18 age group.A.4.9 Change in Skill Supply and Industry Specialization: DifferentMeasuresTable A.10 repeats the analysis in Table 1.5, but replaces ∆HighSch.Enrollit−1 with ∆HighSch.Shareit−1(i.e., change in the share of population with some high school education over the period 1990to 2000), and replaces ∆College.Enrollit−1 with ∆College.Shareit−1 (i.e., change in theshare of the population with some college education over the period 1990 to 2000). Un-like school enrollment, the share of workers with different educational attainment is a stockvariable measuring skill supply.The findings in Table A.10 are consistent with those in Table 1.5, regardless of the mea-sures of industry share. The 2SLS estimates in Column (2) suggest that a 10 percentagepoint increase in CollegeShareit−1 raises a high-skill industry’s share of total employment inthe manufacturing sector by 3.37 percentage points, and reduces a low-skill industry’s shareof total employment in the manufacturing sector by 3.79 percentage points.124Table A.2: Educational Provision and Export Export Exposure∆ ln(FiscalExpEdu) ∆ ln(NbrMidSchTeacher)(1) (2) (4) (5)PLI 0.216 0.029(0.136) (0.156)PHI 2.603 -3.617(2.217) (3.559)∆ExportLS 0.042* 0.021(0.021) (0.046)∆ExportHS -0.015 0.005(0.009) (0.013)Province(×Year) Y Y Y YN 680 680 340 340R2 0.389 0.388 0.226 0.225Notes: All regressions control for changes in log population of age groups 7-15, 16-18 and 19-22.Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, * p<0.1 *** p<0.01,** p<0.05, * p<0.1Table A.3: Migration and Export Shocks∆IMRLS ∆IMRHS ∆EMRLS ∆EMRHS ∆ IMLSEMLS∆ IMHSEMHS(1) (2) (3) (4) (5) (6)∆ExportLS 0.023*** -0.001 -0.005** 0.008** 24.153** 0.357(0.005) (0.002) (0.002) (0.003) (10.008) (0.361)∆ExportHS -0.002* -0.001 0.002 0.000 -2.433* -0.029(0.001) (0.001) (0.001) (0.002) (1.321) (0.039)Province Y Y Y Y Y YN 340 340 340 340 340 340R2 0.471 0.573 0.626 0.485 0.535 0.614Notes: Robust standard errors in the parenthesis. *** p<0.01, ** p<0.05, * p<0.1125Table A.4: Changes in Share of Immigrants, Change in School Enrollmentand Export Shocks: Different Samples by Migration StatusAge 16-18 Age 19-22∆ IMPop ∆Enroll ∆Enroll ∆IMPop ∆Enroll ∆Enroll ∆EnrollAll All Non-migrants All All Non-migrants Out-migrants(1) (2) (3) (4) (5) (6) (7)∆ExportLS 0.017 -0.029** -0.019* 0.007 -0.027*** -0.013* -0.013**(0.011) (0.012) (0.011) (0.009) (0.006) (0.008) (0.06)∆ExportHS -0.001 0.004** 0.006*** 0.002 0.006*** 0.004** 0.003**(0.002) (0.002) (0.001) (0.003) (0.002) (0.002) (0.001)Province Y Y Y Y Y Y YControls Y Y Y Y Y Y YInitial Conditions Y Y Y Y Y Y YPHI and PLI Y Y Y Y Y Y YN 340 340 340 340 340 340 340R2 0.594 0.694 0.633 0.718 0.662 0.703 0.687Notes: All regressions are weighted by the start-of-the-period prefecture’s age group population. Controls include change in averageage, change in sex ratio, change in share of Han ethnic group, change in share of population with urban Hukou, change in log fiscalexpenditure on education, and change in log GDP per capita. Initial conditions include the start of period school enrollment rate andlog GDP per capita. Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, * p<0.1126Table A.5: Changes in School Enrollment and Export Shocks: by SamplesBoy Girl Urban Rural(1) (2) (3) (4)Panel A: Age 16-18∆ExportLS -0.029*** -0.041*** -0.027* -0.036***(0.009) (0.012) (0.014) (0.011)∆ExportHS 0.004 0.008*** 0.008*** 0.007***(0.002) (0.002) (0.002) (0.002)Province×Year Y Y Y YControls Y Y Y YInitial Conditions Y Y Y YPHI and PLI Y Y Y YN 673 673 665 672R2 0.640 0.716 0.729 0.707Panel B: Age 19-22∆ExportLS -0.029*** -0.030*** -0.028** -0.015***(0.006) (0.005) (0.011) (0.004)∆ExportHS 0.009** 0.005* 0.018*** 0.005***(0.003) (0.003) (0.005) (0.001)Province×Year Y Y Y YControls Y Y Y YInitial Conditions Y Y Y YPHI and PLI Y Y Y YN 673 673 673 673R2 0.734 0.715 0.770 0.769Notes: All regressions are weighted by the start-of-the-period prefecture’s age group population. Controlsinclude change in average age, change in sex ratio, change in share of Han ethnic group, change in share ofpopulation with urban Hukou, change in log fiscal expenditure on education, and change in log GDP per capita.Initial conditions include the start of period school enrollment rate and log GDP per capita. Standard errorsare clustered at the province level. *** p<0.01, ** p<0.05, * p<0.1127Table A.6: Dropping One Province/Two Provinces at a TimeAge 16-18 Age 19-22min(βˆ) max(βˆ) min(βˆ) max(βˆ)(1) (2) (3) (4)Panel A: Dropping one province at a time∆ExportLS -0.050** -0.026*** -0.034** -0.030***(0.020) ( 0.008) (0.013) (0.005)∆ExportHS 0.005** 0.007*** 0.006*** 0.008***(0.002) (0.002) (0.002) (0.002)Panel B: Dropping two provinces at a time∆ExportLS -0.060*** -0.024*** -0.040*** -0.021*(0.019) (0.007) (0.013) (0.012)∆ExportHS 0.004** 0.009*** 0.006** 0.008***(0.002) (0.002) (0.002) (0.002)Notes: All regressions are weighted by the start-of-the-period prefecture’s share of the cohort population.All regressions include the controls in the specification (4) of Table 1.2. Standard errors are clustered at theprovince level. *** p<0.01, ** p<0.05, * p<0.1Table A.7: Change of School Enrollment and Export Shocks:Ages under Compulsory SchoolingAge 7-12 Age 13-15(1) (2) (3) (4) (5) (6)∆ExportLS 0.001 0.001 0.001 -0.005 -0.006* -0.012(0.002) (0.002) (0.004) (0.003) (0.003) (0.008)∆ExportHS -0.000 -0.000 -0.000 0.003*** 0.002*** 0.004*(0.000) (0.000) (0.001) (0.001) (0.001) (0.002)Province Y Y Y Y Y YControls Y Y Y Y Y YInitial Conditions Y Y Y Y Y YPHI and PLI Y YPrefecture Y YN 673 673 673 673 673 673R2 0.956 0.956 0.995 0.858 0.859 0.980Notes: All regressions are weighted by the start-of-the-period prefecture’s age group population.Controls include change in average age, change in sex ratio, change in share of Han ethnic group,change in share of population with urban Hukou, change in log fiscal expenditure on education,and change in log GDP per capita. Initial conditions include the start of period school enrollmentrate and log GDP per capita. Standard errors are clustered at the province level. *** p<0.01, **p<0.05, * p<0.1128Table A.8: Changes of Market Employment, Home Production,Unemployment Rate and Export ShocksAge 16-18 Age 19-22∆Market ∆Home ∆Unemp. ∆Market ∆Home ∆Unemp.Empl. Prod. Rate Empl. Prod. Rate(1) (2) (3) (4) (5) (6)∆ExportLS 0.036*** 0.001 -0.011* 0.039*** -0.001 -0.012***(0.009) (0.002) (0.006) (0.004) (0.001) (0.003)∆ExportHS -0.006** 0.000 0.002 -0.009*** 0.001 0.002(0.002) (0.000) (0.003) (0.002) (0.000) (0.001)Province Y Y Y Y Y YControls Y Y Y Y Y YInitial Conditions Y Y Y Y Y YPHI and PLI Y Y Y Y Y YN 673 673 673 673 673 673R2 0.713 0.279 0.545 0.781 0.300 0.635Notes: Notes: All regressions are weighted by the start-of-the-period prefecture’s age group population. Con-trols include change in average age, change in sex ratio, change in share of Han ethnic group, change in shareof population with urban Hukou, change in log fiscal expenditure on education, and change in log GDP percapita. Initial conditions include the start of period school enrollment rate and log GDP per capita. Standarderrors are clustered at the province level. *** p<0.01, ** p<0.05, * p<0.1129Table A.9: Changes in School Enrollment, Export Shocks and Import Shocks:Disaggregated Education GroupsAge 16-18 Age 19-22(1) (2) (3) (4) (5) (6)∆ExportLS -0.019** -0.025*** -0.036 -0.017*** -0.027*** -0.023*(0.007) (0.009) (0.023) (0.004) (0.005) (0.012)∆ExportMS -0.005 -0.007 -0.022* -0.004 -0.004 -0.019(0.007) (0.007) (0.012) (0.006) (0.006) (0.012)∆ExportHS 0.007*** 0.006** 0.011** 0.011*** 0.007*** 0.012**(0.002) (0.002) (0.004) (0.002) (0.002) (0.004)Province Y Y Y Y Y YControls Y Y Y Y Y YInitial Conditions Y Y Y Y Y YPHI and PLI Y YPrefecture Y YN 673 673 673 673 673 673R2 0.682 0.693 0.925 0.696 0.762 0.920Notes: All regressions are weighted by the start-of-the-period prefecture’s age group population. Controlsinclude change in average age, change in sex ratio, change in share of Han ethnic group, change in shareof population with urban Hukou, change in log fiscal expenditure on education, and change in log GDPper capita. Initial conditions include the start of period school enrollment rate and log GDP per capita.Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, * p<0.1130Table A.10: Changes in Skill Supply and Change in Industry Specialization:Different Measures∆ShareM ∆ShareEmployment Output EmploymentOLS 2SLS OLS 2SLS OLS 2SLS(1) (2) (3) (4) (5) (6)LS ×∆HighSch.Sharet−1 -0.005 0.253** -0.005 0.322* -0.002 0.134**(0.021) (0.099) (0.033) (0.187) (0.005) (0.061)MS ×∆HighSch.Sharet−1 0.013 -0.040 0.005 -0.196 0.002 0.024(0.022) (0.061) (0.031) (0.214) (0.003) (0.016)HS ×∆HighSch.Sharet−1 -0.009 -0.214** 0.000 -0.111 0.000 -0.060*(0.016) (0.089) (0.032) (0.212) (0.004) (0.033)LS ×∆College.Sharet−1 -0.058 -0.379*** -0.055 -0.388** -0.017 -0.245***(0.037) (0.140) (0.049) (0.156) (0.013) (0.087)MS ×∆College.Sharet−1 -0.006 0.042 0.010 0.075 -0.010 -0.074***(0.033) (0.079) (0.045) (0.191) (0.006) (0.028)HS ×∆College.Sharet−1 0.064** 0.337*** 0.055 0.313* 0.003 0.087*(0.028) (0.125) (0.050) (0.175) (0.007) (0.051)∆ShareEmpMt−1 Y Y Y Y∆ShareEmpt−1 Y YProvince×Ind Y Y Y Y Y YN 8,721 8,721 9,180 9,180 8,721 8,721Notes: All regressions control for the start-of-the-period industry employment/output share. Standard errors areclustered at the prefecture level. *** p<0.01, ** p<0.05, * p<0.1131A.5 CalibrationFigure A.1: Regions in ChinaNortheastNorth MunicipalitiesNorth CoastCentral CoastSouth CoastCentralSouthwestNorthwestNotes: Regions include Northeast (provinces Heilongjiang, Jilin and Liaoning), North Municipalities(provinces Beijing and Tianjin), North Coast (provinces Hebei and Shangdong), Central Coast (provincesShanghai, Jiangsu and Zhejiang), South Coast (provinces Guangdong, Fujian and Hainan), Central (provincesHenan, Shanxi, Anhui, Jiangxi, Hubei and Hunan), Southwest (provinces Guangxi, Chongqing, Sichuan,Guizhou, Yunnan and Tibet), and Northwest (provinces Inner Mongolia, Shannxi, Gansu, Qinghai, Ningxiaand Xinjiang).132Table A.11: Calibration and Data SourcesVariables and Parameters DataIncome(Yi), bilateral tradeflows(λij,k) China Regional IO Table, 2007sectoral output(Yi,k), & consumption share(βk) & WIOD, 2007Employment share of skilled China 2005 Mini Censusand unskilled labor(hi,k and li,k) & WIOD SEA, 2007Income share of skilled workers(αk) China 2005 Mini CensusSchool enrollment rate(pii,s) China 2005 Mini CensusBarro and Lee, 2005Trade elasticity (ε = 4.14) Simonovska and Waugh(2014)Skill distribution parameter (κ = 3.36) China 2005 Mini CensusTable A.12: Calibrated αk and βkSector αk βkAgriculture, Hunting, Forestry and Fishing 0.01 0.06Mining and Quarrying 0.16 0.05Food, Beverages and Tobacco 0.14 0.05Textiles, Apparels, Footwear and Leather Products 0.06 0.04Wood and Products of Wood and Cork 0.04 0.01Pulp, Paper, Paper, Printing and Publishing 0.14 0.02Chemicals and Chemical Products 0.24 0.11Non-Metallic Products 0.07 0.03Basic Metals and Fabricated Metal 0.13 0.10Machinery, n.e.c. 0.24 0.06Transportation Equipment 0.24 0.05Electrical and Optical Equipment 0.34 0.07Manufacturing, Nec 0.06 0.02Other Non-manufacturing 0.35 0.34133Appendix BAppendix for Chapter 2B.1 Data AppendixB.1.1 Administration Division: Consistent PrefecturesEach prefecture is assigned a four-digit code in the censuses. The codes can change over years,usually due to the urbanization of the rural prefectures (“Diqu”) to urban prefectures (“Shi”),which does not necessarily mean re-demarcation. The changing boundary of prefectures is athreat to the consistency of our defined local economies over time. To cope with the problem,we construct a concordance mapping the counties in 1990, 2000 and 2010 to the prefectureswhere they belong in 2005. By construction, we have consistent 340 prefectures over years.The municipalities Beijing, Chongqing, Shanghai and Tianjin are treated as prefectures inthis paper.B.1.2 Industrial ClassificationsOur dataset is complied from multiple sources which adopt different industrial classifications.We map the data on employment share, emission intensities, exports, tariffs, and so on toconsistent 3-digit CSIC codes. The details are provided as follows: (1) CSIC employed by ourdata sources has three versions, CSIC1984, CSIC1994 and CISC2002. We firstly build theconcordances which map 4-digit CSIC of different versions to consistent 3-digit CSIC codes.(2) To convert the ISIC data to CSIC, we employ the concordance built by Dean and Lovely(2010) which cross-matches the four-digit CSIC2002 and ISIC Rev.3. (3) For the SIC data,we firstly concord it to ISIC Rev.3 using the concordance provided by the United NationsStatistics Division, and then map it to CSIC2002.B.1.3 Prefecture Level Data on Wind DirectionThe data on wind direction is collected from NOAA Integrated Surface Global Hourly Data.We collapse the hourly data to the daily level and calculate the average wind direction foreach weather station-day observation using the “unit-vector” average method according toNOAA.85 The wind direction is categorized into 37 groups, i.e., wd ∈ {0, pi36 , ..., 35pi36 , No Wind}.We drop the weather station with more than 40% of the daily observations within a decadeare missing. There are 544 and 394 weather stations in our samples for the decade 1990-2000and 2000-2010, respectively.We impute the data on wind direction to each prefecture using the data from its nearest85More details can be found in http://www.ndbc.noaa.gov/wndav.shtml.134weather station.86 For each prefecture, we calculate the share of days that the wind directionis wd for each decade, i.e., 1990-2000 and 2000-2010. It is denoted by sit,wd.B.1.4 Employment Weighted and Wind Direction Weighted NeighboringExport Pollution ShocksWe identify the set of prefectures sharing a border with prefecture i, and denote it asNeighbori. The employment weighted neighboring pollution shock is defined as followsPollutionExportShockp,Ni,t =∑r∈NeighboriψirtPollutionExportShockprt .where ψirt =Lrt−1∑r′∈Neighbori Lr′t−1denotes the employment weight of prefecture r among all theneighboring prefectures of i.The wind direction weighted neighboring export pollution shock is constructed as followsWindPollutionExportShockp,Ni,t =∑wd∈{0, pi36,..., 35pi36,NW}∑r∈Neighboripiirt,wdPollutionExportShockprt .The neighboring prefecture r’s export pollution shock is weighted bypiirt,wd =srt,wdwir,wd∑wd∑r srt,wdwir,wd,where srt,wd denotes the share of days in which the wind direction is wd in neighboring pre-fecture r and decade t. wir,wd captures the differential weighting of different wind directions.It is determined by the relative position between i and r and the wind direction in r, and itis constructed as followswir,wd ={12 [1 + cos(θir,wd)] if wd ∈ {0, pi36 , ..., 35pi36 }0 if wd = No Wind,where θir,wd denotes the absolute value of the angle between neighboring prefecture r’s an-gular position and its wind direction wd. The example in Figure B.4 illustrates how θir,wd iscalculated. The green triangle represents a prefecture, who has three neighboring prefectures(r =1,2 and 3) represented by blue circles. The neighboring prefectures are located in theNorthwest, South and East. (Their angular position to i are 3pi/4, 3pi/2 and 0, respectively.)Suppose the wind directions in the three neighboring prefectures are pi/4, 0 and pi/3, respec-tively. Then the angles between their angular position and wind direction are θ1,pi/4 = pi/2,θ2,0 = pi/2 and θ3,pi/3 = pi/3, respectively. Note that wir,wd = 1 if prefecture i is in the down-wind position of r, i.e., θir,wd = 0, and wir,wd = 0 if prefecture i is in the upwind position ofr, i.e., θir,wd = pi. This weighting scheme is intuitive. Prefecture i receives more cross-borderpollution from prefecture r if i is located downwind of r more often, i.e., srt,wdwir,wd is larger.86The nearest weather station is identified as the one with the shortest distance to the centroid of a prefec-ture. As is discussed in the Section 2.7.1, we test the robustness of the result using the data from the firstand second most nearest weather stations.135B.1.5 Input-Output TablesTo construct the alternative export shocks as described in 2.7.8, we use the 1997 and 2007input-output (IO) tables published by National Bureau of Statistics China. The 1997 IOtable contains information of input-output relationships among 124 industries, 70 of whichbelong to manufacturing sector. The 2007 IO table contains information of input-outputrelationships among 135 industries, 80 of which belong to manufacturing sector. We aggregateand match our trade, employment and pollution intensity data to the industries in the IOtables.B.1.6 Data Quality of Air Pollution: Comparison of Official Data and USEmbassy DataSince 2009, the US Embassy started to monitor and report the hourly concentration ofPM2.5, the particulate matters up to 2.5 micrometers in size, in five major cities in China, i.e.,Beijing, Chengdu, Guangzhou, Shanghai and Shenyang. The data are collected independentlyof Chinese government agencies, and hence provide a benchmark to check the validity of theofficial pollution data. As discussed earlier, environmental protection was unlikely to be amajor factor determining a politician’s career trajectory in the past, so if manipulation of theofficial pollution data existed, it is more likely to have occurred in the later period. Therefore,we believe the comparison is informative, although it is restricted to the later years.Daily Average of Air Quality Index (AQI)MEP publishes AQI and the main pollutant daily for major cities in China. Accordingto MEP, the AQI and the main pollutant are derived following the steps: (1) convertthe pollution readings to IAQIp for each pollution p; (2) construct the overall AQI us-ing the formula AQI = max{IAQI1, IAQI2, ..., IAQIP }; (3) the main pollutant is p ifIAQIp = max{IAQI1, IAQI2, ..., IAQIP }. The individual pollutant index IAQIp is notpublic available. However, we know that AQI = IAQIPM2.5 conditional on the main pollu-tant being PM2.5.We obtain the daily data from MEP for 2014. The hourly data from US Embassy isaggregated to daily data, and converted to IAQIPM2.5 using the conversion table providedby MEP. The summary statistics of the two data series are presented in Table B.1. We findthat, conditional on that the main pollutant being PM2.5, the correlation of the two seriesis very high, ranging from 0.94 for Chengdu to 0.97 for Shanghai. The average IAQIPM2.5 ofMEP data is generally lower than the US embassy data, with largest discrepancy observed forBeijing. The discrepancy is likely be due to the different locations where the data is collectedby the US embassy and MEP. In each city, the US embassy readings are collected from onlyone monitor located in a populous area. On the other hand, the MEP readings are collectedfrom multiple locations, including the suburban area.Annual Average of PM2.5 (PM10) ConcentrationThe cities in China only started to monitor PM2.5 in 2013. Until then, only the concentrationlevel of PM10, the particulate matters up to 10 micrometers in size, was reported. As thePM2.5 belongs to the PM10, the concentration level of the former should should always be136smaller than the latter. It is taken as strong evidence for data manipulation if we detect theconcentration level of PM10 reported by MEP is smaller than that of PM2.5 reported byUS Embassy. As is shown in Table B.2, the official data of PM2.5 track the US Embassydata closely in 2013. In addition, we fail to find any case such that the concentration levelof PM10 reported by MEP is smaller than that of PM2.5 reported by US Embassy.B.1.7 Data Quality of IMR: Comparison of Cohort Size across CensusesInfant mortality rate is measured with error if either number of births or number of deathsis misreported. One may concern that due to the One Child Policy, households may haveincentives to underreport live births or over-report infant mortality to hide unsanctionedbirths. To investigate this possibility, we compare the population size of a newborn cohortacross census years. In particular, we predict the cohort size of the Age 0 ten years later,Pop0,i,t+10, using the information of mortality rate of age groups of 1-4 and 5-10, as followsPop0,i,t+10 ≈ (Bi,t −D0,i,t)(1− D1−4,i,tPop1−4,i,t)4(1− D5−10,i,tPopt−10,i,t)5 ,where Bi,t is the number of births of prefecture i in census year t, Da,i,t and Popa,i,t denoterespectively the number of deaths and population of age group a in prefecture i and censusyear t. In principle, without inter-region migration, the predicted population size of Age 10in census year t + 10, Pop0,i,t+10, should closely track the actual population size if the dataon number of births and deaths are free of measurement error. However, if Bi,t is prevalentlyunderreported (over-reported) or D0,i,t is over-reported (underreported), one should expectthe predicted population size of Age 10 to be always smaller (larger) than the actual data.87Figure B.3 plots the log of predicted Age 10 population Pop0,i,2000+10 (derived from the2000 census), against the log of actual Age 10 population Pop10,i,2010 (obtained from the 2010census).88 The points are closely cluster along the 45 degree line, suggesting the birth anddeath statistics are not systematically misreported. The coefficient of correlation betweenthe predicted and actual log Age 10 population is as high as 0.984. In addition, we find thatthe prefectures that lies significantly below the 45 degree line are the ones receiving large netinflows of immigrants. As a result, the actual population size is larger than the predictedpopulation size.B.2 Additional ResultsB.2.1 Robustness Checks: Effect of Export Shocks on Changes in PM2.5ConcentrationIn this section, we present a variety of robustness tests on the effect of export expan-sion on changes in pollutant concentration. We focus on pollutant PM2.5 because, dueto the data limitation, some of our robustness checks are restricted to the second decade(2000-10) for which we have more observations for PM2.5 concentration. The results are87As it is more difficult to hide a ten-year-old than an infant, we consider that the actual population size ofAge 10 in census year t+ 10 is subject to less misreporting.88The findings are similar when comparing the 1990 census and 2000 census (available on request).137shown in Table B.7. Column (1) include WindPollutionExportShockTSP,N as an addi-tional control. Albeit less precisely estimated, we find neighboring pollution export shockhas a positive effect on PM2.5 concentration. More importantly, the estimated coefficientof PollutionExportShockTSP remains similar to our baseline findings in Table 2.4, whichagain suggest that local pollution export shock affects PM2.5 concentration independentof cross-border spillovers. Columns (2)-(8) repeat the regressions in Table 2.12. We findthat our results are robust to the inclusion of fossil energy production, although the expan-sion in energy production increases local PM2.5 concentration. The estimated coefficientof PollutionExportShockTSP is also robust to the considerations of pollution import shock,export-induced skill demand shock, changes in the ownership structure and TCZ policy.Moreover, we find consistent results when we take into account of the input-output linkageor use output shocks as alternative measures.138Figure B.1: Timeline of Steel SafeguardFigure 1: 2001-03 Steel Safe Guards Timeline2001 2002 2003 20044 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12 1 2USA I F LEUN P F LCHN P F LNote: I-Investigation; P-Provisional Measures; F-Final Measures; L-First Scheduled Liberalization.1Figure B.2: Steel Export and Production Share over 2000-0456789Share of Output11.522.5Share of Export2000 2001 2002 2003 2004YearShare of Export Share of Output139Figure B.3: Predicted and Actual Population Size of Age 10 in 2010Corr=.9847891011121314Predicted ln(Pop Age 10)7 8 9 10 11 12 13 14ln(Pop Age 10)45 degree−linePopulation of Age 10 in 2010Figure B.4: Wind Direction Example2	1	3	θ1=π/2	Θ2=π/2	Θ3=π/3		140Table B.1: Correlation of pollution data of US embassy and MEP (AQI)Conditional on Main Pollutant is PM2.5 UnconditionalCity US Embassy MEP Corr. PM2.5 % days Corr.Beijing 214.995 168.109 0.956 47.945 0.932(9.084) (5.955)Shanghai 105.092 104.453 0.972 37.534 0.916(3.774) (3.706)Guangzhou 103.248 94.765 0.947 32.603 0.641(3.258) (2.931)Chengdu 154.702 140.000 0.938 47.945 0.945(5.015) (5.047)Shenyang 174.187 150.550 0.950 38.356 0.941(7.154) (5.057)Notes: Standard errors in the parentheses. The hourly data for PM2.5 from US Embassy isaggregated to daily average, dropping the daily observations with less than 12 valid readings.Table B.2: Concentration of PM2.5 (PM10) from different sources (mg/m3, 2013)Beijing Shanghai Guangzhou ChengduUS MEP US MEP US MEP US MEPPM2.5 PM2.5 PM2.5 PM2.5 PM2.5 PM2.5 PM2.5 PM2.5(PM10) (PM10) (PM10) (PM10)2013 0.102 0.09 0.060 0.062 0.055 0.053 0.098 0.097(0.108) (0.082) (0.072) (0.15)2012 0.090 (0.109) 0.051 (0.071) 0.058 (0.069)2011 0.099 (0.113)2010 0.104 (0.121)2009 0.102 (0.121)Notes: The hourly data for PM2.5 from US Embassy is aggregated to daily average, dropping the daily ob-servations with less than 12 valid readings. The annual average is calculated based on the daily average data.Chinese Official data comes from Chinese Environmental Yearbooks and the Bulletins published by the prefec-ture Bureau of Environmental Protection. The concentration levels of PM2.5 are only available since 2013. Thedata for PM10 in the parentheses is included for comparison.141Table B.3: Changes in Infant Mortality Rate and Shocks by Causes of Death:TSP, 2SLSCardio- Infant DigestiveRespiratory SpecificDep. Var: ∆IMRC (1) (2) (3) (4) (5) (6)PollExShockTSP 0.313*** 0.496** -0.357 0.356 -0.005 0.009(0.107) (0.199) (0.385) (0.631) (0.019) (0.064)ExShock -2.105 -5.099 0.344(1.737) (4.652) (0.358)Region Y Y Y Y Y Y∑k γpktLikt−1Lit−1 Y Y YN 117 117 117 117 117 117Infectious External MalnutritionCauses(7) (8) (9) (10) (11) (12)PollExShockTSP 0.086*** 0.081 0.110 0.478 -0.019 -0.021(0.032) (0.069) (0.089) (0.418) (0.022) (0.041)ExShock -0.693 -2.470 0.038(0.541) (2.489) (0.197)Region Y Y Y Y Y Y∑k γpktLikt−1Lit−1 Y Y YN 117 117 117 117 117 117Notes: All regressions are weighted by total birth. All regressions start of period overall mortalityrate, change in log GDP per capita, change in hospital beds per capita, and change in agriculturalemployment share. Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, *p<0.1142Table B.4: Changes in Infant Mortality Rate and Shocks by Causes of Death:NO2, 2SLSCardio- Infant DigestiveRespiratory SpecificDep. Var: ∆IMRC (1) (2) (3) (4) (5) (6)PollExShockNO2 1.009*** 2.241** -1.736 -2.471 0.051 -0.038(0.360) (1.002) (1.408) (3.739) (0.083) (0.266)ExShock -4.401 0.302 0.408(2.782) (8.104) (0.496)Region Y Y Y Y Y Y∑k γpktLikt−1Lit−1 Y Y YN 117 117 117 117 117 117Infectious External MalnutritionCauses(7) (8) (9) (10) (11) (12)PollExShockNO2 0.237** 0.459 0.052 1.137 -0.064 -0.168(0.103) (0.395) (0.286) (1.266) (0.072) (0.216)ExShock -1.380 -1.781 0.214(0.842) (2.383) (0.361)Region Y Y Y Y Y Y∑k γpktLikt−1Lit−1 Y Y YN 117 117 117 117 117 117Notes: All regressions are weighted by total birth. All regressions start of period overall mortalityrate, change in log GDP per capita, change in hospital beds per capita, and change in agriculturalemployment share. Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, *p<0.1143Table B.5: Change in Infant Mortality Rate and Shocks: TSP, RobustnessEnergy Import High-skill Share of TCZ IO Adjusted OutputProduction Shocks Shock Ownership ShocksDep. Var: ∆IMR (1) (2) (3) (4) (5) (6) (7)PollExShockTSP 0.261*** 0.497*** 0.381*** 0.216** 0.383*** 0.062**(0.100) (0.145) (0.105) (0.093) (0.107) (0.026)ExShock -1.462*** -0.423 2.496* -1.434*** -0.325 0.242(0.399) (0.486) (1.423) (0.378) (0.528) (0.299)∆EnergyProd 0.307**(0.131)PollImShockTSP -0.124(0.107)HighSkillShock -6.945**(3.407)Share SOE 7.182**(3.420)Share Foreign 0.479(3.506)TCZ 2.688***(0.751)PollOutputShockTSP 0.080***(0.027)OutputShock -0.560**(0.258)Region×Year Y Y Y Y Y Y YInitial Conditions Y Y Y Y Y Y YContemporary Shocks Y Y Y Y Y Y Y∆IMRt−1 & ∆IMR2t−1 Y Y Y Y Y Y Y∑k γpktLikt−1Lit−1Y Y Y Y Y Y YN 340 673 673 340 673 673 340Notes: All regressions are weighted by population of age 0. Initial conditions include start of period GDP per capita, overallmortality rate, agriculture employment share and population density. Contemporaneous shocks include change in log GDPper capita, change in share of boys, change in share of population with middle school education, change in share of populationwith high school education or above, change in number of hospital beds per capita, change in agricultural employment share,a dummy of provincial capital city and distance to the nearest port. Standard errors are clustered at the province level. ***p<0.01, ** p<0.05, * p<0.1144Table B.6: Change in Infant Mortality Rate and Shocks: NO2, RobustnessEnergy Import High-skill Share of TCZ IO Adjusted OutputProduction Shocks Shock Ownership ShocksDep. Var: ∆IMR (1) (2) (3) (4) (5) (6) (7)PollExShockNO2 1.908*** 2.410*** 1.893*** 1.719*** 2.068*** 0.364***(0.458) (0.681) (0.509) (0.437) (0.480) (0.113)ExShock -2.156*** -1.192* 1.873 -2.096*** -1.089* -0.074(0.498) (0.618) (1.226) (0.459) (0.559) (0.263)∆EnergyProd 0.303**(0.135)PollImShockNO2 -0.384(0.291)HighSkillShock -6.991**(2.831)Share SOE 6.821**(3.408)Share Foreign 0.020(3.421)TCZ 2.734***(0.766)PollOutputShockNO2 0.651***(0.249)OutputShock -0.163(0.524)Region×Year Y Y Y Y Y Y YInitial Conditions Y Y Y Y Y Y YContemporary Shocks Y Y Y Y Y Y Y∆IMRt−1 & ∆IMR2t−1 Y Y Y Y Y Y Y∑k γpktLikt−1Lit−1Y Y Y Y Y Y YN 340 673 673 340 673 673 340Notes: All regressions are weighted by population of age 0. Initial conditions include start of period GDP per capita, overallmortality rate, agriculture employment share and population density. Contemporaneous shocks include change in log GDPper capita, change in share of boys, change in share of population with middle school education, change in share of populationwith high school education or above, change in number of hospital beds per capita, change in agricultural employment share,a dummy of provincial capital city and distance to the nearest port. Standard errors are clustered at the province level. ***p<0.01, ** p<0.05, * p<0.1145Table B.7: Change in PM2.5 Concentration and Shocks: RobustnessNeighboring Energy Import High-skill Share of TCZ IO Adjusted OutputShocks Production Shocks Shock Ownership ShocksDep. Var: ∆PM2.5 (1) (2) (3) (4) (5) (6) (7) (8)PollExShockTSP 0.178*** 0.179*** 0.211* 0.184*** 0.178*** 0.162** 0.033*(0.064) (0.067) (0.109) (0.068) (0.065) (0.070) (0.019)ExShock -0.708 -0.585 -0.655 -0.900 -0.613 -0.617 -0.067(0.574) (0.530) (0.553) (1.397) (0.519) (0.513) (0.252)Wind PollExShockTSP,N 0.132(0.128)∆EnergyProd 0.209**(0.092)PollImShockTSP -0.042(0.129)HighSkillShock 0.699(2.690)∆Share SOE 0.615(2.701)∆Share Foreign 1.131(4.027)TCZ 1.516*(0.887)PollOutputShockTSP 0.043**(0.019)OutputShock -0.120(0.259)Region×Year Y Y Y Y Y Y Y YInitial Conditions Y Y Y Y Y Y Y YContemporary Shocks Y Y Y Y Y Y Y Y∆IMRt−1 & ∆IMR2t−1 Y Y Y Y Y Y Y Y∑k γpktLikt−1Lit−1Y Y Y Y Y Y Y YN 340 340 340 340 340 340 340 340Notes: All regressions are weighted by population of age 0. Initial conditions include start of period GDP per capita, overall mortalityrate, agriculture employment share and population density. Contemporaneous shocks include change in log GDP per capita, change in shareof boys, change in share of population with middle school education, change in share of population with high school education or above,change in number of hospital beds per capita, change in agricultural employment share, a dummy of provincial capital city and distance tothe nearest port. Standard errors are clustered at the province level. *** p<0.01, ** p<0.05, * p<0.1146Appendix CAppendix for Chapter 3C.1 Additional Tables and FiguresFigure C.1: Survivor Cohort Size — Mean and Coefficient of Variation across Counties. Size: Std Dev/Mean406080100120Cohort Size: Mean1950 1955 1960 1965Birth Yearmean Std Dev/MeanFigure C.2: Population Size by Year — Mean and Coefficient of Variation across Counties1.481.51.521.541.561.58Population Size: Std Dev/Mean293031323334Population Size: Mean (10,000)1953 1957 1961 1964Yearmean Std Dev/Mean147Figure C.3: China’s Historical Railway Network (CIA, 1967)148Table C.1: Famine Severity and Weather Conditions1959 Spring Temperature 0.0366 1959 Spring Temperature2 -0.0013(0.0728) (0.0030)1959 Spring Precipitation 0.0042 1959 Spring Precipitation2 -0.0001(0.0142) (0.0004)1959 Summer Temperature 0.1871 1959 Summer Temperature2 -0.0042(0.5225) (0.0111)1959 Summer Precipitation 0.0167** 1959 Summer Precipitation2 -0.0003**(0.0066) (0.0001)1960 Spring Temperature -0.0689 1960 Spring Temperature2 0.0014(0.1258) (0.0052)1960 Spring Precipitation 0.0033 1960 Spring Precipitation2 0.0004(0.0124) (0.0006)1960 Summer Temperature -0.3728 1960 Summer Temperature2 0.0104*(0.2783) (0.0061)1960 Summer Precipitation -0.0131 1960 Summer Precipitation2 0.0004(0.0101) (0.0003)1961 Spring Temperature 0.0170 1961 Spring Temperature2 0.0003(0.0803) (0.0035)1961 Spring Precipitation -0.0051 1961 Spring Precipitation2 -0.0000(0.0127) (0.0004)1961 Summer Temperature 0.1689 1961 Summer Temperature2 -0.0054(0.2593) (0.0054)1961 Summer Precipitation -0.0016 1961 Summer Precipitation2 -0.0000(0.0068) (0.0002)Notes: Robust standard errors are clustered at the prefecture level. *** p<0.01, ** p<0.05, * p<0.1149Table C.2: First Stage Results of IV RegressionsCol (2) Table 3.3 Col (4) Table 3.3 Col (4) Table 3.4Dep. Var: Sr Ss Sw Sr Ss Sw ∆Export ∆Export(1) (2) (3) (4) (5) (6) (7) (8)ψr 0.379*** -0.142* -0.161** 0.324*** -0.141* -0.151*** 0.267*** 0.221***(0.060) (0.072) (0.063) (0.051) (0.081) (0.047) (0.044) (0.041)ψs 0.044 0.294 -0.213* 0.054 0.355** -0.219*** 0.090** 0.114**(0.046) (0.184) (0.113) (0.044) (0.152) (0.076) (0.044) (0.044)ψw -0.125** 0.009 0.551*** -0.050 -0.038 0.385*** -0.032 -0.002(0.052) (0.140) (0.147) (0.036) (0.097) (0.109) (0.037) (0.037)Province Y Y Y Y Y Y Y YWeather Controls Y Y Y YN 1,164 1,164 1,164 1,164 1,164 1,164 1,164 1,164R2 0.708 0.591 0.519 0.771 0.645 0.685 0.573 0.641Notes: Weather controls include spring temperature and its squared term, summer temperature and its squared term, spring precip-itation and its squared term, and summer precipitation and its squared term for famine years 1959, 1960 and 1961. Robust standarderrors are clustered at the prefecture level. *** p<0.01, ** p<0.05, * p<0.1150C.2 Data AppendixC.2.1 Yearly County Population SizeWe compile a cross-county panel data on population size from various sources. The majorityof data is from numerous volumes of Local Chronicle, which is collected and assembled by Cao(2005). We supplement this data with information collected from various statistical bookspublished by provincial Statistics Bureau, including Hebei Population Statistics: 1949-1984,Henan Population Statistics: 1949-1988, Hunan Population Statistics: 1949-1991, Jilin Pop-ulation Statistics: 1949-1984, Shandong Population Statistics: 1949-1984, Shanxi PopulationStatistics: 1949-1990, Shannxi Population Statistics: 1949-1990, and Zhejiang PopulationStatistics: 1949-1985.C.2.2 Historical Trade Data from Different SourcesAnother data which records bilateral trade flows between China and other countries in the1950s is the International Monetary Fund’s Direction of Trade Statistics (DOTS). DOTSprovides information on trade flows at Country×Year level and has been used in severalrecent studies including Head et al. (2010) and Berger et al. (2013). In this section, wecompare the China Customs data with that from DOTS. We use the data from Correlatesof War Projects Trade Data Set (CWT), which cleans the data from DOTS. CWT datamainly rely on the figures reported by the importing countries. In the cases of missing orzero importing values, the values reported by exporters will be used.Figure C.4 provides a scatter plots of trade flows recorded by CWT against the onesreported by Chinese Customs for the top 15 trading partners of China in 1957. The scatterpoints cluster along the 45 degree line and there is no pattern that China Customs system-atically over or under report trade flows. In addition, we detect many cases of missing valuesor incorrect zeros in CWT over years 1958 to 1961. For example, as is recorded by CWT,trade flows between China and the USSR were missing in 1958 and 1959, and they were zerosin 1960 and 1961.Figure C.4: Trade Flows in 1967 from Different Sources (Top 15 Trading Partners)234567log Trade Flow: CWT2 3 4 5 6 7log Trade Flow: China CustomsImport Export151


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items