University of Southern California Dissertations and Theses

Essays on health and aging with focus on the spillover of human capital

(USC Thesis Other)

Essays on health and aging with focus on the spillover of human capital

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Essays on Health and Aging with Focus on the Spillover of Human Capital by Mingming Ma A Dissertation Presented to the FACULTY OF THE GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulllment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (Economics) August 2018 Copyright 2018 Mingming Ma Dedication This dissertation is dedicated . . . To my grandmother. ii Acknowledgements I would like to pay special gratitude and appreciation to the persons below who generously con- tributed to the work presented in this dissertation and assisted me at every point to cherish my goal: First and foremost, my advisor, John Strauss for his vital guidance and support. His work and adventure as an economist inspired me to become a researcher in health and development issues. During my endeavor to achieve this goal, he has taught me passionately, guided me persistently and encouraged me ardently. I have beneted not only academically from the hard work under his advisement, but also emotionally from harvesting and sharing the fruits at his house. Without his mentoring and constant motivation, the completion of my Ph.D. work would not have been possible. Professor Cheng Hsiao, for oering me the opportunity to work with him as a research assistant and for providing valuable comments and advice on my research. Professor Jerey Nugent, for being so available and patient all the time. Professor Eileen Crimmins and Professor Geert Ridder, who generously serve on my committee and give me insightful feedback. My co-author, Urvashi Jain, with whom I have had two papers, a great amount of good food and countless hours of awe-inspiring discussions. Life at USC is much more enjoyable with her around. iii Morgan Ponder, Young Miller and Fatima Perez, who have been always there to lend a hand. Wenjing Chu and all my colleges and friends in the Economics Department, for listening to me and supporting me. Last but not least, my parents and my husband, for their unconditional love and support. They have let me be who I am and believed in me when I doubted myself. I owe everything in my life to them. iv Table of Contents Dedication ii Acknowledgements iii List Of Tables viii List Of Figures xiii Abstract xiv Chapter 1: Introduction 1 Chapter 2: Together in Sickness and in Health: Spousal In uence in Health and Health Behaviors of Elderly in England 5 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2.3 Theoretical Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.3.1 Assortative Mating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.3.2 Shared Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.3.3 Spillover Eect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.4 Emprical Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.4.1 Empirical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.4.2 Assumptions and Empirical Methods . . . . . . . . . . . . . . . . . . . . . . . 12 2.5 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.5.1 Outcome Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.5.2 Control Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.5.3 Sample and Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.6 Estimation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.6.1 Health Status . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.6.2 Health-related Behaviors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.7 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.7.1 Dierence GMM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.7.2 Path Dependence and Spousal In uence . . . . . . . . . . . . . . . . . . . . . 41 2.7.3 Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 2.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 v Chapter 3: Does Children's Education Matter for Parents' Health and Cognition in Old Age? Evidence from China 45 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 3.2 Related Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 3.3 Potential Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 3.3.1 Social Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 3.3.2 Social In uence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 3.3.3 Access to Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 3.3.4 Psychological Well-being . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 3.3.5 Labor Supply . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 3.3.6 Existing Evidence of Dierent Pathways . . . . . . . . . . . . . . . . . . . . . 56 3.4 Institutional Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 3.4.1 Demographic Transition, Aging and Old-Age Support in China . . . . . . . . 57 3.4.2 Compulsory Education Reform in China . . . . . . . . . . . . . . . . . . . . . 58 3.5 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 3.5.1 CHARLS Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 3.5.2 Measure of Children's Education . . . . . . . . . . . . . . . . . . . . . . . . . 62 3.5.3 Birth Information of Children . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 3.5.4 Control Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 3.5.5 Health and Cognition Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . 65 3.5.6 Potential Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 3.5.7 Sample and Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . 68 3.6 Baseline Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 3.6.1 Empirical Methods: IV/2SLS . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 3.6.1.1 Identication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 3.6.1.2 Threat to Identication . . . . . . . . . . . . . . . . . . . . . . . . . 77 3.6.2 Basline Model Estimation Results . . . . . . . . . . . . . . . . . . . . . . . . 80 3.6.2.1 First Stage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 3.6.2.2 Cognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 3.6.2.3 Subjective Health Measures . . . . . . . . . . . . . . . . . . . . . . . 84 3.6.2.4 Objective Physical Health . . . . . . . . . . . . . . . . . . . . . . . . 84 3.6.2.5 Mental Health . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 3.6.2.6 Mechanism: Social In uence . . . . . . . . . . . . . . . . . . . . . . 90 3.6.2.7 Mechanism: Social Support . . . . . . . . . . . . . . . . . . . . . . . 90 3.6.2.8 Mechanism: Access to Resources . . . . . . . . . . . . . . . . . . . . 91 3.6.2.9 Mechanism: Other Proximate Pathways . . . . . . . . . . . . . . . . 92 3.7 Dynamic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 3.7.1 Empirical Methods: Lewbel's IV Approach . . . . . . . . . . . . . . . . . . . 97 3.7.2 Dynamic Model Estimation Results . . . . . . . . . . . . . . . . . . . . . . . 99 3.8 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 3.9 Conclusion and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 Chapter 4: Height Shrinkage, Cognition and Health among the Elderly: Compar- isons across England and Indonesia 111 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 4.2 Existing Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 4.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 vi 4.3.1 England . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 4.3.2 Indonesia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 4.4 Shrinkage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 4.4.1 Extent of Shrinkage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 4.4.2 (Erroneous) Height Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 4.4.3 Covariates of Shrinkage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 4.5 Shrinkage and Health . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 4.5.1 Shrinkage and Cognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 4.5.2 Shrinkage and Physical Health . . . . . . . . . . . . . . . . . . . . . . . . . . 138 4.5.3 Shrinkage and Mortality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 4.6 Strengths and Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 Chapter 5: Conclusion 151 Bibliography 154 Appendix A Chapter 2 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 Appendix B Chapter 3 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 Appendix C Chapter 4 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 vii List Of Tables 2.1 Descriptive Statistics of Main Variables for Men . . . . . . . . . . . . . . . . . . . . . 19 2.2 Descriptive Statistics of Main Variables for Women . . . . . . . . . . . . . . . . . . . 20 2.3 Correlation Coecients of Health Status and Health Behaviors Within Couples . . . 21 2.4 Results on Good Self-reported Health (SRH) . . . . . . . . . . . . . . . . . . . . . . 24 2.5 Results on Depressive Symptoms (CES-D 8 Score) . . . . . . . . . . . . . . . . . . . 25 2.6 Results on Smoking Status (= 1 if smokes now) . . . . . . . . . . . . . . . . . . . . . 27 2.7 Results on Smoking Intensity (# cigarettes/day) . . . . . . . . . . . . . . . . . . . . 28 2.8 Results on Drinking Frequency (# days/week drinks) . . . . . . . . . . . . . . . . . 29 2.9 Results on Frequency of Exercise (# days/week participating in vigorous physical activities, VPA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 2.10 Results on Social Contact (= 1 if any weekly contact with relatives in person) . . . . 31 2.11 Dierence GMM Results on Good Self-reported Health (SRH) . . . . . . . . . . . . . 34 2.12 Dierence GMM Results on Depressive Symptoms (CES-D 8 Score) . . . . . . . . . 35 2.13 Dierence GMM Results on Smoking Status (= 1 if smokes now) . . . . . . . . . . . 36 2.14 Dierence GMM Results on Smoking Intensity (# cigarettes/day) . . . . . . . . . . 37 2.15 Dierence GMM Results on Drinking Frequency (# days/week drinks) . . . . . . . . 38 2.16 Dierence GMM Results on Frequency of Exercise (# days/week participating in vigorous physical activities, VPA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 viii 2.17 Dierence GMM Results on Social Contact (= 1 if any weekly contact with relatives in person) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.1 Implementation of Compulsory Education and Regional Schooling Level . . . . . . . 60 3.2 Descriptive Statistics of Major Variables . . . . . . . . . . . . . . . . . . . . . . . . . 70 3.3 First Stage Results: Years of Schooling of the Highest Educated Child . . . . . . . . 81 3.4 Level Eects of Children's Education on Baseline Cognition of Parents . . . . . . . . 83 3.5 Level Eects of Children's Education on Baseline Subjective Health Measures of Parents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 3.6 Level Eects of Children's Education on Baseline Physical Health of Parents . . . . 87 3.7 Level Eects of Children's Education on Baseline Body Weight of Parents . . . . . . 88 3.8 Level Eects of Children's Education on Baseline Mental Health of Parents . . . . . 89 3.9 Level Eects of Child's Education on Baseline Health Behaviors of Parents (Social In uence) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 3.10 Level Eects of Child's Education on Baseline Social Support from Children . . . . . 94 3.11 Level Eects of Child's Education on Baseline Access to Resources of Parents . . . . 95 3.12 Level Eects of Child's Education on Baseline Labor Supply and Psychological Well- being of Parents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 3.13 Incremental Eects of Child's Education on Wave 2 Cognition of Parents . . . . . . 101 3.14 Incremental Eects of Child's Education on Wave 2 Subjective Health Measures of Parents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 3.15 Incremental Eects of Child's Education on Wave 2 Physical Health of Parents . . . 103 3.16 Incremental Eects of Child's Education on Wave 2 Body Weight of Parents . . . . . 104 3.17 Incremental Eects of Child's Education on Wave 2 Mental Health (CES-D 10) of Parents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 4.1 Shrinkage (in cm) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 4.2 Shrinkage (in cm) Covariates in ELSA . . . . . . . . . . . . . . . . . . . . . . . . . . 128 4.3 Shrinkage (in cm) Covariates in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 ix 4.4 Descriptive Statistics: ELSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 4.5 Descriptive Statistics: IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 4.6 Height Shrinkage and Cognition (Word Recall) in ELSA . . . . . . . . . . . . . . . . 134 4.7 Height Shrinkage and Fluency (Animal Naming) in ELSA . . . . . . . . . . . . . . . 135 4.8 Height Shrinkage and Cognition (Word Recall) in IFLS . . . . . . . . . . . . . . . . 136 4.9 Height Shrinkage and Fluency (Animal Naming) in IFLS . . . . . . . . . . . . . . . . 137 4.10 Height Shrinkage and Lung Function (Forced Expiratory Volume) in ELSA . . . . . 141 4.11 Height Shrinkage and Grip Strength in ELSA . . . . . . . . . . . . . . . . . . . . . . 142 4.12 Height Shrinkage and Lung Function (Peak Expiratory Flow) in IFLS . . . . . . . . 143 4.13 Height Shrinkage and Grip Strength in IFLS . . . . . . . . . . . . . . . . . . . . . . 144 4.14 Height Shrinkage and Disability in ELSA . . . . . . . . . . . . . . . . . . . . . . . . 145 4.15 Height Shrinkage and Diculty in Kneeling, Stooping or Crouching in ELSA . . . . 145 4.16 Height Shrinkage and Disability in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . 146 4.17 Height Shrinkage and Diculty in Kneeling, Bowing or Squatting in IFLS . . . . . . 147 4.18 Height Shrinkage and Mortality in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . 148 A1 Results on Self-reported Health (1{3 Scale) . . . . . . . . . . . . . . . . . . . . . . . 168 A2 Results on Frequency of Exercise (=1 if any weekly vigorous physical activities) . . . 169 A3 Results on Chronic Conditions (# of severe diseases) . . . . . . . . . . . . . . . . . . 170 B1 Level Eects of Children's Education on Wave 2 Cognition of Parents . . . . . . . . 171 B2 Level Eects of Children's Education on Wave 2 Subjective Health Measures of Parents172 B3 Level Eects of Children's Education on Wave 2 Physical Health of Parents . . . . . 172 B4 Level Eects of Children's Education on Wave 2 Body Weight of Parents . . . . . . 173 B5 Level Eects of Children's Education on Wave 2 Mental Health of Parents . . . . . . 174 x B6 Level Eects of Children's Education on Baseline Health Behaviors of Parents: Fre- quency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 B7 Level Eects of Children's Education on Transfer of Time: One IV . . . . . . . . . . 176 B8 Level Eects of Children's Education on Labor Supply: Rural Sample . . . . . . . . 177 B9 Robustness with Alternative Instruments: Level Eects of Children's Education on Baseline Cognition of Parents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178 B10 Robustness with Alternative Instruments: Level Eects of Children's Education on Baseline Subjective Health Measures of Parents . . . . . . . . . . . . . . . . . . . . . 178 B11 Robustness with Alternative Instruments: Level Eects of Children's Education on Baseline Physical Health of Parents . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 B12 Robustness with Alternative Instruments: Level Eects of Children's Education on Baseline Body Weight of Parents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 B13 Robustness with Alternative Instruments: Level Eects of Children's Education on Baseline Mental Health (CESD 10) of Parents . . . . . . . . . . . . . . . . . . . . . . 181 B14 Robustness to Control of OCP: Level Eects of Children's Education on Baseline Health and Cognition of Parents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 B15 Robustness to Control of OCP: Level Eects of Children's Education on Baseline Body Weight and Depressive Symptoms of Parents . . . . . . . . . . . . . . . . . . . 183 B16 Instrumenting Own Education: Level Eects of Children's Education on Baseline Health and Cognition of Parents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184 B17 Instrumenting Own Education: Level Eects of Children's Education on Baseline Body Weight and Depressive Symptoms of Parents . . . . . . . . . . . . . . . . . . . 185 B18 Own Education as Exogenous: Incremental Eects of Children's Education on Wave 2 Health and Cognition of Parents . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186 B19 Own Education as Exogenous: Incremental Eects of Children's Education on Wave 2 Body Weight and Depressive Symptoms of Parents . . . . . . . . . . . . . . . . . . 187 B20 Robustness to Adjustment for Multiple Hypotheses Testing . . . . . . . . . . . . . . 187 B21 Robustness with Lewbel IV for Children's Education: Level Eects of Children's Education on Baseline Cognition and Health of Parents . . . . . . . . . . . . . . . . 188 B22 Robustness with Lewbel IV for Children's Education: Level Eects of Children's Education on Baseline Body Weight of Parents . . . . . . . . . . . . . . . . . . . . . 188 xi C1 Limb Length and Shrinkage in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . 194 C2 Height Shrinkage and ADL in ELSA . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 C3 Height Shrinkage and IADL in ELSA . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 C4 Height Shrinkage and ADL in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 C5 Height Shrinkage and IADL in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 C6 Height Shrinkage and Diculty in Lifting in ELSA . . . . . . . . . . . . . . . . . . . 199 C7 Height Shrinkage and Diculty in Getting up from Chair in ELSA . . . . . . . . . . 200 C8 Height Shrinkage and Diculty in Carrying a Heavy Load in IFLS . . . . . . . . . . 201 C9 Height Shrinkage and Diculty in Getting up from Chair in IFLS . . . . . . . . . . 202 xii List Of Figures 3.1 Linking Parents' Health and Children's Education, based on the Social Network Theory of Health of Berkman, Glass and Brissette (2000) . . . . . . . . . . . . . . . 53 3.2 Roll-out of Compulsory Education . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 3.3 Eect of Compulsory Education on Completed Education Level by High/Low School- ing areas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 3.4 Eects of Exposure to Compuslory Education on Years of Schooling . . . . . . . . . 75 4.1 Shrinkage over 4 years in ELSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 4.2 Shrinkage over 8 years in ELSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 4.3 Shrinkage over 7 years in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 4.4 Shrinkage over 17 years in IFLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 4.5 Shrinkage over 4 and 8 years by gender and age in ELSA . . . . . . . . . . . . . . . 124 4.6 Shrinkage over 14 and 17 years by gender and age in IFLS . . . . . . . . . . . . . . . 124 xiii Abstract This dissertation studies the transmission of human capital within family networks and over time. Chapter 2 analyzes the widely documented spousal concordance in health status and health behav- iors, utilizing data from the English Longitudinal Study of Ageing (ELSA). GMM estimations of a dynamic panel model with individual xed eects imply signicant and positive contemporaneous spillover eects in health within elderly couples, with women's mental health being aected more by their spouse. Strong causal eects are also found for spousal smoking, drinking and physical activity, as well as social contact. Gender dierences are found in the spillover eects, including mental health and smoking behaviors. Chapter 3 evaluates the causal impacts of children's education on parents' health and cognition, using data from the rst two waves of the China Health and Retirement Longitudinal Study. Identication is achieved by using the exposure of adult children to the compulsory education reform around 1986 in China and its interaction with enforcement intensity as instruments for children's years of schooling. IV estimation results using the baseline survey data demonstrate that increasing years of education of adult children lead to higher levels of cognitive functions of older adults. Parents with better educated children also have higher subjective survival expectations, improved lung function and greater body weight. Dynamic model results for the follow-up sample indicate positive and signicant incremental eects of children's education on cognitive abilities of older adults when baseline cognition is controlled for. Further evidence suggests that adult xiv children's education might shape parental health in old age by providing social support, aecting parental access to resources as well as in uencing parental labor supply and psychological well- being. Chapter 4 documents the extent of height loss and explores the associations of early life health measured by height, depletion of health endowments measured by height shrinkage and later life health among the English and Indonesian elderly populations. Older women and Indonesian elderly are found to shrink more. Height shrinkage is signicantly associated with lower cognition and reduced physical function for elderly men and women in both countries, even when baseline height is controlled for. Extreme height loss is also found to be a major predictor of mortality among the Indonesian elderly. xv Chapter 1 Introduction There exists a considerable body of literature on the development of human capital and its im- plications for individual's economic and non-economic outcomes. Public policies regarding health and education have been widely adopted and recognized as crucial for reducing inequality and im- proving well-being of the society as a whole. Furthermore, there has been growing interest in the spread of benets of education and health capital that go beyond individuals at one point in time which are usually targeted as the unit of interventions. For instance, human capital in one's community is found to be an informative predictor of individual's human capital. Recent research suggests that life expectancy of people without much post-secondary schooling increases with the share of the college educated population in an area where they live (Cutler, 2018). From a more micro prospective, analyses in eects of peers and social ties in educational attainment and health increasingly demonstrated that there might be causal spillovers underlying the observed transmission of human capital within social networks. In light of these previous ndings, the rst two essays in this dissertation aim to provide further evidence on the spillover eects of human capital among one of the most important social networks of individuals{ their family{within which despite the well-established correlations of health and education, causal studies remain limited. 1 In another strand of human capital literature, a large number of studies have been conducted to understand the eect of early life health and experience on later life outcomes (Almond and Currie, 2011). Among measures of early life health, adult height, is commonly used and found to be an important predictor of human capital formation and income in early and late adulthood (e.g., Thomas and Strauss, 1997; Case and Paxson, 2008a; Smith et al., 2012). Quantifying the lasting eect of early life health on older adults using measured height as a proxy, however, is complicated by the aging process which results in loss of height in old age. A still unsolved question is whether the height-health nexus in old age is driven by height or the loss of height. The third essay of this dissertation thus tries to ll this gap in literature by decomposing the association of height and health of older adults, along with describing the covariates of height loss and its empirical relationship with late life health. Chapter 2 examines whether there are causal spillovers of health status and behaviors among married and cohabiting English older adults. Although spousal concordance in health is widely doc- umented, direct evidence on the existence of spousal spillovers is relatively scarce due to confound- edness from other non-causal factors: assortative mating and shared environment. Disentangling the causal spillover from the other factors is important, because of its relevance for ecient policy designs and comprehensive program evaluations in health domain. We utilize the data from the English Longitudinal Study of Ageing (2000{2014) which is representative of the population aged over 50 in UK and estimate the contemporaneous eect of spousal health status and behaviors in a dynamic panel model which allows for path dependence of one's own health. Using rst dierencing and System GMM to address endogeneity problems, we nd there is strong and positive spousal concordance in health status and health behaviors. GMM results conrm that there exist spillover eects between spouses for English couples in health status measured by self reported health and depressive symptoms, as well as health-related behaviors including smoking, drinking, exercising 2 and social contacting. There exists certain but not major gender dierences. While wives have larger impact on their husband's smoking likelihoods, men have larger eects on their wives' mental health then do their wives on them. In addition to the spillovers of human capital within couples, in Chapter 3, I evaluate the causal eects of the education of children on the health and cognition of parents and what are the un- derlying mechanisms. Most of the existing ndings on this topic are based on correlational studies with focus on physical functioning and longevity of older adults. To rule out the possibility of reverse causality and omitted variables and provide empirical evidence on multiple dimensions of health, I apply an instrumental variables approach to data from the national baseline survey of the China Health and Retirement Longitudinal Study (CHARLS), and exploit as natural experiments the gradual roll-out of the nine-year compulsory education law from 1985 to 1991 across Chinese provinces. The law enforcement timing and the heterogeneous law enforcement intensity generate exogenous variation in years of schooling of children both across birth provinces and cohorts. IV estimation results show that increasing years of education of children lead to substantial improve- ment in a wide array of health and cognition outcomes of parents, including episodic memory, mental intactness, subjective survival expectation, peak ow (which measures lung function) and body weight measures including BMI. Consistent with the social integration theory of health and the tradition of resource pooling among family members, as well as the obligation of children sup- porting parents in old age, I nd evidence for positive and causal in uence of children's years of schooling on the net amount of nancial transfers parents received from children, on log per capita household expenditure of parents, parental access to clean fuels and improved sanitation measured by in-house ushable toilets. For psychological and behavioral factors that have been shown to relate to later life health, I nd that parents with better educated children are more satised with their lives and less likely to work in old age. These ndings suggest that children's education might 3 shape parental health through both social support and resource provision, and proximate behav- ioral and psychological pathways. I also estimate a dynamic model using the 2 year follow-up data of CHARLS to understand the eects of children's education on the progression of health of parents as they age. Lagged baseline health outcomes are included and instrumented using Lewbel's IV approach based on higher moment conditions. The estimations of the dynamic model indicate a positive and signicant incremental (periodic) eect of children's education on the episodic memory of older adults. In other words, besides its level eects, children's education is also benecial for the maintenance of cognitive functions of older adults, even in the short run. Chapter 4, on the other hand, explores the interaction of early life health measured by height, deterioration of health endowment measured by height shrinkage and later life health. We rst document the extent of height loss based on measured height over time among the English and Indonesian elderly populations, where we see that older women and Indonesian elderly shrink more. We then explore associations between height, height loss and cognitive and physical functions. We nd that height shrinkage is signicantly associated with lower cognition and reduced physical function for elderly men and women in both countries, even when baseline height is controlled for. Extreme height loss is also found to be a major predictor of mortality among the Indonesian elderly. Our results, though being purely correlational, indicate that height shrinkage in old age might serve as an informative yet underutilized biomarker for later life health. The remainder of this dissertation is organized as follows: Chapter 2 examines the spousal impact of health status and behaviors; Chapter 3 evaluates the causal eects of children's education on parental health and cognition in old age, along with analysis on the underlying mechanisms; Chapter 4 explores the interaction of height, height loss and health in old age; Chapter 5 summarizes the ndings from the three essays and discusses the policy implications. 4 Chapter 2 Together in Sickness and in Health: Spousal In uence in Health and Health Behaviors of Elderly in England 1 2.1 Introduction A considerable number of studies in social science and medicine have associated health status and health behaviors between spouses (Meyler, Stimpson and Peek, 2007). The general consensus of ndings is that there exists marital concordance of health, both in terms of physical health (Wilson, 2002; Davillas and Pudney, 2017), mental health (Fletcher, 2009) as well as health related behaviors (Fletcher and Marksteiner, 2017; Banks, Kelly and Smith, 2013; Brown, Hole and Roberts, 2014). Theories of the concordance draw upon assortative mating, shared environment and resources, spillover or social learning and aective contagion (e.g., Wilson, 2002; Meyler, Stimpson and Peek, 2007). Although some of the existing studies in the area have dedicated themselves to link one or two of the theories to their correlation ndings, most of them fail to disentangle the eect of dierent channels from each other, especially the causal channels from the non-causal ones. While the rst two channels are of no less importance than the latter two, assortative mating and shared 1 This is a joint work with Urvashi Jain. This analysis uses data or information from the Harmonized ELSA dataset and Codebook, Version E as of April 2017 developed by the Gateway to Global Aging Data. The development of the Harmonized ELSA was funded by the National Institute on Aging (R01 AG030153, RC2 AG036619, 1R03AG043052). For more information, please refer to www.g2aging.org. 5 environment and resources are more of a selection process and re ection problem as in Manski (1993). Understanding the causal pathways (such as spillover or social learning) between spouses generates increasing interest both in observational and experimental studies, because of its strong link with social network and peer eects theories, its relevance for policy designs and program evaluations in health domain and help with benet calculations when there are spillover eects along social ties (Fletcher and Marksteiner, 2017; Brown, Hole and Roberts, 2014). This chapter examines whether there are causal spillover eects of spousal health status and behaviors among married and cohabiting English older adults, by estimating dynamic panel mod- els and controlling for the endogeneity problem due to marriage market matching and shared environment using Generalized Method of Moments (GMM). We use the seven waves of the En- glish Longitudinal Study of Ageing (2000{2014) which is representative of the population aged over 50 in UK. We nd there is strong and positive spousal concordance in health status and health behaviors. GMM results conrm that there exist spillover eects between spouses for En- glish couples in health status measured by self reported health and depressive symptoms, as well as health-related behaviors including smoking, drinking, exercising and social contacting. There are some gender dierences for them but not much. While most studies on spousal health con- cordance have been documenting correlations, this chapter is among the few to provide causal evidence on the spillover eects among spouses, which is based upon similar methodologies ap- plied in Michaud and van Soest (2008)'s study on US elderly couples. Our analysis focuses on the inter-spousal relationships in health domains among a comparable yet dierent population. We also extend the current literature by evaluating various health outcomes and health behaviors at the same time. This chapter is organized as follows. Section 2.2 reviews the related literature. Section 2.3 and 2.4 provides theoretical background, the dynamic model of health and health behaviors and 6 the empirical strategies. Section 2.5 describes the dataset we use. Section 2.6 shows the empirical results. Section 2.7 discusses the robustness of the causal estimation results. Section 2.8 concludes. 2.2 Literature Review While the correlations in health status and health behaviors within couples have been widely recognized in the literature (Meyler, Stimpson and Peek, 2007), relatively little is known about the mechanisms resulting in spousal concordance, and the importance of dierent mechanisms. Wilson (2002) serves as an example by documenting the spousal concordance in physical health and evaluating the causal and non-causal underlying channels. He uses data from Health and Retirement Study (HRS) wave 1992 and nds that assortative mating in the observed characteristics { such as age, education, income, and other socioeconomic and demographic determinants of health status as well as behavioral risk factors { helps explain about 33% to 57% of the overall correlation in health between spouses. The remaining unexplained part of the correlations could either work through non-causal pathways such as sharing environment and assortative mating in unobserved characteristics, for example, health endowment, or causal pathways such as spillover. A recent paper by Davillas and Pudney (2017) explores multi-dementional health data from the UK Understanding Society Panel (origially the British Household Panel Survey) and by estimating a cummulative health exposure model, they nd that health correlation within spouses is unrelated to the length of marriage, which is interpreted as evidence of equal importance of assortative mating and shared in uences. However, these two papers still fail to distinguish the causal spillover eects of spousal health from the shared environment and resources eects. Dierent empirical strategies have been applied to distinguish the causal factors from non- causal ones in the observed spousal concordance in health behaviors. Clark and Etil e (2006) use 9 7 waves of British Household Panel Survey data to examine interactions between spouses in smoking. They show from their bivariate Probit panel estimations that the correlation of smoking status between spouses are mostly due to assortative mating which is captured by the (correlated) indi- vidual random eects instead of endogenous eects rising from bargaining within marriage that is captured by partner's past smoking behaviors or social learning from partner's past health. Brown, Hole and Roberts (2014) and Clark and Etil e (2011) look at spousal correlation in body weight. Clark and Etil e (2011) use life satisfaction as proxy for individual utility and recover in semi-parametric analysis that own Body Mass Index (BMI) has a negative well-being eect but the eect is lower when the BMI of the individuals is larger, which is consistent with social spillover eects in weight. Brown, Hole and Roberts (2014) in their Seemingly Unrelated Regressions with random eects demonstrate the importance of matching in explaining correlated body weight. Their results using single equation random eects model with spousal BMI as a right hand side variable and a rich set of controls are also suggestive of social in uence and spillover eect. Although these ndings have provided insights on the mechanisms of spousal correlations in health and health behaviors, it is still less clear whether there exists causal spillover from spouse in the health domain. In order to have causal interpretations and provide policy implications, Fletcher (2009) studies the spillover eect of mental health between working spouses using data from the Na- tional Survey of Midlife Development in the United States (MIDUS). His results from xed eects es- timation where spousal mental health is instrumented with spousal job problems show that spousal mental health status has an economically signicant eect on individuals' mental health which is even larger than the eect of individuals' own mental health endowment. Michaud and van Soest (2008) which employs 6 waves (1992-2002) of HRS data in the US to study wealth-health gradient in a dynamic panel framework nds in the\Granger" sense a causal impact of the health of elderly men on their wives' mental health two years later by using system GMM. Fletcher and Marksteiner 8 (2017) on the other hand, utilize experimental data from the Lung Health Study and COMBINE study and nd evidence for signicant causal spousal spillover eects from addiction interventions in smoking and drinking. This chapter builds upon Michaud and van Soest (2008) and Powdthavee (2009) in develop- ing a dynamic panel model to further the study in spousal concordance in health, controlling for assortative mating and confounding factors from shared environment and unobservables. Ap- plying the econometric methods by Arellano and Bond (1991), Arellano and Bover (1995) and Blundell and Bond (1998), we are able to test specically the theories on spillover eects. 2.3 Theoretical Background There are three broadly dened channels through which spousal health and health behaviors are correlated: assortative mating which happened when the knot is tied and hence time invariant, shared environment and direct spousal spillover (social learning or aective contagion). 2.3.1 Assortative Mating Becker's marriage market matching theory (Becker, 1973, 1974) predicts that people are married based on their individual characteristics. When the utility of marriage is transferable and com- plementary in individual characteristics, there will be positive assortative mating which implies that people tend to marry people with similar backgrounds and personal traits. Homogamy has been found in education (Mare, 1991), culture and occupation (Kalmijn, 1994), and anthropometric traits (Chiappori, Orece and Quintana-Domeque, 2012). Dohmen et al. (2012) also provide evi- dence for the positive assortative mating on risk and trust preferences. While it is likely that people would select their partners based on their health status, sorting in socio-economic characteristics 9 and lifestyles (e.g., smoking) would also lead to the concordance of health and health behaviors between spouses (Banks, Kelly and Smith, 2013). 2.3.2 Shared Environment Couples who live in the same household, share resources and risk, and make joint decisions on many issues. Thus health behaviors and health status of husbands and wives might be correlated because they are aected by common factors simultaneously. For example, shared exposure to environmental hazard, diet and stress could have similar eects on both partners' health at the same time. This correlated eect might also interrelate with partner selection and be reinforced by the assortative mating between spouses (Davillas and Pudney, 2017). Spouses that are more homogeneous in personal traits and lifestyles are more likely to experience simultaneous shocks. However, it is empirically challenging to estimate shared environment eects because they are normally not observed. Nevertheless, we need to control for this channel in order to causally estimate the spillover eect in health and health behaviors between spouses. 2.3.3 Spillover Eect Social networks have been found to signicantly in uence one's health (Fowler and Christakis, 2008a; Raspe, Hueppe and Neuhauser, 2008) and health behaviors (Christakis and Fowler, 2007; Eisenberg, Golberstein and Whitlock, 2014; Powell, Tauras and Ross, 2005; Clark and Loh eac, 2007; Fletcher, 2010). Marriages or partnerships are arguably the most important social network that individuals form in lifetime, which shape one's knowledge, decisions, behaviors and preferences. Multiple theories indicate the existence of spillover eects (contagion). If spousal health and health behaviors are complementary to each other and endogenous, positive spillover eects exist. Having an obese partner decreases the cost for oneself to be obese and that is why we see the similar 10 body weight trends of couples. On the ip side, smoking has negative externalities to the health of non-smoking spouses, thus an altruistic individual might quit smoking for his or her partner. It is also possible that spouses can learn from or control each other's behavior, and hence be- come more alike. For health status, couples could aect each other directly through transmitting of infectious diseases and caring for each other. There is evidence showing that informal care of a sick family member leads to a reduction in both the physical and mental health of the care- giver (Schulz and Sherwood, 2008). Couples might also directly benet from each other's mental well-being, as shown by Powdthavee (2009) and Fletcher and Frisvold (2009). 2.4 Emprical Strategies 2.4.1 Empirical Model The main goal of this chapter is to identify the causal spillover eects in health and health behaviors within couples. We develop a linear dynamic model for health status and health behaviors, allowing for their persistence and spousal in uence over time. Following Powdthavee (2009) and Michaud and van Soest (2008), let (y it ;y jt ) denote the health status and behaviors of interest of individual i and spouse j as measured by self-reported health, depressive symptoms, behaviors such as smoking, drinking, exercising and social contact respec- tively, in one household at time t. The system of dynamic conditional health demand functions with cross-spouse dependence for a couple can be written as follows: y it = 0 + 1 y it1 +::: + m y itm + 0 y jt +::: + g y jtg + 1 X it + 2 X jt + i +u it (2.1) 11 y jt = 0 + 1 y jt1 +::: + m y jtm + 0 y it +::: + g y itg + 1 X jt + 2 X it + j +u jt (2.2) where i and j are individual-specic and time-invariant characteristics which could be correlated, such as family background, personality, genetic features and endowments. u it andu jt capture time and individual-specic unobservables that have impact on either spouse's outcomes, such as time varying tastes and health shocks which are often common to both spouses, but are not observed by econometricians. X it and X jt are other health covariates (inputs). For simplicity, we are going to omit X it and X jt to better illustrate our empirical strategies. By construction,y it andy jt and all their past and future realizations are related to the individual heterogeneity, which result in endogeneity problems. For instance, individuals born with poor health endowments and inheritable conditions are less likely to be healthy as they grow older. What is more,y it andy jt are correlated with each other, due to correlation in individual heterogeneity, i.e., i and j , which capture the eect from assortative mating. y it and y jt might be correlated with both u jt and u it as a result of shared environments and health shocks. 's and 's are parameters of interest that identify the causal channel, among which 0 and 0 evaluate instantaneous spousal spillover eects. First dierencing of the above system of equations could easily purge out all time-invariant unobserved heterogeneity as well as the eect of assortative mating. y it = 1 y it1 +::: + m y itm + 0 y jt +::: + g y jtg + u it (2.3) y jt = 1 y jt1 +::: + m y jtm + 0 y it +::: + g y itg + u jt (2.4) 2.4.2 Assumptions and Empirical Methods By construction, y it1 and y jt1 are correlated with u it and u jt and so are their current values. For the model to be tractable, we make the following assumptions. 12 Assumption 1. (Dynamics of health and health related behaviors): A 1st order linear Markov process is used to model the state dependence of health and health behaviors, i.e., m = 1. It also implies a separability assumption: health in the previous period serving as a sucient statistic for all health inputs and outcomes for all other prior periods (Strauss and Thomas, 2008). We will relax the baseline assumption and compare the results when m = 1 with cases with m = 2. The empirical result section shows that for certain health status and health behaviors, it is more appropriate to control for further lags based on the specication tests. However, the estimated spillover eects coecients for most outcomes are insensitive to the length of lagged values of own health or behaviors. Assumption 2. Spousal lagged outcomes are irrelevant in the structural model, i.e., g = 0 . This implies the absence of direct lagged spillover eects from spouses. In other words, lagged spousal health or behaviors only aects individuals current health or behaviors through the current level without having any independent eect. We argue that this is a reasonable assumption as the lagged spousal outcomes are 2 years from the current time and we report results of estimations in the reduced form where we only include spousal lags (g = 1) as a robustness check. We will show that for most outcomes of interest, the lagged spillover coecients are insignicant, as evidence for the lack of lagged spousal spillover eects. Assumption 3. Serial uncorrelatedness inu it andu jt . This assumption is standard in time se- ries and dynamic panel literature and can be tested (Arellano and Bond, 1991; Arellano and Bover, 1995). Under Assumption 1 to Assumption 3, the instantaneous spousal spillover eects in the struc- tural model are identiable and can be estimated by using instruments. Valid instruments are available as below: 13 For the rst dierences of lagged outcomes: (y it2 ;y it3 ;:::y i1 ) and (y jt2 ;y jt3 ;:::y j1 ) respec- tively. For the rst dierences of current outcomes: (y it2 ;y it3 ;:::y i1 ) and (y jt2 ;y jt3 ;:::y j1 ) respectively. Inclusion restriction is satised because y it1 and y jt1 are correlated with y it2 and y jt2 and further lags. Exclusion restriction is satised because (y it2 ;y it3 ;:::y i1 ) and (y jt2 ;y jt3 ;:::y j1 ) are uncorrelated with u it and u jt . Blundell and Bond (1998) proposed to use the rst dierences of the endogenous variables to serve as instruments for the level equations, if the following assumption holds as well. Assumption 4. (Stationarity) There is no correlation between the dierences of the explana- tory variables and the unobserved individual xed eects, i.e., E(y it1 i ) = 0 orE(y it i ) is time invariant for alli. This assumption relies on the initial conditions of the data generating process as formalized in Blundell and Bond (1998): the deviations ofy i1 from the steady states are orthogonal to the individual xed eects, i.e., no \regression to the mean". Under the four assumptions, we have the following moment conditions: For the rst dierence equations: s>= 2, t = 3;;T E(y its u it ) = 0; E(y jts u jt ) = 0; E(y jts u it ) = 0; E(y its u jt ) = 0: 14 For the level equations: s = 1, t = 3;;T E(y its ( i +u it )) = 0; E(y jts ( i +u it )) = 0; E(y jts ( j +u jt )) = 0; E(y its ( j +u jt )) = 0: The previous moment conditions outline what are called Dierence and System GMM approach developed by Arellano and Bond (1991), Arellano and Bover (1995) and Blundell and Bond (1998). It can be applied to estimate the \small T , large N" linear dynamic panel model with not strictly exogenous or endogenous independent variables and individual xed eects. Dierence GMM uti- lize the moment conditions for rst dierence equations while System GMM approach increases eciency by imposing the additional moment conditions. System GMM deal with the weak in- strument problem in Dierence GMM if y is close to random walk, and hence the levels bear little information about future changes (Roodman, 2009). System GMM is our preferred estimation strategy because of its improved eciency, however, given that it requires further assumptions, we report Dierence GMM results as a robustness check. Two-step Dierence and System GMM estimates with Windmeijer (2005) corrected robust standard errors are reported. Single equations for men and women are estimated separately. Arellano-Bond test for autocorrelation is applied to test the assumption of the absence of au- tocorrelation in the error term (Assumption 3) which is crucial for the validity of y it2 and further lags as instruments for the rst dierences. After eliminating individual xed eects, the test of 15 rst order autocorrelation in the error term is equivalent to test second order autocorrelation in residuals in dierences since rst order autocorrelation is created by construction. The exogeneity of the instruments is tested by Hansen test of overidentifying restrictions for the whole model and Dierence in Hansen test is conducted for the subset of level equation instruments, given that the additional assumptions in System GMM are non-trivial. 2.5 Data England is one of most developed economies in the world with per capita GDP in PPP international dollars reaching 41,458.7 by 2015. Population in England is older on average than in the U.S. with ratio of elderly over 65 in the national population in 2015 around 18% 2 . We use the English Longitudinal Study of Ageing (ELSA), one of the sister studies of the Health Retirement Study (HRS) in the US, which is representative of the national population aged 50 and above. Respondents with eligible age and their spouses irrespective of age in the survey year are interviewed. The eldwork is conducted biennially since 2002, with refreshment samples added from the wave 1 age- ineligible households in wave 3/4/6/7. We use the harmonized data of the waves 1{7 (2002{2015) of ELSA, which following the RAND version of HRS provides information on the demographics, health and health behaviors of all respondents as well as their spouses 3 . 2.5.1 Outcome Variables We consider two groups of outcomes related to health: health status and health behaviors. Health Status. Health status of interest include self-reported good health measured by a binary variable which equals one if the respondent reports good, very good or excellent health, 2 https://www.census.gov/content/dam/Census/library/publications/2016/demo/p95-16-1.pdf. Accessed on Jan- uary 15, 2017. 3 Gateway to Global Aging https://g2aging.org/ accessed on January 15, 2017. 16 zero if otherwise. In appendix, we also report results using a self-reported health measure on a scale from 1 to 3 which are calculated based on the self-reported health question in wave 3 with European scale 4 and the ones in all other waves with US scale 5 . We regroup the answers to a Likert-scale measure, which equals 1 if the self-report is either excellent, very good or good; 2 if it is fair health; 3 if the respondent reports bad, very bad or poor health. For mental health, we also utilize a depressive symptom index which is based on an 8 item CES-D (total score from 0 to 8) 6 . The health outcomes are chosen such that they cover both general and mental health domains and have been measured relatively consistently and frequently. Health Behaviors. The health related behaviors considered consist of smoking and drinking behaviors for which we examine whether the respondents smoke currently, their intensity as mea- sured by number of cigarettes per day 7 and number of days in a week they drink 8 . Frequency of physical exercise measured in number of days in a week participating in vigorous physical activi- ties 9 is examined as well. We also test in appendix on the extensive margin of exercise, which is measured by a binary indicator for any weekly participation in vigorous physical activities. We also investigate the spousal spillover eects in social activity which is measured by a binary variable indicating whether the respondent has any weekly contact with relatives in person. Biomarkers are not included as they are only available for every alternate wave. Behavior related outcomes are selected as they are important predictors in either physical or mental health. 4 Very good, good, fair, bad and very bad. 5 Excellent, very good, good, fair and poor. 6 In ELSA, there are only 8 questions on CES-D and the harmonized value for each item is 1 if the respondent answers that he or she has the following feeling much of the time last week and 0 otherwise: feeling depressed, feeling that everything was an eort, sleep was restless, felt lonely, felt sad, felt he/she could not get going, felt happy (reverse), felt he/she enjoyed life (reverse). The CES-D 8 score is the sum of the values of the rst 6 items and 1 minus the values of the last 2 items. Thus the higher the score, the more depressive the respondent is. 7 Number of cigarettes per day is recorded as 0 if the respondent does not smoke when interviewed. 8 Number of days in a week of drinking is recorded as 0 if the respondent does not drink when interviewed. 9 e.g., running/jogging, swimming, cycling, aerobics/gym workout, tennis, and digging with a spade. We impute the number of days participating in vigorous physical activities based on respondents' choice from 4 categories. We assign 4 if they chose \more than once per week"; 1 if \once per week"; 0.25 if \one to three times per month"; 0 if \hardly ever". We have experimented with other values for the \more than once per week" category, and the conclusions are not changed. Results are available upon request. 17 2.5.2 Control Variables We control for a wide array of individual and household level characteristics, including: 1) respon- dent and spouse age dummies (5 years interval) 10 , numbers of living daughters and sons, respondent and spouse levels of education which are categorized as less than high school, high school, some college or college and above, with less than high school being the omitted group, as well as length of marriage; 2) logarithm of de ated per capita household wealth 11 In OLS and FE specications, all control variables are treated as exogenous. In GMM estimations, the rst set of covariates is treated as exogenous variables or not estimable when taking rst dierences in Dierence GMM and FE estimations. Household wealth is treated as endogenous and instrumented with valid lags or dierence of lags in GMM estimations. 2.5.3 Sample and Descriptive Statistics In our analysis, we keep couples that remained married or partnered and lived in the same household (i.e., not separated), and both of whom responded with non-missing values in any of the outcome of interest and all control variables, for at least 3 consecutive waves. For dierent outcomes of interest, the number of couples and size of estimation sample varies. Table 2.1 and 2.2 summarize for men and women respectively, the main variables under analysis. We report year-specic descriptive statistics in wave 3 and 5 when the original ELSA cohorts and refreshment cohorts rst became estimable by GMM, as well as wave 7, which is most recent wave with publicly available data. 10 <50, 50-54, 55-59, 60-64, 65-69, 70-74, 75-79, 80 and older. 11 Household wealth is constructed as the sum of the dierent wealth components: Net value of primary residence + Net value of business + Net value of non-housing nancial wealth + net value of the secondary home residence and other property after paying all the debts + total value of other physical assets for those reporting having other land, money owed by others, a trust, a covenant or inheritance, or other assets (Phillips et al., 2017). In practice, the formula for transformation is: log(wealth) =log(wealth + 1) if wealth>= 0 or log(wealth) =log(1wealth) if wealth< 0, following Michaud and van Soest (2008). 18 Across these waves, the levels of self-reported health status of both men and women are fairly stable and similar to each other. The CES-D scores are higher on average for women, which is consistent with the fact that depression in more common among women. There have been a steady decline in the fraction of smokers and intensity of drinking and smoking for both genders. This could be due to either the decline across cohorts or over the years as respondents age. Wives on average are younger and less educated than their husbands. Women also drink and exercise less frequently, although they smoke more intensively than men. Table 2.1: Descriptive Statistics of Main Variables for Men Wave 3 Wave 5 Wave 7 VARIABLES N mean N mean N mean Health Outcomes Have good self reported health 1,510 0.718 1,489 0.765 1,232 0.762 CESD Score 1,508 0.897 1,479 1.060 1,228 0.894 Health Behaviors Smokes now 1,510 0.100 1,512 0.0847 1,269 0.0678 # cigarettes/day 1,456 0.947 1,457 0.681 1,229 0.429 # days/week drinks 1,358 3.140 1,396 3.003 1,150 2.891 Freq vigorous phys activ : # days/week 1,510 1.077 1,519 1.048 1,269 1.162 Any weekly contact w/ relative in person 1,271 0.297 1,303 0.303 1,073 0.285 Demographic & Socioeconomic Characteristics Age (years) 1,510 66.12 1,519 66.81 1,270 69.85 Education 1,510 Fraction 1,519 Fraction 1,270 Fraction Less than high-school 494 0.327 430 0.283 324 0.255 High-school graduate 261 0.173 279 0.184 241 0.190 Some college 402 0.266 409 0.269 351 0.276 College and above 353 0.234 399 0.263 354 0.279 Length of marriage 1,510 37.38 1,519 37.87 1,270 40.66 Log per capita de ated household wealth 1,510 11.54 1,519 11.45 1,270 11.56 Family Composition Number of living son 1,510 1.113 1,519 1.047 1,270 1.072 Number of living daughter 1,510 1.111 1,519 1.078 1,270 1.107 Number of living siblings 1,510 1.705 1,519 1.706 1,270 1.653 Note: Data are restricted to married couples living together, both of who responded with non-missing values in any of the outcome of interest and all of the control variables, for at least 3 consecutive waves. 19 Table 2.2: Descriptive Statistics of Main Variables for Women Wave 3 Wave 5 Wave 7 VARIABLES N mean N mean N mean Health Outcomes Have good self reported health 1,571 0.733 1,526 0.782 1,291 0.778 CESD Score 1,568 1.369 1,518 1.439 1,283 1.242 Health Behaviors Smokes now 1,572 0.108 1,533 0.0978 1,316 0.0783 # cigarettes/day 1,549 1.301 1,505 1.118 1,290 0.791 # days/week drinks 1,418 2.315 1,430 2.235 1,205 2.135 Freq vigorous phys activ : # days/week 1,572 0.907 1,539 0.892 1,315 0.906 Any weekly contact w/ relative in person 1,371 0.373 1,351 0.376 1,144 0.354 Demographic & Socioeconomic Characteristics Age (years) 1,572 63.30 1,540 64.06 1,316 67.17 Education 1,572 Fraction 1,540 Fraction 1,316 Fraction Less than high-school 635.1 0.404 530 0.344 415 0.315 High-school graduate 389.9 0.248 400 0.260 351 0.267 Some college 320.7 0.204 320 0.208 287 0.218 College and above 226.4 0.144 290 0.188 263 0.200 Length of marriage 1,572 37.33 1,540 37.80 1,316 40.63 Log per capita de ated household wealth 1,572 11.52 1,540 11.43 1,316 11.54 Family Composition Number of living son 1,572 1.125 1,540 1.040 1,316 1.067 Number of living daughter 1,572 1.118 1,540 1.062 1,316 1.106 Number of living siblings 1,572 1.866 1,540 1.890 1,316 1.837 Note: Data are restricted to married couples living together, both of who responded with non-missing values in any of the outcome of interest and all of the control variables, for at least 3 consecutive waves. 20 Table 2.3 describes the pairwise correlations between spouses in health status and health be- haviors in wave 3, 5 and 7. Self-reported good health and CES-D scores are signicantly correlated within couples, with Pearson correlation P value well below 1%. The same patterns are found for smoking, drinking, exercising and social contacting behaviors. There is no clear time trend in correlations in health and health behaviors among the ELSA elderly couples, which is also true if we look at ELSA original cohorts and refreshment cohorts separately 12 . The signicant spousal concordance in health and health behaviors can be the result of assortative mating, shared envi- ronment and resources, or causal spillover, as we have discussed previously. The next section will provide further evidence on the spillover channel specically. Table 2.3: Correlation Coecients of Health Status and Health Be- haviors Within Couples VARIABLES Wave 3 Wave 5 Wave 7 Have good self reported health 0.26 0.24 0.27 CESD Score 0.23 0.26 0.30 Smokes now 0.29 0.31 0.27 Intensity of smoking: # cigarettes/day 0.19 0.27 0.18 Intensity of drinking: # days/week drinks 0.55 0.55 0.55 Freq vigorous phys activ : # days/week 0.33 0.29 0.32 Any weekly contact w/ relative in person 0.31 0.31 0.40 Note: All correlation coecients are signicant at 1% level. 2.6 Estimation Results 2.6.1 Health Status Table 2.4 and 2.5 show results on the spousal spillover eects in health outcomes including self- reported health and depressive symptoms (CES-D). OLS and FE results are reported as well. 12 Results are available upon request. 21 Self-reported good health. The self-reported health is strongly and positively correlated between spouses: OLS and FE coecients before spousal self-reported good health are similar in magnitude for both men and women. Column (3) and (8) report the system GMM results when we only include one period lagged value of own health outcome, while in column (4) and (9) we allow for state dependence up to a 2-period lag. Although all coecients for spousal self-reported good health are signicantly positive for both husbands and wives, the specication tests column (3) and (4) imply violations of Assumption 4 and the overall exclusion assumption for male sample. On the other hand, the GMM results for English women indicate that if the husband has good health as apposed to fair or poor health, there is a 27 percentage points increase in the probability for the wife being in good health. A comparison between column (8) and (9), as well as between (3) and (4) also indicate that the lag dependence of self-reported good health seem to dier between men and women. While the estimated instantaneous spillover eects from spouses decline for both genders when we include more lagged values of own health outcomes, the extent of the decline is larger for men and the eect from lagged 2 period own self-reported health is only signicant for husbands. This nding suggests it might be important to consider dierent lag structures for men and women. It is also noteworthy that the estimates of own lagged self-reported good health decline in GMM estimations, compared with OLS results, to the extent that they are smaller than the instantaneous spousal eects. This leads to concern about the credibility of our GMM results especially the exclusion assumption such that the past spousal health only aects one's own health only through the current health of spouses. The dierence between the OLS and GMM estimation results on one's own lagged health is re-stated in reduced form estimation results where we include spousal lagged health instead of their current health. From column (6) we see that even we don't include spousal current health, the coecient before the own previous health barely changes. Comparisons 22 between column (3) and (5), as well between column (8) and (10) in addition imply that the current health of one's spouse matters but spousal past health does not, which is consistent with our Assumption 2. Clark and Etil e (2006) found that spousal past health status does not have signicant impact on their health behaviors (smoking). If the spousal past health does not have eect on ones behavior, it could lead to the case that one's current health status is not strongly aected by their spousal health status. In appendix, we also show as a robustness check the results for the self-reported health measured on a scale from 1 to 3, which are similar to what we nd for a binary outcome for good self-reported health. Depressive Symptoms. Powdthavee (2009) found that there is strong causal impact of happiness of one's partner on that of one's own. We nd that such spillover eects also exist in depressive symptoms. CES-D 8 score of one's spouse is signicantly and positively correlated with one's own CES-D 8 score for both men and women. Furthermore, the causal estimates for the instantaneous spillover eects are larger than the OLS or FE estimates, while the state dependence in depression is much smaller in GMM specications than in OLS. Arrelano-Bond AR(2) tests, Hansen J tests and Dierence in Hansen J tests prove the validity of the instruments used for women irrespective of the lag structure. Reduced form estimation results reveal in column (5) and (10) that spousal lagged depression does not signicantly aects one's current depression level, which is true for both genders. Comparing results between column (3) and (4) points out that mental health from 2 periods ago of men still have strong impact on their current mental health. For women, however, it is only the most recent depressive episodes that matter. For both genders, state dependence on one's own depression is smaller than the instantaneous spousal eect. The magnitudes of system GMM estimates for the spillover eects are larger for women than those for men, implying that women are aected more by their husband's depression than men are by their wives. 23 Table 2.4: Results on Good Self-reported Health (SRH) (1) (2) (3) (4) (5) Men OLS FE System GMM L.(own good SRH) 0.513*** -0.148*** 0.111*** 0.192*** 0.115*** (0.0174) (0.0203) (0.0277) (0.0353) (0.0286) L2.(own good SRH) 0.0943*** (0.0297) spouse's good SRH 0.140*** 0.145*** 0.287*** 0.169* (0.0135) (0.0188) (0.0926) (0.0940) L.(spouse's good SRH) -0.00803 (0.0222) Observations 6718 6718 6718 4838 6759 Number of couples 1881 1881 1881 1606 1886 AR(1) test p val 1.71e-57 1.70e-43 2.01e-57 AR(2) test p val 0.351 0.246 0.279 Hansen J test p val 0.000205 0.0653 0.0000192 Di in Hansen p val for levels 0.0477 0.279 0.0360 (6) (7) (8) (9) (10) Women OLS FE System GMM L.(own good SRH) 0.519*** -0.143*** 0.127*** 0.144*** 0.128*** (0.0174) (0.0196) (0.0255) (0.0331) (0.0262) L2.(own good SRH) 0.0301 (0.0281) spouse's good SRH 0.138*** 0.136*** 0.297*** 0.265*** (0.0126) (0.0174) (0.0760) (0.0774) L.(spouse's good SRH) -0.00753 (0.0189) Observations 6850 6850 6850 4916 6916 Number of couples 1944 1944 1944 1636 1956 AR(1) test p val 1.79e-59 3.10e-41 4.08e-59 AR(2) test p val 0.959 0.832 0.972 Hansen J test p val 0.0898 0.127 0.0243 Di in Hansen p val for levels 0.191 0.220 0.332 Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 24 Table 2.5: Results on Depressive Symptoms (CES-D 8 Score) (1) (2) (3) (4) (5) Men OLS FE System GMM L.(own CESD score) 0.438*** -0.174*** 0.0438* 0.0851** 0.0498** (0.0216) (0.0183) (0.0235) (0.0381) (0.0246) L2.(own CESD score) 0.0556* (0.0295) spouse's CESD score 0.137*** 0.146*** 0.372*** 0.358*** (0.0122) (0.0161) (0.0927) (0.0941) L.(spouse's CESD score) -0.00991 (0.0154) Observations 6641 6641 6641 4771 6680 Number of couples 1872 1872 1872 1597 1877 AR(1) test p val 8.98e-42 4.78e-32 3.46e-46 AR(2) test p val 0.818 0.624 0.577 Hansen J test p val 0.0338 0.0354 0.00109 Di in Hansen p val for levels 0.333 0.566 0.00878 (6) (7) (8) (9) (10) Women OLS FE System GMM L.(own CESD score) 0.457*** -0.143*** 0.111*** 0.142*** 0.118*** (0.0179) (0.0169) (0.0221) (0.0308) (0.0225) L2.(own CESD score) 0.0301 (0.0262) spouse's CESD score 0.204*** 0.201*** 0.529*** 0.454*** (0.0174) (0.0228) (0.107) (0.110) L.(spouse's CESD score) -0.00463 (0.0247) Observations 6773 6773 6773 4857 6829 Number of couples 1931 1931 1931 1626 1940 AR(1) test p val 4.97e-64 3.14e-45 1.21e-63 AR(2) test p val 0.343 0.325 0.121 Hansen J test p val 0.476 0.153 0.00961 Di in Hansen p val for levels 0.793 0.262 0.179 Note: Two-step corrected robust standard errors in the parenthesis. *** p< 0:01, **p< 0:05, * p< 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 25 2.6.2 Health-related Behaviors As health behaviors are closely related to health status, it is natural to ask whether there are also spillover eects in health behaviors among couples. Although we do not attempt to test explicitly whether the spousal spillovers in behaviors also contribute to the spousal concordance in health status, some hypotheses accounting for spousal health concordance suggest that shared environment within couples translates into shared health risks which benet or harm health, depending on the health behaviors of the spouses (Meyler, Stimpson and Peek, 2007). Therefore, estimating the spillover eects in behaviors within couples is not only interesting to its own end, but could also help further the understanding of spousal concordance in health. In this section, we examined 4 important health related behaviors of adults: smoking, drinking, exercising and social contacting. Smoking. Table 2.6 shows the results on smoking status, which is a binary outcome. Since smoking is not just a conscious behavior but a compulsive addiction, state dependence are found to relative high. Moreover, we nd evidence that this addiction can thrive further in case of spousal support. OLS and FE results in Table 2.6 indicate of positive correlation in smoking status between husband and wife. What is more, GMM estimations in Column (3) and (4) imply that there is signicant spousal spillover eect in smoking behaviors for English elderly men. In contrast, the in uence of husband's smoking on their wife is smaller and not statistically signicant as shown in column (8) and (9): having a smoking wife tends to increase the likelihood for the husband to be a smoker by about 14% while it is only around 8% for wife if the husband smokes. A comparison between these columns also suggest the smoking is persistent for both genders and allowing for further path dependence might be important, given the signicance of 2-period lagged smoking status and the rejection of specication tests in columns with fewer lags. 26 Table 2.6: Results on Smoking Status (= 1 if smokes now) (1) (2) (3) (4) (5) Men OLS FE System GMM L.(own smoking status) 0.761*** 0.121*** 0.368*** 0.431*** 0.372*** (0.0183) (0.0394) (0.0635) (0.0596) (0.0604) L2.(own smoking status) 0.118*** (0.0413) spouse's smoking status 0.0675*** 0.102** 0.188*** 0.140* (0.0122) (0.0440) (0.0715) (0.0757) L.(spouse's smoking status) 0.0224 (0.0337) Observations 6981 6981 6981 5048 6969 Number of couples 1926 1926 1926 1665 1918 AR(1) test p val 5.42e-11 1.02e-08 2.76e-11 AR(2) test p val 0.0779 0.110 0.101 Hansen J test p val 0.0900 0.733 0.0437 Di in Hansen p val for levels 0.659 0.989 0.613 (6) (7) (8) (9) (10) Women OLS FE System GMM L.(own smoking status) 0.826*** 0.0513 0.308*** 0.476*** 0.304*** (0.0159) (0.0452) (0.0640) (0.0844) (0.0659) L2.(own smoking status) 0.136*** (0.0385) spouse's smoking status 0.0460*** 0.0839*** 0.0340 0.0871 (0.0107) (0.0272) (0.0614) (0.0619) L.(spouse's smoking status) 0.0386* (0.0201) Observations 7174 7174 7174 5147 7138 Number of couples 2019 2019 2019 1712 1997 AR(1) test p val 6.16e-09 0.0000105 2.86e-08 AR(2) test p val 0.396 0.786 0.472 Hansen J test p val 0.00648 0.163 0.0106 Di in Hansen p val for levels 0.600 0.976 0.624 Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 27 Table 2.7: Results on Smoking Intensity (# cigarettes/day) (1) (2) (3) (4) (5) Men OLS FE System GMM L.(# cigarettes/day smokes) 0.801*** 0.0812 0.246*** 0.297*** 0.226*** (0.0218) (0.0527) (0.0790) (0.107) (0.0846) L2.(# cigarettes/day smokes) 0.0893* (0.0458) # cigarettes/day spouse smokes 0.0660*** 0.0792*** 0.0803 0.0461 (0.0187) (0.0275) (0.0607) (0.0478) L.(# cigarettes/day spouse smokes) 0.0169 (0.0199) Observations 6734 6734 6734 4839 6694 Number of couples 1919 1919 1919 1632 1900 AR(1) test p val 0.0000364 0.00353 0.000151 AR(2) test p val 0.310 0.280 0.527 Hansen J test p val 0.131 0.224 0.0670 Di in Hansen p val for levels 0.768 0.594 0.722 (6) (7) (8) (9) (10) Women OLS FE System GMM L.(# cigarettes/day smokes) 0.732*** 0.167** 0.320*** 0.391*** 0.309*** (0.0325) (0.0745) (0.0870) (0.0899) (0.0902) L2.(# cigarettes/day smokes) 0.137*** (0.0404) # cigarettes/day spouse smokes 0.0679*** 0.0863** 0.154* 0.188** (0.0188) (0.0346) (0.0924) (0.0872) L.(# cigarettes/day spouse smokes) -0.00783 (0.0227) Observations 6528 6528 6528 4701 6524 Number of couples 1824 1824 1824 1569 1819 AR(1) test p val 0.000208 0.00361 0.000317 AR(2) test p val 0.0726 0.478 0.0650 Hansen J test p val 0.316 0.513 0.234 Di in Hansen p val for levels 0.400 0.770 0.384 Note: Two-step corrected robust standard errors in the parenthesis. ***p< 0:01, **p< 0:05, *p< 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 28 Table 2.8: Results on Drinking Frequency (# days/week drinks) (1) (2) (3) (4) (5) Men OLS FE System GMM L.(# days/week drinks) 0.660*** -0.199*** 0.0943*** 0.171*** 0.135*** (0.0175) (0.0239) (0.0336) (0.0426) (0.0355) L2.(# days/week drinks) 0.127*** (0.0398) # days/week spouse drinks 0.217*** 0.203*** 0.484*** 0.468*** (0.0161) (0.0316) (0.103) (0.100) L.(# days/week spouse drinks) 0.0567* (0.0342) Observations 4023 4023 4023 2678 4020 Number of couples 1448 1448 1448 1185 1457 AR(1) test p val 2.06e-36 8.08e-22 8.05e-33 AR(2) test p val 0.612 0.179 0.377 Hansen J test p val 0.136 0.621 0.00334 Di in Hansen p val for levels 0.398 0.786 0.176 (6) (7) (8) (9) (10) Women OLS FE System GMM L.(# days/week drinks) 0.671*** -0.142*** 0.153*** 0.242*** 0.206*** (0.0168) (0.0268) (0.0348) (0.0457) (0.0370) L2.(# days/week drinks) 0.0937*** (0.0313) # days/week spouse drinks 0.191*** 0.198*** 0.563*** 0.582*** (0.0144) (0.0280) (0.101) (0.0923) L.(# days/week spouse drinks) 0.0549* (0.0285) Observations 4083 4083 4083 2697 4090 Number of couples 1483 1483 1483 1203 1496 AR(1) test p val 1.80e-29 8.24e-17 3.78e-34 AR(2) test p val 0.476 0.555 0.0970 Hansen J test p val 0.840 0.982 0.00508 Di in Hansen p val for levels 0.745 0.857 0.228 Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 29 Table 2.9: Results on Frequency of Exercise (# days/week participating in vigorous physical activities, VPA) (1) (2) (3) (4) (5) Men OLS FE System GMM L.(# d/w w/ VPA) 0.441*** -0.105*** 0.126*** 0.217*** 0.130*** (0.0150) (0.0158) (0.0218) (0.0245) (0.0226) L2.(# d/w w/ VPA) 0.120*** (0.0200) # d/w spouse w/ VPA 0.222*** 0.271*** 0.244** 0.261*** (0.0146) (0.0186) (0.0959) (0.101) L.(# d/w spouse w/ VPA) 0.0263 (0.0196) Observations 7020 7020 7020 5077 7013 Number of couples 1934 1934 1934 1672 1931 AR(1) test p val 6.33e-86 7.02e-70 4.39e-87 AR(2) test p val 0.0197 0.184 0.0542 Hansen J test p val 0.000303 0.346 0.000636 Di in Hansen p val for levels 0.387 0.834 0.498 (6) (7) (8) (9) (10) Women OLS FE System GMM L.(# d/w w/ VPA) 0.436*** -0.119*** 0.150*** 0.150*** 0.162*** (0.0164) (0.0168) (0.0210) (0.0270) (0.0210) L2.(# d/w w/ VPA) 0.0216 (0.0225) # d/w spouse w/ VPA 0.181*** 0.234*** 0.131* 0.133* (0.0131) (0.0160) (0.0722) (0.0788) L.(# d/w spouse w/ VPA) 0.00335 (0.0159) Observations 7212 7212 7212 5182 7193 Number of couples 2022 2022 2022 1717 2019 AR(1) test p val 1.55e-73 2.82e-51 2.61e-77 AR(2) test p val 0.183 0.600 0.234 Hansen J test p val 0.376 0.202 0.422 Di in Hansen p val for levels 0.212 0.311 0.259 Note: Two-step corrected robust standard errors in the parenthesis. *** p< 0:01, **p< 0:05, * p< 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 30 Table 2.10: Results on Social Contact (= 1 if any weekly contact with relatives in person) (1) (2) (3) (4) (5) Men OLS FE System GMM L.(wkly contact w/ relativ) 0.411*** -0.137*** 0.147*** 0.175*** 0.154*** (0.0185) (0.0226) (0.0289) (0.0367) (0.0301) L2.(wkly contact w/ relativ) 0.0427 (0.0360) spouse's wkly contact w/ relativ 0.180*** 0.115*** 0.191** 0.205** (0.0155) (0.0236) (0.0885) (0.0902) L.(spouse's wkly contact w/ relativ) 0.0243 (0.0250) Observations 4371 4371 4371 2991 4408 Number of couples 1456 1456 1456 1131 1472 AR(1) test p val 2.42e-50 1.15e-31 4.82e-51 AR(2) test p val 0.325 0.908 0.724 Hansen J test p val 0.216 0.250 0.0282 Di in Hansen p val for levels 0.978 0.925 0.888 (6) (7) (8) (9) (10) Women OLS FE System GMM L.(wkly contact w/ relativ) 0.464*** -0.137*** 0.169*** 0.247*** 0.161*** (0.0177) (0.0226) (0.0296) (0.0325) (0.0294) L2.(wkly contact w/ relativ) 0.139*** (0.0312) spouse's wkly contact w/ relativ 0.183*** 0.117*** 0.212** 0.282** (0.0165) (0.0232) (0.108) (0.126) L.(spouse's wkly contact w/ relativ) 0.0309 (0.0258) Observations 4500 4500 4500 3091 4535 Number of couples 1508 1508 1508 1188 1528 AR(1) test p val 5.92e-57 1.34e-39 2.16e-55 AR(2) test p val 0.00188 0.791 0.0975 Hansen J test p val 0.0133 0.478 0.00678 Di in Hansen p val for levels 0.518 0.813 0.696 Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 31 Table 2.7 reports results from estimations on the intensity of smoking, measured by number of cigarettes per day. On the intensive margin, there exist positive spillover eects in smoking for both men and women, although such eects from wives are not signicant. Neither of the estimated coecients for lagged spousal spillover eects is signicant in the reduced form GMM estimation, and the validity of instruments are rejected in column (5) and (10). Drinking. Table 2.8 shows that there is persistent inter-spousal concordance in the number of days they drink in a week as demonstrated by OLS and FE results for both men and women. This is intuitive since couples share their meals and drinks together. From causal specications, both men and women have strong impact on their partners and this eect is present even in the reduced form results, however, the specication tests are rejected in column (5) and (10). If one's spouse drinks one more day a week, he/she will tend to drink about 0.47/0.58 more days if we allow for further path dependence. Similar to what we have found for health status measures, the instantaneous spillover eect is also larger than the eect from path dependence. Exercise. We also nd strong causal spillover eects in physical exercise for both genders. In Table 2.9, we see that the magnitudes of the causal estimates are similar to those from OLS or FE estimations. Inclusion of more lags might be appropriate for studying exercise, based on the specication tests in column (3) and (8). One more day of vigorous activities of one's spouse in a week leads to 0.26 more days for husbands and 0.13 more days for wives, based on column (4) and (8). In line with ndings for drinking, the instantaneous eects of wives' participation in physical exercise are larger than husbands' own previous habit, though they are not signicant in the reduced form estimations. The same conclusion can be drawn if we examine the spousal spillover eects on whether the respondent participates in vigorous activities at least once a week, as shown in Table A2 in appendix. 32 Social Contact. Table 2.10 demonstrates that for weekly contact with relatives in person, both spouses have strong and positive in uence on each other. The spillover eect of wives on husbands is larger and more signicant. AR(2) tests and Hansen J tests suggest that it might be necessary to allow for further lag dependence of women, though GMM estimates for instantaneous spousal eects are similar across specications. Having a spouse contacting relatives weekly in person implies an increase in the probability of the respondent doing so by 20 percentage points for men and 21 to 28 percentage points for women. 2.7 Robustness Checks In this section, we discuss the potential sources of bias of GMM estimates and perform robustness checks. 2.7.1 Dierence GMM There are concerns about the validity of the additional moment conditions in system GMM. How- ever, most of the Dierence in Hansen tests results are consistent with exclusion restrictions of the rst dierences as instruments in the level equations. In addition, we show that most of our ndings are robust to the methods applied and assumptions imposed. Table 2.11 to 2.16 report the results when we relax the stationarity assumption and utilize moment conditions only for rst dierenced equations. We nd signicant estimates for spousal spillover eects in self-reported health, depres- sive symptoms and exercise for both genders, which are in line with system GMM results. There are also statistically signicant spillover eects from wives' smoking intensity and social contact on their husbands, based on dierence GMM results. However, the estimated spillover eects in smoking and drinking behaviors are insignicant in dierence GMM estimations. 33 Table 2.11: Dierence GMM Results on Good Self-reported Health (SRH) (1) (2) (3) (4) (5) (6) Men Women L.(own good SRH) 0.0486 0.0679 0.0423 0.0941*** 0.0771 0.0896*** (0.0315) (0.0519) (0.0320) (0.0301) (0.0471) (0.0297) L2.(own good SRH) 0.0350 0.00212 (0.0334) (0.0311) spouse good SRH 0.219* 0.186 0.350*** 0.377*** (0.116) (0.125) (0.113) (0.109) L.(spouse good SRH) -0.0190 -0.0252 (0.0243) (0.0231) Observations 6718 4838 6759 6850 4916 6916 Number of couples 1881 1606 1886 1944 1636 1956 AR(1) test p val 1.14e-45 1.26e-25 7.34e-46 2.32e-45 1.45e-25 1.38e-49 AR(2) test p val 0.752 0.327 0.783 0.620 0.828 0.501 Hansen J test p val 0.00836 0.0103 0.00169 0.0342 0.0327 0.0135 Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 34 Table 2.12: Dierence GMM Results on Depressive Symptoms (CES-D 8 Score) (1) (2) (3) (4) (5) (6) Men Women L.(own CESD score) 0.00782 0.00816 0.0151 0.0947*** 0.111*** 0.0777*** (0.0279) (0.0548) (0.0290) (0.0266) (0.0429) (0.0270) L2.(own CESD score) 0.0202 0.0167 (0.0361) (0.0299) spouse CESD score 0.353*** 0.376*** 0.538*** 0.510*** (0.116) (0.124) (0.123) (0.124) L.(spouse CESD score) -0.0252 -0.0257 (0.0187) (0.0293) Observations 6641 4771 6680 6773 4857 6829 Number of couples 1872 1597 1877 1931 1626 1940 AR(1) test p val 5.46e-30 9.52e-19 5.52e-35 5.43e-49 2.27e-31 2.92e-41 AR(2) test p val 0.445 0.570 0.313 0.254 0.391 0.0436 Hansen J test p val 0.00926 0.00213 0.00391 0.109 0.0237 0.00471 Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 35 Table 2.13: Dierence GMM Results on Smoking Status (= 1 if smokes now) (1) (2) (3) (4) (5) (6) Men Women L.(own smoking status) 0.358*** 0.349*** 0.363*** 0.199** 0.195 0.186*** (0.0833) (0.0925) (0.0785) (0.0783) (0.129) (0.0702) L2.(own smoking status) 0.0902** 0.0456 (0.0456) (0.0442) spouse's smoking status 0.156* 0.0440 0.00611 0.0373 (0.0901) (0.0954) (0.0515) (0.0622) L.(spouse's smoking status) 0.0298 0.0304 (0.0375) (0.0253) Observations 6981 5048 6969 7174 5147 7138 Number of couples 1926 1665 1918 2019 1712 1997 AR(1) test p val 2.14e-08 0.0000136 8.25e-09 0.00000273 0.00332 0.00000200 AR(2) test p val 0.0828 0.136 0.102 0.716 0.897 0.780 Hansen J test p val 0.0259 0.0935 0.0158 0.0138 0.0355 0.0252 Note: Two-step corrected robust standard errors in the parenthesis. *** p< 0:01, ** p< 0:05, * p< 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 36 Table 2.14: Dierence GMM Results on Smoking Intensity (# cigarettes/day) (1) (2) (3) (4) (5) (6) Men Women L.(# cigarettes/day smokes) 0.215** 0.159 0.166 0.286*** 0.308** 0.253*** (0.0903) (0.121) (0.103) (0.0912) (0.124) (0.0932) L2.(# cigarettes/day smoked) 0.0441 0.119** (0.0448) (0.0472) # cigarettes/day spouse smokes 0.0192 0.0200 0.0627 0.106* (0.0449) (0.0358) (0.0796) (0.0617) L.(# cigarettes/day spouse smokes) 0.000708 -0.00139 (0.0241) (0.0231) Observations 6734 4839 6694 6528 4701 6524 Number of couples 1919 1632 1900 1824 1569 1819 AR(1) test p val 0.000226 0.00834 0.00222 0.000643 0.0151 0.000879 AR(2) test p val 0.401 0.177 0.665 0.0558 0.534 0.0553 Hansen J test p val 0.0351 0.106 0.0190 0.176 0.223 0.170 Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 37 Table 2.15: Dierence GMM Results on Drinking Frequency (# days/week drinks) (1) (2) (3) (4) (5) (6) Men Women L.(# days/week drinks) 0.0336 0.0639 0.0553 0.129*** 0.212*** 0.179*** (0.0421) (0.0718) (0.0438) (0.0463) (0.0780) (0.0504) L2.(# days/week drinks) 0.0567 0.0628* (0.0469) (0.0379) # days/week spouse drinks 0.255* 0.335** 0.294* 0.491*** (0.153) (0.146) (0.150) (0.158) L.(# days/week spouse drinks) 0.0499 -0.0183 (0.0395) (0.0332) Observations 4023 2678 4020 4083 2697 4090 Number of couples 1448 1185 1457 1483 1203 1496 AR(1) test p val 7.07e-24 1.64e-12 3.05e-22 6.55e-21 1.76e-11 1.26e-17 AR(2) test p val 0.864 0.432 0.894 0.785 0.410 0.456 Hansen J test p val 0.146 0.208 0.0723 0.646 0.870 0.0772 Note: Two-step corrected robust standard errors in the parenthesis. *** p< 0:01, ** p< 0:05, * p< 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 38 Table 2.16: Dierence GMM Results on Frequency of Exercise (# days/week participating in vigorous physical activities, VPA) (1) (2) (3) (4) (5) (6) Men Women L.(# d/w w/ VPA) 0.0961*** 0.193*** 0.107*** 0.146*** 0.129*** 0.155*** (0.0245) (0.0336) (0.0252) (0.0230) (0.0329) (0.0234) L2.(# d/w w/ VPA) 0.111*** 0.00999 (0.0224) (0.0231) # d/w spouse w/ VPA 0.297** 0.283** 0.132 0.135 (0.116) (0.133) (0.0857) (0.0966) L.(# d/w spouse w/ VPA) 0.0260 0.00953 (0.0222) (0.0184) Observations 7020 5077 7013 7212 5182 7193 Number of couples 1934 1672 1931 2022 1717 2019 AR(1) test p val 1.83e-67 1.56e-55 4.07e-69 5.99e-64 1.29e-40 1.00e-68 AR(2) test p val 0.0628 0.124 0.150 0.279 0.562 0.357 Hansen J test p val 0.00000339 0.0500 0.00000413 0.346 0.0935 0.358 Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 39 Table 2.17: Dierence GMM Results on Social Contact (= 1 if any weekly contact with relatives in person) (1) (2) (3) (4) (5) (6) Men Women L.(wkly contact w/ relativ) 0.146*** 0.170*** 0.166*** 0.151*** 0.235*** 0.136*** (0.0352) (0.0523) (0.0372) (0.0344) (0.0497) (0.0343) L2.(wkly contact w relativ) 0.0342 0.135*** (0.0383) (0.0335) spouse's wkly contact w/ relativ 0.175* 0.216** 0.207* 0.285** (0.0939) (0.102) (0.114) (0.144) L.(spouse's wkly contact w/ relativ) 0.0128 0.0202 (0.0307) (0.0295) Observations 4371 2991 4408 4500 3091 4535 Number of couples 1456 1131 1472 1508 1188 1528 AR(1) test p val 1.10e-41 1.04e-24 1.63e-41 3.39e-47 9.28e-25 5.46e-45 AR(2) test p val 0.363 0.858 0.641 0.00335 0.939 0.172 Hansen J test p val 0.00851 0.0132 0.000480 0.000512 0.156 0.0000957 Note: Two-step corrected robust standard errors in the parenthesis. *** p< 0:01, ** p< 0:05, * p< 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 40 2.7.2 Path Dependence and Spousal In uence Comparing OLS, FE and GMM results, we nd that the causal estimates of spousal spillover eects are usually larger than one's own path dependence, except for smoking behaviors. Concerns arise as to whether the GMM estimates of spousal spillover eects are upward biased while those of state dependence are downward biased in the structural model, because of wrong lag structures. We test the robustness of system GMM estimates of state dependence by estimating the reduced form equations and allowing for more own lags. The self 1-period lag coecients are very similar in the reduced form results, while the lagged spousal spillover eects are rarely signicant. With inclusion of further self-lags, the estimated state dependence generally increases, while the instantaneous spousal spillover eect estimate decreases for some outcomes, and increases for others. However, the extent of changes in the coecients are limited such that the results still imply that path dependence is less important than instantaneous spousal in uence. This might be true for our subjective measures of health status which are based on self reports: self reported good health and CES-D 8 index, for which state dependence is relatively low. We also examined the spousal spillover eects in chronic conditions, measured by the number of severe chronic conditions ever diagnosed by doctors 13 , where we show in Table A3 in appendix that instantaneous spousal in uence is insignicant and much smaller than state dependence of one's own health 14 . For health behaviors, except for smoking which is proven to be highly addictive, social in uence and peer eects in exercising, drinking and contacting relatives could possibly be more important than one's own past behaviors. 13 The severe diseases are diabetes, cancer, chronic lung disease, heart attack and stroke. 14 Here the spousal in uence is not necessarily a spillover of health status, but is likely to be a spillover in diagnosis. 41 2.7.3 Sample Selection As the estimation equation requires at least 3 consecutive (when m = 1) waves of observations for each respondent, one potential source of bias for the estimated spousal spillover eects might be sample selection or attrition in the panel data, if attrition is correlated with endogenous regressors. By taking rst dierences, we can remove all selection bias which is related to the time-invariant heterogeneity. However, there could remain selection bias on unobservables or shocks that are time- varying. If couples with stabler relationships have stronger in uence on each other, and more likely to remain in more waves, our GMM estimates might serve as upper bounds of the true spillover eects. Couples might separate due to growing care burden in the event of major illness to one spouse or even death, which might cause our causal estimates to be upward biased as well. On the other hand, if couples that are both healthier are more likely to move and hence attrite, our health spillover estimates might be downward biased. Therefore, a priori, it is not clear what the direction of the sample selection bias is if any. In ELSA, we nd that the cross-wave divorce rate is very low and the main sample attrition is from the failure of follow-up. Therefore, we argue it is unlikely that our analysis is biased from selective attrition due to marital dissolution. 2.8 Conclusion This chapter studies the widely documented spousal concordance in health, albeit in a causal way. We utilize data from the English Longitudinal Study of Aging and estimate in a dynamic panel model the spousal spillover regarding health status and health related behaviors among elderly couples in England. To address endogeneity biases from assortative mating and shared environment, we apply the system GMM to an dynamic model of health with individual xed eects. GMM estimations imply signicant and positive eects on health between elderly couples, with women's 42 mental health being aected more by their spouses. We also nd strong causal eects of spousal smoking, drinking and physical activity, as well as social contact, which could be an important channel leading to the inter-spousal concordance in health outcomes. While wives are a bigger in uence when it comes to their husband's smoking likelihoods, men aect more than women as to how much they smoke, among the English older adults. However, the gender dierences in spillover eects in both health and health behaviors are not substantial. We explored both reduced form and structural estimations, as well as dierent lag structures. We nd that allowing for state dependence with more distant lags is important for some health measures but not for the others, implying a heterogeneity in the persistence of health status and behaviors. The comparison between structural and reduced form estimations suggests the impor- tance of instantaneous spillover eects, while we fail to nd evidence for the lagged spousal eects in most of the health measures and behaviors considered. We nd the path dependence in health and health behaviors (except smoking) is relatively low, in contrast with the correlation coecients and the instantaneous spousal spillover eects. However, the state dependence also vary across dierent dimensions of health status and behaviors. For health behaviors that are often conducted with peers, and mental health, the spillover eects are larger than state dependence. Sensitivity analysis shows that our ndings on the spousal spillover eects in self-reported health, depressive symptoms, and exercise for both genders, as well as social contact for men are robust to less restrictive methods (dierence GMM). This chapter has not explicitly addressed the bias from sample selection or attrition by modeling its relation to health and marital status. A more comprehensive framework on the mechanisms underlying the spillover eects, which also incorporates the heterogeneity in such eects and the sample selection process, is warranted for future research. 43 The analysis of this chapter has important policy implications. Understanding the extent to which couples causally aect each other's health is relevant for comprehensive cost-benet calcu- lations of health policies, since interventions towards improving individual heath have potentially signicant spillovers along social ties. Health policies targeted at couple level might be more e- cient to reduce health disparities in certain health dimensions and behaviors (e.g., mental health, smoking, drinking and exercise), which involves strong social or spousal in uence. Furthermore, with aging population and declining familial support, individuals have to increasingly rely on their spouses in later life, especially in the face of deteriorating health. As the spillovers from spouses accumulate and the burden of care grows over time, targeting elderly couples, which have been shown to manifest signicant spousal concordance in health, might be especially important for better delivering health-care services. 44 Chapter 3 Does Children's Education Matter for Parents' Health and Cognition in Old Age? Evidence from China 3.1 Introduction Social science literature has consistently found positive and robust relationships between education and health (Baker et al., 2011). Compared with income, race, social rank and other individual so- cioeconomic characteristics, education has been shown to be one of the most important predictors for mortality and health (Cutler and Lleras-Muney, 2010). Over time, education plays an increas- ingly important role in explaining independently disparities in health, adding fuel to the interest in education as a policy instrument for enhancing population health (Montez and Friedman, 2015). Studies have also pointed out that the eects of education on health spill over to later generations. Grossman (2006) in a detailed review shows that individuals' schooling is also the most important correlate of the health of their children. A natural question thus arises: is the interaction between parents and children in terms of education and health a downward stream from parents to children? Or, other than individuals' own education, does their children's education also matter for their health and longevity, especially when they are old and dependent on familial support? Can policies aimed at improving educational 45 attainment of one generation benet not only the present and future (generations) but also the past (generations)? However, only a few recent studies have tried to test such hypothesis empirically (e.g., Lee, 2017; Lee et al., 2017; Yang, Martikainen and Silventoinen, 2016; Torssander, 2013, 2014; Lundborg and Majlesi, 2018; Zimmer, Hermalin and Lin, 2002; Zimmer et al., 2007; Kravdal, 2008; Friedman and Mare, 2014; Yahirun, Sheehan and Hayward, 2016, 2017; De Neve and Harling, 2017; De Neve and Fink, 2018). Whether there is indeed an upward spillover eect from children' educa- tion to parental health is still an open question and the importance of such eect might be dierent in dierent contexts (De Neve and Kawachi, 2017). Understanding the relationship and the causal pathways between children's education and in- dividual old-age health might be especially important in developing countries where older adults are more reliant on children and informal support due to strong family liation and/or insucient old-age security network. It might help identify the elderly people that are more susceptible to poor health and help design policy instruments by utilizing the benet of intergenerational eect of socioeconomic status (Yahirun, Sheehan and Hayward, 2016). This chapter contributes to the small but expanding literature on upward spillover eects of children's education on parental health by providing a causal analysis on older adults in an ageing developing country (China). It extends the correlational studies of Lee (2017), Zimmer et al. (2007), Yang, Martikainen and Silventoinen (2016) and Zimmer, Hermalin and Lin (2002) who have also studied Chinese elderly, and De Neve and Fink (2018) and Lundborg and Majlesi (2018), which conducted causal analysis. I employ an instrumental variables (IV) estimation method similar to Du o (2001), Chou et al. (2010), Lundborg and Majlesi (2018), De Neve and Fink (2018) and Huang (2015) to identify the impact of adult children's schooling on health of older adults by exploiting the decentralized gradual implementation of compulsory education in China during the period from 1985 to 1991. The reforms in compulsory education took place in dierent years 46 across Chinese provinces and within each province increased dierentially the amount of formal education of dierent birth cohorts. Constructions of instruments for children's education are based on the number of years exposed to compulsory education of the highest educated child and its interaction with the regional average years of schooling before the compulsory schooling law enforcement (related to program intensity). I include in both stages birth year xed eects, along with province or community xed eects and province-specic trends to control for cohort and regional heterogeneity in schooling levels and the dierent paths of the educational expansion across provinces. Robustness checks were also conducted with regard to alternative instruments, further control of other social programs, the instrumentation of education of older adults themselves and multiple hypotheses testing. The second contribution of this chapter is that a wide array of health and cognition outcomes are studied in order to oer a comprehensive analysis of the health eect of children's education on older adults. These measures are shown to be important predictors of old-age disability and mortality. I nd that children's education is signicantly and positively correlated with health and cognition of Chinese elderly. Strong causal relations based on the IV estimations are found for cognitive abilities, expected survival, lung function and body weight. Third, this chapter adds to the work of Friedman and Mare (2014), De Neve and Fink (2018), Yahirun, Sheehan and Hayward (2017), Lee (2017), and Lundborg and Majlesi (2018), who shed light on the potential pathways through which children's education in uences health of parents in later life, by explicitly discussing and testing the following possible mechanisms and pathways: social in uence on health behaviors, social support, access to resources, labor supply and psycho- logical well-being. I nd evidence for improved social support that parents received in the form of net monetary transfer from children, and increased parental access to resources such as general eco- nomic resources (measured by household expenditure), clean fuels (gas and electricity as apposed to 47 solid fuel) and improved sanitation (private in-house toilets). Parents with better educated children are more satised with their lives and less likely to work in old age, implying that psychological well-being and leisure/labor supply of parents might also play a role in explaining the health eect of education of children. Last, this chapter also estimates a dynamic model for health and cognition of older adults, which controls for lagged baseline outcomes that are treated as endogenous. It oers evidence on the causal incremental eect of children's education on the short-run changes of health and cognition, which also re-inforces the correlation results found by Yahirun, Sheehan and Hayward (2017) and Zimmer et al. (2007). Using the two-year follow-up data, I show in the appendix that the baseline conclusion using a static model still holds for most of the health outcomes. Furthermore, Lewbel IV estimation of the dynamic model reveals that given the baseline level of cognitive abilities, adult children's education has incremental eects on the cognition, especially the episodic memory, of older adults. However, for other health measures of parents, children's education does not seem to have signicant incremental eects. The rest of the chapter is organized as follows. Section 3.2 reviews the related literature. Section 3.3 outlines the possible channels through which the health of older adults may be linked to the education among their children. Section 3.4 introduces the institutional context and compulsory education reform of China. Section 3.5 describes the data. Section 3.6 presents the baseline empirical models, the identication strategy and reports baseline model results for health of older adults and the underlying channels. Section 3.7 presents the dynamic model and the estimated results. Section 3.8 discusses the robustness of the results. Section 3.9 concludes and discusses the chapter. 48 3.2 Related Literature The analysis of this chapter builds upon several strands of economic research in human capital. The rst related line of literature is the study of the eect of education on health whose theoreti- cal foundation was provided by Grossman (1972, 1976) and Strauss and Thomas (2008). Empirical evidence mounted that linked increased education to lower mortality (e.g., Elo and Preston, 1996; Lleras-Muney, 2005), healthier life styles (e.g., Cutler and Lleras-Muney, 2010), and improved mor- bidity indicators including self reported health, mental health, chronic conditions, functional lim- itations and disabilities (e.g., Mackenbach, 2005, for European countries). Although it is still an on-going debate whether such correlation is causal, recent reviews by Grossman (2006), Baker et al. (2011), and Cutler and Lleras-Muney (2012) have shown that at least part of it is. The second strands of literature this chapter relates to are the studies in the intergenerational transmission of human capital, social mobility and eects of family background on child outcomes. Past research has documented the positive in uence of parental socioeconomic status on children's early life health, education and economic outcomes (e.g., Thomas, Strauss and Henriques, 1990, 1991; Strauss and Thomas, 1995; Case, Lubotsky and Paxson, 2002; Currie and Moretti, 2003; Currie, 2009; Oreopoulos, Page and Stevens, 2006; Lundborg, Nilsson and Rooth, 2014). There is also a literature in economics on eects of peers on health. Empirical evidence has shown that socioeconomic status, health and behaviors of social network members, such as spouses, friends and siblings, are important in shaping an individual's health and behaviors (e.g., Wilson, 2002; Brown, Hole and Roberts, 2014; Meyler, Stimpson and Peek, 2007; Monden et al., 2003; Fowler and Christakis, 2008b; Altonji, Cattan and Ware, 2017). Last but not least, this chapter belongs to the small strand of literature that studies the re- lationship between later life health of older adults and the education of their children. Recent 49 works have demonstrated that children's education level is important for their parents' physical functioning (Zimmer, Hermalin and Lin, 2002; Yahirun, Sheehan and Hayward, 2016, 2017), phys- iological dysregulation (Lee, 2017), depressive symptoms trajectories (Lee et al., 2017) and sur- vival (e.g., Zimmer et al., 2007; Kravdal, 2008; Torssander, 2013, 2014; Friedman and Mare, 2014; Yang, Martikainen and Silventoinen, 2016; Lundborg and Majlesi, 2018; De Neve and Harling, 2017; Yahirun, Sheehan and Hayward, 2017; De Neve and Fink, 2018). Zimmer et al. (2007) and Lee (2017) in addition showed the importance of the education of the children on mortality and bi- ological functioning of older adults when controlling for socioeconomic and demographic charac- teristics, and health at baseline. Furthermore, Zimmer et al. (2007) and Lee et al. (2017) nd that children's education seems to have stronger association than older adults' own education with mortality 1 and depressive symptoms 2 of older adults in Taiwan, suggesting dierential ef- fects of education in the progression of health problems. Yahirun, Sheehan and Hayward (2017) also made use of panel data and their results imply a stronger relation of children's education with long-term survival of parents than with the short-term change in functional limitation that may be more contingent on earlier life exposures. These studies cover both countries with devel- oped old-age security networks, e.g., Sweden (Torssander, 2013, 2014; Lundborg and Majlesi, 2018), regions with limited public support, e.g., Africa (De Neve and Harling, 2017; De Neve and Fink, 2018) and Mexico (Yahirun, Sheehan and Hayward, 2016, 2017), countries with less lial obligation, e.g., the United States (Friedman and Mare, 2014), societies with strong lial links, e.g., Taiwan (Zimmer, Hermalin and Lin, 2002; Zimmer et al., 2007), and societies with strong lial obligation but insucient public support network, e.g., mainland China (Yang, Martikainen and Silventoinen, 2016). 1 for those with severe diseases. 2 But the association decreases as older adults age. 50 Though research in this area is growing rapidly, most of the aforementioned results are corre- lational evidence that are more econometrically challenged. Upward bias arises from the reverse causality problem that healthier parents are better able to invest in children's schooling, and the omitted variable problem that unobserved genetic and/or environmental factors could lead to con- cordant levels of human capital of both parents and children. The use of sibling xed eects models as in Torssander (2013) can only partially control for the latter and not at all for the former. Lundborg and Majlesi (2018) and De Neve and Fink (2018) are the only exceptions that esti- mated the causal impact of children's educational attainment on parents' longevity by using expo- sure to educational reforms in Sweden and Tanzania as instruments for years of schooling of children respectively. Lundborg and Majlesi (2018) exploited changes in years of compulsory schooling from seven to nine that took eect at dierent points in time across municipalities in Sweden to construct an instrument for years of schooling of children in estimating the causal eect of children's educa- tion on parental survival, while controlling for state xed eects and municipality-specic cohort trends. While Lundborg and Majlesi (2018) fail to nd any causal eect of children's education on survival until 2013 of parents born between 1899 and 1941 on average, they do nd some hetero- geneity in the eects. Their results imply that the schooling of daughters aects survival of fathers and especially those from low socio-economic background. De Neve and Fink (2018), on the other hand, using data from a much poorer economy and less educated population, nd positive impacts of schooling of children born after 1945 but older than 15 years old on parental survival in 1988 or 2002 for both genders 3 . As a missing piece of puzzle, the study of the eect of children's SES or human capital on parents' health builds upon and enriches the literature in development of human capital and its 3 The authors pooled the data from two sources: United Republic of Tanzania Population Census 1988 and Population and Housing Census 2002. 51 spillover within networks, across generations and dimensions. Without taking into account the upward spillover of human capital from younger to older generations, the social benet of education and other programs that lead to human capital advancement will be underestimated and maybe even misleading. It is also especially important to empirically estimate such spillover eect and the various pathways through which it might work in dierent settings in order to understand when and how the spillover is eective. This chapter is therefore motivated to add to the limited empirical causal studies in this growing area of research with special focus on a developing and rapidly aging country, China. 3.3 Potential Mechanisms Children's education can relate to parents' health in old age in multiple ways, as depicted in the diagram in Figure 3.1. Parents who are healthier have more resources and thus are better able to give birth to healthier children and invest in their children's education. Children who are healthier in early life are also more likely to achieve higher levels of education in adulthood. It is also possible that genetic endowments or unobserved abilities give rise to both parents and ospring who are healthier and better educated. Parents might also invest more in the education of their children in exchange for old age support due to informal contract incentive (Nugent, 1985). On the other hand, children is part of the social network of parents, whose SES might have independent eect on parental health in old age (Berkman, Glass and Brissette, 2000). Children's education being an indispensable family resource could aect parental health in a number of behavioral processes: provision of social support, access to resources, social in uence which work through more proximate behavioral and psychological pathways. 52 Adult Children Parent Knowledge, Behaviors, Income, Job flexibility, Migration, etc. Health Social Influence Access to Resources Social Support Behaviors: Health Behaviors Behaviors: Labor Supply Psychological Factors Unobserved Genetic Endowment/Abilities Childhood Health, Investment Spouse, Friends, etc. Based on the Social Network Theory of Health Informal Contract Education Figure 3.1: Linking Parents' Health and Children's Education, based on the Social Network Theory of Health of Berkman, Glass and Brissette (2000) 3.3.1 Social Support Children's socioeconomic status can directly aect type, frequency, intensity, and extent of sup- port provided to parents. Children with higher educational attainment are better able to provide informational support to parents, including advice or information on health care services, because they have better health knowledge themselves (Torssander, 2013; Friedman and Mare, 2014). On the other hand, as more-educated children are more likely to have a full-time job with higher wages and to migrate, it might be costlier for them to have frequent and intense in person contacts with parents such that the emotional support they provide might be less (Torssander, 2013). In China, for example, people with more schooling are also more likely to migrate (Zhao, 1997) and the possibility of co-residence of parents with higher-educated children is lower than those with less educated children (Lei et al., 2015). While living away from children is shown to constrain Chinese 53 elderly receiving help on daily activities (Sun, 2002) and frequent visits (Lei et al., 2015), parents receive more net transfers from children who are living farther away and in whom they invested more in schooling (Lei et al., 2015). Children with more schooling might also have better health, more skills and greater exibility in their work schedule because of occupational prestige or more job control (Fletcher and Frisvold, 2009) which could allow them to provide more care to their ill parents when needed (Friedman and Mare, 2014). Therefore, children with dierent education levels tend to provide dierent types of support with varying quality (Friedman and Mare, 2014), making the net impact of children's education on parental health in old age less clear theoretically. 3.3.2 Social In uence One's social network plays a key role in forming health-related behaviors among its members. Ed- ucation of one's family members, including children, could have in uence on individual attitudes toward health habits, such as smoking and excessive drinking, when the family as a whole share their health knowledge and values (Torssander, 2013). Additionally, children's education could have indirect impact on their parents' health outcome through their own behaviors. If there is positive spillover eect of health behaviors from children to parents via peer pressure, social in u- ence or social comparison (Eisenberg, Golberstein and Whitlock, 2014; Card and Giuliano, 2013; Berkman, Glass and Brissette, 2000), children with better health-related behaviors that result from more education could also help their parents adopt concordant lifestyles and aect their health in old age. 3.3.3 Access to Resources Economic and environmental resources (risk factors) such as sanitation and clean fuel also play a signicant role in shaping human health (Thomas, Strauss and Henriques, 1990; Lavy et al., 1996; 54 Smith, 2007; Zhang et al., 2010). Social networks help distribute and transfer such resources to indi- viduals when private and public access is constrained. This pathway might be especially important in low income settings, like China, where old-age poverty measured by consumption is prevalent and household air pollution from solid fuel use along with poor sanitation are leading risk fac- tors for disease and remain major public health problems (Carlton et al., 2012; Zhang and Smith, 2007). Better educated children could improve parental health by nancing and boosting parental expenditure, helping them get access to clean fuels and installing sanitation facilities at home. 3.3.4 Psychological Well-being Other than micro-psychosocial and behavioral processes (Berkman, Glass and Brissette, 2000), chil- dren's education could aect parents' health through more \proximate" pathways which might also be results of the previously discussed mechanisms. Steptoe, Deaton and Stone (2015) have shown that psychological well-being and health are closely related to each other for elderly people. Lower physiological stress and higher self-ecacy might lead to better health outcomes and the adop- tion of and adherence to health-promoting behaviors among older adults. (Perkins et al., 2008; Schneiderman, Ironson and Siegel, 2005; de Leon et al., 1996; Duncan and McAuley, 1993). Chil- dren that are more auent in life because of higher educational attainment (Friedman and Mare, 2014; Torssander, 2013) are better able to support parents, protect parents from stressful events and raise parental psychological well-being by reinforcing social class identication (Lee, 2017). 3.3.5 Labor Supply A considerable share of older population in developing countries, especially in rural areas, continues to work in very old age until physically incapacitated (Cameron and Cobb-Clark, 2008; Cai et al., 2012). Although it is less clear in the literature about the eect of retirement on health of older 55 people (van der Heide et al., 2013), it is possible that in developing countries, where older people work out of necessity, retirement could have a benecial impact on their health in old age. Cai et al. (2012) found that rural elderly with more educated household members and pensions are less likely to be working in China. Children's socioeconomic status could aect parental health in old age through aecting the labor supply of parents which could be substituted by informal old-age support (Cameron and Cobb-Clark, 2008). 3.3.6 Existing Evidence of Dierent Pathways Several previous studies have provided suggestive evidence on the potential pathways through which children's education forms the health of parents at older age. De Neve and Fink (2018) show that children with higher education level are more likely to work, get married with better educated spouses, have a larger household size and have better access to basic utilities in Tanzania. Their results are based on information for households of children instead of their parents, and hence are only indirect evidence for the possible causal pathways. Lundborg and Majlesi (2018) nd that for the Swedish population on average, children's schooling has no signicant impact on parental economic resources. However, they suggest that the positive eects of female schooling on parental survival arise through improved health knowledge. Friedman and Mare (2014) nd that for U.S. elderly, the health eect of children's schooling is more pronounced for deaths that are linked to behavioral factors (chronic lower respiratory disease and lung cancer) which appears to be explained by smoking and exercise behaviors of parents. Yahirun, Sheehan and Hayward (2017) nd that for the Mexican elderly, even after controlling for children's nancial status and their nancial transfer to parents, education of children is still strongly correlated with longevity which implies that other mediating factors might be important, such as spillovers of health knowledge and social in uence on behavior and values. Lee (2017) investigates the behavioral and psychological pathways that 56 linked children's education to parental biological functioning in Taiwan and nds that parents with better educated children have higher level of psychological well-being and are more likely to engage in healthy behaviors. As parent-child interactions vary across societies, it is less clear about the relative importance of mechanisms through which children and their educational attainment are working to aect the health of parents in dierent settings. In poor countries where extended family structures and the pooling of resources as informal insurance are prevailing, mechanisms like nancial support and access to resources might be more salient than in developed countries where public support systems are more sucient and upward net transfer is less common (Torssander, 2013; Friedman and Mare, 2014). 3.4 Institutional Background 3.4.1 Demographic Transition, Aging and Old-Age Support in China As one of the fastest growing economies in the world, China has experienced unprecedented demo- graphic transition in the past thirty years. Not only is prominent the slowdown in its population growth, but also is its changing population structure, which makes China an acceleratingly aging society, especially in the rural areas (Cai et al., 2012; Cai and Du, 2015). Meanwhile, the inadequate pension system in China has led to persistent old-age poverty (Cai et al., 2012). The Chinese system of extended families, which is traditionally organized as a corporate unit by pooling economic resources and providing assistance to all family members (Zimmer, Hermalin and Lin, 2002), has been the major form of old-age security for Chinese elderly. Zhao et al. (2013a) showed that private transfers are critical for reducing consumption poverty rates of Chinese elderly households, especially in rural areas. Given the increase of life expectancy, the decrease of family size along with rapid urbanization and migration process in China, the duty 57 of old-age support has been placed increasingly on a smaller network of children who are better educated. Consequently, the socioeconomic status of children might play an increasingly important role in the later life of Chinese elderly. 3.4.2 Compulsory Education Reform in China There are huge gaps in educational opportunities and attainments across cohorts in China because of various education expansion endeavors after 1949. Less than 46% of the population born before 1960 completed at least lower secondary education in China, while the number is 74% and 82% for cohorts born during 1961{1970 and 1971{1980 respectively (Population Census Oce, 2012). Such changes could potentially make children's social in uence, provision of social support and resources more ecient and valuable for parental old-age health (Zimmer, Hermalin and Lin, 2002; Zimmer et al., 2007). This chapter specically exploits variations in compulsory education enforcement across Chi- nese provinces as a natural experiment. With the economic reform and opening up of China in the 1980s, Chinese government started its rst structural educational reform by announcing the Decision of The Central Committee of the Communist Party on Reform of China's Educational Structure (Central Committee of the Communist Party of China, 1985). The reform aimed at the decentralization of free basic education and the implementation of nine-year compulsory education, which is composed of six-year primary education and three-year junior secondary education 4 , in the entire country (Tsang, 1991; Ming, 1986). In addition to the the reform document, the Law on Nine-Year Compulsory Education was passed and took eect on July 1, 1986 5 . Table 3.1 lists 4 After 1981, some regions adopted the \pilot" system with ve-year primary education and four-year junior secondary education which is much less prevalent than the 6+3 system. 5 Other components include the development of vocational and technical education; and changing the student admission and allocation of higher education graduates while increasing \decision-making power" or autonomy of higher education institutions. 58 the timing of the provincial law enforcement of the nine-year compulsory education and rst birth cohorts aected. Based on the information in Table 3.1, Figure 3.2 illustrates the gradual and decentralized roll-out of the nine-year compulsory education graphically 6 . Anhui Beijing Fujian Gansu Guangdong Guangxi Guizhou Hainan Hebei Henan Heilongjiang Hubei Hunan Jilin Jiangsu Jiangxi Liaoning Inner Mongol Ningxia Qinghai Shandong Shanxi Shaanxi Shanghai Sichuan Taiwan Taiwan Taiwan Tianjin Xizang/Tibet Xinjiang Yunnan Zhejiang Chongqing 1985 1986 1987 1988 1989 1991 No data Figure 3.2: Roll-out of Compulsory Education 6 The reform document actually provided guidelines for implementation levels and timetables for three tiers of regions with dierent economic development levels on achieving the goal (Ming, 1986). Economically developed areas and cities (25% of the population) were set to make junior secondary education universal by 1990, followed by economically semi-developed townships and villages (50% of the population) by around 1995 after completing the spread of primary education, while the last category, which consisted of economically under-developed areas (25% of the population), were supposed to take various steps in universalizing primary school education as economic development permits with support from the central government (Ming, 1986). 59 Table 3.1: Implementation of Compulsory Education and Regional Schooling Level Province Law Eect First Eligible Urban Area Mean Rural Area Mean Year Birth Cohort Years of Schooling of Ineligible Cohorts Beijing 1986 1971 11.31 8.62 Tianjin 1987 1972 10.30 7.92 Hebei 1986 1971 9.92 8.25 Shanxi 1986 1971 10.14 8.08 Inner Mongolia 1989 1974 9.91 7.30 Liaoning 1986 1971 10.20 7.88 Jilin 1987 1972 10.21 7.38 Heilongjiang 1986 1971 9.94 7.58 Shanghai 1985 1970 9.93 8.03 Jiangsu 1987 1972 9.75 8.21 Zhejiang 1985 1970 8.63 7.25 Anhui 1987 1972 9.14 7.20 Fujian 1988 1973 8.68 7.01 Jiangxi 1986 1971 9.60 6.86 Shandong 1987 1972 9.80 7.84 Henan 1987 1972 9.98 8.22 Hubei 1987 1972 9.84 7.41 Hunan 1991 1976 10.33 8.03 Guangdong 1987 1972 9.41 7.77 Guangxi 1991 1976 9.90 7.72 Chongqing 1986 1971 9.35 6.87 Sichuan 1986 1971 9.38 7.01 Guizhou 1988 1973 8.58 4.86 Yunnan 1987 1972 8.84 5.53 Shaanxi 1987 1972 10.13 7.44 Gansu 1991 1976 10.19 6.11 Qinghai 1989 1974 9.68 4.52 Xinjiang 1988 1973 10.03 7.23 Note: Implementation information is retrieved from China's National People's Congress and Chinese Laws and Regulations Information Database. Average years of schooling of law ineligible cohorts is calculated by the author using 2000 population census data. 60 The rst municipality to implement the nine-year compulsory education was Shanghai which issued the provincial regulation on July 28, 1985, which became eective on September 1, 1985 even before the Law on Nine-year Compulsory Education was passed in 1986. It was also one of the most developed coastal areas in China then and now. Gansu is one of the poorest provinces in China; it ocially wrote compulsory education into provincial regulations that came into eect from the school year of 1991 7 . The majority of provinces and municipalities make the age of compulsory edu- cation to be six to fteen years old with exceptions in certain poor regions where the eligible age can be set from seven to sixteen 8 . Because of the decentralization of basic education, local governments instituted the compulsory education according to their own economic and educational resources which does not make the roll-out of the law implementation random. Table 3.1 also shows vast inter-province and rural-urban dierence in educational attainment among cohorts that were ineli- gible to compulsory schooling 9 (cohorts that were one to ve years older than the rst compulsory education eligible cohorts). Such dierences in economic developments could induce disparities in health of parents and education of children. I control for the province (or community) level time-invariant heterogeneity to address the endogenous program roll-out (Rosenzweig and Wolpin, 1986), and allow for province-specic birth cohort (of children) trends to capture any deviation from the national trend. 7 Gansu provincial government issued a pilot regulation in 1985 to implement nine-year compulsory schooling without success. Retrieved from http://fgk.chinalaw.gov.cn/article/dfgz/198510/19851090520158.shtml. Accessed on Jan 14, 2017. 8 The actual eligible age of compulsory education is not stated clearly in the Law on Nine-year Compulsory Education and most provincial regulations. However, the provision to the Law in 1992 explained that the age eligibility should be made by provincial governments and hence varied across provinces and regions within provinces. 9 In this chapter, I will use compulsory education and compulsory schooling interchangeably. Exposure to compul- sory education (law) and or compulsory schooling (law) will also be used as the same measure. I will use abbreviation CS for compulsory schooling and CSL for compulsory schooling law. 61 3.5 Data 3.5.1 CHARLS Data I use data from the China Health and Retirement Longitudinal Study (CHARLS) for the analysis of this chapter. CHARLS is a sister study of the US Health and Retirement Study (HRS) and is designed to facilitate multidisciplinary research on population aging in China. The CHARLS national baseline survey was conducted in 2011/2012 and is representative of the Chinese population of age 45 and above, covering 450 random communities from 150 randomly sampled counties in 28 Chinese provinces. Randomly selected age-eligible respondent and the spouse irrespective of age in one household were interviewed and the response rate was around 85% for the national baseline (Zhao et al., 2013b,a). A wide array of information was collected: demographics, family information, health and functioning, health care and insurance, work and retirement, income, consumption and assets. The follow-up surveys happen every 2 years. This chapter utilizes data from the national baseline and wave 2, which are publicly available for all modules. 3.5.2 Measure of Children's Education Children's education is measured in years of completed formal schooling which are based on the number of years \usually" taken to obtain a certain degree (Kemptner, J urges and Reinhold, 2011). Years of schooling is then imputed from the information on the highest completed educational level, ranging from 0 which means no completed formal schooling to 21 which equals the number of years to earn a doctoral degree 10 . In the entire CHARLS child sample before imposing any restriction, 60% completed lower secondary education (9 years) or above while only 28% of their parents have 10 In CHARLS, formal schooling consists of elementary school (6 years), middle school (3 years), high school (3 years), vocational school (3 years), 2 or 3 year college/associate degree (2.5 years), 4 year college/bachelor's degree (4 years), master's degree (3 years) and doctoral degree (3 years). 62 done so. Children with parents living in rural communities have less education than children whose parents live in urban areas. The dierence between the two groups is around 2.4 years. There is also gender gap in educational attainments of children. Female children have about 1 year less in schooling than male children. On average, there are about 3.5 living children for each responding household in the unre- stricted data. Since it is not clear which child's education is the most important for the parents (Zimmer et al., 2007) and it is not possible to include all children's information in instrumental variables estimations, I chose the years of schooling of the highest educated child as the measure of children's education, and hence extract his or her birth information to construct instruments 11 . When there was more than one highest educated children, I chose the birth information of the oldest one. Results from dierent measures of children's education are very similar. I report only results using the highest educated children's information following Zimmer et al. (2007) who found that the association between parents' mortality and highest educational level of all adult children is the strongest among others. Results using other measures of children's education are available upon request. 3.5.3 Birth Information of Children In order to get the exposure of each highest educated child to the compulsory schooling policy, I use their birth year as well as birthplace information. As it is not possible to know where the child lived when they were subject to compulsory education, their birthplace serves as a better proxy than their current Hukou location or their rst Hukou information which has more missing data. Of the children in the entire child sample, 94.2% were born in the same province of the current residence of their parents. Although the CHARLS baseline and wave 2 survey asked the birth province of 11 I experiment with dierent ways of constructing a single variable for children's education, including among oldest child, highest educated child/children and lowest educated child/children. 63 children if they were not born in the province where their parents live, the coding of the province in the baseline is dierent from the primary sampling unit and the province list in the appendix of the questionnaire and this information from wave 2 was not released to the public. Because of data limitations from both waves, I dropped children with unmatchable birth location because they were born in a province/county dierent from parents' current residing province or county which account for about 4.5% of the entire children sample. I also drop children with missing birth location information which are another 1.2% of the unrestricted children sample. 3.5.4 Control Variables Parental and Household Characteristics. I control for a set of variables of each parent that are found to be highly correlated with health and cognition in old age, including gender, a set of age dummies (5 year interval), years of schooling of their own, number of living children of the older adults, height at baseline and childhood health status (Smith et al., 2012). I refrain from adding more mediating factors that are endogenous such as household expenditure, labor supply and living arrangements. Instead I treat them as left hand side variables. All control variables are treated as exogenous including education of older adults. As a robustness check, I also report results from estimations where own education is instrumented in Appendix A. Table 3.2 summarizes the descriptive statistics of those controls. Child Characteristics. Besides years of schooling of the highest educated child, I control for child's gender, child's birth cohort xed eects and birth province xed eects. Regional Characteristics. As found by Smith, Tian and Zhao (2013) that the unmeasured community characteristics of living had a signicant eect on individual's health and cognition in China, residence information of the parent is controlled for in all estimations. Either province or community xed eects are included to control for any regional heterogeneity that are time 64 invariant. 12 When I include only province xed eects, I also control for the average years of schooling of compulsory education ineligible cohorts in the province (which is also the birthplace of the highest educated child), stratied by rural or urban area, as a measure for regional educational level before compulsory schooling enforcement. When community xed eects are included, the regional educational level before compulsory schooling enforcement is absorbed. A dummy variable that equals one if the individual lives in a rural area is controlled for in specications with province xed eects. Since all children in the analysis sample were born in the same province and county where the parents resided at the time of the survey, province and community xed eects absorb child's birth-province xed eects. As in Figure 3.3, there is potential heterogeneity in cohort trends of years of schooling across provinces, which could result from dierential economic and social development, I control for provincial GDP per capita, number of doctors and hospital beds per 10,000 population in the child's birth year as well as province-specic linear trends of child birth cohort in both rst and second stage. 3.5.5 Health and Cognition Outcomes I studied a wide array of health and cognition outcomes of Chinese older adults. Cognition. The rst group of outcomes consists of two cognition measures: mental intact- ness(MI) and episodic memory (EM). Mental intactness of the respondent is dened as the sum of correct answers to questions on the date, day of the week, season of the year (orientation), suc- cessive subtractions of 7 from 100 (numeracy) and picture re-drawing (Lei et al., 2012). Episodic memory of the respondent is measured as the average score of immediate word recall and delayed word recall following McArdle, Smith and Willis (2011). In CHARLS, respondents were asked to 12 Results from estimations with county xed eects are available upon request. 65 recall, both immediately and in 10 minutes, the same ten simple nouns that were read to them just once. Health. I studied both subjective health measures, objective physical health outcomes mea- sured by trained nurses and mental health (depressive symptoms). The rst subjective health measure is self-reported health based on a general health question in CHARLS with responses of 1 (very good), 2 (good), 3 (fair), 4 (poor) and 5 (very poor). I take the actual Likert scale answer such that smaller values indicate better general health. Another subjective heath measure is based on the expectation of survival to age 75 (80, 85, 90, 95, 100, 105, 110, 115) for respondents aged 64 or below (65-69, 70-74, 75-79, 80-84, 85-89, 90-94, 95-99, 100 and above) on a Likert scale from 1 (almost certain) to 5 (almost impossible). A corresponding binary outcome is dened that equals 1 if one has a low expectation (not very likely or impossible), or 0 otherwise. Objective physi- cal health indicators include lung function, which is measured as average of peak expiratory ow readings (L/min) and grip strength, which is the average of readings from the dominant hand mea- sured by dynamometer. Body mass index (BMI) and indicators of being underweight (BMI<18.5) and overweight (BMI25) are also studied as they re ect nutritional status and risk-factors for health problems like diabetes and cardiovascular disease (Kemptner, J urges and Reinhold, 2011). Depressive symptoms are measured by the Center for Epidemiological Studies-Depression (CES-D) 10 question index of depression which ranges from 0 to 30. 3.5.6 Potential Channels Furthermore, I explore the possible working pathways to parental health that children's education operates on. Social In uence. Eects of children's education on health behaviors of parents, including smoking and drinking, are studied to investigate whether children in uences parental health through 66 the social in uence pathway. Smoking and drinking status are considered as binary outcomes. Smoking (drinking) status equals 1 if the respondent smokes now (was drinking over last year). Frequency of smoking dened by number of cigarettes per day and frequency of drinking measured by number of days per week the respondent drank last year are also checked to provide further analysis in Appendix A. Social Support. Frequency of contact with children{or number of days in a week children visit or call, send text messages or mails/emails (co-residence with children accounts for 7 days per week){is the indicator of time transfer from children, which measures amount for emotional support from children. Dierent types of contacts, either in-person or not-in-person, are studied and measured by (1) number of days per week children visit and; (2) contact via phone, mails, text and emails 13 . Net annual amount of cash and in-kind transfer from children (in RMB) is the measure of monetary or nancial support provided by children. To rule out the in uence from outliers, I trim the net transfer to exclude the top 1.5% and bottom 1.5% 14 . Access to Resources. Economic resources, clean fuels and sanitation are considered to test the pathway of access to resources. Following Smith et al. (2012), I use per capita household expenditure as an indicator of long-run economic resources as it has less measurement error than current income. Expenditure includes both food and non-food consumption in the past year. Parental access to in-house ush-able toilets is a binary outcome variable, which measures the level of sanitation. In addition, whether parents used non-solid fuel (mainly gas and electricity) for cooking is studied as another mediating factor that is important for population health in low-income settings (World Health Organization, 2017). 13 Results for dierent types of time transfer are shown in Appendix A 14 The results are similar if I use other trimming criteria, e.g.,top and bottom 1%, 2% or 5% 67 Labor Supply and Psychological Well-being. Furthermore, I explore other proximate pathways through which children's education could shape parental health, such as labor supply and psychological well-being. Labor supply, which could be a result of social support and economic resources provided by children, is a binary variable which equals 1 if the parent is currently working which includes farming and unpaid family work. For the psychological well-being of parents, I use a 5-point scale life satisfaction measure with 1 being most satised and 5 being least satised and examine whether parents with better educated children enjoy life more. 3.5.7 Sample and Descriptive Statistics I restrict the sample to parents who are older than 40 at CHARLS baseline with highest educated child born between 1956 and 1991 such that the children under consideration are at least 20 years old in 2011 and more likely to have completed education 15 , while not too old to be without a living parent. Dierent age and birth cohort restriction criteria have barely changed the main conclusion and results using dierent samples are available upon request. Wave 2 data on children's information are used to complement the baseline analysis because of data limitations 16 . By doing so, I keep parents or parental households that were in both waves 17 . Table 3.2 shows the summary statistics of health, cognition outcomes, pathway variables of parents, parental demographic variables, years of schooling of highest educated adult child and other child-level, household-level or regional control variables. For both the baseline and the follow- up sample, 52% of whom are female. More than 83% of them are married and 64% of them live 15 In practice, I could drop children who are still in school but this does not change our results. 16 Because of the design of the baseline survey, information of co-residing children and non-coresiding children were asked in dierent modules. After careful scrutiny, basic information of many co-residing children were also asked in the module where they should not have been, which created duplications of children and a lot of missing data for other co-residing children in that module. As a result, it is impossible to merge the information or append the two groups to children to create a single child set accurately. In contrast, in wave 2, both types of children were asked in the same family information module and hence, this chapter makes use of wave 2 child information instead of the baseline. 17 Sample selection problem will be discussed in Appendix B. 68 in rural areas. The average age of parents is about 58 at baseline and 60 in the follow-up survey. Sample sizes vary for dierent outcomes of interest. It is worth noting that on average, more than 30% of the parents are overweight and mean BMI is about 24 in the baseline sample, which is in line with the nding that old-age overweight being a prevalent problem in developing countries. Average years of completed schooling are 4.6 years for parents and 10.5 years for their highest educated children. Figure 3.3 shows that there were slight positive cohort trends in years of school- ing among highest educated children ineligible for compulsory education, which were \parallel" in high schooling areas and low schooling areas. In low schooling areas where the initial schooling levels were lower than regional median, the trend in years of schooling among cohorts partially aected by compulsory education is slightly more positive than that in high schooling areas with initial schooling levels higher than regional median. And the discontinuity in years of schooling at the rst compulsory education eligible cohorts is also more conspicuous in low schooling areas. For cohorts which were completely eligible to the compulsory education, children's years of schooling atten out 18 . It manifests that exposure to compulsory education creates variation in children's education, with higher potential gain among law-eligible cohorts in areas with lower educational attainment among the law-ineligible cohorts. Hence the regional educational attainment and its interaction with the compulsory education exposure creates additional variations in the years of schooling of children. 18 A similar graph by provinces also implies the heterogeneity of such trends in dierent areas. For instance, education expansion in provinces such as Anhui, Chongqing and Zhejiang is more aggressive than provinces in the Northeast. The graph is available upon request. 69 Table 3.2: Descriptive Statistics of Major Variables Baseline Wave 2 VARIABLES mean N sd mean N sd Health and Cognition Measures mental intactness score(0-11) 7.427 10,776 3.133 7.044 10,270 3.432 episodic memory(0-10) 3.408 9,988 1.756 3.279 10,270 1.879 self-report of health (very good=1,very poor=5) 3.044 10,808 0.889 2.998 10,174 0.932 peak expiratory ow (average of 3 measurements) 266.3 10,290 112.8 266.4 7,696 114.0 dominant hand grip strength (average of 2 measurements) 30.55 10,610 10.56 29.72 8,035 10.57 CESD 10 score(0-30) 8.432 10,220 6.274 7.901 9,130 5.768 body mass index=kg/m2 23.51 10,769 3.928 23.76 8,093 3.797 BMI <18.5 0.0657 10,769 0.248 0.0605 8,093 0.239 BMI 25 0.312 10,769 0.463 0.346 8,093 0.476 low subjective survival expectation 0.296 9,373 0.457 0.328 7,925 0.469 Channels ever drinks any alcohol last year 0.330 10,812 0.470 # of drinks per week 1.601 9,422 4.101 r smoke now 0.308 10,774 0.462 # cigarettes/day 4.533 9,886 9.733 # of d/w contacting children (co-residence = 7) 4.139 10,704 2.325 # of d/w contacting children in person (co-residence = 7) 5.266 10,675 2.897 # of d/w contacting children by phone/mail/text/email 1.982 8,492 2.379 net monetary transfer from children 628.4 10,714 12,016 log of per capita household expenditure last year 8.445 9,311 0.853 access to clean fuel (gas, electricity) for cooking 0.412 10,699 0.492 access to in-house ush-able toilets 0.329 10,773 0.470 life satisfaction (completely satised=1,not at all=5) 2.939 9,732 0.699 currently working 0.727 10,780 0.445 Characteristics of Parents age in years 58.51 10,811 9.043 60.42 10,480 8.996 height in meters 1.581 10,812 0.0855 bad childhood health 0.0727 10,812 0.260 gender:female = 1 0.531 10,812 0.499 own years of schooling 4.443 10,812 4.516 marital status 10,812 10,480 married (%) 85.6 9,256 83.4 8,743 married, sp abs (%) 4.3 462 4.7 489 widowed (%) 0.4 46 0.3 28 widowed divorced (%) 0.6 65 0.5 57 single separated (%) 9 977 11 1,157 single never married (%) 0.1 6 0.1 6 single Child Characteristics gender of the highest educated child 0.402 10,812 0.490 years of schooling of the highest educated child 10.45 10,812 3.714 exposure to CSL of the highest educated child 0.587 10,812 0.429 birth year of the highest educated child 1,977 10,812 8.395 Household Characteristics lives in rural or urban area 0.659 10,812 0.474 0.660 10,480 0.474 number of living children 2.722 10,812 1.361 2.814 10,480 1.374 Regional Characteristics average years of schooling among ineligible cohorts 8.174 10,812 1.321 # of beds/10000 population in child's birth yr & prov 17.79 10,812 8.449 # of doctors/10000 population in child's birth yr & prov 10.59 10,812 4.929 log per capita provincial GDP in child's birth yr 5.931 10,812 0.750 Note: CHARLS baseline and Wave 2 data. All numbers are unweighted. 70 6 7 8 9 10 11 12 13 14 -12 -9 -6 -3 0 3 6 9 12 15 -12 -9 -6 -3 0 3 6 9 12 15 High Schooling Area Low Schooling Area Cohort mean Left: fitted trend-ineligible cohorts Middle: fitted trend-partially eligible cohorts Right: fitted trend-completely eligible cohorts Mean Years of Schooling Age Difference with First CS Eligible Cohorts* Source: CHARLS and author's own calculation. *Year of birth of the highest educated child minus year of birth of the the first CS eligible cohort in each province. Figure 3.3: Eect of Compulsory Education on Completed Education Level by High/Low Schooling areas 3.6 Baseline Model 3.6.1 Empirical Methods: IV/2SLS A linear regression equation for baseline health and mediating factors is estimated as below: H ijk0 = 0 + 1 ChildEdu ijk + 3 Control ijk +ChildCohort j + k + j +u ijk (3.1) 71 where H ijk0 is the health and cognition outcome or the mediating factor of individual i living in province j whose highest educated adult child was born in year k. ChildEdu ijk denotes the years of schooling of the highest educated adult child of i. 1 is the parameter of interest which mea- sures the eect of increasing 1 year of schooling of the highest educated child on parental health, cognition or mediating factors. Control ijk is a vector of other child, parental level sociodemo- graphic characteristics, household characteristics and observed regional characteristics, including PreLaw ijh , h =frural;urbang, which is the regional average educational attainment of cohorts ineligible for compulsory schooling. It is calculated as the average years of schooling of cohorts born ve or less years before the rst cohort aected by compulsory education in area h in province j. j are the province xed eects, which account for time invariant heterogeneity across provinces. Models with county and community xed eects are also estimated while only the results from province and community xed eects are reported. 19 k are the child birth cohort xed eects to capture the nonrandom birth of children and cohort heterogeneity in educational attainments. I also include the province-specic child birth cohort trends, ChildCohort j , to control for province- specic deviations from the common nationwide trend that is captured by the birth cohort dummies (Kemptner, J urges and Reinhold, 2011; Lundborg and Majlesi, 2018). Ideally, it would be better if I could separately control for children's birth-place xed eects and parental residing province xed eects, but as we only include households whose children were born in the same place where their parents currently live because of data limitations; those two xed eects are actually the same ( j ). As a matter of fact, children born elsewhere constitute less than 4.5% of the entire children sample. 19 Results from models with county xed eects are available upon request. 72 3.6.1.1 Identication As has been discussed,ChildEdu ijk in Equation (1) is endogenous because of omitted variable and reversed causality problems, and hence OLS estimates are biased. Natural experiments such as compulsory education laws and education expansion programs have been widely used to address the endogeneity problem of schooling in studying the return to educa- tion (e.g., Du o, 2001; Kemptner, J urges and Reinhold, 2011; Huang, 2015; Lundborg and Majlesi, 2018). Exposure or eligibility of individuals to such programs and/or its interaction with the pro- gram intensity are assumed to be correlated with educational attainment but uncorrelated with the unobserved omitted variables, and hence they serve as instruments for educational attainments (Chou et al., 2010). Du o (2001) made use of a primary school construction program in Indonesia between 1973-74 and 1978-79 and showed that individuals who entered school later gained more in schooling and so did those from districts with greater program intensity. Because program intensity was determined by the pre-program supply of primary education on the district level that directly aects the educational opportunity of individuals, she dened the interaction between birth cohort and program intensity as the IV for years of schooling to estimate a wage equation for men in which district of birth xed eects were controlled for. In a recent paper by Huang (2015) he rst exploited the variation in the dierent timing of compulsory education adoption across provinces in China and generated a policy eligibility indicator for each birth cohort in dierent provinces as an IV for years of schooling. He then further explored the cross-province variation in the potential increase in education from compulsory schooling policy as in Du o (2001), by hypothesizing that provinces with less nine-year schooling completion rate before the law enforcement should potentially gain more from the law. I follow the IV estimation method of Huang (2015). In addition, to measure the dierential law eectiveness or program intensity, I draw upon more detailed information on 73 regional dierences in educational attainments in cohorts that were not eligible to compulsory ed- ucation both across provinces and between rural and urban areas within provinces as in Rawlings (2015). The highest educated child's education is instrumented using the variation in exposure to compulsory education and its interaction with program intensity across birth-provinces. Formally, a rst stage equation is estimated as follows: ChildEdu ijk = 0 + 1 Exp ijk + 2 ExpPreLaw ijk + 3 Control ijk +ChildCohort j +t k + j +e ijk (3.2) Whether the child was born at or after the rst aected cohort in each province is a natural instrument. However, as the variation in timing of law enforcement is limited, to introduce more variations in the exposure of children to compulsory education born between the rst (marginally) aected cohort (age 15 when the law took eect) and fully aected cohort (age less than 7 when the law took eect), I dene the rst instrument Exp ijk as Exp ijk = 8 > > > > > > < > > > > > > : 0; if ChildCohort<FirstCohort j ; ChildCohortFirstCohort j +1 10 ; if FirstCohort j ChildCohortFullCohort j ; 1; if ChildCohort>FullCohort j : which will range from 0 (for those age 16 or older at law enforcement) to 1 (for those age 6 or younger at law enforcement). The linear extrapolation of the policy exposure variable is based on the years of the child eligible for compulsory education, given the assumption that the more years of the child eligible for compulsory education, the larger the potential eect of the law. The left panel in Figure 3.4 shows the point estimates and condence intervals of the eect of compulsory education on years of schooling of cohorts with dierent level of exposure. They are from a regression of years of schooling on indicators of years of compulsory education exposure with 74 marginally~1 year 1~2 years 2~3 years 3~4 years 4~5 years 5~6 years 6~7 years 7~8 years 8~9 years 9+ years -2 0 2 4 -2 0 2 4 -2 0 2 4 All Low Schooling Area High Schooling Area Note: Point estimates and condence intervals from regressions of years of schooling of the highest educated child on indicators of compulsory education exposure, birth cohort xed eects, province xed eects, province-specic trends and rural residence indicator using CHARLS data. Figure 3.4: Eects of Exposure to Compuslory Education on Years of Schooling 75 birth cohort xed eects, province xed eects and province-specic trends. In comparison with cohorts not exposed to compulsory education, which is the omitted category, cohorts that were 12 years old or were exposed to compulsory education for three to four years when regional regulations took eect have roughly one more year of schooling. Except for the rst three groups, the eects of compulsory education law increase almost monotonically and linearly. Therefore, a linear function of law exposure could capture such pattern reasonably well. A formal rst stage regression will be presented in section 3.6.2 to show the strength of the instruments used in this chapter. As a robustness check, I also present in Appendix A the results using indicator variables for each year of exposure as instruments instead of the linearly extrapolated law exposure. I did not use indicators for each level of compulsory schooling exposure as the preferred instruments because of the resulting problem of \too many" weak instruments, which will be discussed in the robustness check section. The middle and right panels in Figure 3.4 illustrate the heterogeneous eects of compulsory schooling law by regional schooling levels before the law enforcement. For children that were born in regions with average years of schooling less than the regional median among law ineligible co- horts, the gain in years of schooling for each level of law exposure is higher than their counterparts in high schooling areas. Therefore, in addition toExp ijk , I introduce another instrument that is the interaction term of Exp ijk with the average years of schooling of cohorts ineligible for compulsory schooling, PreLaw ijh ;h =frural;urbang. Following Du o (2001) and Huang (2015), this speci- cation allows for the compulsory education law to be more eective in increasing schooling level in areas with lower educational attainments before the law enforcement. On basis of the province 76 where the parents lived at the time of the survey (which is the same as children's birth province) and whether it is in rural or urban communities, I calculate the second instrument as: ExpPreLaw ijk =Exp ijk PreLaw ijh ;h =frural;urbang By adding the second instrument, I can test the exogeneity restriction using HansenJ statistics. All instrumental variable regression results reported in the main text employ both of the instruments. Standard errors are clustered at the province level since the implementation of compulsory education was on province level. Results from using compulsory education exposure as the only instrument are available upon request. 3.6.1.2 Threat to Identication The identication of children's education rests on the assumption that conditional on other co- variates and the xed eects, variations in compulsory education exposure and its interaction with implementation eectiveness aect health, cognition and other outcomes of parents only indirectly through children's education. In other words, changes in education policies are independent of any unobserved determinants of health in the second stage (Kemptner, J urges and Reinhold, 2011). This assumption implies that there are no social programs in uencing the health of parents or omitted children's characteristics other than education underway that are also varying across re- gions and cohorts and coincide with compulsory education enforcement. I will address the potential threat to validity of compulsory education exposure as instruments in four parts. Firstly, investment in the social safety net had been largely neglected until the 2000s when major reforms in social security, including the minimum support reforms, the establishment of New Cooperative Medical Scheme and new Urban Residents Basic Medical Insurance in China, formally 77 started (Daemmrich, 2013; Fan, Kanbur and Zhang, 2011). For example, the New Cooperative Medical Scheme which replaces the rural Cooperative Medical Scheme (1950s-1980s) was piloted in three hundred counties in four pilot provinces in 2003 and then subsequently rolled out nationally with nearly universal (94.3%) coverage among the rural population in 2009. For urban workers, since 1998 health insurance has been provided by the Urban Employee Basic Medical Insurance program. In 2007, the new Urban Residents Basic Medical Insurance started in 79 pilot cities to test coverage of non-working urban residents, which subsequently expanded nationally 20 . Secondly, while it seems unlikely for any major social policy or investment in social infrastruc- ture changes that varied across provinces to have aected individuals' health directly as did the compulsory education laws in 1980s, the provincial GDP per capita and health facilities measured by number of doctors and hospital beds per 10,000 population from the birth year of the highest educated child are included as covariates in all specications to help control for concurrent invest- ment in social infrastructure. The province-specic cohort trends should also help capture eects of other public policies which might be implemented together with the compulsory schooling law. Thirdly, though it would be impossible to test or control for all concurrent or previous policy changes at regional level as argued by Huang (2015), Huang (2015) directly tested the extent to which compulsory education increased the years of schooling and suggested that the positive association between education and the compulsory schooling eligibility and its interaction term with intensity result from compulsory education implementation rather than from other unobserved 20 Other public expenditure in healthcare includes investment in better national disease prevention and surveillance system after the SARS outbreak of 2003. The central government allocated 2.9 billion RMB and additional funding to help every province, city and county, especially the rural area, set up its own disease control and prevention center. The government has also implemented initiatives to control costs of healthcare by providing guidelines to restrict unnecessary prescriptions of doctors in 2004. Reforms are also undergoing since 2005 to develop hospital management and organization and improve quality of health services. The government has set up ambitious plans to improve the national health care infrastructure by 2010 with goals of ve to eight pilot regional health care systems established by the end of 2006, covering digital services, integration with insurance, referral systems, electronic records, etc (Business Consulting Services, International Business Machines, 2006). As part of the scal stimulus package of 2009, construction and renovation of new and existing facilities, and creation of an essential drug list and public hospital nance reform are also announced (Daemmrich, 2013). 78 factors. He also conducted a placebo test to support the exclusion restriction of the IVs by showing the common trend and absence of regression to the mean in education of individuals not aected by the law. In this chapter, I test the robustness of the main baseline results to the further control of one confounding policy: the One Child Policy which is a nationwide program with dierential enforcement strictness across regions and time (Ebenstein, 2010). the One Child Policy is shown to have aected fertility, timing of birth, number of siblings, health, education, economic outcomes and marriage market outcomes for children (e.g., Huang, Lei and Zhao, 2016; Zhang, 2017). Since it became eective in the late 1970s, all children who were completely eligible to compulsory education were also subject to One Child Policy, while some of children with partial exposure or no exposure were not. I re-estimate the model by additionally controlling for provincial level strictness of the One Child Policy measured by nes, bonus, and premium from Ebenstein (2010) to take into account the eects of One Child Policy that is not captured by province-specic trends and other xed eects. Lastly, it is possible that the policy of compulsory education itself could aect children or household characteristics that are related to parents' old-age health. The instruments used in this chapter might be invalid if such children or household characteristics are unobserved. For example, children's early life health predicts their health in adulthood, and hence aects their ability to support their parents. If it is directly aected by the policy of compulsory education, the exclusion assumption of the instruments will not hold. Though information of children's early life health was not asked in CHARLS, Huang (2015) using other datasets from China found that compulsory education law has not improved individual's childhood health/nutritional status which is measured by adult height. This nding lends more credibility to the identication strategy in this chapter. 79 3.6.2 Basline Model Estimation Results 3.6.2.1 First Stage Table 3.3 shows the rst-stage results using the baseline estimation sample for mental intact- ness. First-stage results barely change for samples using dierent outcomes and hence are omitted. Columns (1) and (3) use only the linearly imputed law exposure measure as the excluded instru- ment, while columns (2) and (4) add its interaction term with average years of schooling of law ineligible cohorts, as the second instrument. The rst two columns control for province xed ef- fects, the second two control for community xed eects 21 . All columns control for child birth cohort xed eects and province-specic linear birth cohort trends. Standard errors are clustered at the province level in each column. Exposure to compulsory education is a strong predictor of years of schooling of the highest educated child's, with all coecients positive and signicant at the 1% level. All interaction terms are signicantly negative, which implies the heterogeneity in the impact of compulsory schooling law: children born in areas with lower education levels before law enforcement gain more in years of schooling from the law. The weak identication test F statistics based on the KleibergenPaap rk statistic for the two excludable instruments in the over-identied specication are above 10 in the province xed eects model, indicating the strength of the two instruments (Staiger and Stock, 1997). The weak identication test F statistic is around 8.6 in community xed eects model when the outcome of interest in the second stage is mental intact- ness. It is generally larger than 10 for other health outcome of interests in the second stage, while it tends to be smaller than 10 when sample size is greatly reduced for mechanism analysis. Because the strength of the two instruments varies in dierent samples and might result in weak instrument problems in some specications, for all second-stage results, I report the test statistics or p values 21 Results of county xed eects models are available upon request. 80 for under-identication test 22 , weak-identication F testand Anderson-Rubin F test on the joint signicance of endogenous variables 23 . If the null hypotheses of the tests are rejected, there is more condence in drawing a statistical inference from the IV estimates. Table 3.3: First Stage Results: Years of Schooling of the Highest Educated Child (1) (2) (3) (4) VARIABLES Years of Schooling of Children Exposure to CSL 1.766*** 3.869*** 1.382*** 3.177*** (0.485) (0.734) (0.490) (0.742) Exposure PrelawEdu -0.264*** -0.225*** (0.069) (0.069) Own Education 0.239*** 0.239*** 0.194*** 0.195*** (0.008) (0.008) (0.008) (0.008) Observations 10,776 10,776 10,769 10,769 Child Birth Year FE YES YES YES YES Prov Birth Year Trend YES YES YES YES Province FE YES YES NO NO Community FE NO NO YES YES F stat for Excludable IVs (weak Identication) 7.492 12.02 3.766 8.620 Note: Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 3.6.2.2 Cognition Table 3.4 presents the estimated results for baseline cognitive functions of Chinese older adults. For both cognition measures, I rst show simple OLS results and IV estimation results using two instruments with province xed eects controlled for. Then I report results from specications with community xed eects. For either outcome, the OLS estimates with province and community xed eects are very similar. So is the case with IV regression results. The causal eects of one year 22 The LM version of the KleibergenPaap rk statistic tests whether all of the canonical correlations between endoge- nous variables and the instruments are signicantly dierent from zero. If one or more of the canonical correlations is zero, the model is underidentied or unidentied. The rejection of the null hypothesis implies identication. 23 When the instruments are weak, 2SLS or IV estimates are biased and t tests are unreliable. The null hypothesis of Anderson-rubin test is that the endogenous variables are jointly zero, conditional on that the instruments are exclude-able. When instruments are irrelevant, Anderson-rubin test has no power which makes this test robust to weak instruments problems. 81 increase of education are about a 0.4-point increase in the mental intactness score, which is about four times of the size of OLS estimates, and a 0.2-point increase in the episodic memory score, which is about ve times the size of OLS results. The estimated eects of children's education are not sensitive to the control of community xed eects, though there seems to be a decline of signicance of coecients. Instruments appear weaker in community xed eects model for episodic memory score because weak identication F statistic is smaller than 10. For all IV regressions and outcomes, under-identication is rejected and Anderson-Rubin F statistics show that coecients of children's education are signicantly dierent from zero, robust to the weak instrument problem. Hansen J statistics are also consistent with the exclusion restrictions being satised. A comparison of the coecients of children's education and own education in OLS models implies that own education matters more for both cognition measures, but IV regression results show the opposite. The caveat in interpreting the coecients for older adults' own education is that they are biased because they are endogenous in nature 24 . The estimated causal eect of increase of one year of schooling of the highest educated child is about 6% of the mental intactness score and the episodic memory score, given their averages being 7.4 and 3.4, or about 0.11 standard deviation for both measures. In Huang (2015), which studied a younger Chinese population, one additional year of schooling increased number of words recalled, which corresponds to the episodic memory, by 0.09 standard deviation, and mathematical calculation ability, which is a component of the mental intactness index, by 0.16 standard deviation. Therefore, the estimated eects of children's education on Chinese elderly's cognitive abilities are comparable to that of one's own education but smaller than the estimated causal eects of own education from some developed countries 25 . 24 Similar results using Lewbel's IV approach to instrument own education are reported in Appendix A. 25 For example, Banks and Mazzonna (2012) nd that for English older adults, the estimated causal eect of one additional year in school of one's own education is about an increase of 0.4{0.5 standard deviation for memory. 82 Table 3.4: Level Eects of Children's Education on Baseline Cognition of Parents (1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV VARIABLES Baseline Mental Intactness Baseline Episodic Memory Education of Children 0.107*** 0.423*** 0.102*** 0.477** 0.045*** 0.245*** 0.038*** 0.209** (0.006) (0.153) (0.007) (0.197) (0.006) (0.090) (0.005) (0.095) Own Education 0.246*** 0.171*** 0.244*** 0.172*** 0.099*** 0.053** 0.093*** 0.060*** (0.008) (0.034) (0.009) (0.037) (0.007) (0.021) (0.006) (0.018) Observations 10,776 10,776 10,776 10,769 9,988 9,988 9,988 9,979 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 28 429 422 28 28 429 420 Mean of Dependent Variable 7.427 7.427 7.427 7.427 3.408 3.408 3.408 3.408 LM stat for underid. 10.17 10.34 11.48 10.98 p-val of LM stat 0.00620 0.00568 0.00322 0.00413 F stat for weak-id. 12.02 8.620 15.82 10.31 Anderson-Rubin F stat 4.481 3.640 6.849 4.003 p-val of AR F stat 0.0209 0.0399 0.00393 0.0300 p-val of Hansen J stat 0.549 0.582 0.181 0.0499 Note: Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 83 3.6.2.3 Subjective Health Measures Table 3.5 reports results on subjective health measures of parents. Consistent with Zimmer et al. (2007) who used actual mortality data, OLS results imply that the correlation of low expected survival likelihood with of children's schooling is stronger than that with one's own education level. Although more schooling of adult children is correlated with both improved self-reported health and higher survival expectation, only the latter correlation remains signicant in the IV estimations. The estimated eects of children's education on low expected survival are not sensitive to the xed eects that are controlled. While the OLS estimate for eect of one additional year of children's schooling is 0.9 percentage point, IV results show an 8.5-percentage-point decrease in the likelihood of low expected survival, which is about a 28% reduction compared with the sample mean. For all specications, under-identication is rejected and the the Anderson-Rubin F statistics show that eects of years of schooling of highest educated child on parental expected survival are signicant. Hansen J statistics are consistent with exclusion restrictions being satised. 3.6.2.4 Objective Physical Health Lung function and grip strength. Table 3.6 shows the estimated eects of child's education on two objective physical health measures of parents: peak expiratory ow which measures lung function and grip strength of the dominant hand. For peak ow, the estimated causal eect of the schooling of the highest educated child is positive and signicant. The estimated eects of one year increase in children's schooling on peak ow of parents are not sensitive to dierent xed eects specications, ranging from 19 to 22 L/min, which is about an 8% increase in peak ow over its baseline average. On the other hand, I nd no evidence for causal impacts of children's education on grip strength of parents. For both outcomes, under-identications are rejected for 84 IV regressions with either province or community xed eects and Hansen J statistics prove that exclusion restrictions are satised. Body weight. Education of the highest educated child has a signicant and positive eect on parents' body weight in old age. It seems to matter more than older adults' own education for body weight, as evident from OLS estimates in Table 3.7. IV regression results show that with either province or community xed eects, one additional year of schooling of the highest educated child leads to a 0.83-unit increase in BMI. Over-identication tests are not rejected for any of the IV specications according to Hansen J statistics. Furthermore, columns (6) and (8) of Table 3.7 show that one more year of schooling of the highest educated child leads to better nutrition status of parents, with a decrease of about 6 percentage points in the probability of being underweight (BMI <18.5). The estimates are not sensitive to dierent xed eects specications. Column (10) and (12), on the other hand, report the eect of children's years of schooling on the probability of parents being overweight (BMI 25). A substantial proportion (more than 30%) of individuals in the baseline sample are overweight, and columns (10) and (12) show that the probability of being overweight increases by about 6 percentage points with each additional year of schooling of the highest educated child. However, the Anderson-Rubin F statistic in column (12) is not signicant, though the magnitudes of estimated coecients for children's education are similar in column (10) and (12). The eects of children's education on the probability of parents being overweight might be related to the fact that more auent people are more likely to be overweight and obese in developing countries 26 As to be discussed shortly, parents with better educated children have greater household expenditure and receive more nancial transfer from children, which might translate into higher body weight in older adults. 26 For a review on the association between obesity and socioeconomic status in developing countries, see Dinsa et al. (2012). 85 Table 3.5: Level Eects of Children's Education on Baseline Subjective Health Measures of Parents (1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV VARIABLES Self Reported Health Low Subjective Survival Expectation Education of Children -0.018*** -0.048 -0.017*** -0.059 -0.009*** -0.085** -0.008*** -0.085** (0.003) (0.045) (0.003) (0.051) (0.001) (0.041) (0.001) (0.039) Own Education -0.011*** -0.004 -0.011*** -0.003 -0.008*** 0.010 -0.007*** 0.008 (0.003) (0.012) (0.003) (0.011) (0.001) (0.010) (0.001) (0.008) Observations 10,808 10,808 10,808 10,801 9,373 9,373 9,373 9,367 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 28 429 422 28 28 426 420 Mean of Dependent Var. 3.044 3.044 3.044 3.044 0.296 0.296 0.296 0.296 LM stat for underid. 10.33 10.60 9.360 9.667 p-val of LM stat 0.00572 0.00499 0.00928 0.00796 F stat for weak-id. 12.61 9.276 14.20 9.889 Anderson-Rubin F stat 0.659 1.070 2.673 3.931 p-val of AR F stat 0.525 0.357 0.0873 0.0317 p-val of Hansen J stat 0.900 0.708 0.660 0.888 Note: Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 86 Table 3.6: Level Eects of Children's Education on Baseline Physical Health of Parents (1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV VARIABLES Baseline Peak Flow Baseline Grip Strength Education of Children 1.433*** 19.207*** 1.163*** 22.101*** 0.198*** 0.448 0.130*** 0.651 (0.308) (7.252) (0.246) (8.399) (0.034) (0.400) (0.020) (0.478) Own Education 2.329*** -1.870 1.808*** -2.272 0.124*** 0.065 0.123*** 0.023 (0.371) (1.843) (0.264) (1.654) (0.030) (0.091) (0.019) (0.090) Observations 10,290 10,290 10,290 10,283 10,610 10,610 10,610 10,603 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 28 426 419 28 28 427 420 Mean of Dependent Var. 266.3 266.3 266.3 266.3 30.55 30.55 30.55 30.55 LM stat for underid. 10.43 10.21 9.891 9.273 p-val of LM stat 0.00545 0.00607 0.00712 0.00969 F stat for weak-id. 12.88 8.475 11.29 7.554 Anderson-Rubin F stat 6.071 5.681 1.160 1.371 p-val of AR F stat 0.00665 0.00873 0.329 0.271 p-val of Hansen J stat 0.407 0.659 0.636 0.877 Note: Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 87 Table 3.7: Level Eects of Children's Education on Baseline Body Weight of Parents (1) (2) (3) (4) OLS IV OLS IV VARIABLES Baseline BMI Education of Children 0.076*** 0.826*** 0.064*** 0.829** (0.014) (0.268) (0.010) (0.325) Own Education 0.023** -0.156** 0.004 -0.144** (0.011) (0.062) (0.012) (0.062) Observations 10,769 10,769 10,769 10,763 Fixed Eects Province FE Province FE Community FE Community FE Number of groups 28 28 428 422 Mean of Dependent Variable 23.51 23.51 23.51 23.51 LM stat for underid. 10.07 10.02 p-val of LM stat 0.00652 0.00668 F stat for weak-id. 12.41 8.916 Anderson-Rubin F stat 7.271 3.145 p-val of AR F stat 0.00298 0.0592 p-val of Hansen J stat 0.777 0.401 (5) (6) (7) (8) VARIABLES Baseline Underweight Education of Children -0.004*** -0.063** -0.003*** -0.059** (0.001) (0.027) (0.001) (0.028) Own Education -0.001 0.014** -0.000 0.011* (0.001) (0.007) (0.001) (0.006) Observations 10,769 10,769 10,769 10,763 Fixed Eects Province FE Province FE Community FE Community FE Number of groups 28 28 428 422 Mean of Dependent Variable 0.0657 0.0657 0.0657 0.0657 LM stat for underid. 10.07 10.02 p-val of LM stat 0.00652 0.00668 F stat for weak-id. 12.41 8.916 Anderson-Rubin F stat 6.346 4.183 p-val of AR F stat 0.00551 0.0262 p-val of Hansen J stat 0.393 0.247 (9) (10) (11) (12) VARIABLES Baseline Overweight Education of Children 0.006*** 0.061*** 0.005*** 0.064** (0.002) (0.021) (0.002) (0.029) Own Education 0.002 -0.012** 0.000 -0.011* (0.002) (0.005) (0.002) (0.006) Observations 10,769 10,769 10,769 10,763 Fixed Eects Province FE Province FE Community FE Community FE Number of groups 28 28 428 422 Mean of Dependent Variable 0.312 0.312 0.312 0.312 LM stat for underid. 10.07 10.02 p-val of LM stat 0.00652 0.00668 F stat for weak-id. 12.41 8.916 Anderson-Rubin F stat 3.106 1.784 p-val of AR F stat 0.0611 0.187 p-val of Hansen J stat 0.821 0.859 Note: Standard errors clustered at province level in parentheses. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for child birth year xed eects and birth province-specic cohort trends. 88 3.6.2.5 Mental Health Table 3.8 shows the estimated eects of children's education on parents' depressive symptoms, measured by a CES-D 10-item index. Based on OLS results, education of children is negatively correlated with CES-D score and seems to matter more for Chinese elderly's mental well-being than their own education. However, I nd no evidence for any causal impact as such, as IV coecients are not precisely estimated. Table 3.8: Level Eects of Children's Education on Baseline Mental Health of Parents (1) (2) (3) (4) OLS IV OLS IV VARIABLES Baseline CESD 10 Score Education of Children -0.155*** -0.267 -0.142*** -0.470 (0.028) (0.350) (0.021) (0.472) Own Education -0.134*** -0.108 -0.130*** -0.068 (0.015) (0.085) (0.015) (0.093) Observations 10,220 10,220 10,220 10,216 Child Birth Year FE YES YES YES YES Prov Birth Year Trend YES YES YES YES Province FE YES YES NO NO Community FE NO NO YES YES Number of groups 28 28 426 422 Mean of Dependent Variable 8.432 8.432 8.432 8.432 LM stat for underid. 9.710 9.498 p-val of LM stat 0.00779 0.00866 F stat for weak-id. 10.29 6.961 Anderson-Rubin F stat 0.377 1.010 p-val of AR F stat 0.689 0.378 p-val of Hansen J stat 0.605 0.242 Note: Standard errors clustered at province level in parentheses. *** p < 0:01, ** p< 0:05, * p< 0:1. 89 3.6.2.6 Mechanism: Social In uence Berkman, Glass and Brissette (2000) suggests that the social in uence of children on parents could be an important pathway through which upward spillover happens. I test this hypothesis by estimating the impact of children's education on health behaviors of parents, including smoking and drinking. Table B6 reports both OLS and IV estimation results using the baseline sample. Because only about 6% of elderly women smoke in the sample of analysis, while more 57% of elderly men do, I only report results on smoking behaviors for men. Although men with higher educated children are less likely to smoke, as shown by OLS, IV estimation results fail to provide any signicant causal estimates. For drinking behaviors, I do nd signicantly positive eect of children's education on parents ever drinking in the past year: One additional year of schooling of the highest educated child increases the probability of drinking of parents by 4 to 6 percentage points. As moderate drinking has been shown to be correlated with better old age cognitive functions in the medicine literature (e.g., Lang et al., 2007), I also test whether children's schooling in uences the intensity of drinking of parents in Table B6. I nd no signicant eects of children's education on number of drinks per week (for both men and women) or number of cigarettes per day (for men). Drinking is more prevalent among individuals with better educated children, but the intensity of drinking does not seem to be aected by children's education. 3.6.2.7 Mechanism: Social Support Past research has found that for Chinese elderly, children's educational level is associated with living arrangements, time and nancial transfer within family (Lei et al., 2015). Table 3.10 shows that at least some of the association is causal. For instance, IV estimates for the eect of children's education on time transfer or emotional support, which is measured by the frequency of contact 90 with children, show that better educated children are not signicantly related to reduced time transfer. On average, children's education does not have negative eects on the amount of emotional support provided to parent. On the other hand, children with more education can provide their parents with more nancial support as shown in the last 4 columns in Table 3.10. One additional year of schooling of the highest educated child leads to a 518{578 RMB increase in the annual amount of net monetary transfer from children to parental household. These ndings suggest that children's education signicantly improves the nancial support provided to their parents but does not signicantly in uence the amount of time transfer from children to parents. 3.6.2.8 Mechanism: Access to Resources Access to resources was tested as a channel through which children's education could in uence parents' health, by examining parental household per capita expenditure, access to clean fuel (elec- tricity and gas) and private ush-able toilets. Table 3.11 shows the eect of children's educational attainment on those measures. Simple OLS results imply that household per capita expenditure and access to private ush-able toilets is correlated to children's education, more than to their own education. The estimated causal eects on household expenditure of education of the highest edu- cated child are sensitive to the choice of xed eects. When only province xed eects are included and both instruments are used, the percentage increase in expenditure is about 14%, while it is only 5% when community xed eects are included. As is consistent with Lei et al. (2015), such ndings imply the importance of controlling for community xed eects. The eect of a one-year increase in children's schooling was about an 8{9 percentage-point increase in the likelihood of household having clean fuel, though Anderson-Rubin F statistics are insignicant (marginally). Access to improved sanitation is positively and causally related to the education of children, with the eect 91 of one additional year of schooling of the highest educated child being a 14.5{17.5 percentage-point increase in the probability of parents having access to ush-able in-house toilets. Anderson-Rubin F statistics are not signicant (at 10% level) in the IV regression with community xed eects. In general, there are positive and signicant causal eects of children's education on parental access to economic resources, clean fuels and sanitation, though the nding is not robust to the level of xed eects controlled for. 3.6.2.9 Mechanism: Other Proximate Pathways I check whether children's education also aects working status of parents. Table 3.12 and B8 demonstrates that children's education has a negative and signicant impact on labor supply of older adults for both the entire sample and the rural households sample (in Appendix). IV regression estimates using the entire sample are signicant: A one year increase in schooling of the highest educated child reduces the likelihood of parents working by about 8 percentage points. Hansen J tests only marginally reject the assumption of over-identication, as their p values are very close to 10%. This result is consistent with the hypothesis that children's education has positive eects on the health of older adults through decreasing the their need to work for income and consumption, rather than the reverse interpretation that children's education keep older people healthy so they are better able to work. Table 3.12 also reports results for parental psychological well-being which is proxied by the evaluative subjective well-being measure: life satisfaction. Parents with better educated children are more satised with their life though the estimated coecients for children's education are not robust according to Anderson-Rubin test. 92 Table 3.9: Level Eects of Child's Education on Baseline Health Behaviors of Parents (Social In uence) (1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV VARIABLES Whether Smoke now (men) Whether Drink last year Education of Children -0.012*** 0.046 -0.010*** 0.099 0.002 0.039* 0.002 0.059** (0.002) (0.066) (0.002) (0.106) (0.001) (0.020) (0.001) (0.024) Own Education -0.006*** -0.021 -0.006*** -0.031 0.001 -0.008 0.001 -0.009* (0.002) (0.017) (0.002) (0.025) (0.001) (0.005) (0.001) (0.005) Observations 5,036 5,036 5,036 5,021 10,812 10,812 10,812 10,805 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 28 422 407 28 28 429 422 Mean of Dependent Variable 0.590 0.590 0.590 0.590 0.330 0.330 0.330 0.330 LM stat for underid. 6.877 6.213 10.32 10.50 p-val of LM stat 0.0321 0.0448 0.00573 0.00524 F stat for weak-id. 9.039 4.516 12.53 9.079 Anderson-Rubin F stat 0.404 0.696 2.467 3.868 p-val of AR F stat 0.672 0.507 0.104 0.0333 p-val of Hansen J stat 0.748 0.739 0.276 0.502 Note: Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 93 Table 3.10: Level Eects of Child's Education on Baseline Social Support from Children (1) (2) (3) (4) OLS IV OLS IV VARIABLES Frequency of Visits of Children: # d/w Education of Children -0.039*** 0.043 -0.045*** -0.124 (0.013) (0.268) (0.013) (0.327) Own Education -0.004 -0.027 -0.004 0.015 (0.010) (0.077) (0.010) (0.079) Observations 6,547 6,547 6,547 6,533 Fixed Eects Province FE Province FE Community FE Community FE Number of groups 28 28 429 415 Mean of Dependent Variable 5.299 5.299 5.299 5.299 LM stat for underid. 6.760 7.578 p-val of LM stat 0.0341 0.0226 F stat for weak-id. 8.498 6.878 Anderson-Rubin F stat 1.695 0.591 p-val of AR F stat 0.203 0.561 p-val of Hansen J stat 0.203 0.291 (5) (6) (7) (8) VARIABLES Frequency of Contacts of Children by Mail etc.: # d/w Education of Children 0.040*** 0.204 0.020* 0.249 (0.012) (0.233) (0.010) (0.250) Own Education 0.043*** -0.003 0.041*** -0.015 (0.012) (0.065) (0.014) (0.063) Observations 5,210 5,210 5,210 5,193 Fixed Eects Province FE Province FE Community FE Community FE Number of groups 28 28 423 406 Mean of Dependent Variable 1.948 1.948 1.948 1.948 LM stat for underid. 8.471 6.640 p-val of LM stat 0.0145 0.0361 F stat for weak-id. 10.06 7.235 Anderson-Rubin F stat 5.541 5.194 p-val of AR F stat 0.00963 0.0123 p-val of Hansen J stat 0.0323 0.128 (9) (10) (11) (12) VARIABLES Net Annual Transfer Amount from Children: RMB Education of Children 27.496** 508.294** 33.898*** 577.637** (11.673) (213.655) (11.963) (235.476) Own Education -10.771 -142.124** -7.282 -135.418** (6.570) (60.960) (7.480) (57.684) Observations 6,377 6,377 6,377 6,363 Fixed Eects Province FE Province FE Community FE Community FE Number of groups 28 28 429 415 Mean of Dependent Variable 817.3 817.3 817.3 817.3 LM stat for underid. 5.610 6.974 p-val of LM stat 0.0605 0.0306 F stat for weak-id. 7.151 6.794 Anderson-Rubin F stat 8.033 7.730 p-val of AR F stat 0.00183 0.00222 p-val of Hansen J stat 0.408 0.647 Note: Standard errors clustered at province level in parentheses. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for child birth year xed eects and birth province-specic cohort trends. 94 Table 3.11: Level Eects of Child's Education on Baseline Access to Resources of Parents (1) (2) (3) (4) OLS IV OLS IV VARIABLES Log Per Capita Household Expenditure Education of Children 0.031*** 0.147* 0.023*** 0.062 (0.003) (0.079) (0.003) (0.090) Own Education 0.023*** -0.009 0.021*** 0.011 (0.003) (0.022) (0.003) (0.021) Observations 5,652 5,652 5,652 5,640 Fixed Eects Province FE Province FE Community FE Community FE Number of groups 28 28 425 413 Mean of Dependent Variable 8.439 8.439 8.439 8.439 LM stat for underid. 7.356 7.305 p-val of LM stat 0.0253 0.0259 F stat for weak-id. 8.650 6.303 Anderson-Rubin F stat 2.957 0.836 p-val of AR F stat 0.0690 0.444 p-val of Hansen J stat 0.331 0.390 (5) (6) (7) (8) VARIABLES Access to Clean Fuel for Cooking Education of Children 0.012*** 0.079* 0.005** 0.093* (0.002) (0.048) (0.002) (0.049) Own Education 0.013*** -0.005 0.009*** -0.013 (0.002) (0.014) (0.001) (0.013) Observations 6,562 6,562 6,562 6,549 Fixed Eects Province FE Province FE Community FE Community FE Number of groups 28 28 427 414 Mean of Dependent Variable 0.419 0.419 0.419 0.419 LM stat for underid. 7.477 8.660 p-val of LM stat 0.0238 0.0132 F stat for weak-id. 9.556 7.855 Anderson-Rubin F stat 2.111 2.442 p-val of AR F stat 0.141 0.106 p-val of Hansen J stat 0.686 0.775 (9) (10) (11) (12) VARIABLES Access to Flush-able In-house Toilets Education of Children 0.013*** 0.068* 0.005** 0.069** (0.002) (0.038) (0.002) (0.030) Own Education 0.008*** -0.007 0.005*** -0.011 (0.002) (0.012) (0.001) (0.008) Observations 6,608 6,608 6,608 6,596 Fixed Eects Province FE Province FE Community FE Community FE Number of groups 28 28 428 416 Mean of Dependent Variable 0.332 0.332 0.332 0.332 LM stat for underid. 7.342 8.219 p-val of LM stat 0.0255 0.0164 F stat for weak-id. 9.226 7.613 Anderson-Rubin F stat 8.775 4.476 p-val of AR F stat 0.00116 0.0209 p-val of Hansen J stat 0.223 0.960 Note: Standard errors clustered at province level in parentheses. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for child birth year xed eects and birth province-specic cohort trends. 95 Table 3.12: Level Eects of Child's Education on Baseline Labor Supply and Psychological Well-being of Parents (1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV VARIABLES Whether Work Now Life Satisfaction (1-5) Education of Children -0.004*** -0.076*** -0.001 -0.082** -0.015*** -0.086* -0.013*** -0.108* (0.001) (0.029) (0.001) (0.033) (0.003) (0.046) (0.004) (0.064) Own Education -0.006*** 0.011 -0.003** 0.013** 0.002 0.018* 0.002 0.019 (0.001) (0.007) (0.001) (0.006) (0.003) (0.011) (0.003) (0.013) Observations 10,780 10,780 10,780 10,773 9,732 9,732 9,732 9,726 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 28 429 422 28 28 426 420 Mean of Dependent Variable 0.727 0.727 0.727 0.727 2.939 2.939 2.939 2.939 LM stat for underid. 10.27 10.35 10.55 10.36 p-val of LM stat 0.00589 0.00566 0.00513 0.00563 F stat for weak-id. 12.42 8.888 13.71 8.961 Anderson-Rubin F stat 5.930 5.093 1.866 1.296 p-val of AR F stat 0.00733 0.0133 0.174 0.290 p-val of Hansen J stat 0.0982 0.0845 0.742 0.575 Note: Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 96 3.7 Dynamic Model 3.7.1 Empirical Methods: Lewbel's IV Approach The static model strives to explain the level eect of children's education on tje health of their parents in old age. Whether children's education has any eect on the changes of health of parents is of interest as well. Conditional on the past health of parents, does children's education still matter for the health and cognition of parents? Or in other words, does children's education has incremental eects on parental health? This section provides some tentative answers to this question. Following Zimmer et al. (2007), I estimate a dynamic model of health which follows an AR(1) process using the two-year follow-up data. A static model for the follow-up data is also estimated as a robustness check; the results are in Appendix A. A linear dynamic model of health for the follow-up sample is as follows: H ijk1 =b 0 +cH ijk0 +b 1 ChildEdu ijk +b 3 Control ijk +ChildCohort j +h k +l j +w ijk (3.3) As baseline health is endogenous and related to children's education, instruments are needed for the lagged baseline health so that the incremental eects of children's education are identiable. It is dicult to nd external instruments for lagged health since it requires very restrictive assumptions on them to aect only current health through lagged health. Therefore, I apply an instrumental variable estimation method introduced by Lewbel (2012) which relies on heteroscedasticity of er- rors and higher moment conditions when traditional instruments are weak or exclusion restriction conditions for traditional IVs are violated. Specically, Lewbel-type instruments are generated as the product of heteroscedastic error terms, which are predicted from auxiliary regressions of the baseline health outcomes on only the exogenous independent variables, with the subset of demeaned 97 exogenous variables. Heteroscedasticity in errors from the auxiliary regressions is tested using the Breusch Pagan test and Sargan-Hansen test of overidentication are conducted for the Lewbel IV regressions. (Appendix B describes the Lewbel IV approach in more detail.) Lewbel-type IVs are generated only for lagged health, not for years of schooling of the highest educated adult child, so the results on the incremental eects of children's education are not driven by the use of generated instruments, which are unnecessary and in some cases reduce the power of estimates for children's education. The IV regression with Lewbel's generated instruments are conducted using two-stage least squares method. The subset of exogenous variables used for instrument construction includes age categories and respondent's gender, which are exogenous and strongly correlated with both past and current health of parents, as well as baseline height, which is a measure of childhood health and shown to be strongly related to later life health among Chinese elderly by Smith et al. (2012). By estimating the dynamic model, I provide suggestive evidence on the eects of children's education on the progression of parental health and cognition. With baseline health controlled of, which is arguably a sucient statistic for the accumulated eects of children's education on parents' health, I can estimate the incremental or periodic eects of children's education on the health of parentsI estimate a dynamic model instead of taking rst dierences of outcomes to avoid the magnication of measurement error. In addition, on basis of the data generation process of health in Appendix B, the dynamic model provides estimates for the incremental eects of children's education in a more direct manner, while the estimates for children's education from the rst dierence equation can be interpreted as the persistent eects. For the detailed description of the data generating process, please see Appendix B. An obvious limitation in the estimation strategy is that own education of older adults has been treated as exogenous. This contradicts the fact that both education and health are endogenously determined. With no credible external instruments, I therefore also apply Lewbel's IV approach to 98 generate \internal" instruments for older adults' own education for the main static as a robustness check. For the dynamic models, I use the dynamic specication where parents' education are instrumented using Lewbel's IV approach as the main specication, and the results where older adults' own education is treated as exogenous in the appendix. A growing number of studies in economics have been utilizing Lewbel's IV method in recent years, either as the main identication strategy when no valid external instruments are present (e.g., Brown, 2014; Banerjee, Chatterji and Lahiri, 2017), or as a complement to traditional IV estimation to provide overidentication or robustness test for models that are just identied (e.g., Sabia, 2007; Beleld and Kelly, 2012; Emran and Hou, 2013; Mishra and Smyth, 2014; Islam and Smyth, 2015). To generate the Lewbel-type IVs for own education of older adults, I use parental age categories, gender and baseline height that are strongly correlated with years of schooling of older adults. Tow- stage least squares estimation results using Lewbel IV (for own education) and external instruments (for children's education) are shown in Appendix A. 3.7.2 Dynamic Model Estimation Results Taking advantage of the follow-up data, I show that most results found for the baseline sample also hold for wave 2 sample. Table B1, B2, B3, B4, and B5 in Appendix A report the static model results using CHARLS wave 2 data. Static estimation results using the follow-up data are similar to the baseline results, though some of the IV estimates for children's education are smaller and insignicant for the follow-up sample. I then estimate the dynamic models for the follow-up sample to study whether there is any incremental eect of children's education on parents' health and cognition in old age. I report simple OLS estimates and Lewbel IV estimation results with generated instruments based on heteroscedasticity in lagged baseline health or cognition. 99 Table 3.13 report estimation results for parents' cognitive abilities in two years after the baseline. For the mental intactness index, the Lewbel IV estimate of the incremental eect of years of schooling of highest educated child is insignicant in the province xed eects model. When community xed eects are controlled of, it is similar in magnitude to the Lewbel IV estimate in the province xed eects model. A similar pattern is found for episodic memory index, except that the estimated causal impact of one additional year of schooling of the highest educated child is signicant even in the community xed eects model. Breusch-Pagan tests of homoscedasticity of auxiliary regressions are rejected for all Lewbel IV models. One general take-away from the dynamic estimation results for cognition is that in comparison with static model results, there is a smaller but signicant incremental eect of children's schooling on parents' cognitive abilities, especially their episodic memory. Another nding worth mentioning is that the state dependence of episodic memory (0.38{0.48) is smaller than mental intactness (0.69), and the coecient of the baseline episodic memory is not signicant in Lewbel IV model with community xed eects. In fact, the state dependence is relatively low in comparison with existing literature, even though I have not taken rst dierences to remove individual xed eects. I fail to nd evidence for any incremental eects of children's education on other health mea- sures of older adults, as shown in Table 3.14), Table 3.15, Table 3.16 and Table 3.17). The lack of signicant results and the decline in magnitude of point estimates suggest that while chil- dren's education matters for health of parents in the long run, it does not have signicant im- pact on their changes or maintenance over a short period of time. This nding is consistent with Yahirun, Sheehan and Hayward (2017). However, all dynamic model results should be interpreted with caution, as for most outcomes and specications, under-identication is not rejected though F statistics for weak-identication tests are above than 10 in most columns. Studies of the incremental eects of children's education are warranted in the future using longer-span panel data. 100 Table 3.13: Incremental Eects of Child's Education on Wave 2 Cognition of Parents (1) (2) (3) (4) (5) (6) (7) (8) OLS Lewbel OLS Lewbel OLS Lewbel OLS Lewbel VARIABLES Wave 2 Mental Intactness Wave 2 Episodic Memory Baseline 0.509*** 0.694*** 0.526*** 0.685*** 0.293*** 0.478*** 0.308*** 0.383*** (0.014) (0.063) (0.012) (0.067) (0.010) (0.118) (0.010) (0.107) Education of Children 0.078*** 0.140 0.072*** 0.141 0.036*** 0.125** 0.036*** 0.129** (0.011) (0.115) (0.010) (0.102) (0.006) (0.061) (0.006) (0.060) Own Education 0.130*** -0.094* 0.126*** -0.069 0.079*** 0.023 0.080*** 0.037* (0.010) (0.051) (0.008) (0.046) (0.006) (0.021) (0.006) (0.021) Observations 10,239 10,239 10,239 10,232 9,507 9,507 9,507 9,499 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 28 427 420 28 28 426 418 Mean of Dependent Variable 7.052 7.052 7.052 7.052 3.340 3.340 3.340 3.340 LM stat for underid. 21.74 19.45 18.99 23.51 p-val of LM stat 0.244 0.365 0.392 0.172 F stat for weak-id. 15.88 8.011 8.655 16.60 Anderson-Rubin F stat 41.88 18.64 8.154 7.247 p-val of AR F stat 0 5.62e-11 6.48e-07 2.17e-06 p-val of Hansen J stat 0.620 0.499 0.755 0.673 Breusch Pagan test p-val 1.20e-92 2.01e-48 1.39e-31 4.11e-25 Note: Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 101 Table 3.14: Incremental Eects of Child's Education on Wave 2 Subjective Health Measures of Parents (1) (2) (3) (4) (5) (6) (7) (8) OLS Lewbel OLS Lewbel IV OLS Lewbel OLS Lewbel VARIABLES Wave 2 Self Reported Health Wave 2 Low Subjective Survival Expectation Baseline 0.420*** 0.488*** 0.411*** 0.438*** 0.393*** 0.051 0.385*** 0.017 (0.012) (0.167) (0.012) (0.160) (0.019) (0.195) (0.020) (0.199) Education of Children -0.010*** 0.001 -0.008** 0.003 -0.010*** -0.042 -0.008*** -0.040 (0.003) (0.038) (0.003) (0.035) (0.002) (0.033) (0.002) (0.031) Own Education -0.002 -0.009 -0.002 -0.006 -0.005*** -0.003 -0.004*** -0.001 (0.002) (0.015) (0.002) (0.014) (0.001) (0.007) (0.001) (0.007) Observations 10,170 10,170 10,170 10,163 7,071 7,071 7,071 7,062 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 28 427 420 28 28 424 415 Mean of Dependent Variable 2.998 2.998 2.998 2.998 0.331 0.331 0.331 0.331 LM stat for underid. 21.64 22.50 21.29 20.47 p-val of LM stat 0.248 0.210 0.265 0.307 F stat for weak-id. 6.895 5.606 18.05 15.30 Anderson-Rubin F stat 2.845 2.508 16.06 21.48 p-val of AR F stat 0.00605 0.0134 3.33e-10 0 p-val of Hansen J stat 0.889 0.742 0.322 0.312 Breusch Pagan test p-val 7.83e-23 1.34e-35 1.25e-15 1.45e-11 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 102 Table 3.15: Incremental Eects of Child's Education on Wave 2 Physical Health of Parents (1) (2) (3) (4) (5) (6) (7) (8) OLS Lewbel OLS Lewbel OLS Lewbel OLS Lewbel VARIABLES Wave 2 Lung Function (Peak Flow) Wave 2 Grip Strength Baseline 0.404*** 0.492*** 0.471*** 0.492*** 0.356*** 0.457*** 0.443*** 0.487*** (0.028) (0.037) (0.024) (0.041) (0.030) (0.062) (0.025) (0.067) Education of Children 1.431*** -1.574 1.374*** 3.302 0.101*** 0.691* 0.125*** 0.505 (0.362) (3.512) (0.317) (3.249) (0.027) (0.383) (0.022) (0.429) Own Education 0.838** 4.279** 0.748** 3.245** 0.080*** 0.008 0.047** 0.056 (0.362) (2.118) (0.354) (1.636) (0.021) (0.103) (0.019) (0.118) Observations 7,383 7,382 7,383 7,371 7,892 7,892 7,892 7,883 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 27 411 399 28 28 416 407 Mean of Dependent Variable 268 268 268 268 29.74 29.74 29.74 29.74 LM stat for underid. 18.35 21.46 21.34 20.37 p-val of LM stat 0.433 0.257 0.263 0.312 F stat for weak-id. 16.76 18.29 11.56 8.533 Anderson-Rubin F stat 32.94 79.84 52.51 22.28 p-val of AR F stat 0 0 0 0 p-val of Hansen J stat 0.228 0.213 0.311 0.204 Breusch Pagan test p-val 9.27e-89 8.60e-109 2.70e-111 1.20e-197 Note: Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 103 Table 3.16: Incremental Eects of Child's Education on Wave 2 Body Weight of Parents (1) (2) (3) (4) OLS Lewbel OLS Lewbel VARIABLES Wave 2 BMI Baseline 0.752*** 0.372*** 0.754*** 0.418*** (0.029) (0.069) (0.028) (0.063) Education of Children 0.019*** -0.133 0.013** -0.033 (0.005) (0.190) (0.006) (0.183) Own Education -0.000 -0.037 -0.004 -0.028 (0.005) (0.052) (0.004) (0.056) Observations 8,057 8,057 8,057 8,050 Fixed Eects Province FE Province FE Community FE Community FE Mean of Dependent Variable 23.77 23.77 23.77 23.77 p-val of LM stat 0.328 0.299 F stat for weak-id. 10.08 9.308 Anderson-Rubin F stat 14.09 47.60 p-val of AR F stat 1.54e-09 0 p-val of Hansen J stat 0.208 0.271 Breusch Pagan test p-val 1.3000e-119 0.00000e+00 (5) (6) (7) (8) VARIABLES Wave 2 Underweight Baseline 0.591*** 0.731*** 0.591*** 0.737*** (0.022) (0.044) (0.021) (0.044) Education of Children -0.001** -0.014 -0.001* -0.013 (0.001) (0.017) (0.001) (0.021) Own Education -0.000 0.002 -0.000 0.002 (0.000) (0.004) (0.000) (0.005) Fixed Eects Province FE Province FE Community FE Community FE Mean of Dependent Variable 0.0604 0.0604 0.0604 0.0604 p-val of LM stat 0.239 0.264 F stat for weak-id. 11.32 5.069 Anderson-Rubin F stat 117.2 42.83 p-val of AR F stat 0 0 p-val of Hansen J stat 0.292 0.301 Breusch Pagan test p-val 0.00000e+00 0.00000e+00 (9) (10) (11) (12) VARIABLES Wave 2 Overweight Baseline 0.736*** 0.731*** 0.731*** 0.719*** (0.010) (0.037) (0.010) (0.034) Education of Children 0.001 0.030*** 0.000 0.024 (0.001) (0.011) (0.001) (0.016) Own Education 0.002** 0.001 0.001 0.002 (0.001) (0.006) (0.001) (0.005) Fixed Eects Province FE Province FE Community FE Community FE Mean of Dependent Variable 0.347 0.347 0.347 0.347 p-val of LM stat 0.384 0.303 F stat for weak-id. 17.29 7.414 Anderson-Rubin F stat 18.83 18.65 p-val of AR F stat 5.00e-11 5.61e-11 p-val of Hansen J stat 0.227 0.273 Breusch Pagan test p-val 1.66000e-22 5.81050e-03 Note: Standard errors clustered at province level in parentheses. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for child birth year xed eects and birth province-specic cohort trends. 104 Table 3.17: Incremental Eects of Child's Education on Wave 2 Mental Health (CES-D 10) of Parents (1) (2) (3) (4) OLS Lewbel OLS Lewbel VARIABLES Wave 2 CESD 10 Score Baseline 0.408*** 0.525*** 0.408*** 0.549*** (0.013) (0.066) (0.013) (0.073) Education of Children -0.111*** -0.199 -0.092*** -0.092 (0.021) (0.253) (0.023) (0.225) Own Education -0.029 0.141 -0.017 0.092 (0.017) (0.127) (0.013) (0.129) Observations 8,705 8,705 8,705 8,700 Province FE YES YES NO NO Community FE NO NO YES YES Number of groups 28 28 425 420 Mean of Dependent Variable 7.829 7.829 7.829 7.829 LM stat for underid. 19.06 19.93 p-val of LM stat 0.388 0.337 F stat for weak-id. 7.146 10.60 Anderson-Rubin F stat 44.62 29.68 p-val of AR F stat 0 0 p-val of Hansen J stat 0.842 0.660 Breusch Pagan test p-val 3.88e-34 2.97e-46 Note: Standard errors clustered at province level in parentheses. *** p < 0:01, ** p< 0:05, * p< 0:1. 3.8 Robustness Checks In this section, I test the sensitivity of the baseline results to the use of alternative instruments, the control of the One Child Policy, the instrumentation of own education of older adults using Lewbel IV approach and the adjustment for multiple hypotheses testing. I also test the sensitivity of the dynamic results to the instrumentation of older adults' own education by showing the results when education of older adults themselves is not instrumented. 105 Table B9 to Table B13 in Appendix A report results from estimations on health and cognition of parents using indicator variables for each level of compulsory education exposure instead of linearly imputed exposure and their interactions with regional average years of schooling of cohorts ineligible to compulsory education as instruments. For all outcomes except for baseline body weight, the estimated eects of children's education are similar to those that have been previously shown in Section 3.6.2. However, from results of Hansen J tests in regressions on the baseline mental intactness, subjective survival expectation and CES-D 10 index, the over-identication assumption is rejected. Another major problem with using indicators as instruments for years of schooling of children is that the indicators seem to be weak as implied by the F statistics of the weak- identication test, most of which are well below the rule of thumb value of 10. Therefore, the alternative instruments are used only for robustness check, not in the main analysis. Table B14 and Table B15 show that the baseline results are not sensitive to the further control of exposure to the One Child Policy. But controlling for the One Child Policy reduces the value of the F statistics from rst stage regressions which is consistent with the previous ndings that the One Child Policy has contributed to the increase of schooling in cohorts that were aected. Table B16, Table B17, Table B18 and Table B19 show how sensitive the baseline and dynamic results are to the instrumentation for older adults' education using Lewbel's approach (Lewbel, 2012). Generally speaking, static model results based on the baseline sample still hold and some of the causal estimates from models with community xed eects, that is, grip strength and CES-D 10 index, become signicant when I instrumented the education of older adults themselves. For dynamic models results, I nd that the incremental eect of children's educaion remains signicant for parental episodic memory index when Lewbel IVs are not used for older adults' education. Using Lewbel's IV method to instrument parents' own education leads insignicant estimates of 106 children's education on parental mental intactness, as the comparison of Table B17 and Table 3.13 shows. I also test whether the baseline results are robust to the adjustment for multiple hypotheses testing using Simes (1986)'s method, since I assess the eects of children's education on a wide array of health and cognition outcomes of older adults. I report results of hypothesis testing based on critical p-values adjusted for multiple testing in Table B20. The estimated eects of children's education on baseline mental intactness, episodic memory, peak ow and body weight remain signicant at the overall 5% signicance level. These four sensitivity analyses support the baseline or dynamic ndings by showing that they are robust to the control of other social programs, uses of instruments whose rst stages are less re- strictive, additional instrumentation for own education of older adults, and adjustment for multiple testing. 3.9 Conclusion and Discussion A multiplicity of research in economics has shown that human capital transmits from older gen- erations to younger ones. However, less evidence is available on whether there is also signicant upward spillover from children to parents (De Neve and Kawachi, 2017). Furthermore, as educa- tion is proven to be an important predictor of human health and longevity in a substantial body of literature, whether the education of younger generations also has signicant and causal in uences on the health of older generations is an interesting question that the literature has neglected for the most part. That said, mounting research has been conducted recently, though most of the stud- ies are association studies that did not control for bias arising from omitted variables and reverse causality problems. 107 This chapter reinforces and extends what existing works have found on the relation between parental health in later life and children's education by providing a causal analysis on a rich set of measures on health and cognition of Chinese elderly. The gradual roll-out of compulsory schooling in the 1980s and early 1990s across Chinese provinces is used as a natural experiment to achieve identication of causation because the exposure to compulsory education creates exogenous variations in the schooling of children within the same province. IV estimation results show that increasing years of education of adult children lead to substantial increase of the cognitive functions of parents including both episodic memory and mental intactness. Parents with better educated children have greater body weight, higher subjective survival expectations and better lung function. In addition to its level eect, children's education also seems to help older adults in terms of maintenance of cognitive functions, shown in results of the dynamic model. For Chinese elderly, improved nancial support from children, better access to resources and sanitation as well as less labor supply and higher level of psychological well-being might explain why children's education could shape parental health and cognition in later life. Cognitive decline, lung function and body weight are important biomarkers for old-age health and strong predictors of mortality (e.g., Mannino et al., 2003; Gale et al., 2006; Lee et al., 1993). The analysis on these biomarkers helps explain the relationship that existing literature has found between children's education and parental survival. Working pathways being social support and access to resources, rather than social in uence of children on parental health behaviors, which might be more important in developed countries, is also consistent with the fact that parents rely on children for old-age support and the pooling of resources among family members is common in developing countries. However, a formal mediation analysis is warranted to understand the relevant importance of dierent mechanisms in explaining the eect of education of children on parental health. 108 Although this chapter has not discussed the possible heterogeneity in the spillover eects from children's education to parents' health, it is very likely that parents benet dierentially by their own schooling level (Friedman and Mare, 2014), gender or children's gender (Lundborg and Majlesi, 2018). Furthermore, other than the potential mechanisms that have been examined in this chapter, there could well be other mediating factors that link children's education with parental health and survival. For example, education is also strongly associated with higher levels of social capital or social engagement (Campbell, 2006) which is shown to be associated with health behaviors, self- rated health, survival (Lindstr om, Hanson and Ostergren, 2001; Nieminen et al., 2013; Veenstra, 2000; Sirven and Debrand, 2008; Sundquist et al., 2004), and cognitive functioning and its decline among the elderly (Bielak, 2010; James et al., 2011) by providing an increased sense of belonging and attachment (Berkman, Glass and Brissette, 2000), mutual trust (Hyypp a and M aki, 2003), stimulation, and resistance to depression (Glei et al., 2005). Although I have not been able to nd any signicant causal relation between parental social activity participation and children's education using CHARLS data 27 , some heterogeneity might remain hidden. Policy implications from the ndings in this chapter are twofold. First, the ndings on the eects of children's education on parental health in old age re-emphasize the broader return or benets of both public and private investments in education. Second, as Friedman and Mare (2014) argued, policies targeting one generation of the family may spill over to previous and subsequent generations, and to the broader family unit. A broader network perspective including family and multiple generations on health is warranted, in order to harness the value of social programs for health in the most ecient way. Education as a policy instrument targeted at one generation might not come at the expense of other generations or other social policies. Education policy on younger 27 Results are available upon request. 109 generations might complement policies in public support system, help reduce health disparities, and improve health of the entire population. 110 Chapter 4 Height Shrinkage, Cognition and Health among the Elderly: Comparisons across England and Indonesia 1 4.1 Introduction There is a rich literature at the intersection of biodemography and economics, devoted to the study of human height both as an outcome as well as a predictor of economic factors. Childhood and adult height have been found to be aected by socioeconomic micro and macro level factors, like parental investment and disease environment respectively. e.g. see Thomas, Strauss and Henriques (1990), Thomas and Strauss (1992) and Deaton (2007). Adult height, considered to be a good proxy for childhood height and hence childhood health, has been established as an important co- variate for adult labor market outcomes and health, e.g. see Case and Paxson (2008b). However, as Huang et al. (2013) explain, while height is a marker of early life health, when studying health of the elderly, height shrinkage is a potential marker for later life health. They show that height- health associations for the elderly ought to take into account age-related height loss since many of the old-age health outcomes they look into are signicantly (and negatively) associated with height 1 This is a joint work with Urvashi Jain. This analysis uses data or information from the Harmonized ELSA dataset and Codebook, Version E as of April 2017 developed by the Gateway to Global Aging Data. The development of the Harmonized ELSA was funded by the National Institute on Aging (R01 AG030153, RC2 AG036619, 1R03AG043052). For more information, please refer to www.g2aging.org. 111 shrinkage, in some cases the coecients on height loss being larger and/or signicant as compared to height. This is especially important for studies using self reported height since there exists some literature documenting that respondents overestimate current and recalled heights, and this over- estimation increases with age (Cline et al., 1989; Briot et al., 2010; Cawley, Maclean and Kessler, 2017), leading to a systematic bias in the height-health coecients. There have been attempts to estimate pre-shrinkage height for the elderly populations by using various limb lengths, since limbs do not shrink with age the way height does 2 . In case of cross sectional data, the association between limb length and height among a younger sample (which has not yet begun to experience shrinkage) is estimated, and then this association is applied to measured limb length of the elderly to predict their pre-shrinkage heights, as explained by Huang et al. (2013). However, this approach needs to be careful about cohort eects. Later born cohorts are taller on average due to improv- ing living conditions (Fogel, 1986; Steckel, 1995), hence it is to be assumed that the relationship between height and limb lengths does not change at higher levels of height. Limb length can also be used more directly, as instruments or proxies for pre-shrinkage height or early life health. e.g. see Huang et al. (2008) and Sohn (2015). A simpler, and more direct, way to measure shrinkage would be using long term panel data which measures pre-shrinkage height, i.e. when respondents are younger, and continues to measure their height at regular intervals as they age. We follow the latter approach, taking advantage of aging and household panel surveys in England and Indonesia respectively. While the English population belongs to the taller end of human height spectrum, In- donesians are among the shorter populations of the world 3 . We begin by documenting the extent of height shrinkage among the elderly in these two very dierent populations. England and Indonesia 2 Using lower limb length, i.e. up to the ankle, might be marginally better than knee height since attening of feet could also decrease measured height. 3 https://www.disabled-world.com/artman/publish/height-chart.shtml 112 are at vastly dierent levels of economic development and social welfare support, and have racially dierent populations. We see that women across both England and Indonesia exhibit more height shrinkage than men. This nding is consistent with the medical literature which has established that women experience greater height loss (Sorkin, Muller and Andres, 1999), with one of the reasons being that are more prone to osteoporosis and broken bones in old age than men (Cawthon, 2011). Indonesian elderly women exhibit the most height shrinkage despite being the shortest, ranging from average of 1.1 cm over seven years to 2.2 cm over seventeen years. We then look into age, current SES and childhood SES gradients in height shrinkage. Consistent with the existing medical literature, we nd that height loss increases with age in both England and Indonesia. However, we do not nd any signicant associations with current or childhood SES variables. Since height loss is associated with spinal compression and/or curving (see Krege et al. (2015)) as well as muscle loss, we estimate the extent to which height loss aects cognition and physical functioning (measured functional capacity, reported diculties). Finally, we also explore whether height loss before death is a signicant predictor of mortality. There are important non-linearities in the association between height-shrinkage and health. Among the English elderly, while we do not nd a continuous measure of shrinkage to be signicantly associated with lower cognition (memory), extreme height loss, which is dened as height loss above 3 cm, is negatively associated with word recall scores for men and women. Extreme height loss is also associated with lower uency scores for both men and women. Among the Indonesian elderly, shrinkage is related to lower memory among women but not men, and the association becomes stronger as we look into longer time spans. Interestingly, this association is ipped for genders when we look at measures of verbal uency, with stronger results for men. Since height-shrinkage is a marker for worsening health in old age, it is negatively associated with physical function and increased disability in both English and Indonesian elderly. Finally, we nd 113 that height loss in the years preceding death is a signicant predictor of subsequent mortality among the Indonesian elderly. To summarize, in most of the height-shrinkage-health associations we study, shrinkage is found to be an important component of height-health associations that is signicantly correlated with worse old-age health. Given rapid population aging around the world, there is an increasing interest in later life health among both academia and policy makers. The growing number of aging surveys over the past and present decade, in both developing and developed countries, is testimony to this. We contribute to this growing literature by decomposing the usual height-health associations into height-health and shrinkage-health associations. The rest of this paper is organized as follows. We begin with a review of the existing literature on height, height shrinkage and later life health in Section 4.2. We describe the datasets used, in Section 4.3. We then go on to document the extent of height shrinkage in England and Indonesia in Section 4.4. Section 4.5 explores height-health and shrinkage-health relationships. Section 4.6 dis- cusses strengths and limitations, along with suggestions for future research. We provide concluding remarks in Section 4.7. 4.2 Existing Literature Height has long been established as a key measure of the health component of human capital, a component especially important for labor market outcomes in developing countries. Some of the earliest population level evidence of the association between height and wages is from Brazil (Thomas and Strauss, 1997). See Thomas and Strauss (1998) and Schultz and Strauss (2008) for detailed reviews on the importance and measurement of health as a key component of human capital. The next set of key contributions, linking adult height with cognition, health and labor 114 market outcomes are by Case and Paxson (2008a), who study these among the elderly in United States, utilizing the Health and Retirement Study. Steckel (2009) provides a literature review of these associations. Height in childhood and as adults has been found to be associated with health and economic outcomes across many countries, using multitudes of population surveys. E.g., height and later life cognition in Latin America (Maurer, 2010), England (Guven and Lee, 2013), and specically among the Irish elderly (Mosca and Wright, 2016). Relationship between adult height and employment in Mexico has been recently explored in LaFave and Thomas (2016). Given the various populations in which height and later life cognition has been studied, there have also been comparisons across various countries, e.g. Guven and Lee (2015) across Europe, Weir, Lay and Langa (2014) across United States, China and India, and McGovern (2014) who compares China, Ghana, India, Mexico, Russia and South Africa. Studying the relationship between height and later life health outcomes for the elderly needs a more careful approach though since human beings lose height as they age, particularly after age 40 (Cline et al., 1989), with women being more prone to height loss (Sorkin, Muller and Andres, 1999). The main reason behind age related height loss is vertebral disc degeneration due to disc dehydration and compression. Second important reason for height loss is muscle degeneration in old age, or sarcopenia, which aects spinal curvature. Other reasons are the attening of feet arcs, and in extreme cases, vertebral compression fractures. One of the mechanisms/factors due to which we expect a relation between shrinkage and health is osteoporosis- if someone has severe osteoporosis, they experience more loss in height (Old and Calvert, 2004), osteoporosis also aects functional status and/or ability. Vitamin D deciency increases the risk for osteoporosis and has also been shown in a few studies to increase the risk for cognitive decline, e.g. see Llewellyn et al. (2010). Height shrinkage is a marker for overall aging and frailty (Wannamethee et al., 2006). The 115 associations between shrinkage and cognition/physical-function are mostly driven by behaviors aecting both- e.g., poor diet/nutrition, low levels of physical activity, smoking, alcohol intake. Age related height loss is a commonly known biological fact, and recent height-cognition later life research acknowledges and addresses this. E.g., Maurer (2010) instruments height with knee height, Guven and Lee (2013) use the earliest height measurements in their analysis, Mosca and Wright (2016) restrict their sample to age 51-70 to avoid the eects of shrinkage biasing the height co- ecients. However, very little research has been done in the health-height economics literature regarding how not only height but also height loss could be associated with health in old age. Two notable and recent contributions to the study of height shrinkage in this literature are Huang et al. (2013) and Fernihough and McGovern (2015). Huang et al. (2013) employ techniques from the human biology literature by using the associ- ation between limb lengths and height of a younger cohort (age 45-49) to estimate pre-shrinkage height, and shrinkage among the Chinese elderly cohort (age 60 and above) in the CHARLS Base- line cross section sample. They also elucidate shrinkage-health associations and clearly show how shrinkage, along with pre-shrinkage height, is in fact a signicant predictor of cognition and phys- ical function in old age. Fernihough and McGovern (2015) are able to calculate shrinkage using repeated height measurements for the English elderly, utilizing Waves 0, 2 and 4 of ELSA. They nd a strong age gradient, but no SES gradient in height shrinkage. They also argue that change in physical function (peak ow and grip strength) over time is a predictor of height loss. Finally, they argue that using current height instead of pre-shrinkage height does not bias the health production function estimations for the elderly. There are two main contributions of this paper. First, we extend Fernihough and McGovern (2015)'s work for the English elderly, by looking into how later life cognition is related to height- loss. Second and more importantly, we document height loss and explore the shrinkage-health 116 relationships for another population- the Indonesian elderly. Indonesia is an interesting and valuable comparison for many reasons. First, Indonesians are among the shorter populations in the world, which indicates worse early life health (e.g. see Currie and Vogl (2013) for a detailed review about early life environment and height). Second, since it is still a developing economy with constrained resources, later life health for older adults is also expected to be worse than their elderly counterparts in a developed country. Third, estimates from Indonesia are also more useful for comparisons with China, shrinkage estimates for which have been obtained by Huang et al. (2013). 4.3 Data In order to study height shrinkage among the elderly, longitudinal data with measured height is preferred over panel data that rely on self-reported height since it has been found that reporting error in height increases with age. For example, see Cawley, Maclean and Kessler (2017) who study reporting errors in height and weight among the elderly in the widely used HRS sample. Hence, objective height measurements over a relatively long time period, along with cognition and health variables were the criteria for selecting the datasets for our study. Similarity in cognition and health measures is also desirable to compare eects across countries. 4.3.1 England We use the nationally representative English Longitudinal Study of Aging (ELSA) 4 , which began in 2002 5 . It is one of the international sister studies 6 of the Health and Retirement Study in the United States. ELSA panel respondents are aged 50 and above, and their spouses are also interviewed, 4 https://discover.ukdataservice.ac.uk/series/?sn=200011 5 Respondents for the baseline wave were recruited from the Health Survey of England (HSE) conducted in 1998, 2000 and 2001. This is also referred to as ELSA Wave 0 by Fernihough and McGovern (2015). 6 https://hrs.isr.umich.edu/about/international-sister-studies 117 regardless of age. We use Harmonized versions 7 of Wave 2 (2004-05), Wave 4 (2008-09), Wave 6 (2012-13) and Wave 7 (2014-15). Anthropometric measures (including height, grip strength and lung function) were collected by nurses every alternate wave. Cognition and self-reported health measures are available for every ELSA wave. In addition, we also use Wave 3 (2006-07) which asked retrospective life history questions. As per Cline et al. (1989), height loss becomes apparent only after 60 years of age, a nding also conrmed by Fernihough and McGovern (2015) 8 . Hence, we only keep respondents who are aged 60 and above as in Wave 6 9 . Attrition for men is higher than for women across waves. 72.3% (7,664 observations) of the sample is of age 60 and above, as in Wave 6. In our estimations, we only keep respondents for whom height measures are available across all waves, which is for 47.77% of the 60-plus sample, giving us 3,661 observations, from which we drop a few more outliers which show implausible height gain or loss across waves. 4.3.2 Indonesia We use the Indonesian Family Life Survey (IFLS), which began in 1993 and is representative of 83% of the Indonesian population 10 . We utilize four waves: IFLS-2 (1997) (Frankenberg and Thomas, 2000), IFLS-3 (2000-01) (Strauss et al., 2004), IFLS-4 (2007-08) (Strauss et al., 2009) and IFLS-5 (2014-15) (Strauss, Witoelar and Sikoki, 2016). For each of these waves, anthropometric variables were measured by trained nurses. Cognition measures for the elderly were introduced in IFLS-4, as was grip strength measurement. 13.04 % (9,343 observations) of the IFLS-5 sample is aged 60 and above. As we do for the ELSA sample, we keep respondents who were aged 60 and above in Wave 5. We checked for mean shrinkage levels by age groups, and it is only after age 60 that 7 Gateway to Global Aging https://g2aging.org/ 8 We checked for mean shrinkage levels by age groups in ELSA and conrmed this pattern. 9 This allows us to get pre-shrinkage heights for some respondents (those aged 52 and less in Wave 2), and for others who had already begun to shrink, that is respondents who were above 60 in Wave 2, we can still compute the extent of further shrinkage as they age. 10 http://www.rand.org/labor/FLS/IFLS.html 118 14-year and 17-year mean shrinkage exceeds 1 cm for women. Within the 60-plus sample, height measurement is available across all waves for 29.02%, giving us a sample size of 2,711, from which we drop outliers with extreme values of height change. 4.4 Shrinkage Height shrinkage is dened simply as the dierence in height (in centimeters) from an older wave and the most recent wave. shrinkage t =height t height c (4.1) shrinkage t is height loss over the course of t years. height t is height (cm) measured t years ago. height c is current height, that is height measured in the most recent available wave. These are Wave 6 in ELSA and Wave 5 in IFLS. Since height was measured every alternate wave in ELSA, t2f4,8g. That is, shrinkage over four years is the change in height from Wave 4 (2008-09) to Wave 6 (2012-13), and shrinkage over eight years- from Wave 2 (2004-05) to Wave 6 (2012-13) 11 . Similarly, for IFLS, t2f7,14,17g. That is, shrinkage over seven years is the change in height from Wave 4 (2007-08) to Wave 5 (2014-15), shrinkage over fourteen years is the height dierence from Wave 3 (2000-01) to Wave 5 (2014-15), and shrinkage over seventeen years is the height dierence from Wave 2 (1996) to Wave 5 (2014-15). Note that positive (increasing) shrinkage t indicates height loss, and negative shrinkage t indicates height gain. 11 Due to data access restrictions, we could not link ELSA respondents from Wave 1 onwards to ELSA Wave 0, which would have given us even longer term measurements of shrinkage. 119 4.4.1 Extent of Shrinkage We drop a few respondents for whom either measured height is less than 100cm, or for whom height gain or loss over subsequent waves exceeds 10cm 12 . See Appendix A for details on data cleaning. Table 4.1: Shrinkage (in cm) VARIABLES N mean sd min max N mean sd min max ELSA Male ELSA Female Current Height (cm) 1608 172.42 6.96 152.1 202 2035 158.76 6.55 134.5 179.8 Height 4 years ago 1608 173.09 6.93 151.8 201.6 2035 159.47 6.44 136.5 181 Height 8 years ago 1608 173.63 6.79 153 201.4 2035 160.1 6.33 137.6 181.4 Shrinkage over 4 years 1608 0.67 1.68 -7.2 8.10 2035 0.7 1.65 -7.2 9.4 Shrinkage over 8 years 1608 1.2 1.73 -5.9 9.5 2035 1.33 1.79 -9.8 10 IFLS Male IFLS Female Current Height (cm) 1162 158.3 6.28 126.4 179 1382 146.42 5.73 129.2 166.8 Height 7 years ago 1162 158.9 6.15 126.6 178.9 1382 147.56 5.50 130.9 168.6 Height 14 years ago 1162 159.77 5.95 127.2 178.1 1382 148.62 5.37 132.1 169.5 Height 17 years ago 1162 159.78 5.88 126.5 178.6 1382 148.64 5.31 132 168.7 Shrinkage over 7 years 1162 0.6 1.81 -8.6 10 1382 1.14 1.99 -9.7 10 Shrinkage over 14 years 1162 1.47 1.85 -9 9.3 1382 2.2 2.05 -8.6 9.8 Shrinkage over 17 years 1162 1.48 2.02 -9.2 8.8 1382 2.22 2.27 -8.9 9.9 Note: ELSA Waves 1 to 6, and IFLS Waves 2 to 5. Table 4.1 presents the extent of shrinkage in ELSA and IFLS. Height shrinkage is evident from the decrease in mean height over time in both the samples, with current height being the lowest. On average, men lost 0.67 cm height over the course of four years (baseline average height was 173.1 cm), and 1.2 cm over eight years (baseline average height was 173.63 cm) in ELSA which is a mean height loss of 0.7%. These are consistent with what Fernihough and McGovern (2015) found for previous waves of ELSA. Women lost 0.7 cm height over the course of four years (baseline average height was 159.47 cm) and 1.33 cm over eight years (baseline average height was 160.1 cm) which is a mean height loss of 0.85%. Even though Indonesians are considerably shorter, 12 Wannamethee et al. (2006) look at how height loss over 20 years predicts cardiovascular disease and mortality in a sample of English men. They drop respondents for whom height dierence over time indicates an increase in height of more than 5 cm. Huang et al. (2013) drop any height observations less than 120 cm. 120 they experience greater shrinkage than their English elderly counterparts. Height loss over seven years among Indonesian elderly men was 0.6 cm (baseline height of 158.9 cm). Height loss over fourteen and seventeen years for men was close, 1.47 and 1.48 cm respectively (baseline height of 159.77 cm). Indonesian elderly women experience the greatest shrinkage- 1.14 cm over seven years (baseline height of 147.56 cm), approximately 2.20 and 2.22 cm over fourteen and seventeen years respectively. That is, elderly women in Indonesia lose 1.6% of their height over the course of 17 years. Note that elderly men in Indonesia lose slightly more height over 7 years than the men in England lose over 8 years. In comparison, mean height shrinkage estimates for the Chinese elderly computed by Huang et al. (2013) were 3.3 cm for men and 3.8 cm for women, over a ten-year and higher time period 13 . Gender dierences in height shrinkage are also evident from Figures 4.1 and 4.2 for ELSA, Figures 4.3 and 4.4 for IFLS, where we plot the kernel densities of height shrinkage over time. All of these distributions have a positive mean, which indicates height loss in the population, and becomes more prominent with greater time lags. That is, the distribution shifts towards the right for values of change in height measured over a longer time period. In all four graphs, the density plot for women's height loss is further right than the plot for men. In Figures 4.5 and 4.6, we plot mean height shrinkage by gender and ve-year age categories. We clearly see that gender dierences are more stark in Indonesia than in England. 13 We looked at change in height from CHARLS Wave 1 (2011) to Wave 4 (2015), which gave us mean height loss over 4 years at 0.49 cm for men and 0.84 cm for women, 60 years of age and above as in Wave 4. 121 Figure 4.1: Shrinkage over 4 years in ELSA Figure 4.2: Shrinkage over 8 years in ELSA 122 Figure 4.3: Shrinkage over 7 years in IFLS Figure 4.4: Shrinkage over 17 years in IFLS 123 Figure 4.5: Shrinkage over 4 and 8 years by gender and age in ELSA Figure 4.6: Shrinkage over 14 and 17 years by gender and age in IFLS 124 4.4.2 (Erroneous) Height Gain Kernel density plots of shrinkage (Figures 4.1, 4.2, 4.3 & 4.4 ) indicate that for some respondents, shrinkage is less than zero, that is, they appear to have been gaining height. In the ELSA male sample, 29.73% indicate height gain over four years, and 20.71% over eight years. Corresponding percentages in ELSA female sample are 29.58% and 18.03%. 28.17% of Indonesian male sample show increase in height over seven years, 12.33% over fourteen years, and 16.01% over seventeen years. In the female sample, 19.68%, 6.20% and 9.23% show height increase over seven, fourteen and seventeen years respectively. Our percentages of height gain in ELSA are well within what (Fernihough and McGovern, 2015) found, and IFLS percentages of height gain are even lower. Height gain is biologically implausible among the age groups we study. Fernihough and McGovern (2015) discuss this issue and oer two explanations. First attributes to the nature of population surveys- measurement error by the interviewer. Second attributes to human biology- individual height uctuations within a day, which could be up to 1 cm. We are all taller in the morning than in the evening. Hence, if the nurses measured respondents' heights earlier in the day in the most recent wave, and towards the latter part of the day in the previous wave, this could lead to a reported increase in measured height over time. We check for correlation between the time of height measurement and measured height in the IFLS sample. A simple regression of measured height in Wave 5 on the time of biomarker interview (continuous) as well as on various dummy measures indicating time of biomarker interview to be in the latter part of the day does not indicate any systematic pattern. Therefore, the main reason for supposed gain in height for some respondents is measurement error which can be due to many reasons- the respondents not standing as erect as they could, tilting their head, hair-dos for women and caps for men etc. It is not intuitively obvious whether the expected measurement error is classical. Older respondents with a stooping 125 posture would have greater diculty in standing erect. It might be more dicult for interviewers to measure extremely tall respondents. While this measurement error shows up as height gain over time, it could also lead to misclassication regarding height loss. Hence, we trim the data allowing for equal extent of height gain and loss (10 cm), even though the medical literature would allow us to cap height gain more strictly than height loss in these age groups, for example Wannamethee et al. (2006) drop respondents with height gain of more than 5 cm. We now summarize height loss across these four demographic groups. First, Indonesian elderly women experience greater height shrinkage than English women over a comparable period. Indone- sian elderly men on the other hand experience lesser height loss on an average than English men. Second, women in both England and Indonesia exhibit higher shrinkage than the men in their re- spective countries, a nding consistent with the medical literature which has established that post menopausal women are more prone to osteoporosis and fractures as they experience declining bone density due to lower estrogen levels. Third, Indonesian men are on an average, shorter than, and experience more height shrinkage, than English women. Finally, Indonesian elderly women seem to bear a double burden- they have the lowest heights and experience highest shrinkage. This points to lower early life nutrition and health, as well as worse later life health, in a developing country compared to a developed one. 4.4.3 Covariates of Shrinkage Next, we examine SES covariates for height shrinkage, using the specication below, separately for men and women. shrinkage i;t;tw = +height i;tw + Age i;t +SES i;t + i;t (4.2) 126 Shrinkage i;t;tw is change in height as measured between a previous wave which was w years ago, and the most recent wave,t. Age i;t are ve year dummies on the basis of current age. w2f4,8g for ELSA respondents, w2f7,14,17g for IFLS respondents. SES i;t are current, i.e. in the most recent available wave, socio-economic variables which include respondent's education, marital status, race (white or non-white in England) or ethnicity in Indonesia 14 , self reported health in childhood, household characteristics in childhood, and mother's current or age at death. Most of the SES controls are time-invariant, given the age group, except for marital status. We also include lagged per capita monthly household consumption corresponding to the time gap w. Tables 4.2 and 4.3 present results for ELSA and IFLS respectively. As expected, age is the most signicant predictor of height shrinkage, especially after 70 years among both English and Indone- sian elderly; shrinkage is highest among the oldest group, aged 80 and above. Height shrinkage over the course of w years is signicantly associated with height w years ago. That is, taller respondents exhibit more shrinkage. Since even lagged measures of height could already be subject to shrinkage, especially for oldest respondents, and hence not capturing their maximum, pre-shrinkage, height, we also estimate equation (2) replacing height i;tw with limb length (which do not shrink with age) for the IFLS sample. However, we do not nd either upper arm length or knee length to be signicant covariates of height shrinkage (Appendix Table C1). We do not nd any education gradients for shrinkage in neither the ELSA nor the IFLS, which is in contrast to Huang et al. (2013) who nd signicant correlations of shrinkage with education levels in the Chinese sample. We also do not see any race gradients for shrinkage in the ELSA sample. We do see that Tionghoa (ethnic Chinese migrants) women experience lower shrinkage in Indonesia. We don't nd any of the childhood SES variables to be signicantly associated with height shrinkage in either ELSA or IFLS sample. 14 Dummy variables for Java, Sunda, Bali, Batak, Bugis, Tionghoa, Madura, Sasak, Minang 127 Table 4.2: Shrinkage (in cm) Covariates in ELSA (1) (2) (3) (4) Male sample Female sample Shrinkage Shrinkage Shrinkage Shrinkage VARIABLES over 4 yrs over 8 yrs over 4 yrs over 8 yrs Height 4/8 years ago 0.0372*** 0.0229*** 0.0260*** 0.0198** (0.00783) (0.00778) (0.00764) (0.00788) Age: \65 to 69" 0.0844 0.320** 0.222 0.340** (0.150) (0.147) (0.135) (0.139) Age: \70 to 74" 0.565*** 0.731*** 0.240 0.406*** (0.163) (0.159) (0.149) (0.153) Age: \75 to 79" 0.298* 0.696*** 0.463*** 1.018*** (0.171) (0.167) (0.156) (0.159) Age: \80+" 0.786*** 1.467*** 0.873*** 1.458*** (0.182) (0.176) (0.176) (0.179) High-school graduate 0.00708 0.0137 -0.199 -0.274** (0.152) (0.149) (0.122) (0.125) Some college -0.193 -0.0406 -0.0277 -0.0819 (0.140) (0.137) (0.130) (0.133) College and above -0.0546 0.00141 -0.0642 -0.0378 (0.156) (0.152) (0.165) (0.170) Partnered -0.179 -0.0476 -0.459 -0.296 (0.335) (0.328) (0.288) (0.294) Separated -0.480 -0.792 0.736 0.101 (0.842) (0.821) (0.572) (0.585) Divorced -0.0198 -0.218 0.208 0.103 (0.242) (0.234) (0.158) (0.161) Widowed -0.0803 0.0178 0.0277 -0.0497 (0.203) (0.197) (0.128) (0.130) Never married 0.494* 0.315 -0.121 -0.0103 (0.287) (0.280) (0.274) (0.280) Non-white -0.316 -0.262 0.344 0.639 (0.428) (0.418) (0.542) (0.554) Reported Poor Childhood health -0.282* -0.186 -0.196 -0.0908 (0.170) (0.166) (0.134) (0.137) Number of bedrooms of residence -0.107* -0.0473 -0.0626 0.00107 lived in when aged 10 (0.0640) (0.0626) (0.0547) (0.0559) Mother's age: \60 to 64" 0.353 0.452 0.679** 0.640** (0.330) (0.323) (0.299) (0.306) Mother's age: \65 to 69" 0.269 0.474 0.165 0.391 (0.308) (0.300) (0.266) (0.272) Mother's age: \70 to 74" 0.358 0.367 0.330 0.470* (0.289) (0.283) (0.252) (0.258) Mother's age: \75 to 74" 0.402 0.385 0.359 0.513** (0.279) (0.273) (0.242) (0.248) Per capita Monthly Household -0.0896 -0.0395 Consumption 4 years ago (0.0920) (0.0573) Per capita Monthly Household -0.116 -0.0301 Consumption 8 years ago (0.0964) (0.0686) Constant -5.460*** -2.871** -3.588*** -2.638** (1.441) (1.425) (1.303) (1.337) Observations 1,042 1,042 1,166 1,166 R-squared 0.055 0.087 0.049 0.090 Note: Omitted category in age: \60 to 64", in education: \No School", in marital status: \Unmarried". Robust standard errrors in parenthesis. *** p < 0:01, ** p< 0:05, * p< 0:1. 128 Table 4.3: Shrinkage (in cm) Covariates in IFLS (1) (2) (3) (4) (5) (6) Male sample Female sample Shrinkage Shrinkage Shrinkage Shrinkage Shrinkage Shrinkage VARIABLES over 7 yrs over 14 yrs over 17 yrs over 7 yrs over 14 yrs over 17 yrs Height 7/14/17 years ago 0.0433*** 0.0170* 0.0204 0.0320** 0.0259** 0.0326** (0.0106) (0.00981) (0.0121) (0.0134) (0.0112) (0.0132) Age: \65 to 69" -0.0489 -0.0971 0.132 0.114 0.492*** 0.509*** (0.186) (0.188) (0.177) (0.0983) (0.118) (0.153) Age: \70 to 74" 0.0931 0.630*** 0.578*** 0.543*** 0.917*** 1.104*** (0.140) (0.154) (0.145) (0.179) (0.167) (0.177) Age: \75 to 79" 0.663** 1.140*** 1.492*** 0.291 1.250*** 1.380*** (0.247) (0.251) (0.276) (0.204) (0.180) (0.248) Age: \80 to 100" 0.588* 1.181*** 1.784*** 0.596* 1.740*** 2.037*** (0.338) (0.350) (0.406) (0.329) (0.372) (0.440) Urban -0.0996 -0.267 -0.339*** 0.349** 0.264 0.347 (0.0920) (0.163) (0.104) (0.153) (0.160) (0.218) Education: Primary 0.0388 0.393 0.251 0.125 -0.0749 0.0738 (0.289) (0.287) (0.273) (0.139) (0.156) (0.165) Education: Junior -0.380 0.195 -0.0104 0.0678 -0.218 -0.231 (0.368) (0.365) (0.362) (0.214) (0.222) (0.173) Education: Senior & above -0.127 0.109 0.0515 -0.0500 -0.267 -0.0410 (0.339) (0.295) (0.300) (0.233) (0.223) (0.256) Married 1.314*** 1.471*** 1.025 0.976 0.457 0.518 (0.298) (0.361) (0.828) (0.792) (0.566) (0.491) Separated -0.811 -0.276 0.639 1.778** -0.719 -1.574*** (1.440) (1.493) (0.955) (0.839) (0.590) (0.569) Divorced 2.054*** 1.241 1.430 1.029 0.751 0.447 (0.626) (0.891) (0.937) (0.684) (0.608) (0.586) Widow 1.771*** 1.693*** 1.349* 1.208 0.687 0.744 (0.367) (0.301) (0.774) (0.774) (0.567) (0.510) Java -0.0500 -0.0195 0.0784 -0.129 0.308 -0.00486 (0.161) (0.180) (0.208) (0.170) (0.231) (0.224) Sunda 0.290 0.120 0.0822 0.272 0.516 0.325 (0.233) (0.145) (0.339) (0.217) (0.340) (0.319) Bali -0.443 -0.276 -0.00160 -0.705*** -0.00184 -0.238 (0.265) (0.259) (0.375) (0.196) (0.227) (0.231) Batak 0.0103 -0.103 -0.0368 -0.0690 0.227 -0.309 (0.459) (0.652) (0.709) (0.333) (0.395) (0.452) Bugis -0.0826 0.138 -0.350 -0.366 0.643** -0.316 (0.193) (0.197) (0.420) (0.340) (0.313) (0.314) Tionghoa -2.583** -0.266 -0.920 1.164 -0.522** -0.901** (1.233) (1.050) (0.901) (1.078) (0.226) (0.419) Madura -0.0952 -0.241 0.0471 0.730 0.101 -0.482 (0.343) (0.278) (0.419) (0.571) (0.345) (0.423) Sasak 0.342 0.263 0.456 -0.104 -0.0363 -0.454 (0.294) (0.212) (0.381) (0.259) (0.293) (0.281) Minang -0.379 -0.439 -0.537* 0.0854 0.313 -0.255 (0.248) (0.285) (0.314) (0.180) (0.345) (0.237) Reported poor childhood health 0.0532 -0.107 0.0693 -0.170* -0.130 -0.211* (0.136) (0.0848) (0.107) (0.0940) (0.147) (0.112) Number of rooms of residence 0.0350 0.0121 0.0293 -0.00120 -0.00672 0.000833 lived in when aged 12 (0.0299) (0.0458) (0.0388) (0.00432) (0.00542) (0.00371) Whether electricity in house 0.0839 0.217 0.497** -0.123 -0.0302 -0.0280 lived in age 12 (0.148) (0.328) (0.186) (0.205) (0.276) (0.261) Whether lived with biological 0.0599 0.123 0.282 0.142 0.127 0.184 mother at age 12 (0.204) (0.212) (0.252) (0.164) (0.138) (0.221) Per capita monthly household 0.0119 -0.0674 0.00827 -0.0244 0.0800 0.201** consumption 7/14/17 years ago (0.0325) (0.0729) (0.0753) (0.101) (0.104) (0.0804) Constant -7.902*** -2.388 -3.556 -4.772 -4.042* -6.296*** (1.772) (2.028) (2.563) (3.057) (2.136) (1.998) Observations 970 970 970 1,089 1,089 1,089 R-squared 0.071 0.082 0.102 0.052 0.100 0.112 Omitted category in age:\60 to 64", in education:\No School", in marital status:\Unmarried", in Whether lived with biological mother at age 12:\Not applicable". Estimations include district (Kabupaten) xed eects, standard errors clustered at district level in parenthesis. *** p< 0:01, ** p< 0:05, * p< 0:1. 129 4.5 Shrinkage and Health We now explore whether shrinkage is signicantly associated with later life health, even after con- trolling for baseline height (proxy for childhood health). Among the health variables, we look at current measures of cognition, measured physical strength and reported functional diculties. Ta- bles 4.4 and 4.5 present the summary statistics for health measures in ELSA and IFLS respectively. Mean cognition levels are similar for men and women in England. In Indonesia, women have lower cognition scores. What is striking is the dierence in cognition across the two countries. Both word recall and verbal uency in England are only slightly less than twice than in Indonesia. One of the major potential explanations for these dierences is education levels- the causal eect of which has been showed in many contexts, e.g. see Glymour et al. (2008), Banks and Mazzonna (2012) and Huang and Zhou (2013). A similar pattern exists for other measures of physical health- no remarkable gender dierences in ELSA, but measures for women are worse than men in IFLS; and the English elderly are overall healthier than the Indonesian elderly. Table 4.4: Descriptive Statistics: ELSA VARIABLES N mean sd min max N mean sd min max Male Female Current Age 1608 71.28 7.63 60 90 2035 71.43 7.68 60 90 Word Recall 1608 5.07 1.73 0 10 2035 5.43 1.82 0 10 Fluency (wave 7) 1442 20.95 7.19 0 62 1822 20.41 7.04 0 57 Any ADL/IADL diculty 1608 0.2 0.40 0 1 2035 0.26 0.44 0 1 Some Diculty-Lift/carry 10lbs 1608 0.13 0.34 0 1 2034 0.31 0.46 0 1 Diculty walking 100 yards 1608 0.11 0.31 0 1 2034 0.13 0.34 0 1 Diculty stooping, kneeling or crouching 1608 0.33 0.47 0 1 2034 0.46 0.50 0 1 Diculty getting up from chair 1608 0.21 0.40 0 1 2034 0.29 0.45 0 1 ADL Diculties 1608 0.26 0.72 0 5 2035 0.29 0.75 0 5 IADL Diculties 1608 0.13 0.54 0 5 2035 0.16 0.52 0 5 Lung function (FEV) 1419 2.67 0.79 0 4.9 1752 1.86 0.57 0 3.5 Lung function (HTPEF) 1419 7.61 2.37 0 17.7 1752 5.18 1.63 0 10.6 Grip Strength 1582 35.25 9.18 0 63.7 1969 20.86 6.22 0 40.7 Self Reported Health as Poor 1606 0.27 0.45 0 1 2035 0.27 0.44 0 1 Note: ELSA Waves 1 to 7. 130 Table 4.5: Descriptive Statistics: IFLS VARIABLES N mean sd min max N mean sd min max Male Female Current Age 1162 68.62 7.35 60 95 1382 68.48 6.97 60 94 Word Recall 980 2.99 1.44 1 8 1106 2.76 1.50 1 8.5 Fluency 979 13.95 5.16 0 42 1097 11.99 4.81 0 31 Any ADL/IADL diculty 1155 0.38 0.49 0 1 1373 0.4 0.49 0 1 Some di- carrying heavy load 1155 0.31 0.46 0 1 1373 0.51 0.50 0 1 Some di- walking 1 km 1155 0.3 0.46 0 1 1373 0.48 0.50 0 1 Some di- bowing, squatting or kneeling 1155 0.17 0.37 0 1 1373 0.27 0.44 0 1 Some di- getting up from chair 1155 0.06 0.24 0 1 1373 0.1 0.30 0 1 ADL Diculties 777 0.19 0.66 0 5 1162 0.25 0.76 0 5 IADL Diculties 1155 1.03 1.74 0 6 1373 0.99 1.75 0 6 Lung Function (PEF) 1126 306.37 114.03 0 733.3 1293 201.76 70.40 0 396.7 Grip Strength 1097 24.82 7.32 0 48 1276 15.8 5.35 0 47 Self Reported Health as Poor 1155 0.37 0.48 0 1 1373 0.41 0.49 0 1 Dead in Wave 5 1643 0.23 0.42 0 1 1918 0.16 0.37 0 1 Note: IFLS Waves 2 to 5. Following Huang et al. (2013), we rst regress each of these variables on current height in equation (3), which has been standard in the literature. We then regress current health on height w years ago and height loss over time t-w and t, as in equation (4). health i;t = +height i;t + Z i + i;t (4.3) health i;t = + 0 height i;tw + 1 shrinkage i;t;tw + Z i + i;t (4.4) We also replace the continuous measure of height-shrinkage into four groups, as classied by Wannamethee et al. (2006), dummies for whether shrinkage i;t;tw falls under categories < 1cm, 1-1.9cm, 2-2.9cm, or 3 cm, to explore non-linearities 15 . health i;t = + 0 height i;tw + 3 X c=0 ! c (shrinkage c ) i;t;tw + Z i + i;t (4.5) As much as each panel data allows us to, we can look at short term and long term shrinkage eects on health. That is, ELSA allows us to explore height shrinkage over four and eight years, 15 Using dummy variables also reduces bias due to measurement error in the continuous shrinkage variable. 131 IFLS allows to look at height shrinkage over seven, fourteen and seventeen years. Age categories and SES variables are all used as controls (Z i ) as was done in estimating equation (2). Coecients in equations (4) and (5) do not have a causal interpretation since there is no exogenous variation in shrinkage, because of omitted variables (like health behaviors), and potential reverse causality. We expect an association since the mechanisms related to height shrinkage, like osteoporosis and specic health behaviors, also aect the health outcomes we look at. Note that we limit our sample to respondents for whom both 4-year and 8-year shrinkage can be calculated in ELSA, and to those for whom 7-year, 14-year and 17-year shrinkage can be calculated in IFLS. Since respondents are more likely to drop out of the the sample over subsequent panel rounds due to declining health and/or mortality given the age groups we are looking at, the coecients would be underestimated since we would expect the healthier of respondents to survive over the years. We conrm this in a later section where we do in fact nd that extreme height shrinkage in the years preceding death is signicantly associated with mortality. The strength of these associations might also be underestimated in our sample since we allow equal maximum values for height gain and loss. 4.5.1 Shrinkage and Cognition In order to compare estimates across England and Indonesia, we rst look at word recall which is a measure of short term memory (McArdle, Smith and Willis, 2011). This is a mean of immediate and delayed word recall, hence ranges from 0 to 10 16 . Results for word recall from ELSA are presented in Table 4.6. First, coecients on current height and lagged height (by 8 years) are of similar magnitudes. These have been interpreted in the literature as the eects of early life health on later life cognition. Hence, both current and lagged height capture early life health to a similar extent. Next, we look at the coecients on shrinkage. While shrinkage over both 4 and 8 years 16 Other variables measuring cognition in ELSA- verbal uency and prospective memory task, while asked from Wave 1 to 5, are not available for Wave 6. We use verbal uency from Wave 7. 132 is associated with reduced memory scores, the relationship is not signicant. However, extreme height-shrinkage over eight years, i.e. over 3 cm, is signicantly associated with lower immediate word recall for both men and women. For men, extreme height shrinkage over 8 years is associated with almost one less word recalled 17 . Next, we look at verbal uency (measured by number of animals named within a given time period) in Table 4.7. For English elderly men, both current and lagged heights are signicantly related with higher scores, and the strength of association is stronger for lagged height measures. Only extreme height shrinkage over 8 years is signicantly associated with lower uency. For women, verbal uency is not related with measures of height to begin with. We do see a negative signicant association extreme 4 year shrinkage. While we would expect the associations to be stronger for a longer time measure of shrinkage since height loss occurs over a longer term, losing height in a short time span could indicate more rapidly worsening health. As in ELSA, we look at word recall in the Indonesian sample. IFLS-5 introduced more cognition variables, allowing us to also look at verbal uency, measured similarly as in ELSA. In the IFLS estimations, we also include Kabupaten (district) xed eects, and cluster standard errors at the same level. Among the sample of Indonesian elderly men, we do not see word recall scores to be related with either height (current or lagged) or height-shrinkage (Table 4.8). For women, however, we do see a signicant negative association with height-shrinkage over 17 years. Verbal uency is positively related with current height for both genders (Table 4.9), but the association doesn't hold for lagged height measures. Among the men that we also see a signicant and negative association with 14-year and 17-year shrinkage, which in turn is seemingly driven by measures of extreme height shrinkage. For the women, while shrinkage coecients are negative, there are no signicant non-linearities. 17 0.5 less on memory score translates into one word less on either immediate or delayed recall 133 Table 4.6: Height Shrinkage and Cognition (Word Recall) in ELSA (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Male sample Panel B: Female sample Dependent Variable: Word Recall 4 years 8 years 4 years 8 years Current Height 0.0125* 0.0135* (0.00708) (0.00716) Height 4/8 yrs ago 0.0129* 0.0120* 0.0123* 0.0134* 0.0124* 0.0119 0.0115 0.0121* (0.00715) (0.00712) (0.00721) (0.00718) (0.00727) (0.00727) (0.00730) (0.00732) Height loss over 4/8 yrs -0.0213 -0.0362 -0.0326 -0.0420 (0.0283) (0.0289) (0.0281) (0.0274) Height loss: 1-1.9cm 0.284** 0.140 -0.0674 -0.0745 (0.121) (0.116) (0.109) (0.111) Height loss: 2-2.9cm -0.256* -0.0714 -0.0997 -0.217* (0.147) (0.138) (0.147) (0.130) Height loss: 3cm + -0.234 -0.433*** -0.00195 -0.267* (0.183) (0.149) (0.190) (0.150) Observations 1,040 1,040 1,040 1,040 1,040 1,165 1,165 1,165 1,165 1,165 R-squared 0.238 0.234 0.243 0.234 0.243 0.229 0.226 0.225 0.228 0.229 Note: Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Age, Education, SES controls suppressed. 134 Table 4.7: Height Shrinkage and Fluency (Animal Naming) in ELSA (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Male sample Panel B: Female sample Dependent Variable: Fluency 4 years 8 years 4 years 8 years Current Height 0.0637** 0.0313 (0.0323) (0.0308) Height 4/8 yrs ago 0.0682** 0.0674** 0.0660** 0.0677** 0.0239 0.0237 0.0241 0.0223 (0.0325) (0.0325) (0.0328) (0.0327) (0.0311) (0.0310) (0.0314) (0.0315) Height loss over 4/8 yrs -0.0575 -0.147 -0.245** -0.169 (0.133) (0.135) (0.122) (0.120) Height loss: 1-1.9cm 0.812 0.925* -0.464 -0.497 (0.557) (0.535) (0.463) (0.475) Height loss: 2-2.9cm -0.528 0.487 -0.531 0.111 (0.679) (0.630) (0.638) (0.562) Height loss: 3cm + -1.164 -1.339* -2.335*** -1.058 (0.854) (0.693) (0.834) (0.649) Observations 947 947 947 947 947 1,059 1,059 1,059 1,059 1,059 R-squared 0.170 0.166 0.171 0.166 0.174 0.186 0.189 0.192 0.186 0.188 Note: Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Age, Education, SES controls suppressed. 135 Table 4.8: Height Shrinkage and Cognition (Word Recall) in IFLS (1) (2) (3) (4) (5) (6) (7) Panel A: Male sample Dependent Variable: Word Recall 7 years 14 years 17 years Current Height 0.00474 (0.00808) Height 7/14/17 yrs ago 0.00355 0.00286 0.00279 0.00229 0.00214 0.00147 (0.00752) (0.00731) (0.00862) (0.00858) (0.00868) (0.00888) Height loss over 7/14/17 yrs -0.0149 -0.0209 -0.0237 (0.0258) (0.0264) (0.0217) Height loss: 1-1.9cm 0.0254 -0.0907 0.0837 (0.171) (0.102) (0.159) Height loss: 2-2.9cm 0.165 0.0763 -0.0132 (0.187) (0.115) (0.126) Height loss: 3cm + -0.156 -0.160 0.0147 (0.178) (0.214) (0.182) Observations 821 823 823 823 823 823 823 R-squared 0.144 0.145 0.146 0.145 0.147 0.145 0.145 (1) (2) (3) (4) (5) (6) (7) Panel B: Female sample Dependent Variable: Word Recall 7 years 14 years 17 years Current Height 0.0112 (0.00898) Height 7/14/17 yrs ago 0.0127 0.0130 0.0106 0.0110 0.00558 0.00672 (0.00947) (0.00945) (0.00903) (0.00886) (0.00973) (0.00951) Height loss over 7/14/17 yrs -0.00486 -0.0309 -0.0513*** (0.0242) (0.0279) (0.0172) Height loss: 1-1.9cm -0.114 -0.236* 0.0160 (0.124) (0.130) (0.130) Height loss: 2-2.9cm -0.172 -0.236 -0.320*** (0.145) (0.180) (0.111) Height loss: 3cm + -0.0158 -0.251* -0.236* (0.184) (0.148) (0.124) Observations 828 828 828 828 828 828 828 R-squared 0.211 0.210 0.212 0.207 0.210 0.213 0.217 Note: Robust standard errors in parentheses.*** p < 0:01, ** p < 0:05, * p < 0:1. Age, Education, SES controls suppressed. IFLS results control for district xed eects, cluster errors at the same level. 136 Table 4.9: Height Shrinkage and Fluency (Animal Naming) in IFLS (1) (2) (3) (4) (5) (6) (7) Panel A: Male sample Dependent Variable: Fluency 7 years 14 years 17 years Current Height 0.0467** (0.0202) Height 7/14/17 yrs ago 0.0477** 0.0494** 0.0332 0.0320 0.0297 0.0303 (0.0214) (0.0213) (0.0215) (0.0215) (0.0242) (0.0248) Height loss over 7/14/17 yrs -0.0175 -0.152** -0.239*** (0.0811) (0.0586) (0.0573) Height loss: 1-1.9cm -0.0124 -0.177 -0.325 (0.390) (0.394) (0.376) Height loss: 2-2.9cm -0.0943 -0.498 -1.296*** (0.595) (0.489) (0.333) Height loss: 3cm + -0.742 -1.092*** -1.051*** (0.584) (0.387) (0.347) Observations 966 969 969 969 969 969 969 R-squared 0.162 0.164 0.165 0.167 0.169 0.169 0.171 (1) (2) (3) (4) (5) (6) (7) Panel B: Female sample Dependent Variable: Fluency 7 years 14 years 17 years Current Height 0.0365* (0.0201) Height 7/14/17 yrs ago 0.0412* 0.0397 0.0324 0.0308 0.0305 0.0296 (0.0232) (0.0235) (0.0204) (0.0201) (0.0212) (0.0212) Height loss over 7/14/17 yrs -0.0189 -0.0749 -0.0866** (0.0586) (0.0531) (0.0345) Height loss: 1-1.9cm 0.0446 0.0784 -0.0243 (0.205) (0.302) (0.383) Height loss: 2-2.9cm 0.505 0.208 -0.187 (0.356) (0.376) (0.225) Height loss: 3cm + -0.0175 -0.119 -0.263 (0.422) (0.340) (0.284) Observations 1,082 1,083 1,083 1,083 1,083 1,083 1,083 R-squared 0.219 0.217 0.218 0.219 0.219 0.222 0.221 Note: Robust standard errors clustered at district level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Age, Education, SES controls suppressed. IFLS results control for district xed eects. 137 4.5.2 Shrinkage and Physical Health We rst look at two measured (or objective) health variables: lung function, and grip strength of the dominant hand which are predictors of old age physical health, including disability (Rantanen et al., 1999). While ELSA provides multiple measures of lung function, we use Forced Expiratory Volume since this was measured in all the relevant waves. In IFLS-5, lung capacity was measured using a Vitalograph peak ow meter for all respondents 9 years of age and above. The variable is Peak Expiratory Flow (L/min), which measures how fast the respondent can exhale. We use an average of the three measurements. For grip strength in both ELSA and IFLS, we use an average of the second and third measurements. Current height, as well as longer term height-shrinkage (that is eight years in ELSA sample) are signicantly associated with lung function for both genders (Table 4.10). Note that the absolute value of height-shrinkage coecients are of similar magnitudes as those of the height coecients. Patterns for grip strength are very similar (Table 4.11). Hence, taller people have better lung function and higher grip strength, but height-shrinkage is even more strongly related with reduced physical functioning capacity. Patterns for lung function and grip strength are very similar in the IFLS samples (Tables 4.12 & 4.13). The extent to which shrinkage is associated with reduced grip strength is higher for women than men since they have lower grip strength to begin with. Finally, we also look at manifested functional diculties, measured by reported disability, and diculty in activities requiring muscle strength. We use a binary measure of disability- an indicator variable which equals one if the respondent reports some or more diculty with any of the 5 ADL's or 6 IADL's. In both the ELSA and IFLS sample (Tables 4.14 & 4.16), this measure is very weakly (and negatively) related with height. Height-shrinkage is, however, signicantly associated with higher disability among English men, and both Indonesian men and women. We also explore 138 these separately for ADL and IADL Indexes, which measure the number of activities with which respondents report some or more diculty (Appendix Tables C2 to C5). The patterns are similar to those for the binary disability variable, hence we do not report these in the main tables. We also look at diculty in kneeling and related actions, which indicates decreased muscle strength. We do not nd signicant relation with height-shrinkage in the ELSA sample (Table 4.15). For Indonesian men, we see that long term shrinkage is associated with higher likelihood of reported diculty with squatting or kneeling (Table 4.17). For Indonesian women, we see an even stronger association, for all measures of shrinkage. The coecients are from 1.65 to 1.85 percentage points, which is of non trivial magnitude given that 27% women report this diculty 18 . Based on the above discussion, we can see that height-shrinkage is systematically associated with old-age specic health measures, even when these measures are often not correlated with height. 4.5.3 Shrinkage and Mortality Since height-shrinkage is a marker for worsening later life health, an association conrmed by Sec- tions 4.5.1 and 4.5.2, we explore whether it is also a predictor of mortality in these age groups. There is some evidence about this association in the medical literature. For example, Wannamethee et al. (2006) look at how height loss over 20 years predicts cardiovascular disease and mortality in a sam- ple of English men. They nd total mortality risk to be higher for men who experience height loss of 3 cm or more. Auyeung et al. (2010) nd that all cause mortality is higher for men experiencing height loss exceeding 2 cm over 4 years, but not for women, in a sample of Chinese elderly aged 65 and above. Masunari et al. (2012) nd a similar association in a Japanese sample- while they also 18 We also checked for diculty with carrying a heavy load, or diculty with getting up from a chair after sitting- also measures of old age muscle strength. We found very similar patterns, hence those results are reported in Appendix Tables C6, C7, C8 and C9. 139 dene marked height loss as above 2 cm, their sample is younger at baseline, starting from age 47. Hillier et al. (2012) nd that height loss exceeding 5 cm over 15 years is associated with increased mortality among American women aged 65 years and above. We are able to investigate this hypothesis in the IFLS sample, by using whether the respondent was reported dead in wave 5 as the outcome variable. We keep respondents who were of age 53 or higher in Wave 4, and either are or would have been of age 60 or higher by Wave 5. We use change in height from wave 4 (which would be the last available measurement of height for respondents dead in wave 5) and wave 3 or wave 2, giving measures of height-shrinkage over 7 or 10 years, going back in time 7 years or less depending on when the Respondent died after 2008 and before 2014-15. Note that this sample size would be bigger than the sample we used for studying health associations 19 . While a continuous measure of height-shrinkage is not signicantly correlated with mortality, height loss exceeding 3 cm over 7 as well as 10 years is a positive and signicant predictor for both men and women (Table 4.18). Coecient of 7-year pre death shrinkage is higher than that on 10-year shrinkage for men, suggesting that extreme height loss over a shorter time period is an indicator of more rapidly worsening health. The magnitude of the coecients is large, at 5 percentage points or higher, when looking at mean mortality- 23% for men and 16% for women. In these estimations, we also control for whether the respondent self reported his/her health as poor in the previous waves, which has been found to be an important predictor of mortality among the elderly (Benjamins et al., 2004). 19 since health measurements are available for respondents alive in Wave 5, a subset of the mortality sample. 140 Table 4.10: Height Shrinkage and Lung Function (Forced Expiratory Volume) in ELSA (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Male sample Panel B: Female sample Dependent Variable: FEV 4 years 8 years 4 years 8 years Current Height 0.0338*** 0.0230*** (0.00332) (0.00256) Height 4/8 yrs ago 0.0344*** 0.0338*** 0.0339*** 0.0341*** 0.0234*** 0.0235*** 0.0229*** 0.0228*** (0.00334) (0.00334) (0.00337) (0.00337) (0.00259) (0.00258) (0.00261) (0.00261) Height loss over 4/8 yrs -0.00771 -0.0342** -0.0131 -0.0222** (0.0138) (0.0139) (0.0104) (0.0102) Height loss: 1-1.9cm 0.102* -0.0289 0.0263 0.0137 (0.0572) (0.0552) (0.0386) (0.0394) Height loss: 2-2.9cm -0.0696 -0.0469 -0.0942* -0.0320 (0.0704) (0.0660) (0.0528) (0.0465) Height loss: 3cm + -0.0561 -0.235*** -0.0692 -0.126** (0.0865) (0.0714) (0.0723) (0.0552) Observations 850 850 850 850 850 893 893 893 893 893 R-squared 0.340 0.343 0.347 0.340 0.344 0.295 0.293 0.296 0.293 0.294 Note: Robust standard errors in parentheses. *** p< 0:01, **p< 0:05, *p< 0:1. Age, Education, SES controls suppressed. FEV indicates the volume of air exhaled under forced conditions. 141 Table 4.11: Height Shrinkage and Grip Strength in ELSA (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Male sample Panel B: Female sample Dependent Variable: Grip Strength 4 years 8 years 4 years 8 years Current Height 0.334*** 0.198*** (0.0366) (0.0261) Height 4/8 yrs ago 0.339*** 0.346*** 0.335*** 0.338*** 0.198*** 0.198*** 0.198*** 0.200*** (0.0367) (0.0367) (0.0371) (0.0370) (0.0264) (0.0263) (0.0265) (0.0265) Height loss over 4/8 yrs -0.154 -0.405*** -0.177* -0.214** (0.145) (0.147) (0.104) (0.102) Height loss: 1-1.9cm -0.682 -0.364 -0.591 0.108 (0.620) (0.594) (0.392) (0.406) Height loss: 2-2.9cm -1.558** -1.053 -0.995* -0.742 (0.744) (0.704) (0.537) (0.465) Height loss: 3cm + -1.584* -2.650*** -1.065 -1.530*** (0.939) (0.763) (0.695) (0.555) Observations 1,007 1,007 1,007 1,007 1,007 1,097 1,097 1,097 1,097 1,097 R-squared 0.315 0.316 0.320 0.316 0.319 0.239 0.240 0.242 0.237 0.242 Note: Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Age, Education, SES controls suppressed. 142 Table 4.12: Height Shrinkage and Lung Function (Peak Expiratory Flow) in IFLS (1) (2) (3) (4) (5) (6) (7) Panel A: Male sample Dependent Variable: PEF 7 years 14 years 17 years Current Height 3.263*** (0.657) Height 7/14/17 yrs ago 3.230*** 3.277*** 3.029*** 3.002*** 3.297*** 3.301*** (0.696) (0.709) (0.620) (0.615) (0.633) (0.614) Height loss over 7/14/17 yrs: -2.882* -3.802 -3.204* (1.567) (2.322) (1.862) Height loss: 1-1.9cm -5.688 -1.538 9.796 (9.880) (8.172) (9.035) Height loss: 2-2.9cm -28.54** -12.62 -13.57 (13.10) (8.503) (10.36) Height loss: 3cm + -25.29** -18.25 -21.68** (12.35) (12.21) (9.975) Observations 947 949 949 949 949 949 949 R-squared 0.228 0.230 0.234 0.234 0.234 0.229 0.236 (1) (2) (3) (4) (5) (6) (7) Panel B: Female sample Dependent Variable: PEF 7 years 14 years 17 years Current Height 1.924*** (0.458) Height 7/14/17 yrs ago 1.855*** 1.854*** 1.735*** 1.724*** 1.729*** 1.755*** (0.500) (0.505) (0.423) (0.413) (0.485) (0.477) Height loss over 7/14/17 yrs: -2.773** -4.048*** -3.612*** (1.041) (1.303) (0.967) Height loss: 1-1.9cm -5.862 -5.053 -4.375 (4.590) (6.096) (4.935) Height loss: 2-2.9cm -11.69** -15.07* -12.12* (5.207) (7.892) (6.727) Height loss: 3cm + -23.64*** -20.80*** -22.11*** (5.525) (5.696) (5.011) Observations 1,014 1,015 1,015 1,015 1,015 1,015 1,015 R-squared 0.171 0.173 0.177 0.173 0.173 0.172 0.174 Note: Robust standard errors clustered at district level in parentheses. *** p < 0:01, ** p < 0:05, * p < 0:1. Age, Education, SES controls suppressed. IFLS results control for district xed eects. PEF is the highest forced expiratory ow measured with a peak ow meter. 143 Table 4.13: Height Shrinkage and Grip Strength in IFLS (1) (2) (3) (4) (5) (6) (7) Panel A: Male sample Dependent Variable: Grip Strength 7 years 14 years 17 years Current Height 0.433*** (0.0408) Height 7/14/17 yrs ago 0.442*** 0.441*** 0.430*** 0.422*** 0.437*** 0.436*** (0.0417) (0.0435) (0.0407) (0.0397) (0.0408) (0.0402) Height loss over 7/14/17 yrs: -0.194* -0.365** -0.370*** (0.0982) (0.145) (0.111) Height loss: 1-1.9cm 0.212 0.662 -0.0272 (0.390) (0.481) (0.518) Height loss: 2-2.9cm -0.373 0.338 -1.172** (0.911) (0.617) (0.472) Height loss: 3cm + -1.991** -2.518*** -2.150*** (0.978) (0.606) (0.716) Observations 890 893 893 893 893 893 893 R-squared 0.296 0.300 0.302 0.297 0.308 0.296 0.299 (1) (2) (3) (4) (5) (6) (7) Panel B: Female sample Dependent Variable: Grip Strength 7 years 14 years 17 years Current Height 0.241*** (0.0283) Height 7/14/17 yrs ago 0.246*** 0.245*** 0.234*** 0.232*** 0.235*** 0.236*** (0.0279) (0.0273) (0.0286) (0.0286) (0.0304) (0.0296) Height loss over 7/14/17 yrs: -0.226*** -0.329*** -0.325*** (0.0738) (0.0932) (0.0716) Height loss: 1-1.9cm -0.284 -0.0156 0.463 (0.357) (0.493) (0.453) Height loss: 2-2.9cm -1.253** -0.223 -0.685** (0.461) (0.381) (0.315) Height loss: 3cm + -1.629*** -1.720*** -1.479*** (0.551) (0.451) (0.380) Observations 1,016 1,017 1,017 1,017 1,017 1,017 1,017 R-squared 0.214 0.209 0.214 0.211 0.215 0.210 0.213 Note: Robust standard errors clustered at district level in parentheses. *** p < 0:01, ** p < 0:05, * p < 0:1. Age, Education, SES controls suppressed. IFLS results control for district xed eects. 144 Table 4.14: Height Shrinkage and Disability in ELSA (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Male sample Panel B: Female sample Dep Var: Any diculty in ADL/IADL 4 years 8 years 4 years 8 years Current Height -0.00176 -0.00325 (0.00186) (0.00199) Height 4/8 yrs ago -0.00142 -0.00117 -0.00153 -0.00170 -0.00234 -0.00227 -0.00233 -0.00220 (0.00187) (0.00187) (0.00188) (0.00188) (0.00201) (0.00201) (0.00202) (0.00202) Height loss over 4/8 yrs 0.00512 0.00964 0.0234*** 0.0213*** (0.00739) (0.00755) (0.00775) (0.00758) Height loss: 1-1.9cm -0.0333 -0.0416 0.0212 -0.0648** (0.0319) (0.0304) (0.0301) (0.0307) Height loss: 2-2.9cm -0.0221 -0.00505 0.104** 0.0239 (0.0385) (0.0362) (0.0407) (0.0357) Height loss: 3cm + 0.0377 0.0916** 0.132** 0.137*** (0.0479) (0.0391) (0.0521) (0.0412) Observations 1,042 1,042 1,042 1,042 1,042 1,166 1,166 1,166 1,166 1,166 R-squared 0.053 0.054 0.055 0.055 0.063 0.086 0.092 0.094 0.091 0.102 Note: Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Age, Education, SES controls suppressed. Table 4.15: Height Shrinkage and Diculty in Kneeling, Stooping or Crouching in ELSA (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Male sample Panel B: Female sample Dep Var: Any diculty in kneeling etc. 4 years 8 years 4 years 8 years Current Height 0.00165 0.00463** (0.00215) (0.00230) Height 4/8 yrs ago 0.00220 0.00201 0.00240 0.00230 0.00577** 0.00604*** 0.00543** 0.00534** (0.00216) (0.00216) (0.00218) (0.00217) (0.00232) (0.00232) (0.00234) (0.00235) Height loss over 4/8 yrs 0.00601 0.0118 0.0191** 0.0122 (0.00854) (0.00873) (0.00894) (0.00877) Height loss: 1-1.9cm -0.00470 -0.0141 0.0667* -0.00311 (0.0369) (0.0352) (0.0347) (0.0357) Height loss: 2-2.9cm 0.0186 -0.0433 0.0491 0.0541 (0.0445) (0.0420) (0.0470) (0.0416) Height loss: 3cm + 0.107* 0.0981** 0.0441 0.0434 (0.0554) (0.0453) (0.0602) (0.0480) Observations 1,042 1,042 1,042 1,042 1,042 1,166 1,166 1,166 1,166 1,166 R-squared 0.068 0.069 0.072 0.069 0.075 0.066 0.076 0.076 0.070 0.071 Note: Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Age, Education, SES controls suppressed. 145 Table 4.16: Height Shrinkage and Disability in IFLS (1) (2) (3) (4) (5) (6) (7) Panel A: Male sample Dep Var: Any diculty in ADL/IADL 7 years 14 years 17 years Current Height -0.00422 (0.00275) Height 7/14/17 yrs ago -0.00337 -0.00314 -0.00295 -0.00259 -0.00244 -0.00250 (0.00260) (0.00267) (0.00272) (0.00268) (0.00255) (0.00256) Height loss over 7/14/17 yrs: 0.0181** 0.0198** 0.0269*** (0.00845) (0.00900) (0.00842) 1-1.9cm -0.0204 -0.0519 0.0626 (0.0445) (0.0452) (0.0392) 2-2.9cm 0.0450 0.0333 0.0776* (0.0782) (0.0418) (0.0416) 3cm + 0.179** 0.131** 0.158*** (0.0767) (0.0544) (0.0397) Observations 967 970 970 970 970 970 970 R-squared 0.054 0.057 0.061 0.058 0.066 0.065 0.066 (1) (2) (3) (4) (5) (6) (7) Panel B: Female sample Dep Var: Any diculty in ADL/IADL 7 years 14 years 17 years Current Height -0.00502** (0.00238) Height 7/14/17 yrs ago -0.00399 -0.00389 -0.00368 -0.00347 -0.00398 -0.00376 (0.00262) (0.00261) (0.00245) (0.00243) (0.00252) (0.00254) Height loss over 7/14/17 yrs: 0.0117 0.0149** 0.0110* (0.00964) (0.00646) (0.00587) Height loss: 1-1.9cm 0.0139 -0.0156 0.0135 (0.0436) (0.0330) (0.0414) Height loss: 2-2.9cm -0.0300 -0.0285 -0.0199 (0.0522) (0.0409) (0.0375) Height loss: 3cm + 0.161** 0.0479 0.0666* (0.0635) (0.0388) (0.0386) Observations 1,088 1,089 1,089 1,089 1,089 1,089 1,089 R-squared 0.094 0.093 0.102 0.093 0.093 0.092 0.094 Note: Robust standard errors clustered at district level in parentheses. *** p < 0:01, ** p < 0:05, * p < 0:1. Age, Education, SES controls suppressed. IFLS results control for district xed eects. 146 Table 4.17: Height Shrinkage and Diculty in Kneeling, Bowing or Squatting in IFLS (1) (2) (3) (4) (5) (6) (7) Panel A: Male sample Dep Var: Any diculty in kneeling etc. 7 years 14 years 17 years Current Height -0.00195* (0.00105) Height 7/14/17 yrs ago -0.00200* -0.00198* -0.00111 -0.000943 -0.00139 -0.00133 (0.00104) (0.00103) (0.00108) (0.00109) (0.00118) (0.00115) Height loss over 7/14/17 yrs: -0.000330 0.00956 0.00909* (0.00418) (0.00665) (0.00533) Height loss: 1-1.9cm 0.0154 0.00259 0.00755 (0.0271) (0.0251) (0.0276) Height loss: 2-2.9cm -0.0555* -0.00265 0.0378 (0.0315) (0.0300) (0.0363) Height loss: 3cm + 0.0190 0.0579 0.0156 (0.0493) (0.0437) (0.0368) Observations 967 970 970 970 970 970 970 R-squared 0.081 0.086 0.088 0.086 0.087 0.084 0.083 (1) (2) (3) (4) (5) (6) (7) Panel B: Female sample Dep Var: Any diculty in kneeling etc. 7 years 14 years 17 years Current Height -0.000518 (0.00230) Height 7/14/17 yrs ago 0.000992 0.00106 0.00143 0.00128 0.00176 0.00166 (0.00245) (0.00245) (0.00238) (0.00234) (0.00220) (0.00216) Height loss over 7/14/17 yrs: 0.0135** 0.0185*** 0.0165** (0.00547) (0.00528) (0.00757) Height loss: 1-1.9cm 0.0713** 0.0476** 0.0251 (0.0305) (0.0209) (0.0269) Height loss: 2-2.9cm 0.0741* 0.0609** 0.0354 (0.0403) (0.0278) (0.0271) Height loss: 3cm + 0.0662 0.144*** 0.133*** (0.0394) (0.0309) (0.0400) Observations 1,088 1,089 1,089 1,089 1,089 1,089 1,089 R-squared 0.093 0.095 0.099 0.099 0.106 0.098 0.105 Note: Robust standard errors clustered at district level in parentheses. *** p < 0:01, ** p < 0:05, * p < 0:1. Age, Education, SES controls suppressed. IFLS results control for district xed eects. 147 Table 4.18: Height Shrinkage and Mortality in IFLS (1) (2) (3) (4) (5) Panel A: Male sample Dependent Variable: Whether dead in wave 5 7 years 10 years Last measured Height (Wave 4) -0.00113 (0.00167) Height 7/10 yrs before death (Wave 3/2) -0.000952 -0.000915 -0.00127 -0.00133 (0.00173) (0.00173) (0.00175) (0.00174) Height loss over 7/10 yrs before death 0.00747 0.00271 (Wave 3/2- Wave 4) (0.00545) (0.00475) Height loss: 1-1.9cm -0.00175 -0.0385 (0.0229) (0.0236) Height loss: 2-2.9cm -0.0150 -0.0353 (0.0300) (0.0300) Height loss: 3cm + 0.118*** 0.0588* (0.0386) (0.0342) Observations 1,642 1,642 1,642 1,642 1,642 (1) (2) (3) (4) (5) Panel B: Female sample Dependent Variable: Whether dead in wave 5 7 years 10 years Last measured Height (Wave 4) -0.00171 (0.00148) Height 7/10 yrs before death (Wave 3/2) -0.00118 -0.00119 -0.00117 -0.00119 (0.00155) (0.00155) (0.00156) (0.00156) Height loss over 7/10 yrs before death 0.00586 0.00504 (Wave 3/2- Wave 4) (0.00393) (0.00365) Height loss: 1-1.9cm -0.00769 0.00934 (0.0198) (0.0202) Height loss: 2-2.9cm 0.0274 0.0290 (0.0243) (0.0246) Height loss: 3cm + 0.0537* 0.0552** (0.0289) (0.0266) Observations 1,918 1,918 1,918 1,918 1,918 Note: Robust standard errors clustered at district level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Age, Education, SES controls suppressed. IFLS results control for district xed eects. 148 4.6 Strengths and Limitations We get estimates of shrinkage based on repeated measurements of height for the same individuals, these would be more accurate than using self-reported height. Most of the medical literature on shrinkage has used clinical data or community level samples, whereas we use nationally represen- tative samples. We are also able to demonstrate the negative association between shrinkage and health using both measured and self reported variables. Both the extent of height loss and the strength of shrinkage-health associations are underes- timated in our study. We are also unable to provide causal estimates of the eect of shrinkage on worsening old-age health. It is also unclear whether early life health environment would aect shrinkage only through aecting height, or whether there exist other biological pathways. Twin studies tracking respondents up until their old age would be very informative. It would be interest- ing to explore whether identical twins experience dissimilar levels of shrinkage, and if so, whether there are any SES or behavioral factors. Another potentially helpful approach would be to study height-shrinkage in a scenario where health insurance coverage is introduced in a phased manner while respondents are in their middle age so that insurance only aects later life health. 4.7 Conclusion In this paper, we document and compare the magnitude of height-shrinkage over time and across English and Indonesian elderly. We nd that women suer from greater height loss, which is consistent with the medical literature. We also nd that Indonesian elderly, even though shorter to begin with, lose more height than the elderly in England. This suggests worse early life as well as later life health. While we nd strong age gradients in height-shrinkage, we do not nd 149 any relation with current or childhood SES variables. We then decompose the standard height- health associations into height-health and shrinkage-health associations. We use both continuous measures and categories of the extent loss, to address measurement error in height, as well as to explore non-linearities in the relationship between shrinkage and later-life health. Extreme height shrinkage is signicantly associated with lower cognitive scores measuring memory and uency in both Indonesian and English older adults. We also nd negative associations between height shrinkage and physical function. Finally, extreme height loss is found to be a predictor of mortality in the Indonesian elderly sample. The association between shrinkage and later life health is expected to be underestimated in our study samples. Overall, shrinkage is found to be an equal or in some cases, even more important, covariate than baseline height for the older-age health outcomes we study. Our results suggest that shrinkage is an informative, and easy to measure, biomarker for worsening old-age health at the population level where more detailed measures might not be available. 150 Chapter 5 Conclusion This dissertation contributes to the broad literature in human capital development, with focus on its transmission along familial ties and over time. Although peer eects, intergenerational transmission of socioeconomic status and the early childhood determinants of later life outcomes have been widely documented in previous studies, the causal relationship of health and education between spouses and from younger generations to older ones as well the interaction of early life health, aging process and late life health are still insuciently explored. With rapid population aging across the globe, policymakers and academics have been paying greater attention to health disparities in old age. Thus this dissertation adds to the growing literature in economics of aging by studying the transmission of human capital within family networks among older adults, and bio-demographic measures of late life health in both developing and developed countries. Empirical analyses in this dissertation utilize aging and household survey data from multiple countries which dier in levels of economic development, population aging and social security system. Chapter 2 uses data from the English Longitudinal Study of Ageing (2000{2014) and examine in a dynamic panel model the causal spillovers of health among English elderly couples. Chapter 3 investigates the causal impacts of children's education on health and cognition of Chinese older adults, by applying an instrumental variables approach to the data from the rst two waves 151 of China Health and Retirement Longitudinal Study (CHARLS). Chapter 4 draws upon both the English Longitudinal Study of Ageing (ELSA) and the Indonesian Family Life Survey (IFLS) in comparing the associations of height, height loss and old age health in UK and Indonesia. These three datasets provide comparable measures in multiple dimensions of health and family networks, allowing for systematic analyses on various components of the inter-relationship between them. Results from Chapter 2 suggest that there are signicant spousal impact in the health domain of human capital among elder couples in England. By controlling for endogeneity problems due to assortative mating and shared environment, we nd evidence for strong spillover eects from spousal self-reported health, depressive symptoms, smoking behaviors, drinking frequency, physical activities and social contact. We nd some gender dierences. For instance, women's mental health and smoking intensity seem to be more aected by their spouses, while the opposite is true for men's smoking status. The causal analysis from Chapter 3 shows that health and cognition of Chinese older adults benet from increasing human capital of their children, measured in years of education of the highest educated adult child. Signicant causal eects of children's education are found on parents' episodic memory, mental intactness, subjective survival expectation, lung function and body weight measures including BMI. Further estimations imply that children might shape parental health by providing nancial support, material resources and by improving parental psychological well-being and reducing old age labor supply. In addition to its level eects, children's education is also benecial for the maintenance of cognitive functions of older adults, as indicated by the estimation of a dynamic model on the 2-year follow-up data, where lagged baseline health outcomes are included and instrumented using Lewbel's IV approach. By incorporating the loss of height which is a common proxy for early life health, Chapter 4 demonstrates that height shrinkage is signicantly associated with lower cognition, reduced physical function and mortality for elderly men and women in both Indonesia and England, when baseline height is controlled for. Overall, 152 shrinkage is found to be an equal or in some cases even more important than baseline height for the late life health outcomes we study. There are important policy implications from the analyses in this dissertation. First, ndings from the rst two essays underscore the benet of health and education policies and investments in improving population health. Second, interventions at individual level may spread along familial ties. Incorporating the potential spillover eects on spouses and multiple generations, might be essential for policy designs aiming to reduce health disparities. Third, in developing countries like China which experience unprecedented population aging, human capital policies on family or social networks might be promising in complementing the public support system, which is still in its infancy. Last, the results from Chapter 4 suggest that policymakers and public health professionals could use height shrinkage as an informative and supplemental biomarker for old-age health at the population level where it is hard to get more detailed measures because of logistic constraints. 153 Bibliography Almond, Douglas, and Janet Currie. 2011. \Killing me softly: The fetal origins hypothesis." Journal of economic perspectives 25(3): 153{72. Altonji, Joseph G., Sarah Cattan, and Iain Ware. 2017. \Identifying Sibling In uence on Teenage Substance Use." Journal of Human Resources 52(1): 1{47. Arellano, Manuel, and Olympia Bover. 1995. \Another look at the instrumental variable estimation of error-components models." Journal of Econometrics 68(1): 29 { 51. Arellano, Manuel, and Stephen Bond. 1991. \Some Tests of Specication for Panel Data: Monte Carlo Evidence and an Application to Employment Equations." The Review of Economic Studies 58(2): 277{297. Auyeung, Tung Wai, Jenny Shun Wah Lee, Jason Leung, Timothy Kwok, Ping Chung Leung, and Jean Woo. 2010. \Eects of height loss on morbidity and mortality in 3145 community-dwelling Chinese older women and men: a 5-year prospective study." Age and Ageing 39(6): 699{704. Baker, David, Juan Leon, Emily Greenaway, John Collins, and Marcela Movit. 2011. \The Educa- tion Eect on Population Health: A Reassessment." Population Development Review 37(2): 307{ 332. Banerjee, Souvik, Pinka Chatterji, and Kajal Lahiri. 2017. \Eects of psychiatric disorders on labor market outcomes: A latent variable approach using multiple clinical indicators." Health Economics 26(2): 184{205. Banks, James, and Fabrizio Mazzonna. 2012. \The eect of education on old age cognitive abilities: Evidence from a regression discontinuity design." The Economic Journal 122(560): 418{448. Banks, James, Elaine Kelly, and James P. Smith. 2013. \Spousal Health Eects - the Role of Selection." National Bureau of Economic Research Working Paper 19438. Becker, Gary S. 1973. \A Theory of Marriage: Part I." Journal of Political Economy 81(4): 813{846. Becker, Gary S. 1974. \A Theory of Marriage: Part II." Journal of Political Economy 82(2): S11{ S26. Beleld, Clive R., and Inas Rashad Kelly. 2012. \The benets of breast feeding across the early years of childhood." Journal of Human Capital 6(3): 251{277. Benjamins, Maureen Reindl, Robert A. Hummer, Isaac W. Eberstein, and Charles B. Nam. 2004. \Self-reported health and adult mortality risk: An analysis of cause-specic mortality." Social Science & Medicine 59(6): 1297{1306. 154 Berkman, Lisa F., Thomas Glass, and Ian Brissette. 2000. \From social integration to health: Durkheim in the new millennium." Social Science & Medicine 51(6): 843{857. Bielak, Allison A. M. 2010. \How can we not `lose it' if we still don't understand how to `use it'? Unanswered questions about the in uence of activity participation on cognitive performance in older age|a mini-review." Gerontology 56(5): 507{519. Blundell, Richard, and Stephen Bond. 1998. \Initial conditions and moment restrictions in dynamic panel data models." Journal of Econometrics 87(1): 115 { 143. Briot, Karine, Erik Legrand, Denis Pouchain, St ephanie Monnier, and Christian Roux. 2010. \Ac- curacy of patient-reported height loss and risk factors for height loss among postmenopausal women." Canadian Medical Association Journal 182(6): 558{562. Brown, Heather, Arne Risa Hole, and Jennifer Roberts. 2014. \Going the same weigh: spousal correlations in obesity in the United Kingdom." Applied Economics 46(2): 153{166. Brown, Timothy Tyler. 2014. \How eective are public health departments at preventing mortal- ity?" Economics & Human Biology 13(Supplement C): 34{45. Business Consulting Services, International Business Machines. 2006. Healthcare in China. Somers, NY: Institute for Business Value, IBM Global Services. https://www- .935.ibm.com/services/us/imc/pdf/g510-6268-healthcare-china.pdf. Cai, Fang, and Yang Du. 2015. \The social protection system in ageing China." Asian Economic Policy Review 10(2): 250{270. Cai, Fang, John Giles, Philip O'Keefe, and Dewen Wang. 2012. The elderly and old age support in rural China: challenges and prospects (English). Washington, DC: World Bank. http://documents.worldbank.org/curated/en/769231468215685476/The-elderly-and-old- age-support-in-rural-China-challenges-and-prospects. Cameron, Lisa A., and Deborah Cobb-Clark. 2008. \Do coresidency and nancial transfers from the children reduce the need for elderly parents to work in developing countries?" Journal of Population Economics 21(4): 1007{1033. Campbell, David Edward. 2006. \What is education's impact on civic and social engagement?" In Measuring the Eects of Education on Health and Civic Engagement: Proceedings of the Copen- hagen symposium, , 25{126. Paris: Organisation for Economic Cooperation and Development. Card, David, and Laura Giuliano. 2013. \Peer eects and multiple equilibria in the risky behavior of friends." Review of Economics and Statistics 95(4): 1130{1149. Carlton, Elizabeth J., Song Liang, Julia Z. McDowell, Huazhong Li, Wei Luo, and Justin V. Remais. 2012. \Regional disparities in the burden of disease attributable to unsafe water and poor sanitation in China." Bulletin of the World Health Organization 90(8): 578{587. Case, Anne, and Christina Paxson. 2008a. \Height, Health, and Cognitive Function at Older Ages." American Economic Review 98(2): 463{67. Case, Anne, and Christina Paxson. 2008b. \Stature and Status: Height, Ability, and Labor Market Outcomes." Journal of Political Economy 116(3): 499{532. 155 Case, Anne, Darren Lubotsky, and Christina Paxson. 2002. \Economic Status and Health in Child- hood: The Origins of the Gradient." American Economic Review 92(5): 1308{1334. Cawley, John, Johanna Catherine Maclean, and Asia Sikora Kessler. 2017. \Reporting error in weight and height among older adults: Implications for estimating healthcare costs." The Journal of the Economics of Ageing 9(Supplement C): 122{144. Cawthon, Peggy M. 2011. \Gender Dierences in Osteoporosis and Fractures." Clinical Or- thopaedics and Related Research 469(7): 1900{1905. Central Committee of the Communist Party of China. 1985. Decision of The Central Committee of the Communist Party of China on the Reform of China's Educational Structure. Beijing: Foreign Language Press. Chiappori, Pierre-Andr, Sonia Orece, and Climent Quintana-Domeque. 2012. \Fatter Attraction: Anthropometric and Socioeconomic Matching on the Marriage Market." Journal of Political Economy 120(4): 659{695. Chou, Shin-Yi, Jin-Tan Liu, Michael Grossman, and Ted Joyce. 2010. \Parental education and child health: Evidence from a natural experiment in Taiwan." American Economic Journal: Applied Economics 2(1): 33{61. Christakis, Nicholas A., and James H. Fowler. 2007. \The Spread of Obesity in a Large Social Network over 32 Years." New England Journal of Medicine 357(4): 370{379. PMID: 17652652. Clark, Andrew E., and Fabrice" Etil e. 2006. \Dont give up on me baby: Spousal correlation in smoking behaviour." Journal of Health Economics 25(5): 958 { 978. Clark, Andrew E., and Fabrice" Etil e. 2011. \Happy house: Spousal weight and individual well- being." Journal of Health Economics 30(5): 1124 { 1136. Clark, Andrew E., and Youenn Loh eac. 2007. \\It" wasnt me, it was \them!" Social in uence in risky behavior by adolescents." Journal of Health Economics 26(4): 763 { 784. Cline, Martha G., Keith E. Meredith, John T. Boyer, and Benjamin Burrows. 1989. \Decline of height with age in adults in a general population sample: estimating maximum height and distin- guishing birth cohort eects from actual loss of stature with aging." Human Biology 61(3): 415{ 425. Currie, Janet. 2009. \Healthy, Wealthy, and Wise: Socioeconomic Status, Poor Health in Child- hood, and Human Capital Development." Journal of Economic Literature 47(1): 87{122. Currie, Janet, and Enrico Moretti. 2003. \Mother's education and the intergenerational transmis- sion of human capital: Evidence from college openings." The Quarterly Journal of Economics 118(4): 1495{1532. Currie, Janet, and Tom Vogl. 2013. \Early-Life Health and Adult Circumstance in Developing Countries." Annual Review of Economics 5(1): 1{36. Cutler, David M. 2018. \The school-rst solution." Accessed March 13, 2018. https://www.politico.com/agenda/story/2018/01/10/cost-eective-health-care-education- 000607. 156 Cutler, David M., and Adriana Lleras-Muney. 2010. \Understanding dierences in health behaviors by education." Journal of Health Economics 29(1): 1{28. Cutler, David M., and Adriana Lleras-Muney. 2012. \Education and Health: Insights from Inter- national Comparisons." National Bureau of Economic Research Working Paper 17738. Daemmrich, Arthur. 2013. \The political economy of healthcare reform in China: negotiating public and private." SpringerPlus 2(1): 448. Davillas, Apostolos, and Stephen Pudney. 2017. \Concordance of health states in couples: Analysis of self-reported, nurse administered and blood-based biomarker data in the UK Understanding Society panel." Journal of Health Economics 56: 87 { 102. de Leon, Carlos F. Mendes, Teresa E. Seeman, Dorothy I. Baker, Emily D. Richardson, and Mary E. Tinetti. 1996. \Self-ecacy, physical decline, and change in functioning in community-living elders: a prospective study." The Journals of Gerontology Series B: Psychological Sciences and Social Sciences 51(4): S183{S190. De Neve, Jan-Walter, and Gnther Fink. 2018. \Childrens education and parental old age survival { Quasi-experimental evidence on the intergenerational eects of human capital investment." Journal of Health Economics 58: 76 { 89. De Neve, Jan-Walter, and Guy Harling. 2017. \Ospring schooling associated with increased parental survival in rural KwaZulu-Natal, South Africa." Social Science & Medicine 176(Supple- ment C): 149{157. De Neve, Jan-Walter, and Ichiro Kawachi. 2017. \Spillovers between siblings and from ospring to parents are understudied: A review and future directions for research." Social Science & Medicine 183(Supplement C): 56{61. Deaton, Angus. 2007. \Height, health, and development." Proceedings of the National Academy of Sciences 104(33): 13232{13237. Dinsa, G. D., Y. Goryakin, E. Fumagalli, and M. Suhrcke. 2012. \Obesity and socioeconomic status in developing countries: a systematic review." Obesity Reviews 13(11): 1067{1079. Dohmen, Thomas, Armin Falk, David Human, and Uwe Sunde. 2012. \The Intergenerational Transmission of Risk and Trust Attitudes." The Review of Economic Studies 79(2): 645{677. Du o, Esther. 2001. \Schooling and Labor Market Consequences of School Construction in Indone- sia: Evidence from an Unusual Policy Experiment." American Economic Review 91(4): 795{813. Duncan, Terry E., and Edward McAuley. 1993. \Social support and ecacy cognitions in exercise adherence: A latent growth curve analysis." Journal of Behavioral Medicine 16(2): 199{218. Ebenstein, Avraham. 2010. \The Missing Girls of China and the Unintended Consequences of the One Child Policy." Journal of Human Resources 45(1): 87{115. Eisenberg, Daniel, Ezra Golberstein, and Janis L. Whitlock. 2014. \Peer eects on risky behaviors: New evidence from college roommate assignments." Journal of Health Economics 33: 126{138. Elo, Irma T., and Samuel H. Preston. 1996. \Educational dierentials in mortality: United States, 1979{1985." Social science & medicine 42(1): 47{57. 157 Emran, M. Shahe, and Zhaoyang Hou. 2013. \Access to markets and rural poverty: Evidence from household consumption in China." Review of Economics and Statistics 95(2): 682{697. Fan, Shenggen, Ravi Kanbur, and Xiaobo Zhang. 2011. \China's regional disparities: Experience and policy." Review of Development Finance 1(1): 47{56. Fernihough, Alan, and Mark E. McGovern. 2015. \Physical stature decline and the health status of the elderly population in England." Economics & Human Biology 16: 30{44. Fletcher, Jason M. 2009. \All in the family: Mental health spillover eects between working spouses." The BE journal of economic analysis & policy 9(1). Fletcher, Jason M. 2010. \Social interactions and smoking: evidence using multiple student cohorts, instrumental variables, and school xed eects." Health Economics 19(4): 466{484. Fletcher, Jason M., and David E. Frisvold. 2009. \Higher education and health investments: Does more schooling aect preventive health care use?" Journal of Human Capital 3(2): 144{176. Fletcher, Jason M., and Ryne Marksteiner. 2017. \Causal Spousal Health Spillover Eects and Implications for Program Evaluation." American Economic Journal: Economic Policy 9(4): 144{ 66. Fogel, Robert. 1986. \Physical Growth as a Measure of the Economic Well-being of Populations: The Eighteenth and Nineteenth Centuries." In Human growth: a comprehensive treatise, edited by Falkner Frank and James M. Tanner, 3: 263{81. New York: Plenum Press. Fowler, James H, and Nicholas A Christakis. 2008a. \Dynamic spread of happiness in a large social network: longitudinal analysis over 20 years in the Framingham Heart Study." BMJ 337. Fowler, James H., and Nicholas A. Christakis. 2008b. \Estimating peer eects on health in social networks: A response to Cohen-Cole and Fletcher; Trogdon, Nonnemaker, Pais." Journal of Health Economics 27(5): 1400{1405. Frankenberg, Elizabeth, and Duncan Thomas. 2000. \The Indonesia Family Life Survey (IFLS): Study Design and Results from Waves 1 and 2." DRU-2238/1-NIA/NICHD. Santa Monica: RAND. Friedman, Esther M., and Robert D. Mare. 2014. \The schooling of ospring and the survival of parents." Demography 51(4): 1271{1293. Gale, Catharine R., Christopher N. Martyn, Cyrus Cooper, and Avan Aihie Sayer. 2006. \Grip strength, body composition, and mortality." International Journal of Epidemiology 36(1): 228{ 235. Glei, Dana A., David A. Landau, Noreen Goldman, Yi-Li Chuang, Germ an Rodr guez, and Maxine Weinstein. 2005. \Participating in social activities helps preserve cognitive function: an analysis of a longitudinal, population-based study of the elderly." International journal of epidemiology 34(4): 864{871. Glymour, M. M., I. Kawachi, C. S. Jencks, and L. F. Berkman. 2008. \Does childhood schooling aect old age memory or mental status? Using state schooling laws as natural experiments." Journal of Epidemiology & Community Health 62(6): 532{537. 158 Grossman, Michael. 1972. \On the concept of health capital and the demand for health." Journal of Political Economy 80(2): 223{255. Grossman, Michael. 1976. \The Correlation between Health and Schooling." In Household Produc- tion and Consumption, edited by Nestor E. Terleckyj, 147{224. Cambridge: National Bureau of Economic Research. Grossman, Michael. 2006. \Education and Nonmarket Outcomes." In Handbook of the Economics of Education, edited by Eric Hanushek and Finis Welch, 1: 577{633. Amsterdam: North-Holland. Guven, Cahit, and Wang Sheng Lee. 2013. \HEIGHT AND COGNITIVE FUNCTION AT OLDER AGES: IS HEIGHT A USEFUL SUMMARY MEASURE OF EARLY CHILDHOOD EXPERI- ENCES?" Health Economics 22(2): 224{233. Guven, Cahit, and Wang-Sheng Lee. 2015. \Height, aging and cognitive abilities across Europe." Economics & Human Biology 16: 16{29. Hillier, Teresa A., Li-Yung Lui, Deborah M. Kado, E. S. LeBlanc, Kimberly K. Vesco, Douglas C. Bauer, Jane A. Cauley, Kristine E. Ensrud, Dennis M. Black, Marc C. Hochberg, and Steven R. Cummings. 2012. \Height loss in older women: Risk of hip fracture and mortality independent of vertebral fractures." Journal of Bone and Mineral Research 27(1): 153{159. Huang, T. L., M. C. Carlson, A. L. Fitzpatrick, L. H. Kuller, L. P. Fried, and P. P. Zandi. 2008. \Knee height and arm span: A re ection of early life environment and risk of dementia." Neu- rology 70(19 Part 2): 1818{1826. Huang, Wei. 2015. \Understanding the Eects of Education on Health: Evidence from China." Bonn: Institute for the Study of Labor IZA Discussion Papers 9255. Huang, Wei, and Yi Zhou. 2013. \Eects of education on cognition at older ages: Evidence from China's Great Famine." Social Science & Medicine 98(Supplement C): 54{62. Huang, Wei, Xiaoyan Lei, and Yaohui Zhao. 2016. \One-child policy and the rise of man-made twins." Review of Economics and Statistics 98(3): 467{476. Huang, Wei, Xiaoyan Lei, Geert Ridder, John Strauss, and Yaohui Zhao. 2013. \Health, Height, Height Shrinkage, and SES at Older Ages: Evidence from China." American Economic Journal: Applied Economics 5(2): 86{121. Hyypp a, Markku T, and Juhani M aki. 2003. \Social participation and health in a community rich in stock of social capital." Health Education Research 18(6): 770{779. Islam, Asadul, and Russell Smyth. 2015. \Do Fertility Control Policies Aect Health in Old Age? Evidence from China's One-Child Experiment." Health Economics 24(5): 601{616. James, Bryan D., Robert S. Wilson, Lisa L. Barnes, and David A. Bennett. 2011. \Late-life social activity and cognitive decline in old age." Journal of the International Neuropsychological Society 17(6): 998{1005. Kalmijn, Matthijs. 1994. \Assortative Mating by Cultural and Economic Occupational Status." American Journal of Sociology 100(2): 422{452. 159 Kemptner, Daniel, Hendrik J urges, and Steen Reinhold. 2011. \Changes in compulsory school- ing and the causal eect of education on health: Evidence from Germany." Journal of Health Economics 30(2): 340{354. Kravdal, ystein. 2008. \A broader perspective on education and mortality: Are we in uenced by other people's education?" Social Science & Medicine 66(3): 620{636. Krege, John H., David Kendler, Kelly Krohn, Harry Genant, Jahangir Alam, Pierre-Yves Berclaz, Barbara Coey, and Corina Loghin. 2015. \Relationship Between Vertebral Fracture Burden, Height Loss, and Pulmonary Function in Postmenopausal Women With Osteoporosis." Journal of Clinical Densitometry 18(4): 506{511. LaFave, Daniel, and Duncan Thomas. 2016. \Height and cognition at work: Labor market produc- tivity in a low income setting." Economics & Human Biology 25: 52{64. Lang, Iain, Robert B. Wallace, Felicia A. Huppert, and David Melzer. 2007. \Moderate alcohol consumption in older adults is associated with better cognition and well-being than abstinence." Age and ageing 36(3): 256{261. Lavy, Victor, John Strauss, Duncan Thomas, and Philippe De Vreyer. 1996. \Quality of health care, survival and health outcomes in Ghana." Journal of Health Economics 15(3): 333{357. Lee, Chioun. 2017. \Adult Children's Education and Physiological Dysregulation Among Older Parents." The Journals of Gerontology, Series B: Psychological Sciences and Social Sciences 00(00): 1{12. Lee, Chioun, Dana A. Glei, Noreen Goldman, and Maxine Weinstein. 2017. \Children's educa- tion and parents' trajectories of depressive symptoms." Journal of Health and Social Behavior 58(1): 86{101. Lee, I-Min, Joann E. Manson, Charles H. Hennekens, and Ralph S. Paenbarger. 1993. \Body Weight and Mortality." JAMA 270(23): 2823{2828. Lei, Xiaoyan, John Strauss, Meng Tian, and Yaohui Zhao. 2015. \Living arrangements of the elderly in China: evidence from the CHARLS national baseline." China Economic Journal 8(3): 191{214. Lei, Xiaoyan, Yuqing Hu, John J. McArdle, James P. Smith, and Yaohui Zhao. 2012. \Gender dierences in cognition among older adults in China." Journal of Human Resources 47(4): 951{ 971. Lewbel, Arthur. 2012. \Using Heteroscedasticity to Identify and Estimate Mismeasured and En- dogenous Regressor Models." Journal of Business & Economic Statistics 30(1): 67{80. Lindstr om, Martin, Bertil S. Hanson, and Per-Olof Ostergren. 2001. \Socioeconomic dierences in leisure-time physical activity: the role of social participation and social capital in shaping health related behaviour." Social Science & Medicine 52(3): 441 { 451. Lleras-Muney, Adriana. 2005. \The relationship between education and adult mortality in the United States." The Review of Economic Studies 72(1): 189{221. Llewellyn, David J., Iain A. Lang, Kenneth M. Langa, Graciela Muniz-Terrera, Caroline L. Phillips, Antonio Cherubini, Luigi Ferrucci, and David Melzer. 2010. \Vitamin d and risk of cognitive decline in elderly persons." Archives of Internal Medicine 170(13): 1135{1141. 160 Lundborg, Petter, and Kaveh Majlesi. 2018. \Intergenerational transmission of human capital: Is it a one-way street?" Journal of Health Economics 57: 206{220. Lundborg, Petter, Anton Nilsson, and Dan-Olof Rooth. 2014. \Parental Education and Ospring Outcomes: Evidence from the Swedish Compulsory School Reform." American Economic Jour- nal: Applied Economics 6(1): 253{78. Mackenbach, Johan P. 2005. Health inequalities: Europe in prole. London: Department of Health. Mannino, D. M., A. S. Buist, T. L. Petty, P. L. Enright, and S. C. Redd. 2003. \Lung function and mortality in the United States: data from the First National Health and Nutrition Examination Survey follow up study." Thorax 58(5): 388{393. Manski, Charles F. 1993. \Identication of endogenous social eects: The re ection problem." The review of economic studies 60(3): 531{542. Mare, Robert D. 1991. \Five Decades of Educational Assortative Mating." American Sociological Review 56(1): 15{32. Masunari, Naomi, Saeko Fujiwara, Fumiyoshi Kasagi, Ikuno Takahashi, Michiko Yamada, and Toshitaka Nakamura. 2012. \Height loss starting in middle age predicts increased mortality in the elderly." Journal of Bone and Mineral Research 27(1): 138{145. Maurer, J urgen. 2010. \Height, education and later-life cognition in Latin America and the Caribbean." Economics & Human Biology 8(2): 168{176. McArdle, John J., James P. Smith, and Robert Willis. 2011. \Cognition and Economic Outcomes in the Health and Retirement Survey." In Explorations in the Economics of Aging, edited by David A. Wise, 209{233. Chicago: University of Chicago Press. McGovern, Mark E. 2014. \Comparing the relationship between stature and later life health in six low and middle income countries." The Journal of the Economics of Ageing 4: 128{148. Meyler, Deanna, Jim P. Stimpson, and M. Kristen Peek. 2007. \Health concordance within couples: a systematic review." Social Science & Medicine 64(11): 2297{2310. Michaud, Pierre-Carl, and Arthur van Soest. 2008. \Health and wealth of elderly couples: Causality tests using dynamic panel data models." Journal of Health Economics 27(5): 1312 { 1325. Ming, Cheng Kai. 1986. \China's recent education reform: the beginning of an overhaul." Com- parative Education 22(3): 255{269. Mishra, Vinod, and Russell Smyth. 2014. \Technological Change and Wages in China: Evidence from Matched Employer{Employee Data." Review of Development Economics 18(1): 123{138. Monden, Christiaan W. S., Frank Van Lenthe, Nan Dirk De Graaf, and Gerbert Kraaykamp. 2003. \Partner's and own education: does who you live with matter for self-assessed health, smoking and excessive alcohol consumption?" Social Science & Medicine 57(10): 1901{1912. Montez, Jennifer Karas, and Esther M. Friedman. 2015. \Educational attainment and adult health: Under what conditions is the association causal?" Social Science & Medicine 127: 1{7. 161 Mosca, Irene, and Robert E. Wright. 2016. \Height and cognition at older ages: Irish evidence." Economics Letters 149: 98{101. Nieminen, Tarja, Ritva Pr att al a, Tuija Martelin, Tommi H ark anen, Markku T. Hyypp a, Erkki Alanen, and Seppo Koskinen. 2013. \Social capital, health behaviours and health: a population- based associational study." BMC Public Health 13(1): 613{624. Nugent, Jerey B. 1985. \The old-age security motive for fertility." Population and development review 11(1): 75{97. Old, Jerry L., and Michelle Calvert. 2004. \Vertebral Compression Fractures in the Elderly." Amer- ican Family Physician 69(1): 111{116. Oreopoulos, Philip, Marianne E. Page, and Ann Hu Stevens. 2006. \The intergenerational eects of compulsory schooling." Journal of Labor Economics 24(4): 729{760. Perkins, Jessica M, Kristi S Multhaup, H Wesley Perkins, and Cole Barton. 2008. \Self-ecacy and participation in physical and social activity among older adults in Spain and the United States." The Gerontologist 48(1): 51{58. Phillips, Drystan, Yu-Chen Lin, Jenny Wight, Sandy Chien, and Jinkook Lee. 2017. Harmonized ELSA Documentation. Version E. Los Angeles: Center for Global Aging Research, University of Southern California. https://g2aging.org/?section=pagepageid=18. Population Census Oce. 2012. Tabulation on the 2010 Population Census of the People's Republic of China. Beijing: China Statistics Press. Powdthavee, Nattavudh. 2009. \I cant smile without you: Spousal correlation in life satisfaction." Journal of Economic Psychology 30(4): 675 { 689. Powell, Lisa M., John A. Tauras, and Hana Ross. 2005. \The importance of peer eects, cigarette prices and tobacco control policies for youth smoking behavior." Journal of Health Economics 24(5): 950 { 968. Rantanen, T., J.M. Guralnik, D. Foley, K. Masaki, S. Leveille, J. D. Curb, and L. White. 1999. \Midlife hand grip strength as a predictor of old age disability." JAMA 281(6): 558{560. Raspe, Heiner, Angelika Hueppe, and Hannelore Neuhauser. 2008. \Back pain, a communicable disease?" International Journal of Epidemiology 37(1): 69{74. Rawlings, Samantha B. 2015. \Parental education and child health: Evidence from an education reform in China." Unpublisted Manuscript. Roodman, David. 2009. \How to do xtabond2: An introduction to dierence and system GMM in Stata." Stata Journal 9(1): 86{136. Rosenzweig, Mark R., and Kenneth I. Wolpin. 1986. \Evaluating the eects of optimally distributed public programs: Child health and family planning interventions." American Economic Review 76(3): 470{482. Sabia, Joseph J. 2007. \The eect of body weight on adolescent academic performance." Southern Economic Journal 73(4): 871{900. 162 Schneiderman, Neil, Gail Ironson, and Scott D. Siegel. 2005. \Stress and health: Psychological, behavioral, and biological determinants." Annual Review of Clinical Psychology 1: 607{628. Schultz, T. Paul, and John Strauss, ed. 2008. Handbook of Development Economics. Amsterdam: North-Holland Press. Schulz, Richard, and Paula R Sherwood. 2008. \Physical and mental health eects of family care- giving." Journal of Social Work Education 44(sup3): 105{113. Simes, R. John. 1986. \An improved Bonferroni procedure for multiple tests of signicance." Biometrika 73(3): 751{754. Sirven, Nicolas, and Thierry Debrand. 2008. \Social participation and healthy ageing: An interna- tional comparison using SHARE data." Social Science & Medicine 67(12): 2017{2026. Smith, James P. 2007. \The impact of socioeconomic status on health over the life-course." Journal of Human Resources 42(4): 739{764. Smith, James P., Meng Tian, and Yaohui Zhao. 2013. \Community eects on elderly health: Evi- dence from CHARLS national baseline." The journal of the economics of ageing 1{2: 50{59. Smith, James P., Yan Shen, John Strauss, Yang Zhe, and Yaohui Zhao. 2012. \The eects of childhood health on adult health and SES in China." Economic Development and Cultural Change 61(1): 127{156. Sohn, Kitae. 2015. \Is leg length a biomarker of early life conditions? Evidence from a historically short population." American Journal of Human Biology 27(4): 538{545. Sorkin, John D., Denis C. Muller, and Reubin Andres. 1999. \Longitudinal Change in Height of Men and Women: Implications for Interpretation of the Body Mass Index: The Baltimore Longitudinal Study of Aging." American Journal of Epidemiology 150(9): 969{977. Staiger, Douglas, and James H. Stock. 1997. \Instrumental Variables Regression with Weak Instru- ments." Econometrica 65(3): 557{586. Steckel, Richard H. 1995. \Stature and the Standard of Living." Journal of Economic Literature 33(4): 1903{1940. Steckel, Richard H. 2009. \Heights and human welfare: Recent developments and new directions." Explorations in Economic History 46(1): 1{23. Steptoe, Andrew, Angus Deaton, and Arthur A. Stone. 2015. \Subjective wellbeing, health, and ageing." The Lancet 385(9968): 640{648. Strauss, John, and Duncan Thomas. 1995. \Human resources: Empirical modeling of household and family decisions." In Handbook of development economics, edited by Jere Behrman and T.N. Srinivasan, 3A: 1883{2023. Amsterdam: North-Holland Press. Strauss, John, and Duncan Thomas. 2008. \Health over the Life Course." In Handbook of De- velopment Economics, edited by T. Paul Schultz and John Strauss, 4: 3375{3474. Amsterdam: North-Holland Press. 163 Strauss, John, Firman Witoelar, and Bondan Sikoki. 2016. \The Fifth Wave of the Indonesia Family Life Survey (IFLS5): Overview and Field Report." WR-1143/1-NIA/NICHD. Santa Monica: RAND. Strauss, John, Firman Witoelar, Bondan Sikoki, and Anna Marie Wattie. 2009. \The Fourth Wave of the Indonesian Family Life Survey (IFLS4): Overview and Field Report." WR-675/1- NIA/NICHD. Santa Monica: RAND. Strauss, John, Kathleen Beegle, Bondan Sikoki, Agus Dwiyanto, Yulia Herawati, and Firman Witoelar. 2004. \The Third Wave of the Indonesia Family Life Survey (IFLS): Overview and Field Report." WR-144/1-NIA/NICHD. Santa Monica: RAND. Sun, Rongjun. 2002. \Old age support in contemporary urban China from both parents' and chil- dren's perspectives." Research on Aging 24(3): 337{359. Sundquist, Kristina, Martin Lindstr om, Marianne Malmstr om, Sven-Erik Johansson, and Jan Sundquist. 2004. \Social participation and coronary heart disease: a follow-up study of 6900 women and men in Sweden." Social Science & Medicine 58(3): 615{622. Thomas, Duncan, and John Strauss. 1992. \Prices, infrastructure, household characteristics and child height." Journal of Development Economics 39(2): 301{331. Thomas, Duncan, and John Strauss. 1997. \Health and wages: Evidence on men and women in urban Brazil." Journal of Econometrics 77(1): 159{185. Thomas, Duncan, and John Strauss. 1998. \Health, Nutrition, and Economic Development." Jour- nal of Economic Literature 36(2): 766{817. Thomas, Duncan, John Strauss, and Maria-Helena Henriques. 1990. \Child survival, height for age and household characteristics in Brazil." Journal of Development Economics 33(2): 197{234. Thomas, Duncan, John Strauss, and Maria-Helena Henriques. 1991. \How does mother's education aect child height?" Journal of human resources 26(2): 183{211. Torssander, Jenny. 2013. \From child to parent? The signicance of children's education for their parents' longevity." Demography 50(2): 637{659. Torssander, Jenny. 2014. \Adult children's socioeconomic positions and their parents' mortality: A comparison of education, occupational class, and income." Social Science & Medicine 122: 148{ 156. Tsang, Mun C. 1991. \The structural reform of secondary education in China." Journal of Educa- tional Administration 29(4): 65{83. van der Heide, Iris, Rogier M. van Rijn, Suzan J. W. Robroek, Alex Burdorf, and Karin I. Proper. 2013. \Is retirement good for your health? A systematic review of longitudinal studies." BMC Public Health 13(1): 1180{1190. Veenstra, Gerry. 2000. \Social capital, SES and health: an individual-level analysis." Social Science & Medicine 50(5): 619{629. 164 Wannamethee, S., Shaper A., Lennon L., and Whincup P. H. 2006. \Height loss in older men: Associations with total mortality and incidence of cardiovascular disease." Archives of Internal Medicine 166(22): 2546{2552. Weir, David, Margaret Lay, and Kenneth Langa. 2014. \Economic development and gender inequal- ity in cognition: A comparison of China and India, and of SAGE and the HRS sister studies." The Journal of the Economics of Ageing 4: 114{125. Wilson, Sven E. 2002. \The health capital of families: an investigation of the inter-spousal corre- lation in health status." Social Science & Medicine 55(7): 1157{1172. Windmeijer, Frank. 2005. \A nite sample correction for the variance of linear ecient two-step GMM estimators." Journal of Econometrics 126(1): 25 { 51. World Health Organization. 2017. \Sanitation. Fact sheet." Accessed August 1, 2017. http://www.who.int/mediacentre/factsheets/fs392/en/. Yahirun, Jenjira J., Connor M. Sheehan, and Mark D. Hayward. 2016. \Adult Children's Education and Parents' Functional Limitations in Mexico." Research on Aging 38(3): 322{345. Yahirun, Jenjira J., Connor M. Sheehan, and Mark D. Hayward. 2017. \Adult children's education and changes to parents' physical health in Mexico." Social Science & Medicine 181: 93{101. Yang, Lei, Pekka Martikainen, and Karri Silventoinen. 2016. \Eects of individual, spousal, and ospring socioeconomic status on mortality among elderly people in China." Journal of Epidemi- ology 26(11): 602{609. Zhang, Junfeng, Denise L. Mauzerall, Tong Zhu, Song Liang, Majid Ezzati, and Justin V. Remais. 2010. \Environmental health in China: progress towards clean air and safe water." The Lancet 375(9720): 1110{1119. Zhang, Junfeng Jim, and Kirk R. Smith. 2007. \Household Air Pollution from Coal and Biomass Fuels in China: Measurements, Health Impacts, and Interventions." Environmental Health Per- spectives 115(6): 848{855. Zhang, Junsen. 2017. \The Evolution of China's One-Child Policy and Its Eects on Family Out- comes." Journal of Economic Perspectives 31(1): 141{159. Zhao, Yaohui. 1997. \Labor migration and returns to rural education in China." American Journal of Agricultural Economics 79(4): 1278{1287. Zhao, Yaohui, Albert Park, John Strauss, John Giles, Shangyi Mao, Eileen Crimmins, Shangyi Mao, Peifeng Hu, Xiaoyan Lei, Yan Shen, Xinzheng Shi, James P. Smith, Yafeng Wang, Xiaoyu Wu, Gonghuan Yang, and Xiangjun Yin. 2013a. \Challenges of Population Aging in China: Evidence from the National Baseline Survey of the China Health and Retirement Longitudinal Study (CHARLS)." National Baseline Survey Report. Beijing: National School of Development, Peking University. http://online.wsj.com/public/resources/documents/charls0530.pdf. Zhao, Yaohui, John Strauss, Gonghuan Yang, John Giles, Peifeng Hu, Yisong Hu, Xiaoyan Lei, Man Liu, Albert Park, James P. Smith, and Yafeng Wang. 2013b. China Health and Retirement Longitudinal Study|2011-2012 National Baseline Users' Guide. Beijing: National 165 School of Development, Peking University. http://charls.pku.edu.cn/uploads/document/2011- charls-wave1/application/CHARLS nationalbaseline users guide.pdf. Zimmer, Zachary, Albert I. Hermalin, and Hui-Sheng Lin. 2002. \Whose education counts? The added impact of adult-child education on physical functioning of older Taiwanese." Journals of Gerontology Series B: Psychological Sciences and Social Sciences 57(1): S23{S32. Zimmer, Zachary, Linda G. Martin, Mary Beth Ofstedal, and Yi-Li Chuang. 2007. \Education of adult children and mortality of their elderly parents in Taiwan." Demography 44(2): 289{305. 166 Appendix A Chapter 2 Appendix Tables 167 Table A1: Results on Self-reported Health (1{3 Scale) (1) (2) (3) (4) (5) Men OLS FE System GMM L.(own SRH) 0.569*** -0.125*** 0.171*** 0.224*** 0.177*** (0.0176) (0.0206) (0.0311) (0.0426) (0.0335) L2.(own SRH) 0.0823** (0.0357) spouse's SRH 0.122*** 0.147*** 0.344*** 0.287*** (0.0138) (0.0203) (0.0890) (0.0907) L.(spouse's SRH) 0.00146 (0.0253) Observations 6718 6718 6718 4838 6758 Number of couples 1881 1881 1881 1606 1886 AR(1) test p val 2.32e-48 1.13e-33 4.79e-46 AR(2) test p val 0.486 0.395 0.263 Hansen J test p val 0.000601 0.00481 0.00000797 Di in Hansen p val for levels 0.0255 0.00627 0.00410 (6) (7) (8) (9) (10) Women OLS FE System GMM L.(own SRH) 0.597*** -0.121*** 0.146*** 0.178*** 0.146*** (0.0182) (0.0229) (0.0285) (0.0384) (0.0288) L2.(own SRH) 0.0294 (0.0302) spouse's SRH 0.115*** 0.124*** 0.324*** 0.310*** (0.0118) (0.0169) (0.0662) (0.0669) L.(spouse's SRH) -0.00642 (0.0229) Observations 6849 6849 6849 4915 6915 Number of couples 1944 1944 1944 1636 1956 AR(1) test p val 1.28e-49 1.48e-35 2.74e-49 AR(2) test p val 0.727 0.949 0.786 Hansen J test p val 0.0351 0.177 0.000901 Di in Hansen p val for levels 0.0744 0.0898 0.00455 Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 168 Table A2: Results on Frequency of Exercise (=1 if any weekly vigorous physical activities) (1) (2) (3) (4) (5) Men OLS FE System GMM L.(weekly vig phy activ) 0.433*** -0.112*** 0.141*** 0.220*** 0.144*** (0.0141) (0.0153) (0.0207) (0.0229) (0.0217) L2.(weekly vig phy activ) 0.116*** (0.0201) spouse's weekly vig phy activ 0.201*** 0.232*** 0.289*** 0.315*** (0.0133) (0.0170) (0.0730) (0.0833) L.(spouse's weekly vig phy activ) 0.0400** (0.0188) Observations 7020 7020 7020 5077 7013 Number of couples 1934 1934 1934 1672 1931 AR(1) test p val 2.75e-97 2.71e-75 2.16e-94 AR(2) test p val 0.00825 0.419 0.0127 Hansen J test p val 0.0000129 0.116 0.00000267 Di in Hansen p val for levels 0.263 0.681 0.185 (6) (7) (8) (9) (10) Women OLS FE System GMM L.(weekly vig phy activ) 0.425*** -0.107*** 0.177*** 0.187*** 0.189*** (0.0152) (0.0169) (0.0215) (0.0269) (0.0217) L2.(weekly vig phy activ) 0.0469** (0.0214) spouse's weekly vig phy activ 0.168*** 0.204*** 0.124* 0.112 (0.0120) (0.0152) (0.0666) (0.0716) L.(spouse's weekly vig phy activ) -0.00964 (0.0162) Observations 7212 7212 7212 5182 7193 Number of couples 2022 2022 2022 1717 2019 AR(1) test p val 4.72e-88 1.63e-62 5.94e-89 AR(2) test p val 0.00457 0.280 0.00395 Hansen J test p val 0.00481 0.0215 0.00438 Di in Hansen p val for levels 0.00279 0.00821 0.00519 Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 169 Table A3: Results on Chronic Conditions (# of severe diseases) (1) (2) (3) (4) (5) Men OLS FE System GMM L.(own # of severe diseases) 0.984*** 0.497*** 0.964*** 0.969*** 0.964*** (0.00540) (0.0172) (0.0119) (0.0132) (0.0120) L2.(own # of severe diseases) 0.00476 (0.0174) spouse's # of severe diseases 0.00776 0.0252 -0.0148 -0.00933 (0.00622) (0.0168) (0.0159) (0.0145) L.(spouse's # of severe diseases) -0.0169 (0.0155) Observations 7044 7044 7044 5097 7044 Number of couples 1936 1936 1936 1675 1936 AR(1) test p val 1.93e-49 4.90e-39 1.86e-49 AR(2) test p val 0.659 0.928 0.643 Hansen J test p val 0.202 0.157 0.213 Di in Hansen p val for levels 0.206 0.189 0.371 (6) (7) (8) (9) (10) Women OLS FE System GMM L.(own # of severe diseases) 0.981*** 0.475*** 0.978*** 0.982*** 0.978*** (0.00460) (0.0188) (0.0121) (0.0137) (0.0121) L2.(own # of severe diseases) -0.0174 (0.0155) spouse's # of severe diseases 0.00824* 0.0287** -0.00161 0.000775 (0.00433) (0.0117) (0.0107) (0.00974) L.(spouse's # of severe diseases) -0.00251 (0.0106) Observations 7237 7237 7237 5193 7237 Number of couples 2028 2028 2028 1719 2028 AR(1) test p val 7.58e-44 2.31e-32 7.70e-44 AR(2) test p val 0.513 0.510 0.517 Hansen J test p val 0.0282 0.0556 0.0302 Di in Hansen p val for levels 0.00125 0.00490 0.00164 Note: Two-step corrected robust standard errors in the parenthesis. *** p < 0:01, ** p < 0:05, * p < 0:1. All columns control for 1) respondent's and spouse's age dummies, number of living daughters and sons, number of living siblings, respondent's and spouse's education dummies and length of marriage; 2) log per capita de ated household wealth. 170 Appendix B Chapter 3 Appendix Tables Table B1: Level Eects of Children's Education on Wave 2 Cognition of Parents (1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV VARIABLES Wave 2 Mental Intactness Wave 2 Episodic Memory Education of Children 0.131*** 0.454*** 0.124*** 0.483* 0.049*** 0.283*** 0.048*** 0.263** (0.011) (0.160) (0.010) (0.247) (0.005) (0.094) (0.005) (0.129) Own Education 0.256*** 0.178*** 0.255*** 0.185*** 0.110*** 0.054** 0.109*** 0.067** (0.010) (0.041) (0.010) (0.048) (0.007) (0.026) (0.006) (0.028) Observations 10,270 10,270 10,270 10,263 10,270 10,270 10,270 10,263 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 28 427 420 28 28 427 420 LM stat for underid. 9.793 9.275 9.793 9.275 p-val of LM stat 0.00747 0.00968 0.00747 0.00968 F stat for weak-id. 10.66 7.103 10.66 7.103 Anderson-Rubin F stat 4.667 2.164 6.465 3.884 p-val of AR F stat 0.0182 0.134 0.00508 0.0329 p-val of Hansen J stat 0.387 0.487 0.383 0.258 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 171 Table B2: Level Eects of Children's Education on Wave 2 Subjective Health Measures of Parents (1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV VARIABLES Wave 2 Self Reported Health Wave 2 Low Subjective Survival Expectation Education of Children -0.018*** -0.046 -0.015*** -0.055 -0.014*** -0.097*** -0.012*** -0.102** (0.003) (0.057) (0.003) (0.065) (0.002) (0.032) (0.002) (0.050) Own Education -0.007*** 0.000 -0.007*** 0.001 -0.008*** 0.012 -0.007*** 0.010 (0.002) (0.014) (0.002) (0.013) (0.002) (0.008) (0.001) (0.010) Observations 10,174 10,174 10,174 10,167 7,925 7,925 7,925 7,914 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 28 427 420 28 28 427 416 LM stat for underid. 9.035 8.114 9.513 6.710 p-val of LM stat 0.0109 0.0173 0.00859 0.0349 F stat for weak-id. 9.907 6.369 11.15 5.100 Anderson-Rubin F stat 0.753 0.682 4.506 2.492 p-val of AR F stat 0.481 0.514 0.0205 0.102 p-val of Hansen J stat 0.456 0.469 0.668 0.831 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Table B3: Level Eects of Children's Education on Wave 2 Physical Health of Parents (1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV VARIABLES Wave 2 Lung Functation (Peak Flow) Wave 2 Grip Strength Education of Children 1.888*** 20.326 1.692*** 19.416 0.179*** 1.031 0.175*** 0.696 (0.368) (12.509) (0.330) (15.647) (0.031) (0.761) (0.019) (1.207) Own Education 1.835*** -2.731 1.654*** -1.862 0.114*** -0.096 0.095*** -0.009 (0.364) (3.184) (0.377) (3.089) (0.025) (0.186) (0.021) (0.232) Observations 7,696 7,695 7,696 7,685 8,035 8,035 8,035 8,026 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 27 414 403 28 28 418 409 LM stat for underid. 4.966 3.111 4.827 2.253 p-val of LM stat 0.0835 0.211 0.0895 0.324 F stat for weak-id. 4.314 2.075 3.954 1.313 Anderson-Rubin F stat 5.492 2.785 1.731 0.412 p-val of AR F stat 0.0102 0.0802 0.196 0.667 p-val of Hansen J stat 0.199 0.425 0.497 0.580 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 172 Table B4: Level Eects of Children's Education on Wave 2 Body Weight of Parents (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) OLS IV OLS IV OLS IV OLS IV OLS IV OLS IV VARIABLES Wave 2 BMI Wave 2 Underweight Wave 2 Overweight Education of Children 0.082*** 0.559** 0.064*** 0.399 -0.003*** -0.045 -0.003*** -0.038 0.007*** 0.042 0.005*** 0.025 (0.014) (0.219) (0.012) (0.362) (0.001) (0.039) (0.001) (0.051) (0.002) (0.035) (0.002) (0.055) Own Education 0.014 -0.104** -0.001 -0.067 -0.001 0.009 -0.001 0.006 0.003 -0.006 0.001 -0.003 (0.013) (0.051) (0.012) (0.071) (0.001) (0.010) (0.001) (0.010) (0.002) (0.009) (0.002) (0.011) Observations 8,093 8,093 8,093 8,085 8,093 8,093 8,093 8,085 8,093 8,093 8,093 8,085 Child Birth Year FE YES YES YES YES YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES NO NO YES YES Number of groups 28 28 418 410 28 28 418 410 28 28 418 410 LM stat for underid. 4.753 2.564 4.753 2.564 4.753 2.564 p-val of LM stat 0.0929 0.277 0.0929 0.277 0.0929 0.277 F stat for weak-id. 4.161 1.680 4.161 1.680 4.161 1.680 Anderson-Rubin F stat 3.806 1.632 1.674 0.702 0.973 0.772 p-val of AR F stat 0.0350 0.214 0.206 0.505 0.391 0.472 p-val of Hansen J stat 0.297 0.142 0.509 0.474 0.274 0.259 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 173 Table B5: Level Eects of Children's Education on Wave 2 Mental Health of Parents (1) (2) (3) (4) OLS IV OLS IV VARIABLES Wave 2 CESD 10 Score Education of Children -0.166*** -0.285 -0.142*** -0.257 (0.027) (0.596) (0.024) (0.708) Own Education -0.091*** -0.063 -0.078*** -0.055 (0.018) (0.149) (0.015) (0.146) Observations 9,130 9,130 9,130 9,123 Child Birth Year FE YES YES YES YES Prov Birth Year Trend YES YES YES YES Province FE YES YES NO NO Community FE NO NO YES YES Number of groups 28 28 427 420 LM stat for underid. 7.746 7.195 p-val of LM stat 0.0208 0.0274 F stat for weak-id. 5.945 4.205 Anderson-Rubin F stat 0.814 0.117 p-val of AR F stat 0.454 0.890 p-val of Hansen J stat 0.272 0.692 Standard errors clustered at province level in parentheses. *** p < 0:01, ** p< 0:05, * p< 0:1. 174 Table B6: Level Eects of Children's Education on Baseline Health Behaviors of Parents: Frequency (1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV VARIABLES # of cigarettes per day (men) # drinks per week Education of Children -0.226*** 0.871 -0.219*** 1.234 -0.012 0.201 0.004 0.370 (0.069) (1.354) (0.075) (2.110) (0.011) (0.210) (0.015) (0.340) Own Education -0.106* -0.387 -0.099* -0.438 -0.025** -0.075 -0.017* -0.087 (0.054) (0.363) (0.054) (0.508) (0.010) (0.048) (0.010) (0.063) Observations 4,300 4,300 4,300 4,279 9,422 9,422 9,422 9,410 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 28 420 399 28 28 429 417 Mean of Dependent Variable 9.862 9.862 9.862 9.862 1.601 1.601 1.601 1.601 LM stat for underid. 9.145 8.004 10.36 9.878 p-val of LM stat 0.0103 0.0183 0.00564 0.00716 F stat for weak-id. 13.31 7.252 10.41 8.155 Anderson-Rubin F stat 0.243 0.192 1.207 1.113 p-val of AR F stat 0.786 0.826 0.315 0.343 p-val of Hansen J stat 0.771 0.917 0.427 0.650 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 175 Table B7: Level Eects of Children's Education on Transfer of Time: One IV (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) OLS IV 1 IV 1+2 OLS IV 1 IV 1+2 OLS IV 1 IV 1+2 OLS IV 1 IV 1+2 VARIABLES Frequency of Visits of Children: # d/w Frequency of Contact with Children not in person: # d/w Education of Children -0.039*** -0.197 0.043 -0.045*** -0.384 -0.124 0.040*** 0.580* 0.204 0.020* 0.578** 0.249 (0.013) (0.398) (0.268) (0.013) (0.514) (0.327) (0.012) (0.307) (0.233) (0.010) (0.292) (0.250) Own Education -0.004 0.040 -0.027 -0.004 0.078 0.015 0.043*** -0.109 -0.003 0.041*** -0.096 -0.015 (0.010) (0.112) (0.077) (0.010) (0.122) (0.079) (0.012) (0.090) (0.065) (0.014) (0.077) (0.063) Observations 6,547 6,547 6,547 6,547 6,533 6,533 5,210 5,210 5,210 5,210 5,193 5,193 Child Birth Year FE YES YES YES YES YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES YES YES YES YES Province FE YES YES YES NO NO NO YES YES YES NO NO NO Community FE NO NO NO YES YES YES NO NO NO YES YES YES Number of groups 28 28 28 429 415 415 28 28 28 423 406 406 LM stat for underid. 3.126 6.760 2.237 7.578 3.672 8.471 2.800 6.640 p-val of LM stat 0.0771 0.0341 0.135 0.0226 0.0553 0.0145 0.0942 0.0361 F stat for weak-id. 7.458 8.498 3.732 6.878 7.100 10.06 4.630 7.235 Anderson-Rubin F stat 0.243 1.695 0.637 0.591 8.781 5.541 9.943 5.194 p-val of AR F stat 0.626 0.203 0.432 0.561 0.00629 0.00963 0.00393 0.0123 p-val of Hansen J stat 0.203 0.291 0.0323 0.128 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 176 Table B8: Level Eects of Children's Education on Labor Supply: Rural Sample (1) (2) (3) (4) OLS IV 1+2 OLS IV 1+2 VARIABLES Whether work now (rural) Education of Children -0.006*** -0.181** -0.003** -0.348 (0.001) (0.085) (0.001) (0.225) Own Education -0.005*** 0.040* -0.003** 0.075 (0.002) (0.022) (0.001) (0.052) Observations 5,053 5,053 5,053 5,037 Child Birth Year FE YES YES YES YES Prov Birth Year Trend YES YES YES YES Province FE YES YES NO NO County FE NO NO NO NO Community FE NO NO YES YES Number of groups 28 28 423 407 LM stat for underid. 4.639 3.309 p-val of LM stat 0.0983 0.191 F stat for weak-id. 4.624 2.009 Anderson-Rubin F stat 7.677 8.696 p-val of AR F stat 0.00229 0.00122 p-val of Hansen J stat 0.138 0.473 Standard errors clustered at province level in parentheses. *** p < 0:01, ** p< 0:05, * p< 0:1. 177 Table B9: Robustness with Alternative Instruments: Level Eects of Children's Education on Baseline Cognition of Parents (1) (2) (3) (4) (5) (6) (7) (8) OLS IV 1+2 OLS IV 1+2 OLS IV 1+2 OLS IV 1+2 VARIABLES Baseline Mental Intactness Baseline Episodic Memory Education of Children 0.107*** 0.288*** 0.102*** 0.328*** 0.045*** 0.202*** 0.038*** 0.214*** (0.006) (0.105) (0.007) (0.119) (0.006) (0.070) (0.005) (0.074) Own Education 0.246*** 0.203*** 0.244*** 0.201*** 0.099*** 0.063*** 0.093*** 0.059*** (0.008) (0.026) (0.009) (0.024) (0.007) (0.017) (0.006) (0.015) Observations 10,776 10,776 10,776 10,769 9,988 9,988 9,988 9,979 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 28 429 422 28 28 429 420 LM stat for underid. 55.94 42.46 53.69 44.20 p-val of LM stat 2.97e-05 0.00241 6.43e-05 0.00142 F stat for weak-id. 2.864 2.146 2.714 2.235 Anderson-Rubin F stat 2.167 2.033 1.551 1.963 p-val of AR F stat 0.00188 0.00418 0.0551 0.00628 p-val of Hansen J stat 0.0214 0.0457 0.354 0.112 Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Table B10: Robustness with Alternative Instruments: Level Eects of Children's Education on Baseline Subjective Health Measures of Parents (1) (2) (3) (4) (5) (6) (7) (8) OLS IV 1+2 OLS IV 1+2 OLS IV 1+2 OLS IV 1+2 VARIABLES Baseline Self Reported Health Baseline Low Subjective Survival Expectation Education of Children -0.018*** -0.010 -0.017*** -0.037 -0.009*** -0.040** -0.008*** -0.048** (0.003) (0.033) (0.003) (0.039) (0.001) (0.019) (0.001) (0.021) Own Education -0.011*** -0.013 -0.011*** -0.007 -0.008*** -0.001 -0.007*** 0.001 (0.003) (0.008) (0.003) (0.008) (0.001) (0.005) (0.001) (0.004) Observations 10,808 10,808 10,808 10,801 9,373 9,373 9,373 9,367 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 28 429 422 28 28 426 420 LM stat for underid. 57.21 44.17 50.63 43.37 p-val of LM stat 1.91e-05 0.00143 0.000180 0.00182 F stat for weak-id. 2.930 2.234 2.602 2.200 Anderson-Rubin F stat 0.804 0.788 2.288 1.646 p-val of AR F stat 0.712 0.731 0.000888 0.0347 p-val of Hansen J stat 0.665 0.755 0.0102 0.223 Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 178 Table B11: Robustness with Alternative Instruments: Level Eects of Children's Education on Baseline Physical Health of Parents (1) (2) (3) (4) (5) (6) (7) (8) OLS IV 1+2 OLS IV 1+2 OLS IV 1+2 OLS IV 1+2 VARIABLES Baseline Peak Flow Baseline Grip Strength Education of Children 1.433*** 13.277*** 1.163*** 13.699*** 0.107*** 0.288*** 0.102*** 0.328*** (0.308) (3.990) (0.246) (4.270) (0.006) (0.105) (0.007) (0.119) Own Education 2.329*** -0.469 1.808*** -0.635 0.246*** 0.203*** 0.244*** 0.201*** (0.371) (0.977) (0.264) (0.870) (0.008) (0.026) (0.009) (0.024) Observations 10,290 10,290 10,290 10,283 10,776 10,776 10,776 10,769 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES Number of groups 28 28 426 419 28 28 429 422 LM stat for underid. 56.55 43 55.94 42.46 p-val of LM stat 2.40e-05 0.00204 2.97e-05 0.00241 F stat for weak-id. 2.891 2.165 2.864 2.146 Anderson-Rubin F stat 1.706 1.567 2.167 2.033 p-val of AR F stat 0.0255 0.0512 0.00188 0.00418 p-val of Hansen J stat 0.383 0.628 0.0214 0.0457 Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 179 Table B12: Robustness with Alternative Instruments: Level Eects of Children's Education on Baseline Body Weight of Parents (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) OLS IV OLS IV OLS IV OLS IV OLS IV OLS IV VARIABLES Baseline BMI Baseline Underweight Baseline Overweight Education of Children 0.076*** 0.354** 0.064*** 0.287* -0.004*** -0.014 -0.003*** -0.005 0.006*** 0.026 0.005*** 0.027 (0.014) (0.153) (0.010) (0.172) (0.001) (0.011) (0.001) (0.012) (0.002) (0.017) (0.002) (0.020) Own Education 0.023** -0.043 0.004 -0.039 -0.001 0.002 -0.000 -0.000 0.002 -0.003 0.000 -0.004 (0.011) (0.038) (0.012) (0.035) (0.001) (0.003) (0.001) (0.002) (0.002) (0.004) (0.002) (0.004) Observations 10,769 10,769 10,769 10,763 10,769 10,769 10,769 10,763 10,769 10,769 10,769 10,763 Child Birth Year FE YES YES YES YES YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES YES YES YES YES Province FE YES YES NO NO YES YES NO NO YES YES NO NO Community FE NO NO YES YES NO NO YES YES NO NO YES YES Number of groups 28 28 428 422 28 28 428 422 28 28 428 422 LM stat for underid. 56.54 42.13 56.54 42.13 56.54 42.13 p-val of LM stat 2.41e-05 0.00266 2.41e-05 0.00266 2.41e-05 0.00266 F stat for weak-id. 2.894 2.129 2.894 2.129 2.894 2.129 Anderson-Rubin F stat 1.241 1.195 1.919 1.815 1.092 1.057 p-val of AR F stat 0.209 0.247 0.00804 0.0143 0.350 0.390 p-val of Hansen J stat 0.369 0.295 0.0110 0.0118 0.458 0.462 Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 180 Table B13: Robustness with Alternative Instruments: Level Eects of Children's Education on Baseline Mental Health (CESD 10) of Parents (1) (2) (3) (4) OLS IV OLS IV VARIABLES Baseline CESD 10 Score Education of Children -0.155*** 0.129 -0.142*** -0.040 (0.028) (0.254) (0.021) (0.281) Own Education -0.134*** -0.201*** -0.130*** -0.150*** (0.015) (0.062) (0.015) (0.056) Observations 10,220 10,220 10,220 10,216 Child Birth Year FE YES YES YES YES Prov Birth Year Trend YES YES YES YES Province FE YES YES NO NO Community FE NO NO YES YES Number of groups 28 28 426 422 LM stat for underid. 50.79 39.24 p-val of LM stat 0.000171 0.00622 F stat for weak-id. 2.605 1.992 Anderson-Rubin F stat 1.877 1.421 p-val of AR F stat 0.0102 0.100 p-val of Hansen J stat 0.0114 0.0835 Robust standard errors in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 181 Table B14: Robustness to Control of OCP: Level Eects of Children's Education on Baseline Health and Cognition of Parents (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) IV VARIABLES Mental Intactness Episodic Memory Peak Flow Grip Strength Low Expected Survival Education of Children 0.436** 0.488** 0.239** 0.195* 20.817*** 24.242*** 0.483 0.731 -0.093* -0.089** (0.184) (0.232) (0.098) (0.102) (7.594) (9.386) (0.440) (0.544) (0.049) (0.046) Own Education 0.168*** 0.170*** 0.054** 0.063*** -2.250 -2.688 0.057 0.007 0.012 0.009 (0.041) (0.043) (0.023) (0.020) (1.935) (1.854) (0.101) (0.103) (0.012) (0.010) Observations 10,776 10,769 9,988 9,979 10,290 10,283 10,610 10,603 9,373 9,367 Child Birth Year FE YES YES YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES YES YES Province FE YES NO YES NO YES NO YES NO YES NO Community FE NO YES NO YES NO YES NO YES NO YES Number of groups 28 422 28 420 28 419 28 420 28 420 LM stat for underid. 9.643 9.357 11.21 10.20 10.08 9.545 9.255 8.138 8.386 8.545 p-val of LM stat 0.00806 0.00929 0.00368 0.00611 0.00647 0.00846 0.00978 0.0171 0.0151 0.0139 F stat for weak-id. 8.015 5.967 10.86 7.140 9.390 6.562 7.344 5.066 7.390 6.259 Anderson-Rubin F stat 3.901 3.272 6.068 3.549 7.527 6.075 1.455 1.436 2.470 3.318 p-val of AR F stat 0.0325 0.0534 0.00666 0.0428 0.00252 0.00663 0.251 0.255 0.103 0.0515 p-val of Hansen J stat 0.460 0.507 0.189 0.0640 0.288 0.539 0.506 0.995 0.772 0.816 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 182 Table B15: Robustness to Control of OCP: Level Eects of Children's Education on Base- line Body Weight and Depressive Symptoms of Parents (1) (2) (3) (4) (5) (6) (7) (8) IV VARIABLES BMI Underweight Overweight CESD Education of Children 0.877*** 0.924** -0.070** -0.067** 0.064*** 0.069** -0.254 -0.475 (0.313) (0.377) (0.031) (0.032) (0.024) (0.031) (0.377) (0.516) Own Education -0.168** -0.163** 0.015* 0.012* -0.012** -0.012* -0.111 -0.067 (0.073) (0.072) (0.008) (0.007) (0.006) (0.006) (0.091) (0.101) Observations 10,769 10,763 10,769 10,763 10,769 10,763 10,220 10,216 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES NO YES NO YES NO YES NO Community FE NO YES NO YES NO YES NO YES Number of groups 28 422 28 422 28 422 28 422 LM stat for underid. 9.620 9.190 9.620 9.190 9.620 9.190 8.884 8.523 p-val of LM stat 0.00815 0.0101 0.00815 0.0101 0.00815 0.0101 0.0118 0.0141 F stat for weak-id. 8.394 6.294 8.394 6.294 8.394 6.294 6.662 5.127 Anderson-Rubin F stat 5.675 3.054 6.544 4.420 2.978 1.863 0.376 1.180 p-val of AR F stat 0.00876 0.0637 0.00482 0.0218 0.0678 0.175 0.690 0.323 p-val of Hansen J stat 0.934 0.522 0.527 0.358 0.707 0.965 0.523 0.193 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 183 Table B16: Instrumenting Own Education: Level Eects of Children's Education on Baseline Health and Cognition of Parents (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) IV+Lewbel VARIABLES Baseline Mental Intactness Baseline Episodic Memory Baseline Peak Flow Low Expected Survival Baseline Grip Strength Education of Children 0.252*** 0.163*** 0.457*** 0.378** 17.929** 16.923** -0.068* -0.065** 0.179 0.878* (0.081) (0.062) (0.146) (0.161) (7.867) (7.604) (0.036) (0.030) (0.445) (0.456) Own Education 0.130*** 0.123*** 0.180*** 0.147*** 3.485* 2.384 -0.001 0.002 -0.087 -0.055 (0.035) (0.031) (0.044) (0.041) (1.883) (1.754) (0.009) (0.008) (0.156) (0.134) Observations 9,988 9,979 10,776 10,769 10,290 10,283 9,373 9,367 10,610 10,603 Child Birth Year FE YES YES YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES YES YES Province FE YES NO YES NO YES NO YES NO YES NO Community FE NO YES NO YES NO YES NO YES NO YES LM stat for underid. 14.16 14.23 13.50 14.12 13.40 12.74 12.35 12.89 13.79 14.77 p-val of LM stat 0.0778 0.0761 0.0959 0.0786 0.0990 0.121 0.136 0.116 0.0875 0.0639 F stat for weak-id. 4.560 5.413 4.337 5.656 3.754 4.119 3.651 4.889 5.445 7.953 Anderson-Rubin F stat 2.569 5.149 1.657 2.069 4.191 3.310 2.084 2.635 1.163 4.645 p-val of AR F stat 0.0279 0.000425 0.149 0.0698 0.00179 0.00758 0.0679 0.0248 0.357 0.000891 p-val of Hansen J stat 0.323 0.165 0.696 0.546 0.471 0.389 0.359 0.388 0.583 0.429 Breusch Pagan test p-val 2.13e-22 1.11e-25 4.96e-30 3.51e-33 1.54e-25 1.43000e-30 2.60e-20 5.26e-21 1.42e-28 2.68e-31 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 184 Table B17: Instrumenting Own Education: Level Eects of Children's Education on Baseline Body Weight and Depressive Symptoms of Parents (1) (2) (3) (4) (5) (6) (7) (8) IV+Lewbel VARIABLES Baseline BMI Baseline Underweight Baseline Overweight Baseline CESD 10 Score Education of Children 0.679** 0.523* -0.057*** -0.037** 0.046* 0.036 -0.474 -0.700** (0.278) (0.285) (0.021) (0.016) (0.025) (0.028) (0.304) (0.298) Own Education 0.004 -0.046 -0.003 0.002 -0.001 -0.004 -0.224* -0.201* (0.093) (0.080) (0.007) (0.005) (0.007) (0.007) (0.127) (0.109) Observations 10,769 10,763 10,769 10,763 10,769 10,763 10,220 10,216 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES NO YES NO YES NO YES NO Community FE NO YES NO YES NO YES NO YES LM stat for underid. 13.57 13.82 13.57 13.82 13.57 13.82 13.88 15 p-val of LM stat 0.0936 0.0865 0.0936 0.0865 0.0936 0.0865 0.0850 0.0591 F stat for weak-id. 4.543 5.809 4.543 5.809 4.543 5.809 3.270 5.247 Anderson-Rubin F stat 2.142 1.766 4.426 3.280 1.250 1.801 1.478 1.429 p-val of AR F stat 0.0610 0.122 0.00124 0.00798 0.308 0.114 0.206 0.225 p-val of Hansen J stat 0.801 0.461 0.375 0.306 0.644 0.541 0.510 0.562 Breusch Pagan test p-val 7.41000e-30 8.10000e-33 7.41000e-30 8.10000e-33 7.41000e-30 8.10000e-33 3.58000e-24 2.09000e-26 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 185 Table B18: Own Education as Exogenous: Incremental Eects of Children's Education on Wave 2 Health and Cognition of Parents (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) IV+Lewbel VARIABLES Mental Intactness Episodic Memory Peak Flow Grip Strength Low Expected Survival Baseline 0.579*** 0.593*** 0.478*** 0.383*** 0.508*** 0.505*** 0.469*** 0.503*** 0.004 -0.033 (0.072) (0.080) (0.118) (0.107) (0.038) (0.039) (0.068) (0.076) (0.253) (0.245) Education of Children 0.268** 0.269 0.125** 0.129** -5.625 -2.952 0.279 -0.025 -0.042 -0.023 (0.130) (0.181) (0.061) (0.060) (4.176) (4.713) (0.691) (0.916) (0.055) (0.056) Own Education 0.065** 0.070** 0.023 0.037* 2.295** 1.535* 0.016 0.068 -0.002 -0.006 (0.031) (0.028) (0.021) (0.021) (1.069) (0.873) (0.157) (0.170) (0.011) (0.009) Observations 10,239 10,232 9,507 9,499 7,382 7,371 7,892 7,883 7,071 7,062 Child Birth Year FE YES YES YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES YES YES Province FE YES NO YES NO YES NO YES NO YES NO Community FE NO YES NO YES NO YES NO YES NO YES Number of groups 28 420 28 418 27 399 28 407 28 415 Mean of Dependent Variable 7.052 7.052 3.340 3.340 268 268 29.74 29.74 0.331 0.331 LM stat for underid. 13.32 12.06 13.32 23.51 15.54 16.34 10.12 11.18 13.33 9.896 p-val of LM stat 0.207 0.281 0.206 0.172 0.114 0.0903 0.430 0.343 0.206 0.450 F stat for weak-id. 3.790 2.573 2.847 16.60 12.85 12.31 1.585 1.739 3.885 1.944 Anderson-Rubin F stat 8.802 7.758 3.256 7.247 25.11 26.40 7.521 6.982 4.118 3.598 p-val of AR F stat 2.25e-06 7.44e-06 0.00606 2.17e-06 0 0 9.91e-06 1.94e-05 0.00132 0.00325 p-val of Hansen J stat 0.625 0.820 0.630 0.673 0.281 0.0735 0.461 0.871 0.121 0.152 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 186 Table B19: Own Education as Exogenous: Incremental Eects of Children's Education on Wave 2 Body Weight and Depressive Symptoms of Parents (1) (2) (3) (4) (5) (6) (7) (8) Lewbel+IV VARIABLES Wave 2 BMI Wave 2 Underweight Wave 2 Overweight Wave 2 CESD Baseline 0.359*** 0.372*** 0.740*** 0.742*** 0.719*** 0.711*** 0.527*** 0.555*** (0.077) (0.080) (0.047) (0.047) (0.040) (0.036) (0.070) (0.081) Education of Children -0.355 -0.562 -0.004 -0.003 0.029** 0.020 -0.280 -0.195 (0.341) (0.462) (0.028) (0.040) (0.015) (0.025) (0.303) (0.356) Own Education 0.108 0.116 0.001 0.000 -0.005 -0.003 0.032 0.027 (0.088) (0.093) (0.007) (0.008) (0.004) (0.005) (0.077) (0.074) Observations 8,057 8,050 8,057 8,050 8,057 8,050 8,705 8,700 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES NO YES NO YES NO YES NO Community FE NO YES NO YES NO YES NO YES Number of groups 28 410 28 410 28 410 28 420 Mean of Dependent Variable 23.77 23.77 0.0604 0.0604 0.347 0.347 7.829 7.829 LM stat for underid. 12.71 9.665 10.81 12.10 12.75 7.151 12.80 15.15 p-val of LM stat 0.240 0.470 0.373 0.279 0.238 0.711 0.235 0.127 F stat for weak-id. 2.739 1.320 2.049 2.970 4.353 0.827 2.918 4.231 Anderson-Rubin F stat 4.918 5.431 45.25 34.27 7.687 9.613 20 9.986 p-val of AR F stat 0.000358 0.000164 0 0 8.10e-06 9.49e-07 3.44e-10 6.49e-07 p-val of Hansen J stat 0.299 0.393 0.412 0.489 0.125 0.207 0.635 0.713 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Table B20: Robustness to Adjustment for Multiple Hypotheses Testing Province FE Community FE Baseline Outcome P Val New Critical Val Sig at 5%* P Val New Critical Val Sig at 5%* Mental Intactness 0.00572 0.015 1 0.0153 0.015 1 Episodic Memory 0.00638 0.02 1 0.0273 0.025 1 Self Reported Health 0.291 0.045 0 0.251 0.045 0 Low Expected Survival 0.0388 0.035 0 0.0281 0.03 1 Peak Flow 0.00808 0.025 1 0.0085 0.005 1 Grip Strength 0.263 0.04 0 0.174 0.04 0 CESD 10 Index 0.446 0.05 0 0.32 0.05 0 BMI 0.00208 0.005 1 0.0108 0.01 1 Underweight 0.0175 0.03 1 0.0322 0.035 1 Overweight 0.004 0.01 1 0.0272 0.02 1 *Uncorrected overall signicance level. New (corrected) critical p-values are based on adjustment method in Simes (1986). 187 Table B21: Robustness with Lewbel IV for Children's Education: Level Eects of Children's Education on Baseline Cognition and Health of Parents (1) (2) (3) (4) (5) (6) (7) (8) IV+Lewbel VARIABLES Mental Intactness Episodic Memory Peak Flow Low Survival Education of Children 0.438*** 0.412*** 0.200*** 0.093 11.751*** 11.060** -0.032 -0.032 (0.114) (0.120) (0.075) (0.065) (3.736) (4.756) (0.025) (0.024) Own Education 0.167*** 0.184*** 0.063*** 0.082*** -0.109 -0.120 -0.003 -0.002 (0.028) (0.025) (0.019) (0.014) (1.111) (0.968) (0.006) (0.005) Observations 10,776 10,769 9,988 9,979 10,290 10,283 9,373 9,367 Child Birth Year FE YES YES YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES YES YES Province FE YES NO YES NO YES NO YES NO Community FE NO YES NO YES NO YES NO YES LM stat for underid. 12.75 14.16 13.16 14.49 13.24 13.79 13.46 14.56 p-val of LM stat 0.238 0.166 0.215 0.152 0.211 0.183 0.199 0.149 F stat for weak-id. 5.438 3.856 5.481 4.471 5.043 3.865 12.45 7.360 Anderson-Rubin F stat 7.031 7.111 6.451 3.686 3.377 4.897 1.288 1.296 p-val of AR F stat 2.50e-05 2.26e-05 5.24e-05 0.00332 0.00568 0.000470 0.286 0.282 p-val of Hansen J stat 0.741 0.805 0.154 0.235 0.730 0.475 0.715 0.792 Breusch Pagan test p-val 3.88e-35 1.58e-72 1.66e-28 1.50e-61 2.79e-36 1.95e-76 2.24e-26 7.36e-51 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. Table B22: Robustness with Lewbel IV for Children's Education: Level Eects of Children's Education on Baseline Body Weight of Parents (1) (2) (3) (4) (5) (6) IV+Lewbel VARIABLES Baseline BMI Baseline Underweight Baseline Overweight Education of Children 0.752*** 0.771*** -0.041*** -0.034** 0.057*** 0.064*** (0.191) (0.211) (0.016) (0.015) (0.017) (0.021) Own Education -0.138*** -0.133*** 0.008** 0.005* -0.011** -0.011** (0.049) (0.046) (0.004) (0.003) (0.004) (0.004) Observations 10,769 10,763 10,769 10,763 10,769 10,763 Child Birth Year FE YES YES YES YES YES YES Prov Birth Year Trend YES YES YES YES YES YES Province FE YES NO YES NO YES NO Community FE NO YES NO YES NO YES LM stat for underid. 12.37 13.60 12.37 13.60 12.37 13.60 p-val of LM stat 0.261 0.192 0.261 0.192 0.261 0.192 F stat for weak-id. 4.952 3.695 4.952 3.695 4.952 3.695 Anderson-Rubin F stat 4.788 5.901 1.832 1.434 3.368 2.861 p-val of AR F stat 0.000555 0.000110 0.103 0.219 0.00577 0.0144 p-val of Hansen J stat 0.997 0.993 0.550 0.539 0.899 0.750 Breusch Pagan test p-val 1.25e-35 4.56e-72 1.25e-35 4.56e-72 1.25e-35 4.56e-72 Standard errors clustered at province level in parentheses. *** p< 0:01, ** p< 0:05, * p< 0:1. 188 Lewbel's IV Approach Based on Heteroskedasticity Lewbel (2012) introduces an instrumental variable estimation method for linear regression models with endogenous independent variables when there are no plausible external instruments available or when traditional instruments are not sucient. For a traditional IV estimation, the instruments must satisfy three conditions at the same time: (1) being exogenous or orthogonal to the error term of the model, (2) being suciently correlated with the endogenous variables, and (3) being excludeable from the model such that it does not directly relate to outcome of interest. When condition (3) is violated, traditional IV estimation method is not valid and the model remains unidentied. Lewbel (2012) hence proposes to generate instruments that satisfy the three conditions by having heteroskedasticity in errors and regressors orthogonal to the product of heteroskedastic errors in structural models to achieve identication. Generally, in a structural model without excludeable instruments as: Y 1 =X 0 1 +Y 2 1 +e 1 ; Y 2 =X 0 2 +Y 1 2 +e 2 : Lewbel (2012) shows that, under (1) heteroskedasicity, i.e., E(X;e 2 j )6= 0;j = 1; 2, as well as (2) Cov(Z;e 1 e 2 ) = 0, whereZ is a subset ofX, and (3)E(X;e j ) = 0;j = 1; 2, the structural parameters are identied. Instead of the exclusion restriction, heteroskedasicity in (1) that is a feature of many models and higher moment conditions in (2) are required. Lewbel (2012) notes that the assumptions can be approximately validated by a Breusch-Pagan type test of heteroskedasticity (scale-related) for (1) and Sargan-Hansen test of overidentication for (2). In context of estimating a dynamic health production model with an endogenous lagged de- pendent variable as in Brown (2014), it is dicult to nd any excludeable instruments that only indirectly relate to current health through past health without having to make very strong assump- tions. In light of this, Lewbel's IV approach can help correct for the endogeneity bias arising from lagged dependent variable with less restrictive assumptions that are standard in many econometric models according to Lewbel (2012). Therefore, Lewbel's IV approach is applied to the dynamic model to supplement traditional IV estimation. Specically, I have a triangular system: H 1 =X 0 1 +e 1 ; H 2 =X 0 2 +H 1 2 +e 2 : To empirically implement Lewbel's IV estimation method, auxiliary regressions of the baseline health outcomes with only exogenous independent variables are conducted to generate residuals, ^ e 1 , which are then tested against homoskedasticity. Instruments are constructed by multiplying Z, a subset of mean-centered exogenous variables, with the residuals, i.e., (Z Z)^ e 1 . It implies that zero mean condition of traditional IV is fullled asE((Z Z)^ e 1 ) = 0 because covariance of ^ e 1 with Z is zero by construction. Lewbel (2012) type IVs are only generated for lagged health, not for years of schooling of highest educated adult child such that the results on the eect of children's education from Lewbel IV approach is not driven by the use of generated instruments 1 . Z includes age categories and respondent's gender which are exogenous and strongly correlated with both past and current health of parents which is implied by the rst stage. I also include in Z the baseline height which is a measure of childhood health and shown to be strongly related to later life health 1 Lewbel's IV approach results when children's education is also instrumented by generated instruments are avail- able upon request. 189 among Chinese elderly by Smith et al. (2012) and conrmed by the rst stage. The generated instruments for lagged health as well as the external instruments for children's years of schooling are used to estimate the dynamic models using IV/2SLS method when both are treated endogenous. Breusch-Pagan tests on residuals from auxiliary regressions and Hansen J test of overidentication of external and generated instruments are reported. 190 Model Parental Health as an AR(1) Process Assume H 0 is the health endowment of older adults. Children's education aects parental health in a fashion as below: H 1 =H 0 + 1 ChildEdu +e 1 H 2 =H 1 + 2 ChildEdu +e 2 H t =H t1 + t ChildEdu +e t where the health status of parents follows an AR(1) process. It assumes that lagged 1 period health is a sucient statistics for impacts of all past health inputs (Strauss and Thomas, 2008). Children's education has a direct and periodic impact on parental health in each period which is captured by t . The periodic eect, t , is the incremental eect to be estimated in the dynamic model. Also by iteration, H t = t H 0 + ( t1 1 + t2 2 +::: + t )ChildEdu + ( t1 e 1 + t2 e 2 +::: +e t ) (B.1) or, we have H t =a t +b t ChildEdu +u t (B.2) Therefore, at time t, estimating equation (5) gives the level (cumulative) eect of children's education,b t , which is a weighted average of the periodic/incremental eect of children's education in the current and all previous periods. By taking rst dierence of equation (4), H t+1 H t =a t+1 a t + (b t+1 b t )ChildEdu +u t+1 u t (B.3) For simplicity, assume t = . Therefore, estimation of the rst dierence equation actually gives the change in the level eects of children's education, which is t , instead of the incremental eect. They will be the same if equals 1. However, as what I have shown in the dynamic model results, the state dependence is relative low and not close to 1, which implies that estimating a rst dierence equation will not provide what is the main eect of interest in this paper. The change in the level eect , t , can be also interpreted as the persistent eect of children's education since it measures how the initial incremental eect persists into future time t. 191 Sample Selection As mentioned in the main text, I conducted the analysis on a relative balanced panel of parents who or whose household responded to both the baseline survey and its rst follow-up. While the restriction of sample is due to data limitations on child information in the baseline, it could result in biased estimates for the health eects of children's education if such sample selection is not random and correlated with parents' health and/or their children's education. Although there is no clear evidence as to the association between children's education and attrition, it is very possible that healthier parents are better able to survive to the follow-up survey (positive selection on health); or it is possible that healthier parents to be able to move and therefore attrite (negative selection on health). If there is positive selection on health, and the eect of children's education is stronger on parents with worse health, the sample selection might lead to underestimation for the causal eects of children's education, or vice versa, overestimation if the eect of children's education is stronger on parents with better eect. In theory, the direction of biases due to sample selection is unclear. To test whether there is selection on baseline health, I run an OLS regression of whether the parent remained in wave 2 on their baseline health measures that are found to be signicantly aected by children's education in the baseline static model, including mental intactness, episodic memory, peak ow, low expected survival and BMI, as well as the same set of control variables as in the IV estimation model. Tests of joint signicance on these health measures in both province xed eects model (F(4,27) = 0.63, p value = 0.6429) and community xed eects model (F(4,27) = 1.82, p value = 0.1547) show that they are not jointly signicant. Therefore, there seems no severe sample selection on baseline health when one includes multiple health measures. These results lend more credibility to the main results of this paper. 192 Appendix C Chapter 4 Appendix Data Cleaning We set the lower limit for measured height at 100cm. We did not need to drop any observations in ELSA as per this rule. Measurement error seemed highest in ELSA Wave 4, and lowest in Wave 6 (Latest Height Measurement), as seen by the extent of outliers for both men and women. In the IFLS, measurement error seemed highest in Wave 4 (Height 7 years ago). We dropped 17 respondents from Wave 4 and 2 respondents from Wave 5 for whom measured height is less than 100cm. For the extent of change in height, we dropped respondents for whom height gain over the years is more than 10cm (shrinkage<-10cm), or height loss is more than 10cm, trimming away the outliers. We dropped 18 respondents in the ELSA sample as per this rule. In the IFLS sample, we dropped the following respondents: 30 for whom height gain over 7 years is more than 10cm, 2 each for whom height gain over 14 and 17 years is more than 10cm. 45, 38 and 20 respondents exhibit height shrinkage more than 10cm over 7, 14 and 17 years respectively- we dropped these as well. 193 Tables Table C1: Limb Length and Shrinkage in IFLS Male sample (1) (2) (3) (4) (5) (6) shrinkage shrinkage shrinkage shrinkage shrinkage shrinkage VARIABLES over 7 yrs over 7 yrs over 14 yrs over 14 yrs over 17 yrs over 17 yrs Upper arm length 0.0295 0.00726 0.00169 (0.0181) (0.0300) (0.0253) Knee Height 0.0263 0.0127 0.0107 (0.0207) (0.0196) (0.0207) Observations 970 970 970 970 970 970 R-squared 0.054 0.055 0.079 0.080 0.099 0.099 Female sample (1) (2) (3) (4) (5) (6) shrinkage shrinkage shrinkage shrinkage shrinkage shrinkage VARIABLES over 7 yrs over 7 yrs over 14 yrs over 14 yrs over 17 yrs over 17 yrs Upper arm length -0.0565* -0.00245 -0.00265 (0.0331) (0.0311) (0.0375) Knee Height -0.0279 -0.0210 -0.00522 (0.0203) (0.0239) (0.0256) Observations 1,086 1,086 1,086 1,086 1,086 1,086 R-squared 0.049 0.046 0.096 0.097 0.107 0.107 Note: Robust standard errors clustered at district level in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed. IFLS results control for district xed eects. 194 Table C2: Height Shrinkage and ADL in ELSA (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Male sample Panel B: Female sample Dependent Variable: ADL Index 4 years 8 years 4 years 8 years Current Height -0.00138 -0.00701** (0.00309) (0.00323) Height 4/8 yrs ago 0.000107 0.000737 0.000489 0.000472 -0.00590* -0.00575* -0.00541* -0.00521 (0.00313) (0.00313) (0.00315) (0.00315) (0.00327) (0.00326) (0.00328) (0.00328) Height loss over 4/8 yrs 0.0287** 0.0338*** 0.0352*** 0.0338*** (0.0124) (0.0126) (0.0126) (0.0123) Height loss: 1-1.9cm -0.0492 -0.0235 0.0724 -0.0923* (0.0534) (0.0511) (0.0488) (0.0498) Height loss: 2-2.9cm -0.0291 0.00762 0.186*** 0.0374 (0.0645) (0.0608) (0.0661) (0.0580) Height loss: 3cm + 0.189** 0.188*** 0.116 0.219*** (0.0803) (0.0656) (0.0847) (0.0669) Observations 1,042 1,042 1,042 1,042 1,042 1,166 1,166 1,166 1,166 1,166 R-squared 0.065 0.050 0.052 0.053 0.056 0.058 0.064 0.065 0.065 0.075 Note: Robust standard errors in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed. 195 Table C3: Height Shrinkage and IADL in ELSA (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Male sample Panel B: Female sample Dependent Variable: IADL Index 4 years 8 years 4 years 8 years Current Height -0.00334 -0.00689*** (0.00226) (0.00220) Height 4/8 yrs ago -0.00285 -0.00287 -0.00223 -0.00222 -0.00649*** -0.00653*** -0.00595*** -0.00570** (0.00228) (0.00229) (0.00229) (0.00229) (0.00223) (0.00223) (0.00224) (0.00224) Height loss over 4/8 yrs 0.00999 0.0299*** 0.0176** 0.0235*** (0.00902) (0.00918) (0.00859) (0.00839) Height loss: 1-1.9cm 0.00261 0.00743 0.0348 -0.0479 (0.0390) (0.0371) (0.0333) (0.0341) Height loss: 2-2.9cm -0.00827 0.0643 0.0953** 0.00790 (0.0471) (0.0442) (0.0451) (0.0397) Height loss: 3cm + 0.101* 0.149*** 0.125** 0.139*** (0.0586) (0.0477) (0.0578) (0.0458) Observations 1,042 1,042 1,042 1,042 1,042 1,166 1,166 1,166 1,166 1,166 R-squared 0.049 0.041 0.043 0.052 0.052 0.073 0.075 0.078 0.077 0.083 Note: Robust standard errors in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed. 196 Table C4: Height Shrinkage and ADL in IFLS (1) (2) (3) (4) (5) (6) (7) Panel A: Male sample Dependent Variable: ADL Index 7 years 14 years 17 years Current Height -0.00249 (0.00302) Height 7/14/17 yrs ago -0.00290 -0.00291 -0.000798 -0.000887 -9.57e-05 -0.000368 (0.00302) (0.00300) (0.00308) (0.00309) (0.00330) (0.00335) Height loss over 7/14/17 yrs: 0.0120 0.0257 0.0382** (0.0121) (0.0167) (0.0143) Height loss: 1-1.9cm 0.0475 0.0263 0.0959 (0.0652) (0.0618) (0.0642) Height loss: 2-2.9cm 0.0165 0.128 0.194** (0.0768) (0.0811) (0.0856) Height loss: 3cm + 0.146 0.177* 0.214** (0.103) (0.0923) (0.0861) Observations 612 614 614 614 614 614 614 R-squared 0.063 0.062 0.064 0.066 0.073 0.075 0.081 (1) (2) (3) (4) (5) (6) (7) Panel B: Female sample Dependent Variable: ADL Index 7 years 14 years 17 years Current Height -0.00613 (0.00422) Height 7/14/17 yrs ago -0.00537 -0.00543 -0.00521 -0.00488 -0.00356 -0.00334 (0.00439) (0.00442) (0.00424) (0.00418) (0.00394) (0.00382) Height loss over 7/14/17 yrs: 0.0160 0.0195 0.0289** (0.0132) (0.0131) (0.0141) Height loss: 1-1.9cm 0.0162 -0.0590 0.0187 (0.0491) (0.0625) (0.0514) Height loss: 2-2.9cm 0.151* -0.102* -0.00226 (0.0818) (0.0564) (0.0521) Height loss: 3cm + 0.118 0.145 0.169** (0.0935) (0.0888) (0.0671) Observations 896 897 897 897 897 897 897 R-squared 0.032 0.033 0.038 0.033 0.050 0.037 0.041 Note: Robust standard errors clustered at district level in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed. IFLS results control for district xed eects. 197 Table C5: Height Shrinkage and IADL in IFLS (1) (2) (3) (4) (5) (6) (7) Panel A: Male sample Dependent Variable: IADL Index 7 years 14 years 17 years Current Height -0.0126 (0.00795) Height 7/14/17 yrs ago -0.00846 -0.00790 -0.00691 -0.00606 -0.00478 -0.00486 (0.00773) (0.00808) (0.00824) (0.00825) (0.00814) (0.00800) Height loss over 7/14/17 yrs: 0.0888*** 0.0739*** 0.0927*** (0.0256) (0.0247) (0.0259) Height loss: 1-1.9cm 0.126 0.0605 0.143 (0.143) (0.142) (0.108) Height loss: 2-2.9cm 0.263 0.149 0.266* (0.248) (0.126) (0.141) Height loss: 3cm + 0.751*** 0.517*** 0.541*** (0.249) (0.184) (0.111) Observations 967 970 970 970 970 970 970 R-squared 0.083 0.093 0.096 0.090 0.094 0.095 0.097 (1) (2) (3) (4) (5) (6) (7) Panel B: Female sample Dependent Variable: IADL Index 7 years 14 years 17 years Current Height -0.0240*** (0.00783) Height 7/14/17 yrs ago -0.0188** -0.0186** -0.0188** -0.0185** -0.0168** -0.0164** (0.00814) (0.00809) (0.00845) (0.00837) (0.00726) (0.00744) Height loss over 7/14/17 yrs: 0.0721** 0.0767** 0.0818*** (0.0275) (0.0287) (0.0224) Height loss: 1-1.9cm 0.0868 -0.0248 0.00571 (0.0814) (0.100) (0.108) Height loss: 2-2.9cm 0.240 -0.0320 0.0894 (0.144) (0.126) (0.0912) Height loss: 3cm + 0.745*** 0.411*** 0.479*** (0.218) (0.136) (0.127) Observations 1,088 1,089 1,089 1,089 1,089 1,089 1,089 R-squared 0.090 0.096 0.111 0.096 0.103 0.100 0.103 Note: Robust standard errors in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed. IFLS results control for district xed eects, cluster errors at the same level. 198 Table C6: Height Shrinkage and Diculty in Lifting in ELSA (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Male sample Panel B: Female sample Dep Var: Diculty in Lifting 4 years 8 years 4 years 8 years Current Height -0.00320** -0.00250 (0.00149) (0.00210) Height 4/8 yrs ago -0.00286* -0.00304** -0.00274* -0.00277* -0.00190 -0.00179 -0.00186 -0.00181 (0.00150) (0.00150) (0.00151) (0.00151) (0.00213) (0.00212) (0.00214) (0.00215) Height loss over 4/8 yrs 0.0122** 0.0147** 0.0154* 0.0136* (0.00594) (0.00607) (0.00820) (0.00801) Height loss: 1-1.9cm 0.0258 0.0163 0.0266 -0.0333 (0.0256) (0.0245) (0.0318) (0.0326) Height loss: 2-2.9cm 0.0621** 0.000873 0.115*** 0.0244 (0.0309) (0.0292) (0.0430) (0.0380) Height loss: 3cm + 0.105*** 0.0902*** -0.0375 0.0709 (0.0385) (0.0315) (0.0551) (0.0438) Observations 1,042 1,042 1,042 1,042 1,042 1,166 1,166 1,166 1,166 1,166 R-squared 0.067 0.066 0.071 0.068 0.071 0.091 0.094 0.098 0.094 0.096 Note: Robust standard errors in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed. 199 Table C7: Height Shrinkage and Diculty in Getting up from Chair in ELSA (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Panel A: Male sample Panel B: Female sample Dep Var: Diculty in getting up 4 years 8 years 4 years 8 years Current Height -0.000262 0.00104 (0.00186) (0.00211) Height 4/8 yrs ago 0.000220 -5.69e-05 -6.18e-05 -0.000164 0.00191 0.00229 0.00251 0.00244 (0.00187) (0.00187) (0.00189) (0.00188) (0.00213) (0.00213) (0.00214) (0.00214) Height loss over 4/8 yrs 0.00693 0.00708 0.0200** 0.0256*** (0.00741) (0.00757) (0.00820) (0.00800) Height loss: 1-1.9cm -0.00128 -0.0296 0.0803** 0.0663** (0.0319) (0.0305) (0.0319) (0.0326) Height loss: 2-2.9cm 0.0372 -0.0317 0.0196 0.0852** (0.0385) (0.0364) (0.0431) (0.0379) Height loss: 3cm + 0.125*** 0.0695* 0.0506 0.123*** (0.0480) (0.0392) (0.0552) (0.0438) Observations 1,042 1,042 1,042 1,042 1,042 1,166 1,166 1,166 1,166 1,166 R-squared 0.046 0.044 0.051 0.046 0.052 0.041 0.047 0.048 0.053 0.053 Note: Robust standard errors in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed. 200 Table C8: Height Shrinkage and Diculty in Carrying a Heavy Load in IFLS (1) (2) (3) (4) (5) (6) (7) Panel A: Male sample Dep Var: Diculty in carrying 7 years 14 years 17 years Current Height -0.00445 (0.00295) Height 7/14/17 yrs ago -0.00409 -0.00440 -0.00278 -0.00243 -0.00345 -0.00335 (0.00305) (0.00301) (0.00320) (0.00325) (0.00315) (0.00316) Height loss over 7/14/17 yrs: 0.00846 0.0184** 0.0139* (0.00872) (0.00897) (0.00691) Height loss: 1-1.9cm -0.00560 -0.00712 0.00830 (0.0321) (0.0298) (0.0377) Height loss: 2-2.9cm 0.108** -0.00597 -0.0136 (0.0463) (0.0384) (0.0425) Height loss: 3cm + 0.147** 0.126** 0.0867** (0.0673) (0.0501) (0.0335) Observations 967 970 970 970 970 970 970 R-squared 0.068 0.072 0.079 0.076 0.080 0.072 0.073 (1) (2) (3) (4) (5) (6) (7) Panel B: Female sample Dep Var: Diculty in carrying 7 years 14 years 17 years Current Height -0.00500* (0.00257) Height 7/14/17 yrs ago -0.00387 -0.00405 -0.00262 -0.00245 -0.00269 -0.00258 (0.00270) (0.00268) (0.00282) (0.00282) (0.00270) (0.00273) Height loss over 7/14/17 yrs: 0.0147* 0.0277*** 0.0218*** (0.00869) (0.00831) (0.00647) Height loss: 1-1.9cm 0.0772** -0.00222 -0.0736** (0.0356) (0.0404) (0.0344) Height loss: 2-2.9cm 0.130** -0.0317 0.0182 (0.0617) (0.0365) (0.0372) Height loss: 3cm + 0.151** 0.137*** 0.0944* (0.0655) (0.0410) (0.0478) Observations 1,088 1,089 1,089 1,089 1,089 1,089 1,089 R-squared 0.082 0.083 0.093 0.088 0.094 0.086 0.092 Note: Robust standard errors clustered at district level in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed. IFLS results control for district xed eects. 201 Table C9: Height Shrinkage and Diculty in Getting up from Chair in IFLS (1) (2) (3) (4) (5) (6) (7) Panel A: Male sample Dep Var: Diculty in getting up 7 years 14 years 17 years Current Height -0.000526 (0.00102) Height 7/14/17 yrs ago -0.000714 -0.000703 -0.000250 -0.000184 9.32e-06 -2.68e-05 (0.000956) (0.000949) (0.000913) (0.000904) (0.000904) (0.000884) Height loss over 7/14/17 yrs: 0.000794 0.00401 0.00810** (0.00312) (0.00383) (0.00329) Height loss: 1-1.9cm 0.0115 -0.00941 0.0334** (0.0212) (0.0175) (0.0149) Height loss: 2-2.9cm -0.0296 0.00932 0.0342 (0.0255) (0.0185) (0.0252) Height loss: 3cm + 0.0212 0.0329 0.0367* (0.0352) (0.0288) (0.0183) Observations 967 970 970 970 970 970 970 R-squared 0.057 0.054 0.057 0.057 0.060 0.060 0.061 (1) (2) (3) (4) (5) (6) (7) Panel B: Female sample Dep Var: Diculty in getting up 7 years 14 years 17 years Current Height -0.00335* (0.00171) Height 7/14/17 yrs ago -0.00284 -0.00290 -0.00233 -0.00235 -0.00203 -0.00192 (0.00193) (0.00194) (0.00190) (0.00183) (0.00195) (0.00191) Height loss over 7/14/17 yrs: 0.00736* 0.0127** 0.0134*** (0.00428) (0.00587) (0.00459) Height loss: 1-1.9cm 0.0266 -0.000244 -0.00393 (0.0175) (0.0150) (0.0174) Height loss: 2-2.9cm 0.0496** -0.0159 0.00361 (0.0234) (0.0194) (0.0205) Height loss: 3cm + 0.0828** 0.0873*** 0.0799*** (0.0379) (0.0306) (0.0250) Observations 1,088 1,089 1,089 1,089 1,089 1,089 1,089 R-squared 0.050 0.054 0.061 0.055 0.068 0.057 0.062 Note: Robust standard errors clustered at district level in parentheses. *** p< 0.01, ** p<0.05, * p<0.1. Age, Education, SES controls suppressed. IFLS results control for district xed eects. 202

Abstract (if available)

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Three essays on health & aging

PDF

Essays on health economics

PDF

Essays on development and health economics

PDF

Essays on health insurance programs and policies

PDF

Essays on human capital accumulation -- health and education

PDF

Intergenerational transfers & human capital investments in children in the era of aging

PDF

Essays on family planning policies

PDF

Essays on development economics

PDF

Three essays on health economics

PDF

The determinants and measurement of human capital

PDF

Essays on the empirics of risk and time preferences in Indonesia

PDF

Three essays on cooperation, social interactions, and religion

PDF

Three essays on human capital and family economics

PDF

Long-term impacts of childhood adversity on health and human capital

PDF

Three essays on the evaluation of long-term care insurance policies

PDF

Essays on effects of reputation and judicial transparency

PDF

Essays on development and health economics: social media and education policy

PDF

Three essays on economics of early life health in developing countries

PDF

Essays in health economics and provider behavior

PDF

Biomarkers of age-related health changes: associations with health outcomes and disparities

Asset Metadata

Creator Ma, Mingming (author)

Core Title Essays on health and aging with focus on the spillover of human capital

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Publication Date 07/18/2018

Defense Date 03/26/2018

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag aging,China,family network,health,human capital,OAI-PMH Harvest,upward spillover

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Strauss, John (committee chair), Crimmins, Eileen (committee member), Hsiao, Cheng (committee member), Nugent, Jeffrey (committee member)

Creator Email melody.mingming@gmail.com,mingminm@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-17069

Unique identifier UC11671957

Identifier etd-MaMingming-6411.pdf (filename),usctheses-c89-17069 (legacy record id)

Legacy Identifier etd-MaMingming-6411.pdf

Dmrecord 17069

Document Type Dissertation

Format application/pdf (imt)

Rights Ma, Mingming

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA