UNEMPLOYMENT AND HEALTH IN CONTEXT AND COMPARISON: A STUDY OF CANADA, GERMANY, AND THE UNITED STATES OF AMERICA by CHRISTOPHER BRUCE McLEOD B.A. University of Victoria, 1998 M.A. McMaster University, 2000 A THESIS SUBMITTED FOR THE PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in THE FACULTY OF GRADUATE STUDIES (Health Care and Epidemiology) THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver) November, 2009 © Christopher Bruce McLeod 2009 ii Abstract This thesis explores how societal-level factors influence the relationship between unemployment and health. Using the Varieties of Capitalism (VOC) framework, hypotheses are developed that specify how this relationship may vary across high-income countries. Economies of high-income countries are grouped into coordinated market (CMEs) and liberal market (LMEs) economies that have different production specializations, but similar economic growth and aggregate levels of wealth and which are supported by different economic and labour market institutions. I hypothesize that these institutional differences give rise to different risks, types and durations of unemployment. After controlling for these differences, it is hypothesized that the higher levels of unemployment protection in CMEs will mediate the effect of unemployment on health compared to LMEs and that there will also be an interaction between skill level and unemployment and health. Two empirical studies are conducted to test these hypotheses using longitudinal micro- data from representative LME (Canada and the United States) and CME (Germany) countries. The first study examines the relationship between unemployment and mortality for Germany and the United States. The risk of death for the unemployed is higher in the United States compared to Germany, especially for the minimum- and medium-skilled unemployed. In Germany the risk of death for the unemployed is concentrated among East Germans. The second study examines the relationship between unemployment and self-reported health status for Canada, Germany and the United States. Across all countries unemployment is associated with poorer self-reported health status, but there is marked effect modification by educational status and by receipt of unemployment compensation. In particular, there is no association for the high-skilled unemployed in the United States, but for minimum- and medium-skilled unemployed those not receiving unemployment compensation have the highest risk of poorer self-reported health status. Policy makers should consider the effect on the health of the unemployed when designing programmes for the unemployed. Future research needs to examine the role that social programmes and in particular public transfers have in reducing health inequalities, not only among the unemployed, but also among workers in other work arrangements that may be harmful to their health. iii Table of Contents Abstract .......................................................................................................................................... ii Table of Contents ......................................................................................................................... iii List of Tables ................................................................................................................................. v List of Figures ................................................................................................................................ x List of Symbols and Abbreviations ............................................................................................ xi Acknowledgements ..................................................................................................................... xii Dedication ................................................................................................................................... xiii Chapter 1: Introduction and Scope of Dissertation .............................................................. 1 1.1: Motivation ....................................................................................................................... 1 1.2: Method of Inquiry ........................................................................................................... 3 1.3: Plan of Thesis .................................................................................................................. 4 Chapter 2: Unemployment and Health in a Comparative Perspective ............................... 7 2.1: Introduction ..................................................................................................................... 7 2.2: Comparative Health Research ......................................................................................... 7 2.3: The Varieties of Capitalism Framework ....................................................................... 13 2.4: Unemployment as a Social Determinant of Health ....................................................... 24 2.5: Unemployment and Health in Context ......................................................................... 29 2.6: Hypotheses for Empirical Studies ................................................................................. 37 2.7: Concluding Remarks ..................................................................................................... 38 Chapter 3: Description of Survey Data, Cohort and Variable Development ................... 45 3.1: Introduction ................................................................................................................... 45 3.2: Description of the Survey Data ..................................................................................... 46 3.3: Derivation of the Study Cohorts ................................................................................... 49 3.4: Development of the Variables Used in the Studies ...................................................... 54 3.5: Assessment of Data Quality and Comparability of Study Cohorts ............................... 74 3.6: Concluding Remarks ..................................................................................................... 77 Chapter 4: Unemployment and Mortality: A Study of Germany and the United States 86 4.1: Introduction ................................................................................................................... 86 4.2: Research Objectives ...................................................................................................... 86 4.3: Methods......................................................................................................................... 87 4.4: Results ........................................................................................................................... 93 4.5: Discussion ..................................................................................................................... 99 iv 4.6: Conclusion .................................................................................................................. 107 Chapter 5: Unemployment and Self-rated Health: A Study of Canada, Germany and the United States 127 5.1: Introduction ................................................................................................................. 127 5.2: Research Objectives .................................................................................................... 127 5.3: Methods....................................................................................................................... 128 5.4: Results ......................................................................................................................... 134 5.5: Discussion ................................................................................................................... 141 5.6: Conclusion .................................................................................................................. 147 Chapter 6: Synthesis & Conclusion .................................................................................... 172 6.1: Introduction ................................................................................................................. 172 6.2: Summary of Main Findings ........................................................................................ 173 6.3: Synthesis of Findings across Studies .......................................................................... 174 6.4: Strengths and Limitations ........................................................................................... 176 6.5: Implications for Policy ................................................................................................ 177 6.6: Towards an Ongoing Research Agenda ...................................................................... 179 Bibliography .............................................................................................................................. 182 Appendix A: Ethics Certificate ................................................................................................ 200 Appendix B: Cohort Studies that Examine the Relationship between Unemployment and Mortality by LME and CME ................................................................................................... 201 Appendix C: Supplementary Tables for the Study Cohorts and Variable Development .. 221 Appendix D: Supplementary Tables for the Unemployment and Mortality Analysis ....... 224 Appendix E: Supplementary Tables for the Unemployment and Self-reported Health Analysis ...................................................................................................................................... 228 v List of Tables Table 2.1: Household income replacement rate scenarios for the unemployed by selected family type for Canada, Germany and the United States ................................................................. 41 Table 2.2: Scope of employment protection regulation ................................................................ 42 Table 2.3: Summary the relationship between unemployment and all cause mortality for cohort studies conducted in Coordinated Market Economies .......................................................... 43 Table 2.4: Summary the relationship between unemployment and all cause mortality for cohort studies conducted in Liberal Market Economies .................................................................. 44 Table 3.1: Derivation of GSOEP mortality cohort (1984-2005) .................................................. 78 Table 3.2: Derivation of American mortality cohort (1984-2005) ............................................... 79 Table 3.3: Derivation of the German SRHS cohort (1994-2005) ................................................. 80 Table 3.4: Derivation of the Canadian SRHS cohort (1996-2005) ............................................... 81 Table 3.5: Derivation of the American SRHS cohort (1984-1997) .............................................. 82 Table 3.6: Cross-tabulation between self-rated health status and health satisfaction for the years 1992, 1994-2005 for the German cohort ............................................................................... 82 Table 3.8: Assignment of monthly labour force status based on the monthly labour force status question for 1984-1992 ......................................................................................................... 83 Table 3.7. Stylized cumulative labour force status example depicting the transition from employed to not working in the mortality cohort. ................................................................ 83 Table 3.9: Highest degree of education based on a modified CASMIN classification at baseline for the SRHS cohort .............................................................................................................. 84 Table 3.10: Creation of the occupational variable used in the SRHS study across the three surveys .................................................................................................................................. 85 Table 4.1: Descriptive statistics at baseline stratified by current labour force status and study country ................................................................................................................................ 113 Table 4.2: Unemployment, unemployment compensation, income and public transfers by number of months unemployed by study cohort ................................................................. 115 Table 4.3: Relative risk of dying by labour force status at the time of the survey for the German cohort, 1986-2004 ............................................................................................................... 116 Table 4.4: Relative risk of dying by labour force status at the time of the survey for the American cohort, 1986-2004 ............................................................................................................... 118 Table 4.5: Relative risk of dying by labour force status at the time of the survey, adjusted for potential confounders, German and American cohorts, 1986-2004 ................................... 120 Table 4.6: Relative risk of dying by labour force status in the year prior to the survey, adjusted for potential confounders, German and American cohorts, 1986-2004 .............................. 120 Table 4.7: Relative risk of dying by cumulative labour force status, adjusted for potential confounders, German and American cohorts, 1986-2004 .................................................. 120 vi Table 4.8: Relative risk of dying by labour force status at time of survey stratified by sex, adjusted for potential confounders, German and American cohorts, 1986-2004 ............... 121 Table 4.9: Relative risk of dying by labour force status in the year prior to the survey stratified by sex, adjusted for potential confounders, German and American cohorts, 1986-2004 ... 121 Table 4.10: Relative risk of dying by cumulative labour force status stratified by sex, adjusted for potential confounders, German and American cohorts, 1986-2004 ................................... 122 Table 4.11: Relative risk of dying, by labour force status at time of survey stratified by educational skill level, adjusted for potential confounders, German and American cohorts, 1986-2004, health (t-1) model ............................................................................................ 122 Table 4.12: Relative risk of dying by labour status in the year prior to the survey stratified by educational skill level, adjusted for potential confounders, German and American cohorts, 1986-2004, health (t-1) model ............................................................................................ 123 Table 4.13: Relative risk of dying by cumulative labour force status stratified by educational skill level, adjusted for potential confounders, German and American cohorts, 1986-2004, health (t-1) model .......................................................................................................................... 123 Table 4.14: Relative risk of dying by labour force status at time of survey with exclusions for baseline health and labour force status, adjusted for potential confounders, German and American cohorts, 1986-2004, health (t-1) model .............................................................. 124 Table 4.15: Relative risk of dying by labour status in the year prior to the survey with exclusions for baseline health and labour force status, adjusted for potential confounders, German and American cohorts, 1986-2004, health (t-1) model .............................................................. 124 Table 4.16: Relative risk of dying by cumulative labour force status stratified with exclusions for baseline health and labour force status, adjusted for potential confounders, German and American cohorts, 1986-2004, health (t-1) model .............................................................. 124 Table 4.17: Relative risk of dying by all three labour force status variables stratified by East and West Germany, adjusted for potential confounders, German cohort, 1986-2004, health (t-1) model ................................................................................................................................... 125 Table 4.18: Relative risk of dying by all three labour force status variables stratified by race, adjusted for potential confounders, American cohort, 1986-2004, health (t-1) model ....... 125 Table 4.19: Predicted hazard of dying in a given year by current labour force status, skill level and age, based on the skill stratified estimates ................................................................... 126 Table 5.1: Descriptive statistics, by study cohort and labour force status .................................. 153 Table 5.2: Unemployment, unemployment compensation, income and public transfers by number of months unemployed by study cohort .............................................................................. 155 Table 5.3: Odds ratio of self-reported health status (poor/fair vs. good/very good/excellent) for current labour force status, static and dynamic health models, German, Canadian and American cohorts ................................................................................................................ 157 Table 5.4: Odds ratio of self-reported health status (poor/fair/very good vs. very good/excellent) for current labour force status, static and dynamic health model, German, Canadian and American cohorts ................................................................................................................ 159 vii Table 5.5: Odds ratio of self-reported health status (poor/fair vs. good/very good/excellent) for all three labour force status measures, dynamic health model, German, Canadian and American cohorts. ............................................................................................................... 161 Table 5.6: Odds ratio of self-reported health status (poor/fair/good vs. very good/excellent) for all three labour force status measures, dynamic health model, German, Canadian and American cohorts ................................................................................................................ 162 Table 5.7: Odds ratio of self-reported health status (poor/fair vs. good/very good/excellent) for all three labour force status measures, excluding those in poor or fair health at baseline, dynamic health model, German, Canadian and American cohorts ..................................... 163 Table 5.8: Odds ratio of self-reported health status (poor/fair vs. good/very good/excellent) for all three labour force status measures, excluding those unemployed or not working at baseline, dynamic health model, German, Canadian and American cohorts ..................... 164 Table 5.9: Odds ratio of self-reported health status (poor/fair vs. good/very good/excellent) for all three labour force status measures, men only, dynamic health model, German, Canadian and American cohorts ......................................................................................................... 165 Table 5.10: Odds ratio of self-reported health status (poor/fair vs. good/very good/excellent) for all three labour force status measures, women only, dynamic health model, German, Canadian and American cohorts ......................................................................................... 166 Table 5.11: Odds ratio of self-reported health status (poor/fair vs. good/very good/excellent) for all three labour force status measures, minimum skilled only, dynamic health model, German, Canadian and American cohorts .......................................................................... 167 Table 5.12: Odds ratio of self-reported health status (poor/fair vs. good/very good/excellent) for all three labour force status measures, medium skilled only, dynamic health model, German, Canadian and American cohorts ......................................................................................... 168 Table 5.13: Odds ratio of self-reported health status (poor/fair vs. good/very good/excellent) for all three labour force status measures, high skilled only, dynamic health model, German, Canadian and American cohorts ......................................................................................... 169 Table 5.14: Odds ratio of self-reported health status (poor/fair vs. good/very good/excellent) for all three labour force status measures, stratified by East and West German, dynamic health model, German cohort ........................................................................................................ 170 Table 5.15: Odds ratio of self-reported health status (poor/fair vs. good/very good/excellent) for all three labour force status measures, stratified by race (Black and White/Other), dynamic health model, American cohort ........................................................................................... 171 Table 6.1: Summary of thesis hypotheses and results from the empirical studies ..................... 181 viii Table B1: Detailed summary of cohort studies that examine the relationship between unemployment and mortality by country for Coordinated Market Economies 201 Table B2: Summary of population-based cohort studies that examine the relationship between unemployment and mortality by country for Liberal Market Economies .......................... 212 Table C1: Mortality experience by sample population for individuals between the ages of 18 and 64 at baseline for the German cohort .................................................................................. 221 Table C2: Hierarchy of monthly labour force status variables by study year ............................ 222 Table C3: Highest degree of education based on a modified CASMIN classification at ........... 223 Table D1: Comparison of the health model with and without accounting for survey design for all three labour force statuses, German cohort, 1986-2004 ..................................................... 224 Table D2: Comparison of health model with and without accounting for survey design for all three labour force status, American cohort, 1986-2004 ...................................................... 226 Table E1: Testing proportional odds assumption of the ordered logit model for SRHS with current labour force status as the principal dependent variable, German cohort, 1995-2005 ............................................................................................................................................. 228 Table E2: Testing proportional odds assumption of the ordered logit model for SRHS with current labour force status as the principal dependent variable, dynamic health model, 1996- 2005, Canadian cohort (SLID) ............................................................................................ 231 Table E3: Testing proportional odds assumption of the ordered logit model for SRHS with current labour force status as the principal dependent variable, American cohort, 1985-1997 ............................................................................................................................................. 234 Table E4: Odds ratio of self-reported health status for current labour force status, dynamic health model, German cohort, with and without a random effect controlling for survey design .. 237 Table E5: Odds ratio of self-reported health status for current labour force status, dynamic health model, American cohort, with and without a random effect controlling for survey design 240 Table E6: Odds ratio of self-reported health status (poor/fair/good vs. very good/excellent) for all three labour force status measures, males only, dynamic health model, German, Canadian and American cohorts ......................................................................................................... 243 Table E7: Odds ratio of self-reported health status (poor/fair/good vs. very good/excellent) for all three labour force status measures, females only, dynamic health model, German, Canadian and American cohorts ......................................................................................................... 244 Table E8: Odds ratio of self-reported health status (poor/fair/good vs. very good/excellent) for all three labour force status measures, minimum skilled only, dynamic health model, German, Canadian and American cohorts ......................................................................................... 245 Table E9: Odds ratio of self-reported health status (poor/fair/good vs. very good/excellent) for all three labour force status measures, medium skilled only, dynamic health model, German, Canadian and American cohorts ......................................................................................... 246 Table E10: Odds ratio of self-reported health status (poor/fair/good vs. very good/excellent) for all three labour force status measures, high skilled only, dynamic health model, German, Canadian and American cohorts ......................................................................................... 247 ix Table E11: Odds ratio of self-reported health status (poor/fair/good vs. very good/excellent) for all three labour force status measures, excluding those in poor or fair health at baseline, dynamic health model, German, Canadian and American cohorts ..................................... 248 Table E12: Odds ratio of self-reported health status (poor/fair/good vs. very good/excellent) for all three labour force status measures, excluding those unemployed or not working at baseline, dynamic health model, German, Canadian and American cohorts ..................... 249 Table E13: Odds ratio of self-reported health status (poor/fair/good vs. very good/excellent) for all three labour force status measures, stratified by East and West German, dynamic health model, German cohort ........................................................................................................ 250 Table E14: Odds ratio of self-reported health status (poor/fair/good vs. very good/excellent) for all three labour force status measures, stratified by Race (Black and White/Other), dynamic health model, American cohort ........................................................................................... 251 x List of Figures Figure 2.1: Percentage of unemployment insurance beneficiaries to total unemployed for Canada, Germany and the United States, 1976-2006 ......................................................................... 39 Figure 2.2: Social protection and predicted skill profiles ............................................................. 40 Figure 2.3: Pathways through which unemployment could influence health ............................... 40 Figure 4.1: Survivor function for the German and American cohorts by years followed .......... 109 Figure 4.2: Survivor function for the German and American cohorts by years followed stratified by labour force status at baseline (t-2) ................................................................................ 109 Figure 4.3: Survivor function for the German and American cohorts by years followed stratified by gender ............................................................................................................................. 110 Figure 4.4: Survivor function for the German and American cohorts by years followed stratified by educational status ........................................................................................................... 110 Figure 4.5: Survivor function for the German and American cohorts by years followed stratified by health status at baseline (t-1, t-2) ................................................................................... 111 Figure 4.6: Summary of the relative risks of dying for current unemployed for the German and American cohorts ................................................................................................................ 111 Figure 4.7: Predicted hazard of dying by current labour force status and age for the German and American cohorts (adjusted for all covariates including t-1 health status) ......................... 112 Figure 5.1 Self-reported health status at baseline by labour force status for Germany, Canada and the United States ................................................................................................................. 148 Figure 5.2: Proportion of unemployment by months unemployed for Germany (1994-2005), Canada (1996-2005) and the United States (1984-1997) ................................................... 148 Figure 5.3: Percent of individuals in receipt of unemployment compensation by months unemployed for Germany (1994-2005), Canada (1996-2005) and the United States (1984- 1997) ................................................................................................................................... 149 Figure 5.4: Ratio of household income of the unemployed to the continuously employed by months unemployed for Germany (1994-2005), Canada (1996-2005) and the United States (1984-1997) ......................................................................................................................... 149 Figure 5.5: Ratio of unemployment compensation to unemployed household income by months unemployed for Germany (1994-2005), Canada (1996-2005) and the United States (1984- 1997) ................................................................................................................................... 150 Figure 5.6: Summary of the odds ratios of being in poor or fair health to those currently unemployed in Germany, Canada and the United States .................................................... 151 Figure 5.7: Summary of the odds ratios of being in poor or fair health for the currently unemployed, by receipt of unemployment benefits for Germany, Canada and the United States ................................................................................................................................... 152 xi List of Symbols and Abbreviations AIC – Akaike Information Criterion BIC – Bayesian Information Criterion BHPS – British Household Panel Survey CASMIN – Comparative Analysis of Social Mobility in Industrial Nations CME – Coordinated Market Economy CNEF – Cross National Equivalent File GSOEP – German Socioeconomic Panel HR – Hazard Ratio ILO – International Labour Office ISCO88 – International Standard Classifications of Occupations (1988) LFS – Labour Force Survey (Canada) LME – Liberal Market Economy OECD – Organisation for Economic Co-operation and Development OPCS – Office of the Population Censuses and Surveys OR – Odds Ratio PSID – Panel Study of Income Dynamics PY – Person Years RE – Random Effect RR – Relative Risk SES – Socio-economic Status SLID – Survey of Labour Income Dynamics SRHS – Self-reported Health Status VOC – Varieties of Capitalism xii Acknowledgements I have many to thank. While dissertation research is meant to be singular endeavour, I have benefited from the support, encouragement, guidance, and generosity of many people. This research was supported by a Social Science and Humanities Research Council Doctoral Fellowship and by a Western Regional Training Centre for Health Services Research Studentship and would not have been completed without the commitment of public funds to support this kind of research. The comparative research conducted here also could not have taken place without the longstanding commitment of data providers and researchers in Germany (the German Institute for Economic Research (DIW)), Canada (Statistics Canada and particularly the Research Data Centre at UBC) and the United States (the Panel Study of Income Dynamics at the University of Michigan and Department of Policy Analysis and Management (PAM) at Cornell University). The staff at Statistics Canada‘s Research Data Centre, especially Lee Grenon and Cheryl Fu, were instrumental in providing access to and supporting the analysis conducted with the Canadian data. I also benefited greatly from two workshops offered by DIW in Bremen, Germany and by PAM at Cornell University that were key in understanding the German and American data. But the support underlying this thesis is largely personal. I am indebted to the longstanding support of my thesis committee, Clyde Hertzman, John Lavis and Ying MacNab. Their guidance and mentorship, going back in some cases a decade, has been invaluable, not only to this research, but to my career as an academic. The ideas underlying this research captured me during my tenure at the Institute for Work and Health in Toronto and I am grateful for the initial opportunities and the ongoing interest in this research from my colleagues with whom I started my research career. At UBC there have been many who have helped me in large and small ways. I am grateful for the support I received from the Centre for Health Services and Policy Research, my home these past six years at UBC, and the help of my colleagues and friends Dawn Mooney, Mieke Koehoorn, Kim McGrail, Paul Demers and Jeannie Shoveller. Kim and Jeannie provided feedback on the thesis proposal and various chapters and Dawn graciously created the figures in the thesis (and did some editing too). This thesis would not have been written without the seclusion of Paul‘s cabin at Gates Lake. And Mieke‘s constant encouragement, trust and partnership enabled me to complete this thesis while maintaining my other responsibilities at UBC. xiii Dedication For my parents, Gary and Elaine McLeod. This has been a long journey, but you have travelled it with me. 1 Chapter 1: Introduction and Scope of Dissertation 1.1: Motivation Comparative health research is playing an increasing role in understanding the distribution of health inequalities within and between societies. Much of this research has looked at the relationship between the institutional and political organisation of societies and the corresponding average level of population health (Borrell et al. 2007; Espelt et al. 2008) and has attempted to characterize the structural or contextual features of a society that would lead to the best health outcomes within a society (Chung and Muntaner 2007; Navarro and Shi 2001). This research has emphasized the importance of welfare-regime type as the principal independent variable in explaining variation in health inequalities among countries. Welfare- regime typologies classify countries on the basis on which the state provides social and economic protection to its citizens. While there are many welfare-regime typologies (Bambra 2007), Esping-Andersen‘s (1990) typology that classifies countries into social democratic (universal provision), corporatist (class-based provision) and liberal or residual (mean-tested provision) regime clusters is the most common. Yet research using this typology has yielded mixed results in explaining the differences in health inequalities among high-income countries. Others have advocated that rather than focusing on broad based classifications of politics and institutions, comparative health research should focus on the role that society plays in providing resources to people and the effect that these resources, whether through public programmes or through cash transfers, have on reducing health inequality (Fritzell and Lundberg 2007; Lundberg 2008). The main goal of this inquiry is to explore how societal-level factors can influence the relationship between unemployment and health. In particular, this research has three main objectives: (1) to develop a set of hypotheses, based principally on the Varieties of Capitalism (VOC) framework (Hall and Soskice 2001), on how macroeconomic and institutional factors could affect the individual-level relationship between unemployment and health; (2) to conduct a comparative study of the relationship between unemployment and mortality in Germany and the United States; and (3) to conduct a comparative study of the relationship between unemployment and 2 self-reported health status in Canada, Germany, and the United States. The determinants of population health have both contextual and compositional aspects in which the distribution of individual vulnerabilities and health inequalities are determined by and mediated through social, economic and physical environments (Dunn et al. 2006). While there is a long tradition of study into the relationship between unemployment and health, unemployment has almost always been conceptualized as an individual-level risk factor. Where context has been considered it has been to investigate whether the effect of unemployment on health is different during times or places with high unemployment compared with low unemployment (Beland, Birch, and Stoddart 2002; Martikainen, Maki, and Jantti 2007; Novo, Hammarstrom, and Janlert 2000). The unemployment rate has been viewed as a social mediator in which the experience of unemployment is different when large groups of individuals are unemployed and as a test of health selection in which the least healthy workers are more at risk of unemployment in times of low unemployment, while in times of high unemployment the risk of unemployment is generalized to healthy workers. The role of context in the unemployment and health relationship goes far beyond the business cycle. Unemployment may influence health through material (e.g., loss of income) and psychosocial (e.g., loss of individual and social identity) pathways. These pathways are embedded in and influenced by societal context at every point, from determining who is unemployed (and who are labour market participants), the meaning of unemployment, the material effect of unemployment, and the future employment consequences of unemployment. Unemployment is not just an individual-level experience, but at its core a socially mediated one. In order to create a coherent framework that integrates both the contextual and- individual-level factors in understanding how unemployment affects health Hall and Sockice‘s (2001) Varieties of Capitalism (VOC) framework is used. This framework groups the economies of high-income countries into two variants of capitalism – coordinated market economies (CMEs) and liberal market economies (LMEs) – which have different economic and labour market institutions. Accordingly, the focus of this thesis is principally on how the contextual and institutional environment mediates the unemployment and health relationship through the material pathway. 3 1.2: Method of Inquiry This research adopts a comparative study design and uses longitudinal individual-level data from three high-income countries – Germany, Canada, and the United States – that are representative of both CME and LME countries. The surveys used in this study are designed to study labour market, income, family and educational dynamics across the lifecourse and have produced a rich body of research relating these factors to health. In the United States, the Panel Study of Income Dynamics (PSID) has been used to examine the association between working conditions, marital transitions and income dynamics, on the one hand, and both mortality (Amick et al. 2002; Lillard and Waite 1995; McDonough et al. 1997; McDonough et al. 1999) and self-reported health outcomes (Haas 2006; Smith 2005), on the other. The German Socio-economic Panel (GSOEP) has been used to study the effect of unemployment on health impairments (Elkeles and Seifert 1993), the effect of income on health satisfaction and life expectancy (Frijters, Haisken- DeNew, and Shields 2005a; Frijters, Haisken-DeNew, and Shields 2005b) and differences in health inequalities between East and West Germans (Nolte and McKee 2004). The Survey of Labour and Income Dynamics (SLID) has been used to study health transitions in older Canadians (Buckley et al. 2006; Buckley et al. 2004) and the effect of contingent work on health (Tompa, Scott-Marshall, and Fang 2008). German and American data have also provided a rich resource for comparative research on earnings and income dynamics (Burkhauser and Poupore 1997; DiPrete and McManus 1996; McManus and DiPrete 2000), and educational attainment (Daly, Buchel, and Duncan 2000; Szydlik 2002). There is little comparative research, however, that incorporates the Canadian data (Valleta 2005). Only a few studies have used these data to conduct comparative research into health inequalities (Burkhauser and Daly 1998; Rodriguez 2001; Sacker et al. 2007). A challenge of conducting comparative longitudinal research is developing an analytic approach that is valid within and across study countries. There are more potential biases to consider in multi-country studies using cross national data than in single country studies. Indeed, the comparability of the study cohorts and the unemployment measures and the specification of the institutional environments across these countries lie at the crux of this research. Accordingly, developing the comparative study design and assessing the comparability of the study cohorts and measures warrants special attention. 4 The strength of comparative health research is that it allows examination of contextual-level and institutional-level determinants of health that do not (and in some cases cannot) vary within populations. It is only through comparative research that it can be determined how the distribution of health inequalities varies across populations. This thesis is one of the first studies into unemployment and health that uses both a comparative and longitindinal approach. It adopts a longitudinal cohort design that can account for temporality in the unemployment and health relationship, consider exposure to unemployment throughout the lifecourse, and distinguish between the reciprocal effects of health on unemployment (health selection) and unemployment on health (social causation). 1.3: Plan of Thesis Chapter 2 develops the conceptual framework and study hypotheses and places the relationship between unemployment and health into a comparative perspective. It begins by providing an overview of the emerging body of comparative health research that has sought to establish a link between societal and macroeconomic factors in explaining the variation in health inequalities among societies. The organizing typology – the VOC framework – that explains why institutional variation persists in the labour market among capitalist societies is introduced and extended to describe how variation in economic and social institutions among Canada, Germany and the United States may mediate the unemployment and health relationship. The pathways through which unemployment could influence health are explored and the methodological issues key to modelling this relationship are summarized. Applying the VOC lens to the unemployment and health relationship, the existing research on unemployment and mortality is reviewed to determine if this relationship is different by CME and LME. Finally, integrating the above sections, a set of hypotheses that specify how the differences in institutional factors among Canada, Germany and the United States could affect the unemployment and health relationship is developed. It is hypothesized that the effect of unemployment on health is mediated in CMEs compared to LMEs and that there will be differences in this relationship by educational status across CMEs and LMEs due to the greater degree of economic and social stratification by educational attainment in liberal market economies. The receipt of unemployment insurance, however, will mediate the relationship between unemployment and health. Unemployment insurance will be a stronger mediator in LMEs as the receipt of unemployment 5 insurance is a marker for strong labour force attachment and lowers the opportunity cost of waiting for a suitable job; the effect of unemployment compensation on the health of the unemployed in CMEs will be muted as unemployment benefits are also designed as income support for the long-term unemployed. Chapter 3 describes the development of the data and study cohorts across the three study countries. The primary purpose of this chapter is to show that the cross-national survey data can be used to conduct comparative individual-level studies and that the cohort and variables developed will lead to valid inferences both within and across study countries. This chapter provides the justification that the data can support the analytic studies that follow in later chapters. It also describes the development of and key decisions in creating the cohorts and of the principal study variables, and concludes with an assessment of overall comparability and strengths and limitations of the data. Chapter 4 presents a comparative study of unemployment and mortality between Germany and the United States for the period of 1984 to 2005. Using a discrete failure time model and controlling for a range of demographic, socioeconomic and health variables, the relationship between three measures of unemployment – current unemployment, months unemployed, and cumulative lifetime unemployment – and mortality is tested within each country. This chapter also explores if the relationship varies by educational status and gender and a series of baseline exclusions are applied that account for health selection into unemployment. The results of models across country cohorts are then interpreted in light of the hypotheses developed in Chapter 2 which specify how the relationship between unemployment and mortality could vary by institutional context. Chapter 5 presents a comparative study of unemployment and self-reported health status among Canada, Germany and the United States. It covers the time periods of 1996 to 2005 for Canada, 1994 to 2005 for Germany, and 1984 to 1997 for the United States. The period covered varies among cohorts due to the differences in the years surveyed and the availability of the self- reported health status measure. This study is similar in conceptualization to the unemployment and mortality study, but with the greater statistical power of the self-reported health status variable it is possible to examine if the receipt of unemployment compensation modifies the relationship between unemployment and health. Random-effects logistic estimation is used to 6 examine the relationship between unemployment and self-reported health status. Both static and dynamic health models are tested with the inclusion of the same covariates and with the same education and gender stratifications and baseline exclusions as in the unemployment and mortality analysis. The results across the three countries are interpreted in light of the contextual- level hypotheses. Chapter 6 concludes by synthesising and integrating the findings from the two empirical studies. The overall strengths and limitations of the study findings with respect to the robustness of the unemployment and health relationship are discussed. Whether the findings support the hypotheses that the unemployment and health relationship will be mediated in CMEs compared to LMEs and whether this mediation occurs through higher levels of unemployment protection and closer skill-occupation coordination in CMEs is assessed. The relevance of the findings to policy makers is also discussed and the potential of an ongoing research agenda in comparative health inequalities using cross-national survey data is presented. 7 Chapter 2: Unemployment and Health in a Comparative Perspective 2.1: Introduction This chapter is about integrating two perspectives, that of unemployment as a risk factor for poor health and that of a comparative perspective, in describing how societal-level factors can influence the relationship between unemployment and health. This is accomplished by integrating the Varieties of Capitalism (VOC) framework that explains the persistence of different ways of coordinating the economies of high-income capitalist societies with research that examines the individual-level relationship between unemployment and health. From this, two related sets of hypotheses that are tested in the analytic chapters are developed: a set of compositional hypotheses focused on the individual-level relationship between unemployment and health; and a set of contextual hypotheses focused on how this relationship may vary across VOC economies. 2.2: Comparative Health Research 2.2.1: Comparative Health Research and Welfare Regime Typologies This chapter starts with a review of the Three Worlds of Capitalism typology developed by Esping-Anderson (1990). This starting point is chosen because this is the typology most often used in international comparative studies interested in contextual influences on health. A grounding in this approach and the empirical literature behind it helps to highlight differences with the Varieties of Capitalism framework and the reasons the latter was chosen as the main theoretical framework for this research. Esping-Andersen‘s classic welfare state typology groups high-income countries into three welfare regime clusters along the dimensions of decommodification (i.e., the degree to which individuals must rely on labour income for their own welfare) and social stratification (i.e., the extent to which this is based on class or gender (Esping-Andersen 1990; Esping-Andersen 1999)). 8 Means-tested assistance, modest universal transfers, or modest social-insurance plans dominate in countries labelled ―liberal welfare states‖. Benefits are targeted to those at the bottom of the income and class spectrum. The consequence is that this type of welfare regime has minimal decommodification (i.e., people must work to rise above a minimal subsistence level). Canada, the United States, the United Kingdom, and Ireland are examples of countries that fall within this cluster. In ―conservative welfare states‖, benefit entitlements are more generous but are organized around the preservation of status differentials and the role of the family as the primary provider. Benefit entitlement is different for different occupational classes (e.g., manual workers versus civil servants). Social insurance typically excludes non-working wives, and family benefits encourage traditional family roles. France, Germany, and Austria are examples of countries that fall within this cluster. The principles of universalism and decommodification of social rights are extended beyond the marginal worker to the middle and upper class in ―social democratic welfare states‖. Benefit entitlement is not contingent on labour-market participation and is distributed equally among occupational classes. Manual workers tend to enjoy rights similar to those of white-collar workers, with benefit levels tied to earning levels. Denmark, Sweden and to a lesser extent the Netherlands are examples of countries that fall within this cluster. Other authors have advanced typologies that revise the criteria on which welfare state regimes are defined to include measures of social inclusion, benefit replacement and poverty rates, political tradition and the expenditure on services and social transfers (Bambra 2007). These typologies have also been expanded to include the former communist eastern and the ex-fascist southern European countries and high-income countries in Asia. Irrespective of the typology used, the essential argument is that regimes that provide greater access to decommodifying social benefits and services and have lower social and economic stratification will have better overall health outcomes and shallower health gradients (social democratic and to a lesser extent conservative or corporatist welfare regimes 1 ), while regimes characterized by residual or means-tested welfare benefits and high levels of economic or social 1 There is a profusion of terms used to describe welfare states. Throughout the chapter I use the terminology adopted by Esping-Andersen, but denote the alternative terminology in brackets where it applies to other research. 9 stratification will have poorer overall health outcomes and steeper health gradients (liberal welfare and southern and eastern European regimes) (Dahl et al. 2006; Fritzell and Lundberg 2007). While ecological comparative studies have supported this ranking especially for child health outcomes (Chung and Muntaner 2006; Chung and Muntaner 2007; Navarro and Shi 2001; Navarro et al. 2006; Wennemo 1993), the findings from cross-sectional studies are mixed. Bambra, Eikemo and colleagues in a group of analyses using the cross-sectional European Social Survey (ESS) found no consistent pattern in health inequalities among conservative (or Bismarckian), social democratic and liberal welfare regimes, although southern and eastern regimes tended to have the greatest health inequalities. Income-related (Eikemo et al. 2008a) and education-related (Eikemo et al. 2008b) health inequalities were smallest in conservative regimes, but social democratic regimes had greater education-related health inequalities than liberal regimes, with the ranking reversed for income-related health inequalities; unemployment- related health inequalities were greatest in liberal regimes (Bambra and Eikemo 2009) 2 , and no difference was observed among these three regimes in the likelihood of reporting poor self- reported health status (Eikemo et al. 2008c), or when self-reported health status was stratified by gender (Bambra et al. 2009). Olsen and Dahl (2007) found that liberal (Anglo-Saxon), conservative (continental), and southern European regimes were not associated with lower levels of self-reported health, when compared to social democratic (Nordic) regimes, but that Eastern European regimes did have lower levels. Espelt and colleagues (2008) found no systematic differences in inequalities in self- reported health by social class across the same cluster of regime types, except for women in late democracies. Health inequalities by socio-economic status defined by overall family consumption, however, were higher for adolescents in liberal and Mediterranean welfare state countries, but not in conservative welfare state countries compared to social democratic countries (Zambon et al. 2006). There are no longitudinal cohort studies examining health variations among high-income countries that have explicitly used a welfare regime typology. A number of studies have conducted longitudinal analyses of heath inequalities among European countries using the European Community Household Panel (Hernandez-Quevedo et al. 2006) including 10 unemployment-related health inequality (Cooper, McCausland, and Theodossiou 2006; Cooper, McCausland, and Theodossiou 2008). 2 But among these studies, no consistent pattern in income, education- or unemployment-related health inequalities by welfare regime type is found. Dahl and colleagues (2006) provide some insight into why the observed ranking of health inequalities by welfare regime or country may diverge from the hypothesized ranking: - social democratic countries have higher quality data, which could lead to a more precise ascertainment of the health inequalities; 3 - socioeconomic constructs may have different meanings and also represent different social stratifications across countries (e.g. a lower skilled manual labourer in Sweden and in Portugal); - lower absolute risks in health outcomes will necessarily lead to higher relative risks given comparable risk differences; 4 - the welfare regime typology applied at a given point in time does not take into account the timing and historical development of a country‘s welfare state; and - the greater of degree of decommodification is still not enough to counteract the negative health effects of relative deprivation due to psychosocial forces (i.e., relative deprivation gives rise to negative health outcomes through stress and dissatisfaction of being lower down in the social hierarchy). Others have questioned the utility of broad-based regime clusters to explain variations in health inequalities among countries (Bambra 2007; Lundberg 2008). Lundberg states: So while the country clusters may be helpful for descriptive purposes, they are much less useful if we really want to open the black box and analyse what aspects of the welfare states that are of importance. Especially, if we are interested in linking welfare state characteristics to public health outcomes it will be much more fruitful to study aspects of like coverage and generosity in specific programmes and how these co-vary with public health outcomes rather than to merely relate country-cluster averages to each other (Lundberg 2008 p.2). 2 These studies are reviewed in more detail in Section 2.4.4. 3 This point falls under the general rubric of surveillance bias. 4 The point is of particular relevance in that relative measures within a country may obscure large variations and differences in underlying health risks within groups across countries. 11 Lundberg outlines four features – coverage and generosity of cash transfer programmes (e.g., unemployment benefits, pensions) and the availability and quality of services (e.g., health care, education) provided – that may matter in reducing health inequalities. He argues that good cash programmes are not necessarily coordinated with good services programmes. Building on Sen‘s capability theory of inequality (Sen 1999; Sen 1992), the role of the state is to provide control over resources (Fritzell and Lundberg 2007) and act as an enabler in providing equality of opportunity (Siddiqi et al. 2007). Moreover, for comparative health research to be policy relevant the specific mechanisms or pathways through which these factors can influence the health of populations and individuals must be identified. 2.2.2: Health Differences between Canada and the United States The institutional variation between Canada and the United States is often lost in the larger cross- country comparisons of health outcomes that use established welfare typologies. Canada and the United States are grouped together as liberal welfare states, 5 but there are large differences in aggregate levels of health status and in health inequalities between these two countries. A large body of research has made direct comparisons between the health of populations within the United States and the health of the populations within other countries. Many of these studies, motivated, in part by the more than two-year longer life expectancy in Canada (United Nations 2008), are comparative studies between Canada and the United States (Devereaux et al. 2002; Guyatt et al. 2007; Huguet, Kaplan, and Feeny 2008; Kunitz and Pesis-Katz 2005; Lasser, Himmelstein, and Woolhandler 2006; Manuel and Mao 2002; McGrail et al. 2009; Siddiqi and Hertzman 2007; Willson 2009). Most of these studies have focused on the differences in the organisation, financing, coverage, and provision of health care in the two countries as potential explanations. Kunitz and Pesis-Katz (2005) conducted a review of studies that examined differences in mortality rates between Canada and the United States for those health conditions that can be 5 Some typologies place Canada in a different welfare regime cluster than the United States based on a consideration of the provision of universal health care coverage (Bambra 2005). Scruggs, in a recent reanalysis of Esping- Andersen‘s decommodification index argues that Canada was basically misclassified (Scruggs and Allan 2006). In their analysis, Canada‘s decommodification score lies close to the mean and according to Esping-Andersen‘s criteria would have been classified as conservative regime rather than a liberal regime. These scores, however, are based on benefit levels in 1980 and so would not necessarily be reflective of the present. 12 attributed to the receipt of, or quality of, health care. With a few exceptions mortality outcomes were found to be better in Canada compared to the United States. This study also shows that the difference in life expectancy between Canada and the United States cannot be attributed to the lower life expectancy of African Americans compared to white Americans, and that for mortality from causes amenable to health care a persistent and growing gap emerged between white Americans and all Canadians in the early 1970s concurrent with the introduction of universal health insurance in Canada. Research using cross-sectional data from the Joint Canada/United States Survey of Health found that income-related inequalities were greater in the United Sates compared to Canada and that much of the difference can also be explained by differential access to health insurance by income in the United States (Huguet, Kaplan, and Feeny 2008; McGrail, van Doorslaer, Ross, and Sanmartin 2009). Other authors (Siddiqi and Hertzman 2007) have argued that it is the overall differences in the nature and degree of each countries‘ social safety net (the health care system being but one aspect) that matter. In their analysis, the divergence of health outcomes in Canada and the United States over the last half century can be attributed to the slow and, at times, invisible development of institutional and societal factors that have led to a more equitable distribution of health-related resources within Canada compared to the United States. 6 This idea is further supported by a recent cross-sectional study using the Canadian National Population Health Survey and the American Panel Study of Income Dynamics that found that low levels of income and education were more predictive of a highly preventable disease (cardiovascular disease) compared to a less preventable disease (cancer) in the United States, but not in Canada (Willson 2009). Income inequality may be a marker for the societal-level resources that affect population health status. Ecological studies of the effect of income inequality on population-level health have found a consistent relationship between income inequality and mortality in the United States, but not in Canada (Ross et al. 2005; Ross et al. 2000; Sanmartin et al. 2003); while cross-sectional and cohort studies have found mixed results for the effect of income inequality on individual- level health in Canada (Hou and Myles 2005; McLeod et al. 2003; Xi et al. 2005) and a consistent relationship in the United States (Lynch et al. 2004). 6 They reference Pierson‘s idea that institutional change is big, slow-moving and invisible and can only be seen looking backward over a long period of time (Pierson 2003). 13 A comparison between cash transfers received by Canadians and Americans found that the expansion of social transfers and specifically the introduction of the social retirement benefit (the Guaranteed Income Supplement) explained most of the divergence of poverty rates between Canada and United States during the period 1974 to 1994 (Zuberi 2004). Differences in the unemployment insurance system also contributed to the reduction of poverty rates in Canada, but not in the United States. 7 A qualitative study of low income hotel workers in Vancouver, British Columbia and Seattle, Washington found that greater access to health care was only one of many differences associated with better health outcomes (Zuberi 2006). Low income hotel workers in Vancouver had more secure and better paying jobs, stable housing, access to more extensive publicly provided services, as well as cash transfers including transit and recreational opportunities, unemployment benefits, workers compensation, and child benefits. These studies suggest that the differences in the provision of a broad range of public programmes and cash transfers between Canada the United States matter in explaining the difference in health status between Americans and Canadians. Nonetheless research establishing a definitive link between the availability of social programmes or the receipt of cash benefits in explaining health differences across these two countries has yet to be conducted using individual- and contextual- level data. 2.3: The Varieties of Capitalism Framework The Varieties of Capitalism framework characterizes the different labour market institutions among high-income countries. This theorectical framework can be used to specify how the consequences of unemployment can vary by skill-profile, labour market attachment and other measures of socio-economic status across institutional settings. Hall and Soskice (2001) assert that the economies of developed countries can be grouped into two distinct types of equilibria – liberal market economies and coordinated market economies – which reflect allocatively efficient production processes and have led to similar levels of economic growth and aggregate wealth among high-income capitalist societies. In other words, based on historical and institutional considerations, convergence to an Anglo-American style of organisation of production is neither inevitable nor optimal. Their approach is predicated on the 7 Due to the reforms to the unemployment insurance system in Canada (see Figure 2.1) between 1990 and 1996 that reduced entitlements this may no longer be the case. 14 ideas of path dependency 8 and comparative economic advantage, 9 that together lead to a theory of comparative institutional advantage in which it is not only factor endowments, but also historically-dependent institutions that create comparative advantage. Ebbinghaus and Manow (2001) outline the three central characteristics of this approach: (1) it is a systematic account of the functioning of the institutional components of economic systems, (2) it distinguishes national models of production and maps their comparative advantage, and (3) it seeks a micro-foundation of how institutions shape actors‘ behaviour and reinforce existing institutional structures (p.3). In this approach the firm and production processes are placed at the centre of the model, and firms respond to the historical, social and institutional structures within which they operate in order to maximize the allocative efficiency of their production processes. This in turn creates a virtuous cycle as the firm now relies on these social institutions, including the type and quality of worker produced, in order to maintain its comparative advantage. The convergence on multiple equilibria has led to two distinct types of capitalist economies, liberal market economies in which production is coordinated through market mechanisms, and coordinated market economies in which production is organized through coordinating mechanisms like trade organisations and quasi-governmental bodies. Liberal market economies (LMEs) are characterized by flexibility and innovation in both production processes and the labour market; employees who have general and transferable skills are most highly valued. Coordinated market economies (CMEs) are characterized by stable but more complex production processes; workers who have skills specialized to specific areas of production are highly valued. The United States and Germany are considered the archetypical liberal market economy and coordinated market economy, respectively. 10 Canada is considered to be a variant of a liberal market economy, although it has higher levels of both employment and unemployment protection (see section 2.3.2) than the United States. 8 In other words, history matters not only to the development of institutions but also to the scope in which they are allowed to change when faced with similar fiscal and policy pressures (Pierson 2000). 9Comparative advantage refers to the gains in trade that can be made through product specialization due to the lower marginal cost of producing some goods compared to others (Krugman and Obstfeld 2003). 10 Of high-income OECD countries, Hall and Sockice classify the United States, the United Kingdom, Australia, Canada, New Zealand, and Ireland as LMEs and Germany, Japan, Switzerland, the Netherlands, Belgium, Sweden, Norway, Demark, Finland and Austria as CMEs. France, Italy, Spain, Portugal, Greece, and Turkey represent an indeterminate case with some coordinating institutional features, but a deregulated labour market. 15 Estevez-Abe and colleagues (2001) show that in CMEs firms will be better able to induce employees to specialize in firm- or industry-specific skills when there is a high degree of both employment and unemployment protection as unemployment poses a greater danger (in terms of future losses) to workers that have skills that are not readily transferable to other production processes. This implies that in CMEs there is a higher liklihood of structural unemployment. 11 Conversely, for firms in LMEs, low levels of employment and unemployment protection will be optimal as firms require the flexibility to hire and lay-off employees according to the dictates of the market. Employees, too, will place less value on robust employment and unemployment protection as their skills are more readily transferable among firms. Unemployment is more likely to be short in duration and frictional or cyclical in nature. 2.3.1: Social Protection and Skill Profiles There are four different types of skill-production profiles based on the degree of employment and unemployment protection, each with attendant reinforcing institutional arrangements (Figure 2.2) (Estevez-Abe, Iversen, and Soskice 2001). Employment protection (i.e, restrictions on terminating workers, even in the face of an economic downturn) encourages and protects firm- specific skill investments, while unemployment protection (i.e., measures that provide adequate earnings replacement until a suitable job within the same industry is found) encourages and protects industry-specific skill investment. In the absence of both high levels of employment and unemployment protection the optimal investment in training is in general skills that are readily transferable across firms and industries. In the VOC framework high levels of employment and unemployment protection enhance a firm‘s ability to maintain its comparative advtantage in CMEs but not LMEs. This is in contrast to the neo-classical macoeconomic critique of high levels of employment and unemployment protection in which high levels of employment protection retards job creation by reducing the willingness of employers to hire workers as those workers will be difficult to terminate (OECD 2006), while high levels of unemployment protection raise workers‘ reservation wage (i.e., the lowest wage for which workers would be willing to offer their services) and increase the length of unemployment (Atkinson and Micklewright 1991). 11 Structural unemployment is defined as unemployment that occurs because workers do not have the skills that are in demand; frictional unemployment occurs when a job exists for a worker but the worker cannot access the job due to geographic or other constraints; cyclical unemployment occurs when there is an excess supply of labour due to a downturn in the economy. 16 The clustering of skill and institutional arrangements around production processes leads to two distributional outcomes, one relating to wage inequality and the other to industry and occupational gender segregation. In CMEs the returns to vocational training (a principal form of firm and industry-specific skills) are greater than in LMEs. In LMEs there is a greater return to a high level of a general skilled education (i.e, college or a professional degree), while those without these qualifications tend to be relegated to lower wage service sector jobs. Thus in CMEs wage inequality across occupations and skill profiles will be less than in LMEs. Conversely, occupational gender segregation will be greater in CMEs compared to LMEs given the lower return on investment to both female workers and firms of firm- and industry-specific skills compared to male workers, given the likelihood of a career interruption related to raising a family. As a result in CMEs females are more likely to invest in general skills compared to males (Estevez-Abe 2005). Szydlik (2002) provides a complementary perspective underscoring how the demand for different modal skill types in CMEs and LMEs creates an additional mechanism for economic and social stratification. He argues that in coordinated market economies 12 there will be a better educational-occupation match compared to liberal market economies due to the higher degree of coordination between vocational education and employment. There is a greater degree of labour market segmentation by income and job requirements in LMEs with a primary labour market characterized by a high general skill requirement (i.e., a university degree) and well paying jobs, and a larger secondary labour market characterized by low skill requirements and low paying jobs. The economic returns to education are greater in a LME, but there is also a greater risk of not achieving those returns through relegation to the secondary labour market. In CMEs, the primary labour market centres on the need for firm- and industry-specific vocational skills with smaller secondary and tertiary labour markets for the low skilled and the high general skilled. The return to high general skills will be smaller in CMEs compared to LME, but wage inequality will be less given the larger demand for medium- (but specific-) skilled workers. Szydlik‘s research on returns to education and skill-occupation mismatch using the GSOEP and PSID provides support for these hypotheses, finding that there is a higher level of skill-occupational fit in Germany compared to the United States and that there is a greater earnings penalty for both the overqualified (i.e. high skilled but in a low skilled job) and those 12 Szydilk makes the distinction between flexibility coordinated economies, deregulated economies and planned economies. For consistency I continue use the coordinated market and liberal market nomenclature. 17 with low skills in a low skill job compared with the medium and high skilled with an suitable skill-occupation fit in the United States (Szydlik 2002). 2.3.2: Unemployment and Employment Protection in Germany, Canada and the United States Unemployment and employment protection are the central institutional mechanisms that may mediate the unemployment and health relationship. This may occur through influencing who is exposed to unemployment (i.e., who is in the labour market and their risk of unemployment) and through mediating the direct effects of unemployment on health (i.e., the material and career consequences of unemployment). Across high-income countries, there is large variation in the levels of unemployment and employment protection; in this section they are reviewed in detail for the three study countries. 188.8.131.52: Unemployment Protection Unemployment protection relates to coverage of the unemployed (i.e., the proportion of the unemployed who receive unemployment compensation), the generosity of unemployment benefits in terms of net replacement rates of pre-unemployment income and in the duration of benefits, and whether the unemployed are required to take any job available or can wait until a ‗suitable‘ job can be found (Estevez-Abe, Iversen, and Soskice 2001). More broadly, other state support and public transfers can also be viewed as unemployment protection, since income replacement for the unemployed can depend on other public transfers including social assistance, benefits for children including maternity or paternity benefits, and one-time payments for extraordinary expenses. Favourable tax treatment while unemployed, other public transfers and other forms of tax entitlements (e.g., refundable tax credits) can also increase the incomes of the unemployed. In Germany there are two forms of unemployment benefits, namely unemployment insurance benefits and means-tested unemployment assistance that is now coupled with social assistance (Schneider 2004). Workers must be registered at the local unemployment office and actively engage in a job search in order to receive benefits. They can work up to 15 hours a week and earn a nominal sum without losing their benefit entitlement. Unemployment insurance benefits are available to workers with at least 360 days of insured employment in the previous three years and are payable for a maximum of 360 days for workers under 45 years of age, 18 increasing by step-wise age increments to 960 days for workers over 57 years of age. Benefit replacement is 67% of net income for workers with children and 60% for workers without children to a maximum of 94% and 84% of the average wage in 2001. Unemployment assistance is a means-tested benefit for workers with at least 150 days of unemployment in the past year, but without enough employment to qualify for unemployment insurance benefits or who have exhausted their unemployment insurance benefits. Benefits levels are 57% and 53% of net pre- unemployment earnings for workers with or without children, respectively. Prior to 2005 individuals not eligible for unemployment benefits or who have a household income below a minimum threshold could also receive social assistance. After 2005 unemployment assistance and social assistance were merged into one benefit and are no longer contingent on employment income (OECD 2006). Unemployment and social assistance benefits are not subject to income tax or social security contributions. Low-income unemployed in Germany are also eligible for a variety of other public transfers including a means-tested housing benefit and a universal child benefit (Adema, Gray, and Kahl 2003). In Canada unemployment benefits are available through the federal Employment Insurance programme that provides benefits to eligible unemployed workers, parents on maternity or paternity leave and to some seasonal workers such as those who work in the fishing industry. 13 While the programme is federal in nature, eligibility requirements and maximum benefit durations vary by the regional unemployment rate with fewer qualifying hours required for eligibility and maximum benefit durations longer in high unemployment regions. In 2001, the replacement rate was 55% of the average weekly wage in the preceding 26-week period to a ceiling of $418 a week; benefits could be received for a maximum of 45 weeks depending on the local unemployment rate, number of qualifying hours worked and previous receipt of unemployment benefits. The maximum replacement rate in 2001 was 58% of the average wage and unemployment benefits were considered taxable income. There is also a small child supplement for low income families with children. Low-income unemployed may also be 13 Between 1990 and 1996 a number of changes to Canada‘s unemployment insurance system took place that restricted benefit entitlement. These included increasing the number of weeks an individual had to work in low regions of unemployment in order to qualify for benefits, making workers who quit or were fired from their job ineligible for benefits, reducing the replacement rate to 50% for some users, and a 100% claw back of benefits for high-income repeat users. The system also changed from Unemployment Insurance to Employment Insurance as it also included maternity and paternity benefits. For consistency with Germany and the United States, I continue to use the term unemployment insurance. 19 eligible for means-tested tax credits and cash transfers for families with children, provincial social assistance benefits and federal and provincial sales tax refunds (OECD 1999). In the United States unemployment benefits are jointly administered and funded at the federal and state level. Subject to federal guidelines, states set their own eligibility requirements and benefit levels and durations. States require a minimum level of earnings or number of weeks worked in the qualifying period. In general, benefit levels are around 50% of the workers‘ qualifying earnings to a maximum of 50% of the State‘s average wage; maximum benefit duration is usually 26 weeks but can be as long as 39 weeks in high unemployment areas (ORDP 1997). In Michigan, the state the OECD uses in international comparisons, the maximum replacement rate in 2001 was 46% of the average wage. Like Canada, unemployment benefits are considered taxable income. Low income unemployed may also be eligible for social assistance, cash transfers and tax credits for children, and food stamps, but may lose the earned income refundable tax credit 14 for low wage earners. Taken together, the net income replacement rates for the unemployed from total public household transfers vary substantially across the three study countries (Table 2.1) (OECD 2009). Replacement rates are lower in the United States across all income and family types, especially for families with children due to both Canada and Germany having more generous programmes that provide tax credits and cash transfers to families with children. For example, a single person with no children earning 50% of the average wage prior to unemployment would have a net income replacement rate 15 of around two-thirds across the three countries (71% in Germany, 66% in Canada, and 64% in the United States). The net replacement rate rises to 95% in Germany and to 79% in Canada for a similarly-waged single person with two children, but falls to 61% in the United States due to the loss of the earned income tax credit. Germany‘s replacement rates are higher at lower levels of income than Canada‘s due to a low income housing subsidy and more generous welfare rates, while they are similar at average levels of income across all family types ranging from 60% to 91% for Germany and 64% to 85% for Canada. At 150% the average wage Germany maintains a replacement rate ranging from 60% to 14 A refundable tax credit is credit that is paid to individuals even if they owe no income tax, and acts like a direct cash transfer. 15 Net income replacement rate is the ratio of pre-unemployment post tax and transfer household income to post tax and transfer income in the first month in receipt of unemployment benefits. These scenarios assume a single earner in the household. 20 87%, while Canada‘s and the United States‘ replacement rates falls to between 46% to 70% and to 37% to 60%, respectively The above paragraphs describe the maximum benefits levels that are available to the unemployed should they be eligible for unemployment insurance, but these scenarios may not reflect the experience of the typical unemployed who may not be eligible for unemployment benefits or who may receive less than the maximum entitlement. Individuals with weak labour force attachment – those working part-time or with short-term contracts – may be more likely to become unemployed, but less likely to receive unemployment compensation. Figure 2.1 depicts unemployed insurance beneficiaries as a percentage of the total unemployed drawn from the three study countries‘ labour force surveys. About 70% to 80% of the German unemployed received benefits during 1976 to 2000, while in the United States 30% to 40% of the unemployed received benefits. Canada presents a contrast; prior to the reforms to the unemployment insurance programme benefit levels were similar to those in Germany at around 70%, but after the reforms coverage levels are similar to those in the United States. 16 For the decade of the 1990s, Vroman and Brusentsev (2005) report that the coverage of unemployed was 76% for Germany, 60% for Canada, and 34% for the United States and that the average replacement rate was 48% for Germany, 45% percent for Canada, and 34% percent for the United States. 184.108.40.206: Employment Protection Employment protection relates to how difficult it is to dismiss a worker. Employment protection is defined by the OECD as regulatory and legislative requirements pertaining to job separations for regular workers, the use of fixed-term contracts or temporary workers, and collective dismissals (OECD 2004b). The number of regulations is diverse spanning over 18 measures, which are summarized in Table 2.2. For regular employees (i.e., those not on a fixed-term or temporary contract), Germany has higher levels of employment protection than Canada and the United States (OECD 2004a). There are stricter standards for notification including the need to notify the local work council which can contest the dismissal in court, longer required notice periods, and the requirement to retrain or reassign employees. Workers unfairly dismissed may also be eligible for compensation or reinstatement. Germany has limitations on length of and type of work that can be covered under fixed-term or temporary contracts, while there are no 16 Unfortunately Canadian surveys that include both unemployment and health measures are only available from 1994 (NPHS) and SLID (1996), so I am not able to examine whether the change in unemployment benefit coverage had an effect on the health of the unemployed within Canada. 21 limitations in Canada and the United States. All three countries have requirements to notify employee organisations and governmental authorities in the case of collective dismissal, but the threshold for dismissals to be considered a collective dismissal is lower in Germany and the local work council has the right to contest the dismissal. Notably, Canada has more generous severance requirements than Germany. The difference between Canada and the United States is more modest, but Canada has slightly stricter notice and severance pay requirements compared to the United States, which has none. Unfair dismissal in both countries, in general refers to dismissal due to discrimination, but there is greater recourse to compensation and the possibility of reinstatement in Canada. 17 2.3.3: Post-unemployment Trajectories in CME and LME Countries Post-unemployment labour market trajectories may also be a key pathway through which unemployment can affect health. The literature that has examined post-unemployment career and earning trajectories is collectively known as the labour market scarring literature (Jacobson, Lalonde, and Sullivan 1993; Kuhn 2002; Ruhm 1991; Topel 1990). Unemployment can lead not only to the loss of immediate employment income but also may harm a worker‘s future career and earning prospects though reemployment in a job that pays less or is less desirable (Brand 2006) and through the increased likelihood of future unemployment (Eliason and Storrie 2006). Long periods of unemployment may also lead to skill deterioration and the loss of productivity. Post-unemployment career trajectories may also differ by institutional environment and by the strength of employment and unemployment protection (DiPrete and McManus 1996; Gangl 2006; Gangl 2004). The theoretical underpinnings of how unemployment and employment protection could affect post-unemployment labour market trajectories lead to mixed predictions on whether these protections protect future career and earning losses. High unemployment protection reduces the immediate income effects of unemployment and enables the unemployed to wait for a ‗suitable‘ job (i.e., a job with an acceptable occupational-skill match and of comparable pay to the pre- unemployment job). Conversely, high unemployment protection may reduce the search intensity for a new job increasing the time unemployed, thereby leading to skill deterioration and loss of worker productivity. High levels of employment protection may mean that in order to downsize 17 Based on the 18 employment indicators the OCED constructs a weight scale from 0 (lowest) to 6 (highest) reflecting the strength of employment protection. On this scale the United States scores 0.8, Canada 1.1 and Germany 2.5. No country scores higher than 3.5 (Portugal) (OECD 2004c). 22 or close a plant, employers may be required to partner with local authorities, unions and workers to facilitate or mitigate the consequences of the plant closure and this may lead to a smoother transition for the terminated workers (Kuhn 2002). But high levels of employment protection may reduce the outflows from unemployment as employers will be less likely to expand employment given the potential for future dismissal costs (OECD 2004b). High levels of employment protection may also increase the likelihood that when an employer terminates a worker, it is permanent. Post-unemployment earning deficits have been found in economies of all high-income countries (Eliason and Storrie 2006; Gangl 2006; Gangl 2004; Jacobson, Lalonde, and Sullivan 1993; Kuhn 2002). No countries have adopted strategies or developed institutional arrangements that entirely mitigate the negative effects of post-unemployment earning trajectories. There is, however, a growing body of evidence that shows that labour market scarring is worse in LME countries, and in particular the United States, compared to CME countries. In a series of studies that examined post-unemployment trajectories across ten high-income countries, including Germany, Canada and the United States, Kuhn (2002) reports that the likelihood of unemployment after job displacement was lower in countries with high levels of employment protection compared to the United States, while earning losses were greater in Canada, the United States and the United Kingdom (LME countries, but also countries with high wage inequality) for workers with long-standing tenure. 18 Gangl (2004), in a comparative analysis of Germany and the United States, finds that the receipt of unemployment benefits significantly improves post-unemployment earnings in both the Germany and the United States. Overall, the greater coverage and benefit generosity of unemployment compensation explains the lower levels of labour market scarring in Germany compared to in the United States, but the unemployed in the United States who receive benefits (albeit a minority) have better post- unemployment outcomes that their German counterparts. In a second analysis Gangl (2006), drawing on data from 12 European countries and the United States, finds that once the temporarily laid-off (e.g. short-term plant closures) are removed from the data, the unemployed 18 In contrast to standard the macroeconomic critique of strict labour market regulation (e.g., (OECD 2004b) (OECD 2006)), Kuhn and colleagues conclude that strong employment protection laws appear to reduce the incidence of an unemployment spell for those who lose their job involuntarily. However, they caution that high levels of protection may only protect those who already have a job and may retard the earnings and employment prospects of new workers or those with weak labour force attachment (e.g., younger workers and women). 23 in LME countries have similar durations of unemployment as those in CME countries, but poorer employment and earning outcomes. 2.3.4: Critiques and Alternatives The Varieties of Capitalism (VOC) approach is not the only framework that seeks to explain the persistence of institutional variation among developed economies. Indeed, this approach, situated at the nexus of economic theory on the nature of economic growth and production arrangements and political science theory on development and persistence of institutions, has been critiqued by both economists and social theorists. Neo-classical (Watson 2003) and transaction cost (Allen 2004) perspectives take issue with the development and persistence of multiple allocatively efficient equilibria 19 among high-income countries, while the structuralist critique focuses on the lack of emphasis on class relations and power structures in explaining institutional variation (Coates 2005). It is worthwhile to contrast Esping-Andersen‘s (1990) Three Worlds and VOC typologies as it is Esping-Andersen‘s typology that has been used most often to examine labour market structures and attendant government- and firm-level supports (Berthoud and Iacvou 2002; Gallie and Paugam 2000; Muffels and Fouarge 2002). While both are grounded in a historical institutional approach, in the Three Worlds typology it is the legacy of a historical class-based struggle in establishing various degrees of social protection that explains different societal trajectories. In the case of the VOC typology it is the complementarities and reinforcing comparative advantages that arise among different ways of organizing the production process and the social welfare state that explain the different trajectories. There are distinct ontological differences between the two approaches. VOC is a rational choice- based approach which places firms and individuals at the centre of the model and the distribution of skill and wage inequality across CMEs and LMEs is largely an epiphenomenon that results from the two different equilibria. In contrast, inequality, class differences, and the state‘s response are at the core of the Three Worlds approach. Accordingly, the application of the VOC 19 Rather the argument is that the imposition of state-mandated institutions on the market introduces a set of distortionary effects that lead to sub-optimal economic outcomes and that in the absence of state interference in the market these institutions would cease to exist and the convergence to a liberal-market economy equilibrium would occur. 24 framework may lead to a different understanding and interpretation of how social and economic processes affect health inequalities. The VOC approach has been chosen to be the principle organizing framework for two reasons. There is increasing empirical support for its principal conclusion relating to the existence and persistence of two equilibria (Allen, Funk, and Tuselmann 2006; Hall and Gingerich 2004). In addition, it is one of the more tractable frameworks for developing a set of testable hypotheses on how context might influence the relationship between the labour market and health. 2.4: Unemployment as a Social Determinant of Health 2.4.1: Linking Unemployment to Health Building on Evans‘ and Stoddart‘s (2003) health production framework, Figure 2.4 motivates how institutional context could affect the individual-level relationship between unemployment and health. At the individual level, unemployment may affect health through three individual- level pathways or dimensions: 20 material, in which the loss of income reduces the ability to invest in health (e.g., access to health care, housing, education, nutrition, physical activity); psychosocial, in which the loss of status and identity, increased feelings of insecurity, and family or role conflict lead to increased psychological stress which in turn affects health largely through the activation of physiological and nervous system responses (Mustard, Lavis, and Ostry 2006); and, indirectly through the diminishment of future economic or labour market success (i.e., a spell of unemployment leading to an increased likelihood of taking a job that has worse income or working conditions or a job that introduces or magnifies a skill-occupation mismatch). Additionally different types of unemployment may have a differential impact on the other components of the model. Specifically it is expected that cyclical unemployment (i.e. related to the macroeconomic cycle) would have less effect on health than structural unemployment (i.e., related to the mismatch between skills demanded by firms and skills that unemployed workers are able to supply). Indeed in the case of industries characterized by periods of regular and anticipated cyclical unemployment it may be that unemployment would not, at least for those working voluntarily in these arrangements, have any negative effect on health. 20 I use the term dimension to connote that these pathways do not operate in isolation with another, but rather in concert or in interaction on health. Further, these dimensions operate within a temporal one which enables us to consider how unemployment influences health over the life course, either at a point in time (at a specific age) or through cumulative exposure. 25 This individual-level model is embedded in a societal (i.e., economic, institutional and cultural) context which has the potential to mediate (either mitigate or magnify) any of the individual- level pathways. For example, the provision of universal health care insurance in Canada and Germany could, at least partially, mitigate the material pathway between unemployment and health, while the largely firm-contingent provision of health insurance in the United States for those covered neither by Medicaid (the poor), nor Medicare (the elderly) could magnify the material pathway between unemployment and health through the direct loss of health insurance and access to health care or through the loss of additional income resulting from the need to purchase health insurance or health care. 2.4.2: Varieties of Capitalism and Unemployment-related Health Inequalities Building on Figure 2.4 the Varieties of Capitalism framework enables us to specify how the institutional environment can affect the unemployment and health relationship. Hall and Soskice concisely summarize the features of CME and LME that could give rise to differences in health inequalities. In liberal market economies, the adult population tends to be engaged more extensively in paid employment and levels of income inequality are high. In coordinated market economies, working hours tend to be shorter for more of the population and incomes more equal. With regard to the distributions of well- being, of course, these differences are important (Hall and Soskice 2001 p.21). The compression of wage-inequality and the skill-occupational equilibrium that targets the modal medium-skilled worker implies that overall socio-economic gradients will be shallower in CME countries, while the skill-occupation equilibrium that reinforces economic inequality and social stratification may lead to the steepening of these gradients in LME countries. 21 Moreover, the institutional supports in CMEs for the unemployed and the low-waged may further attenuate the socio-economic gradients in health. Taken together, these insights lead to two main hypotheses of how the institutional environments across CME and LME may mediate or magnify the effect of unemployment on health: - Higher levels of employment and unemployment protection will mitigate the effect of unemployment on health in CMEs compared to LMEs. These institutional supports in CMEs provide direct material support to the unemployed and reduce the negative long- term effects of unemployment on career earnings. This also leads to the supplementary 21 In that the potential for gains are greater, but then so is the potential for loss. 26 hypothesis that the receipt of unemployment benefits will also mediate the effect of unemployment on health within countries. - The different occupational-skill equilibria will lead to effect modification in the unemployment and health relationship by skill level. There will be a steeper education- health gradient in LMEs compared to CMEs and the health-risks of unemployment will also be greater for those of lower skill. Further once the CME differences in compositional (individual-level) characteristics are accounted for, the effects of unemployment on health will be lowest in the medium skilled in CMEs as the institutional environment is targeted towards these workers. In CMEs, it is unclear how effective the institutional environment will be in mediating the unemployment and health relationship in the long-term unemployed. While there are institutional supports for the long-term unemployed that mitigate some of the material effects of unemployment, long-term unemployment has the potential for permanent exclusion from the active labour market. Accordingly, the psychosocial effects of unemployment on health may dominate. In LMEs, on the other hand, there would be both material and psychosocial effects on the health of the long-term unemployed, but fewer long-term unemployed given the incentive for the long-term unemployed to return to employment or exit out of the labour force. 2.4.3: Health Selection versus Social Causation The framework also clearly indicates the dual relationship between health and unemployment; unemployment may determine health, but health is also a determinant of labour market success. The debate around health selection into unemployment is not new (Bartley 1988), but it is only recently, using longitudinal study designs and appropriate statistical methodology, that the causal arrows between unemployment and health have begun to be disentangled (Burgard, Brand, and House 2007; Elkeles and Seifert 1993; Gerdtham and Johannesson 2003; Korpi 2001; Leigh 1987). There are sound theoretical reasons and empirical evidence to support the contention that health selection into unemployment will account for some of the association between unemployment and health. Poor initial health or the experience of a negative health shock is strongly associated with unemployment or labour force exit (Arrow 1996; Riphahn 1999). The healthy worker effect (e.g. health selection into employment) is a also a well established phenomenon in occupational health research in that workers report better health outcomes when compared to general populace (Dahl 1993). Good health (or at least some minimal level 27 of health) is a requirement for productive employment. Health selection into unemployment occurs because firms base layoff decisions, in part, on marginal productivity in which the least productive and the least healthy workers (in so far as productivity is related to health) are the first terminated with firms progressively laying off more productive and more healthy workers as demand for labour contracts. Comparisons of the unemployment and health relationship during times of high unemployment and low unemployment (Iversen et al. 1987; Martikainen and Valkonen 1996; Martikainen, Maki, and Jantti 2007; Novo, Hammarstrom, and Janlert 2000) or across areas of high and low unemployment (Beland, Birch, and Stoddart 2002; Lavis 1998) and natural experiments such as plant closure studies (Eliason and Storrie 2006; Hamilton et al. 1993; Keefe et al. 2002; Sullivan and Wachter 2007) are an attempt to account for this form of health selection. Mass unemployment and plant closures may have different effects on health than singular or small scale job loss. Plant closures and mass layoffs require a longer period of notice in most jurisdictions and there are often additional measures to mitigate the effect of job loss including extra severance and buyout provisions, retraining and labour activation measures. Governments may also step in with additional measures to create jobs or otherwise mitigate the impact of job loss. Mass unemployment may create solidarity among those losing a job and reduce the stigma of job loss. Mass unemployment can also have contextual effects at the community level rather than at the individual level as it can represent the loss of financial and other resources in the community (e.g. a major employer closing down in a one-industry town) which leads to lower aggregate community income through the loss of the working-age population (the workers downsized, but also the associated services to support them) as they leave to search for other economic opportunities. Plant closure studies and studies comparing the effect of unemployment on health between places or times of high of unemployment compared to those of low unemployment may not be an accurate test of the health selection hypothesis across all types of unemployment. At the same time there is also health selection out of unemployment. The relationship between unemployment and poor health status at a given point in time can be partially explained in that poorer health is related to a longer duration of unemployment (Korpi 2001; Stewart 2001). Research that has looked at the effect of unemployment on health while controlling for health 28 selection has yielded mixed results with some studies finding that the effect of unemployment on health is robust to controls for health selection (Eliason and Storrie 2007; Gerdtham and Johannesson 2003; Kerkhofs and Lindeboom 1997; Kiuila and Mieszkowski 2007; Korpi 2001; Leigh 1987; Rodriguez 2001) and others finding that the relationship is attenuated and no longer statistically significant (Ahs and Westerling 2006; Elkeles and Seifert 1993; Frijters, Haisken- DeNew, and Shields 2005a; Martikainen, Maki, and Jantti 2007). Both health selection into unemployment and health selection out of unemployment may be modified by institutional arrangements. High levels of unemployment protection creates an incentive for individuals in ill health to remain unemployed rather than exiting the labour force or returning to employment, while in LMEs lower levels of unemployment protection means that the ability for individuals in ill health to remain unemployed is circumscribed and they may be more likely to exit the labour force or (if able) return to work. 22 An argument can be made that higher employment protection will also have an effect on health selection into and out of unemployment. Health selection into unemployment may be less likely in countries with high levels of employment protection as firms are constrained in their ability to lay off the least productive workers. This constraint will also reduce the firm‘s willingness to re- employ or hire new workers. As such the standard theory of labour demand would lead to the conclusion that the effects of high employment protection would be ambiguous on health selection into or out of unemployment. The application of the VOC framework provides a different interpretation given that in CMEs labour demand and labour supply are coordinated through non-market mechanisms. Because firms require specific skills from labour to engage in the specialized production processes characterized by CMEs the difference in productivity between the penultimate and last (marginal) worker will be small. When firms dismiss workers they are choosing between similarly skilled workers for whom there will be small differences in productivity. And as such the effect of high employment protection on productivity-related health selection into and out of unemployment will be less. 22 The same incentive effects of high unemployment protection also exists for the unemployed who are healthy, but I contend that the ill unemployed are more likely to take advantage of the decommodification effects of high unemployment protection than the healthy unemployed. Moreover, even if the ill unemployed are unable to work they may still be able to meet the job search and activation requirements in order to continue to receive unemployment benefits. In LMEs, this incentive is much smaller, and so the unemployed unable to work due to ill health would be more likely to exit the labour force. 29 Thus in CME countries there will be less health selection into unemployment given high levels of employment protection, but there may be greater health selection out of unemployment given the high levels of unemployment protection. Moreover, if the institutional environment in CMEs is effective in mitigating the effect of unemployment on health (i.e., the social causation hypothesis), the residual association between unemployment and health will be due to selection. 2.5: Unemployment and Health in Context The above sections have articulated the ways in which institutional context could affect the individual-level relationship between unemployment and health. This section reviews the comparative studies that have examined differences in unemployment-related health inequalities by welfare-regime type or CME and LME countries. It also categorizes by CME and LME all cohort studies that have examined the relationship between unemployment and mortality to test whether the relationship is mediated by institutional setting. 2.5.1: Comparative Studies of Unemployment and Health Bambra in a cross-sectional analysis (Bambra and Eikemo 2009) applied Ferrera‘s (1996) welfare regime typology to examine whether the relationship between self-reported health status and limiting longstanding illness and unemployment varies by welfare regime cluster using the European Social Survey. In age-standardised models they report that the unemployed in liberal (Anglo-Saxon) welfare regimes tended to have the highest odds ratio of poor or fair self-reported health for men (OR 3.0 95% CI: 1.9-4.6) and for women (OR 2.8 95% CI: 1.6-4.7), but that men (OR 2.7 95% CI: 2.2-3.4) in conservative (Bismarckian) regimes and women (OR 3.0 95% CI: 2.3-4.0) in social democratic (Scandinavian) regimes also had high risks. The risks in the Southern and Eastern welfare regimes were the lowest for both men and women and across health outcomes. Two other studies using longitudinal data have also examined how the relationship between unemployment and health varies across European countries (Cooper, McCausland, and Theodossiou 2006; Cooper, McCausland, and Theodossiou 2008). Cooper and colleagues conducted two related studies examining the relationship between current unemployment and duration of good health (defined as not reporting any physical or mental health problems or illness or disability) among 14 European countries using the European Community Household 30 Panel (ECHP) 23 . Using an accelerated-failure time (AFT) model and a discrete-failure time (DFT) model they find a statistically significant risk of exiting good health due to unemployment for most countries and marked variation across countries in the magnitude of the risk estimates. There is no discernable pattern to their results when a welfare regime or political economy lens is applied to their results and there is considerable variation in the risk estimates and ranking of study countries across statistical methodologies. In the AFT analysis Denmark (HRR 4.1) and the Netherlands (3.6) – CME or social democratic – countries have the highest hazard ratio of the unemployed exiting good health, while France (1.1) , Belgium (1.2), Italy (1.2) and UK (1.5) – a mixture of LME, CME or conservative and liberal welfare state regimes – have the lowest or not statistically significant hazard ratios. In contrast in the DFT model Greece (2.0) and Austria (2.1) have the highest odds ratio of the unemployed exiting good health, while Netherlands (0.92), Denmark (0.97), and Belgium (1.1) and Finland (1.1) have the lowest risk. Germany in both models tends to represent a middle case with a hazard ratio of 1.5 and odds ratio of 1.2. 24 It is difficult to draw comparisons across the Bambra and Cooper studies given that they have different study designs and health measures. Cooper‘s health measure is constructed to account for health selection into unemployment and they account for a broad range of confounders, while Bambra‘s study is cross-sectional and standardized for age. Given the heterogeneity of design, method and results across the three studies the evidence for variation in the relationship between unemployment and health across European welfare regimes is inconclusive. 2.5.2: Studies of Unemployment, Unemployment Compensation and Health Only a few studies have examined whether the direct receipt of unemployment benefits and public transfers (such as welfare or social assistance) ameliorate the effect of unemployment on health in single country and cross country studies (Bolton and Rodriguez 2009; Rodriguez 2001; Rodriguez, Frongillo, and Chandra 2001; Rodriguez, Lasch, and Mead 1997; Strandh 2001). Rodriguez and colleagues have conducted a number of studies using American health and household surveys including the PSID to examine whether the receipt of unemployment benefits moderates the effect of unemployment on self-reported health status, depression, BMI and 23 The ECHP is a longitudinal household panel survey covering 14 West European countries spanning 1994 to 2001. Germany‘s contribution to this survey was taken from the GSOEP. 24 A case could be made based on the DFT results that social democratic countries collectively have lower odds ratio, but this is not supported in the AFT analysis. While the authors do not reconcile the results across the two analyses, the DFT model adjusts for group-specific unobserved heterogeneity and the AFT model does not. 31 health-related behaviours. In a cross-sectional analysis using the 1987 wave of National Survey of Families and Households, they found that those currently unemployed and in receipt of welfare benefits reported higher levels of depression and more days depressed in the week compared to employed controls. Those unemployed and not receiving any welfare or unemployment benefits reported a smaller, but still statistically significant increase in the depression measures, while those unemployed and in receipt of unemployment benefits did not report different levels of depression compared to the employed (Rodriguez, Lasch, and Mead 1997). In a second longitudinal analysis, women in receipt of unemployment benefits in 1987 reported lower levels of depression compared to the employed in 1992, but no other significant differences in depression were observed among the unemployment and benefit groups compared to the employed (Rodriguez, Frongillo, and Chandra 2001). More recently Bolton and Rodriguez used the 1999 and 2001 waves of the PSID to study whether the receipt of unemployment benefits moderated the effect of prior unemployment on changes in BMI and smoking and drinking in a group of re-employed individuals compared to a continuously employed control group (Bolton and Rodriguez 2009). They found that the unemployed who did not receive unemployment benefits were 1.8 times more likely to report an increase in alcohol consumption and 1.7 times more like to report a decline in BMI, but that no associations were observed for the unemployed who received unemployment benefits. Using an early version of the cross-national equivalent file 25 (CNEF), Rodriguez (2001) also examined whether the receipt of means-tested (i.e., welfare or social assistance) and unemployment benefits mediated the relationship between unemployment and self-assessed health 26 in Germany, United Kingdom and the United States over a three year period (1985-1987 for the PSID and 1991-1993 for Germany and the United Kingdom). Regular unemployment benefits moderated the relationship between unemployment and health in the United Kingdom (OR 1.3; 95% CI: 0.8-2.1), Germany (OR: 1.1 95% CI: 0.9-1.4), and the United States (OR 1.7 95% CI: 1.0-2.9) compared to the unemployed in receipt of means-tested benefit (UK OR 1.6 95%CI:1.1-2.4; GER OR 2.2 95% CI: 1.1-4.4; USA OR 2.4 95% CI: 1.4-4.1). An association 25 The CNEF is a set of harmonized files spanning the PSID, GSOEP, SLID and other surveys. See section 3.2.4 for more detail. 26 The self-assessed health outcome was defined as fair or poor self-reported health status in BHPS and the PSID States, and with a similar variable derived from the health satisfaction variable in the GSOEP as self-reported health status was not asked until 1994 in the GSOEP. 32 was also observed between the unemployed not in receipt of any benefits in the United States (OR 1.6 95% CI: 1.0-2.4), but not in the United Kingdom and Germany. The studies conducted by Rodriguez and colleagues have a number of strengths in that almost all were longitudinal cohorts, enabling the appropriate temporal sequencing from exposure to unemployment to health outcomes, as well as consideration of baseline or prior health status. The studies also controlled for a range of other variables that may confound the relationship between unemployment and health including age sex, marital status, socio-economic status and prior employment history. However the longitudinal studies had only one period of follow-up which meant that unmeasured individual-level effects could not be modelled and as such residual or unmeasured confounding cannot be ruled out. Further the different groups of unemployed tended to be small in size, ranging from between 35 to 400, implying that formal statistical testing differences among unemployed groups would not likely yield statistically significant differences. 27 Sweden, 28 like Germany, has two forms of unemployment compensation: a more generous benefit for individuals who have paid into an unemployment insurance fund and have worked five of the prior twelve months that pays up to 75% pre-unemployment earnings and a secondary, less generous, fixed cash benefit of about one-third the maximum payout of the more generous benefit. Strandh (2001) studied the effect of the two benefit systems and participation in labour market activation programmes on the mental health of unemployed in Sweden using a longitudinal survey of a national random sample of unemployed individuals. He found that in both cross-sectional and longitudinal analyses the receipt of the more generous unemployment benefit led to higher levels of mental health (defined by the 12-item version of the General Health Questionnaire) 29 compared to the receipt of no unemployment benefits, but the receipt of 27 The argument here is that for effect of unemployment benefits on the unemployed to be definitive it is the contrast between the two unemployed groups that matters and not just significance of the difference between unemployed groups and the employed control group. Otherwise the interpretation of results can be driven by small differences in similar effect sizes and confounded by differences in sample size among the exposed groups. In the case of the United States, the unemployed group not in receipt of unemployment benefits is always larger than the unemployed group in receipt of unemployment benefits and thus given a similar effect size the results on the former group is more likely to be statistically significant. 28 In Sweden unemployment and income protection is greater than in Germany particularly at the lower end of the income distribution. For example a couple with two children earning 50 of the average wage pre-unemployment would experience no difference in their household income while unemployed as both the low waged and the unemployed are provided income subsidies to reach a minimum income (OECD 2009). 33 the less generous cash benefit did not confer any advantage. Moreover, labour market activation related to workplace participation (volunteer work experience at a regular workplace) also conferred an advantage, but other types of volunteer work experience and vocational training did not. Strandh‘s study supports the hypothesis that unemployment affects health through both the material and psychosocial pathways, both of which may be amenable to intervention. 2.5.3: Unemployment and Mortality in CME and LME Countries Population-based cohort studies that examined the relationship between unemployment and mortality in coordinated and liberal market economies were reviewed to investigate if the extant literature supported a difference in the relationship between unemployment and mortality by CME and LME. Studies were identified through a search strategy that built on a systematic review of published studies to 1998 (Lavis et al. 2001) and supplemented by a review of known study references, Web of Science citations, 30 Medline and Google Scholar searches. Studies were included if they were a population-based cohort study or a plant closure study that looked at the individual-level relationship between unemployment and mortality and were published in a peer- reviewed format (journal or book), or were a research report or working paper from a university or research institution (e.g., National Bureau of Economic Research). Studies that looked at aggregate-level relationships (i.e., the relationship between the unemployment rate and mortality) or those of clinical or patient populations were not included. Studies that grouped other labour force statuses with the unemployed were also not included (e.g. studies that included those on disability pensions in the unemployed (Johansson and Sundquist 1997) or those otherwise not working (Franks, Clancy, and Gold 1993)). Studies were classified into CME and LME clusters and were reviewed for study methodology and statistical method, cohort construction, time period and follow-up, measurement of unemployment, adjustment for confounders or covariates and study results. Tables 2.3 and 2.3 provide a high-level summary of the study results. Results are summarized for the entire cohort (men and women together) and for men and women and younger and older workers separately. The unemployment measures used in the studies were grouped five ways: - Current unemployment (CU) defined as being unemployed on the day of survey or census; 30 I used the Web of Science function that links a paper to all articles that cite that paper. 34 - Short-term unemployed (STU) defined as being unemployed one to three months at or previous to baseline; - Long-term unemployed (LTU) defined as being unemployed longer than four months at or prior to baseline or over more than one measurement period; - Plant closure unemployment defined as unemployment due to plant closure or mass downsizing; and, - Ever unemployed (EU) defined as any unemployment not otherwise specified. Risk ratios or effect sizes are summarized as ‗none‘ (no statistically significant relationship between unemployment and mortality), ‗low‘ (a statistically significant relationship between 1.0 to 1.5), ‗medium‘ (a statistically significant relationship between 1.5 and 2.0) and ‗high‘ (a statistically significant relationship greater than 2.0). Terminology and classification cut-points were chosen to reflect the prospect that statistically significant risk ratios close to one may be due to unmeasured confounding (Fewell, Davey Smith, and Sterne 2007). This is particularly the case for some studies in this review as they did not control for variables that are likely confounders in the unemployment-mortality association such as socioeconomic status or prior health status. Full results, including detailed information on cohort, data and follow-up period, analytic approach, measure of unemployment, covariate adjustment, and study results are found in appendix tables B1 for CME countries and table B2 for LME countries. The focus here is on all cause mortality, but the review includes two American studies that examined all injury-related mortality are also included (Cubbin, LeClere, and Smith 2000; Kiuila and Mieszkowski 2007); studies that examined detailed cause-specific analysis are not included in the synthesis (e.g., studies that examine the relationship between unemployment and suicide (Kposowa 2001; Lewis and Sloggett 1998; Norstrom 1988) ), although these results are still included in the appendix tables. Another study from Italy (Costa and Segnan 1987) is also not included in the synthesis as Italy is not considered either a CME or LME country. Overall 36 studies were included in the synthesis, 19 from CME countries and 17 from LME countries. Twenty-nine of the 36 studies were conducted in four countries (LME: 9 USA, 6 UK; CME: 8 Sweden, 6 Finland). Two studies were conducted in New Zealand (LME) and two in Denmark (CME). The remaining three studies were conducted in three CME countries – Germany, The Netherlands, and Switzerland. 35 Studies largely used population-representative survey data, census data, or linked administrative data. In the USA, four of the studies used data from the National Health Interview Survey, two used data from the Current Population Survey, two used other survey sources (the PSID and the National Longitudinal Survey of Older Men) and one was a plant closure study that used firm administrative data and unemployment records. In the UK, four of the six studies were conducted using the Office of the Population Censuses and Survey (OPCS) longitudinal study and two studies were drawn from other survey data (the British Panel Household Panel Study and the British Regional Heart Study). In Sweden, five of the nine studies were conducted using the Swedish Survey of Living conditions, two used a twin-cohort drawn from all twins born between 1928 and 1958, and the final study was a firm-closure study that linked firm administrative data to mortality records. In Finland, two studies used census data linked to mortality records, while four used census data and administrative employment records linked to mortality records. Of the five studies from the four other countries, one was conducted with census records and four used survey data. Cox proportional hazards estimation (16 studies) was the most common estimation approach. A relationship between unemployment and mortality was found in both CME and LME countries. This relationship remained for both county clusters after controlling for health selection. Across country clusters, the pattern within groups is similar with there being a consistent relationship between unemployment and mortality for men and younger workers and less so for women and older workers. For CME countries and based on the results from the fully- adjusted models, five of the ten studies (50%) found an association in the full (non-stratified) cohort, for men all eleven studies (100%) found an association, for women six of nine studies (67%) found an association, and for younger workers all three studies reported an association (100%), while for older workers only one of three studies (33%) found an association. Across LME countries two of the four studies (50%) found an association in the full (non-stratified) cohort, for men all eight studies (100%) found an association, for women one of four studies (25%) found an association, for younger workers all five studies reported an association (100%), while for older workers five of seven studies (71%) found an association. There are two critical points of divergence that make it difficult to draw conclusions about the strength of the unemployment-mortality relationship across the country clusters. First, in studies from LME countries, unemployment is almost always measured as current 36 unemployment (14 studies), while in CME countries, unemployment is measured in some studies as current unemployment (7 studies) and in others as long-term unemployment (7 studies). In CME countries, fewer studies that use current unemployment find an association with mortality (2 of 6 studies that use a full cohort) compared to studies that use long-term unemployment (3 of 4 that use a full cohort). While long-term unemployment appears to a risk factor for mortality in CME countries, no conclusions on the effect of long-term unemployment on mortality can be drawn from studies of LME countries as this construct is not measured in studies conducted in these countries. How unemployment is measured may reflect important institutional differences that introduce surveillance and ascertainment biases in the measurement of unemployment between CME and LME countries. In section 2.3 it was argued that long-term unemployment is more likely in CME countries and long-term unemployment and its health effects may be of more concern to researchers and policy makers in these countries. Moreover, countries like Sweden and Finland have detailed administrative employment registries that facilitate the tracking of workers‘ unemployment history. The longer duration of unemployment benefits in CMEs also creates both reporting (i.e. individuals are more likely recall being unemployed) and inertial (i.e., individuals are more likely to remain unemployed rather than take a different job or exit the labour force) incentives. In contrast, in LME countries unemployment is seen as more short-term and of a frictional nature and there are fewer incentives for the long-term unemployed to remain unemployed. The second point of divergence between studies conducted in CME and LME countries relates to study design and quality. Most of the LME studies from the United Kingdom (four of the seven) compared standardized mortality rates between unemployed and unemployed controls (Bethune 1996; Moser, Fox, and Jones 1984), but all the CME studies used multivariate estimation techniques making it difficult to compare effect sizes across these studies. Indeed it is challenging to find groups of studies that would be directly comparable. For instance, while the three plant closure studies have similar study designs they implement different statistical methods to create comparable employed controls (Eliason and Storrie 2007; Keefe, Reid, Ormsby, Robson, Purdie, and Baxter 2002; Sullivan and Wachter 2007). And among cohorts drawn from population-representative surveys there is also considerable variation in follow-up period (from one year to 24 years), confounder control (from age-only to a full range of 37 covariates including health status and behaviours), and estimation approach (from logistic regression to parametric duration models). This review of cohort studies from CME and LME countries has found a relationship between unemployment and mortality across both country clusters, but differences in study design, measures and statistical methods prevent the drawing of definitive conclusions about whether this relationship varies by country cluster. Moreover, this literature review did not use a rigorous systematic review methodology including applying consistent search terms across multiple literature databases, exhaustively searching the grey literature and scoring studies using quality criteria. As such, the review may have missed some eligible studies or have drawn different conclusions about the research evidence if quality criteria were applied. Nevertheless, the findings from this review provide some guidance to forming the hypotheses and empirical studies of studies described in Chapters 4 and 5. These findings underscore the need to develop comparable cohorts and measures across studies. They also show that there may be age and gender modification in the relationship between unemployment and mortality particularly in CME countries. The evidence of effect modification is less clear in LME countries as a relationship was found for older workers, while there were too few studies of women to draw conclusions. Further, in CMEs, duration of unemployment appears to matter. There is mixed evidence for an association between current unemployment and mortality, and a more robust association for long-term unemployment. Given the higher likelihood of structural unemployment in CMEs and the potential for permanent exclusion from employment, long-term unemployment may represent a risk to health that is magnified by the CME institutional environment rather than mediated by it. 2.6: Hypotheses for Empirical Studies The purpose of this chapter has been to specify how the institutional environment can affect the health of the unemployed. Using the Varieties of Capitalism framework I have explored how differences in employment and unemployment protection, skill-occupational fit, and health selection out of unemployment may mediate the unemployment and health relationship across CME and LME countries. This leads to the following hypotheses for the empirical studies that are described in Chapters 4 and 5: 1. The association between unemployment and health will be smaller in Germany 38 compared to United States given the higher levels of unemployment and employment protection; Canada will emerge as a middle case. The receipt of unemployment compensation will mediate the effect of unemployment on health within countries. The higher prevalence of the long term unemployed in Germany compared to the LME countries, however, may confound this comparison. 2. There will also be a distinct pattern of effect modification by educational status. The relationship between unemployment and health will be smaller for the minimum skilled and medium skilled in Germany compared to the LME countries, with the minimum skilled in the United States being especially disadvantaged. The effect of unemployment for the high general skilled in the United States and Canada will be smaller compared to those in lower skill categories, but there is no a priori expectation that higher skilled workers in Germany should have a different unemployment-health relationship than those with lower skills. 3. Controlling for health selection will account for some but not all of the relationship between unemployment and health. Further, more of the relationship in Germany will be accounted for by health-selection into unemployment compared to Canada and the United States. 4. The direction of effect modification for men and women is indeterminate, but the ranking across countries will be consistent by gender with the higher associations being in the United States compared to Canada and Germany. 2.7: Concluding Remarks The analytical studies in Chapter 4 and 5 compare unemployment-related health inequalities in Canada and the United States to Germany. The introduction of Canada as a study county offers a contrast to both the Germany and the United States as it represents a middle case between them. Germany and the United States represent ‗pure‘ or archetypical types of institutional and structural variation whether one uses a welfare-regime or varieties of capitalism typology. Canada, on the other hand, while generally included with the United States in these typologies, shares some programme features more commonly associated with European welfare states (e.g., higher levels of employment and unemployment protection, universal access to health care). This enables the comparison of unemployment-related health inequalities in two distinct institutional and cultural contexts and labour markets that have very different approaches to unemployment protection (Germany and the United States), but also within similar institutional contexts where unemployment protection differs (Canada and the United States) and across different 39 institutional contexts with some similarities in unemployment protection (Canada and Germany). Figures and Tables Figure 2.1: Percentage of unemployment insurance beneficiaries to total unemployed for Canada, Germany and the United States, 1976-2006 Sources: Canadian Data: Statistics Canada Cansim II Series V384606, V385120; American Data: US Department of Labor and Bureau of Labor Statistics; German Data: Schneider 2004 Table 5, p 113 40 Figure 2.2: Social protection and predicted skill profiles Employment Protection Low High Unemployment Protection High Low Industry-specific skills Example: Denmark Industry-specific, firm-specific skill mix Example: Germany General skills Example: United States Firm-specific skills Example: Japan (Reproduced from Estevez-Abe, Iversen, and Soskice, 2001 p. 154, Figure 4.1) Figure 2.3: Pathways through which unemployment could influence health 41 Table 2.1: Household income replacement rate scenarios for the unemployed by selected family type for Canada, Germany and the United States Germany Canada United States No children Two children No children Two children No children Two children Single - 50% AW - 100% AW - 150% AW 71% 60% 61% 95% 71% 66% 66% 64% 46% 79% 75% 59% 64% 54% 39% 61% 53% 37% Couple, single earner - 50% AW - 100% AW - 150% AW 85% 60% 60% 91% 75% 70% 81% 66% 48% 85% 77% 60% 72% 55% 38% 66% 55% 39% Couple, dual earner - 50% AW, 67% AW - 100% AW, 67% AW - 150% AW, 67% AW 92% 86% 83% 95% 91% 87% 84% 78% 63% 96% 85% 70% 84% 72% 58% 87% 74% 60% Source: OECD Online Benefit and Wage: Tax-Benefit Calculator (OECD 2009) Notes: 1. Average wages in 2001 for Germany, Canada and the United States were $37,232 CAD, $38,204 EUR and 33,998 USD, respectively. 2. Net replacement rates are derived by dividing net household income in the month after becoming unemployed including unemployment benefits and all other social transfers by net household income in the month prior to becoming unemployment. 3. Germany has a unified tax and benefit structure and the replacement rates are representative across all regions. In Canada and the United States tax rates and benefits vary by province or state and it is not possible to calculate country-wide average rates. The OECD use rates from the province of Ontario and the state of Michigan to be representative of Canada and the United States. 42 Table 2.2: Scope of employment protection regulation Regular contracts Temporary contracts Collective dismissal 1. Degree in flexibility in providing notice to dismiss a worker including the need to notify or seek permission from governmental authorities. 1. Degree of flexibility in a firm’s use of fixed-term contracts. 1. Number of workers that can be dismissed before collective dismissal required come into effect. 2. Amount of time until notice period starts. 2. Permitted maximum number of fixed-term contracts. 2. Degree to which additional notification (e.g., to union or government authorities) is required for collective dismissals. 3.Required length of notice before a worker can be dismissed, based on number of years of tenure. 3. Permitted cumulative duration of fixed-term contracts. 3. Degree to which there are additional notice before for collective dismissal. 4. Amount of required severance pay, based on number of years of tenure. 4. Type of work for which temporary contracts are legal. 4. Other costs to employers including additional training or severance requirements. 5. Definition of justified or unfair dismissal and the degree to which employers must take steps to mitigate the dismissal through training or reassignment. 5. Permitted maximum number of renewals for temporary work arrangements. 6. Length of trial period during which an employee can be dismissed without cause. 6. Permitted cumulative duration of temporary work arrangements. 7. Degree of compensation for unfair dismissal. 8. Right of reinstatement after unfair dismissal. (Revised from OECD, 2004a p.106, Table 2.A1.2.) 43 Table 2.3: Summary the relationship between unemployment and all cause mortality for cohort studies conducted in Coordinated Market Economies Unemploy -ment definition Summary of effect size (None, Low, Medium, High) Adjustments for confounders Full cohort Men Women Younger workers Older workers Sweden (8 studies) Ahs 2006 CU Medium- LUR None-HUR None-HS DEM, SES, HS, UR Gerdtham 2003 CU Medium DEM, SES, HS, Gerdtham 2004 CU None DEM, SES, HS,CON Eliason 2007 PCU Low None High None DEM,SES,HS, CON Nylen 2001 CU, EU High-CU Low-EU Medium-CU Medium-EU DEM, HB, HS, Stefansson 1991 LTU Low High None High Low Age-only Sundquist 1997 LTU Medium None DEM, SES, HS Voss 2004 EU Low Low DEM, SES, HB, HS Finland (6 studies) Blomgren 2007 LTU Medium Medium-HUR Medium-LUR Medium Medium-HUR High-LUR DEM, SES, UR, CON Martikainen 1990 STU, LTU Medium-STU High-LTU DEM, SES, HS Martikainen 1996 LTU High Low-LUR Medium-HUR Medium Medium-LUR None-HUR DEM, SES, HS, UR Martikainen 2007 STU High –LUR Low-HUR None – HS DEM, UR Pensola 2004 STU, LTU High-LTU Medium-STU DEM, SES Saarela 2005 EU High Medium DEM, UR Denmark (2 studies) Iversen 1987 CU Medium Medium Medium DEM,SES Osler 2003 STU, LTU Low-STU High-LTU Medium- STU None- STU SEM, HB, CON The Netherlands Schrijvers 1999 CU None DEM, SES, CON, HS Germany Frijiters 2005 CU None DEM, SES, HS, CON Switzerland Gognalons- Nicolet 1999 EU High DEM, SES, HS 1. ‗CU‘ refers to current unemployment; ‗LTU‘ refers to long-term unemployment; ‗STU‘ refers to short-term unemployment; ‗EU‘ refers to ever unemployed; and ‗PCU‘ refers to unemployment due to a plant closure. 2. ‗None‘ means that there is no statistically significant relationship between unemployment and mortality; ‗Low‘ refers to a statistically significant risk ratio between 1.00 and 1.49; ‗Medium‘ a statistically significant between 1.50 and 1.99; and, ‗High‘ a statistically significant risk ratio greater than 2.0. 3. The adjustments refer to the following sets of covariates: ‗DEM‘ – demographic measures; ‗SES‘ – socioeconomic status measures; ‗HB‘ – health behaviours; ‗HS‘ – health status measures; ‗UR – unemployment rate with ‗HUR‘ and ‗LUR‘ referring to results for high and low unemployment rate time periods or areas; and, ‗CON‘ – contextual measures. 4. A detailed summary, including measures of effect size, can be found in appendix table B1. 44 Table 2.4: Summary the relationship between unemployment and all cause mortality for cohort studies conducted in Liberal Market Economies Unemploy -ment definition Summary of effect size (None, Low, Medium, High) Adjustments for confounders Full cohort Men only Women only Younger workers only Older workers only United States (9 studies) Cubbin 2000 CU High DEM, SES Hayward 1989 CU None DEM, SES, HS Kiuila 2007 CU Medium Medium-HS Low None –HS DEM, SES, HB, HS Lavis 1998b CU High-CU None-LUR None-HUR DEM, SES, UR Rogers 2000 – Chapter 7 CU None DEM, SES, HS Rogers 2000 – Chapter 10 CU Medium DEM, SES, HB, HS Sorlie 1990 CU Medium – Whites High– Blacks None DEM, SES Sorlie 1995 CU None Medium Low DEM,SES Sullivan 2007 PCU Low, Medium High Low-Men None- Women Age, SES, United Kingdom (6 studies) Bethune 1996 CU Low Low Medium Low-Men None- Women Age, SES Gardner 2004 CU, LTU Low DEM, SES, HB, HS Morris 1994 EU Low DEM, SES, HB, HS Moser 1984 CU Low Age, SES Moser 1986 CU Low Age, Region Moser 1987 CU Medium None Age New Zealand (2 studies) Blakely 2002 CU Low None DEM Keefe PCU None DEM 1. ‗CU‘ refers to current unemployment; ‗LTU‘ refers to long-term unemployment; ‗STU‘ refers to short-term unemployment; ‗EU‘ refers to ever unemployed; and ‗PCU‘ refers to unemployment due to a plant closure. 2. ‗None‘ means that there is no statistically significant relationship between unemployment and mortality; ‗Low‘ refers to a statistically significant risk ratio between 1.00 and 1.49; ‗Medium‘ a statistically significant between 1.50 and 1.99; and, ‗High‘ a statistically significant risk ratio greater than 2.0. 3. The adjustments refer to the following sets of covariates: ‗DEM‘ – demographic measures; ‗SES‘ – socioeconomic status measures; ‗HB‘ –health behaviours; ‗HS‘ – health status measures; ‗UR – unemployment rate with ‗HUR‘ and ‗LUR‘ referring to results for high and low unemployment rate time periods or areas; and, ‗CON‘ – contextual measures, 4. A detailed summary, including measures of effect size, can be found in appendix table B2. 45 Chapter 3: Description of Survey Data, Cohort and Variable Development 3.1: Introduction This chapter describes how the cohorts were developed for both analytic studies. An overview of the four data sources (the German Socio Economic Panel, the American Panel Study of Income Dynamics, the Canadian Survey of Income Dynamics, and the Cross National Equivalent File) used in the thesis is first provided, then the development of the German and American cohorts for the unemployment and mortality study, and the German, American and Canadian cohorts for the unemployment and self-reported health study is described. The development of the variables used in both analysis is discussed with a particular focus on the comparability (or lack thereof) of the variables. The chapter concludes with an overall assessment of the comparability of data and variables. One of the challenges in conducting comparative research, particularly individual-level research, is balancing using the best data available within a single data source with conducting analyses that maximize comparability across countries and datasets. This challenge increases with the number of countries, datasets and years used in the research. While there is a rich literature looking at unemployment and health, the heterogeneous nature of this literature makes it difficult to make comparisons across this literature. The review of studies described in Chapter 2 that looked at unemployment and mortality was limited in drawing conclusions about the effect of the relationship across CME and LME countries because of the differences in the study design, definition of unemployment, analytic methods, and inclusion of covariates. This thesis uses three longitudinal household panel surveys from Canada, Germany and the United States as well as a derived cross-national equivalent file based on these and other surveys. These surveys, described in more detail below, are all based on the concept of following households across time and focus on measuring income and labour force dynamics. But there are differences, some by design and some by circumstance that make it challenging to conduct comparative research with these surveys. The American survey, owing to a lack of funding, moved from yearly data collection in 1997 to biennial data collection thereafter. The Canadian 46 survey follows individuals for a maximum of six years, while the German and American surveys follow individuals until death or loss to follow-up. The German survey has higher rates of attrition than both the American and Canadian surveys. Decisions were made with the intent of creating comparable cohorts and variables and in choosing analytic methods that would be appropriate for all three surveys. Creating comparable variables, both in the principal labour force measures and in the covariates, was viewed as most important. Indeed, the comparability of the labour force status measures lies at the crux of this thesis. Where possible the same time period has been used, but this was not always possible because of the change in the PSID to a biennial survey after 1997. Similar models and functional forms of variables were adopted across the surveys, although with the occasional sacrifice of maximizing model fit within a cohort. 31 There were some unique country differences that required the inclusion of specific measures. The reunification of Germany in 1991 and differences in income and labour market outcomes for citizens of the former Democratic Republic of Germany (East Germany) (Nolte, Shkolnikov, and McKee 2000; Nolte and McKee 2004) meant that it was necessary to control for whether a study member was originally from East Germany. Similarly, the established differences in labour market, income and health outcomes between black and white Americans in the United States (Adler and Rehkopf 2008; Kunitz and Pesis-Katz 2005) led to race being an essential demographic covariate in the American cohort, but not for the German or Canadian cohorts. 32 3.2: Description of the Survey Data 3.2.1: The Panel Study of Income Dynamics (PSID)33 The Panel Study of Income Dynamics (Hill 1992) is a longitudinal survey of individuals and families that started in 1968. The PSID was started with the goal of understanding income and labour income dynamics, particularly among low-income families. It started with a population- 31 For example, in the mortality models (Chapter 4) a quadratic specification for age was slightly preferred based on the Akaike Information Criterion (AIC) in the PSID, but a linear specification for age was preferred in the GSOEP. Neither specification affected the coefficients of the labour force status variables so I choose the more parsimonious linear specification. 32 The GSOEP and SLID do not collect information on race/ethnicity in the same way as the PSID. The GSOEP collects information on country of birth, while SLID collects information on ethnicity and language group. 33 Detailed information on the PSID, including downloadable public-use data files, is found at http://psidonline.isr.umich.edu/. 47 representative sample and an additional low income sample. Between 1990 and 1995 a sample of Latino or Hispanic households was also collected. The PSID principally collects information on heads and spouses, but all family members in the original sample of households are followed and new sample households are created when children in the original households become adults and form their own household or when two households are created due to marital dissolution. 34 The survey measures are based on self reports from the principal respondent in the household (usually the head) and proxy interviews for the rest of the household. The survey has grown in size over time from 4,800 families in 1968 to 8,500 families in 1996. In 1997, because of funding constraints, the number of families followed was reduced by about 30% and the survey went from yearly to biennial follow-up, although efforts were made to collect income and labour market information on the non-survey years (PSID 2009). The survey focuses on income and labour force status questions including calendar and retrospective measures, and also includes detailed information on education and training, household assets, and health and activity limitations. Mortality has been ascertained from the beginning of the survey. Self-reported health status was asked of head starting in 1984 and of spouses in 1985. More recently, from 1999 onwards, questions on chronic health conditions have been asked. 3.2.2: The German Socio-economic Panel (GSOEP)35 The German Socio-economic Panel (Haisken-DeNew and Frick 2005), like the PSID, is a longitudinal study that follows households over time with the aim of collecting information on a broad range of economic and social conditions. Measures in the survey are based on self report for all adults in a household or proxy interviews if the respondent is a child. The survey originally was a representative sample of the population of the Federal Republic of Germany (West Germany) starting in 1984 and with the reunification of Germany in 1990 the survey was expanded to include residents of the former Democratic Republic of Germany (East Germany). 34 Men are heads in the PSID for the purposes of following households across time (i.e., households are tracked though headship); women are the heads if single or a lone parent, but become spouses if they marry or live common law. These household following conventions reflect the times when the PSID was created, but are maintained to ensure consistency of follow-up. In contrast, the GSOEP lets respondents self-identify who is the head and spouse for the purposes of follow-up, while in the SLID the head is deemed the individual with the largest labour income or in the case of individuals with equal income, the eldest. . 35 Detailed information on the GSOEP is found at http://www.diw.de/english/sop/. Access to the GSOEP English scientific-use data file is arranged through a research contract with the German Institute for Economic Research (DIW Berlin). The English scientific use data file is a 95% sample designed to perverse confidentiality according to German data protection laws. 48 The GSOEP contains a number of other samples that target specific populations, including a sample of households with heads of Turkish, Greek, Yugoslavian, Spanish or Italian citizenship (Sample B – Foreigners in the FRG) that also started in 1984 and a broader immigrant sample that started in 1994 and 1995 (Sample D – Immigrants). In 1998 the SOEP was supplemented with a refreshment sample (Sample E) in order to maintain the representativeness of the SOEP to the German population. More recently a larger refreshment sample was conducted (Sample F – Innovation) in 2000, and in 2002 a sample of high-income households was conducted (Sample G – High income oversample). 3.2.3: The Survey of Labour and Income Dynamics (SLID)36 The Survey of Labour and Income Dynamics (Statistics Canada 1997) is a longitudinal survey of Canadian households that started in 1994 with the objective of supporting research on family, education, labour and income dynamics in relation to economic well being. Questions on self- reported health status and disability and activity restrictions have been asked since 1996. The SLID is comprised of overlapping longitudinal panels of six years in duration. Unlike the PSID and GSOEP, which follow individuals until death or loss to follow-up the maximum length of follow-up in the SLID is six years. To date, there have been five SLID panels, with panels starting in 1994, 1996, 1999, 2002, and 2005, and each consisting of about 15,000 households and 30,000 individuals. The SLID consists of two interviews every year; a general interview in January covers labour market, personal characteristics and education, and an income-specific interview that occurs in May of every year. Respondents have the option of foregoing the income questionnaire if they provide permission for a linkage to their tax return. Between 50% and 90% of individuals consent to a tax-file linkage (the consent rate increases the longer the person stays in the survey). 3.2.4: The Cross-National Equivalent File (CNEF) The Cross-national Equivalent File (Burkhauser et al. 2000; Burkhauser and Lillard 2005) is a set of data files containing harmonized and equivalent variables derived from contributing country-specific household panel data sets. The GSOEP, the PSID and the SLID, along with the 36 Information on how to access the SLID can be found at http://data.library.ubc.ca/rdc/. 49 British Household Panel Survey (BHPS) are the principal surveys contributing to the CNEF. 37 The CNEF is a collective effort. It is primarily researcher driven and the harmonized variables reflect the research interests of those participating in the project and particular attention has been given to the development of comparable income measures, including pre- and post-tax and transfer household income, individual labour market income and household public transfers. 38 Each country‘s CNEF data files can be merged with the underlying household survey to create a blended data file. Where possible the CNEF variables are used, alhough some harmonized variables specific to this analysis are also derived, as described in section 3.4. 3.3: Derivation of the Study Cohorts This research conducts two sets of analyses, for which two distinct sets of cohorts are derived across the three surveys. The first analysis uses the PSID and GSOEP to examine the relationship between unemployment and mortality for the years 1984 to 2005, while the second uses the PSID, GSOEP and SLID to examine the relationship between unemployment and self-reported health status. In the second analysis, the PSID cohort is followed from 1984 to 1997, the GSOEP cohort from 1994 to 2005, and the SLID cohort from 1996 to 2005. 3.3.1: Mortality Cohorts The mortality cohorts were designed to be as comparable as possible, both in terms of follow-up and in terms of composition across the two surveys. Table 3.1 illustrates the derivation of the German mortality cohort, and Table 3.2 illustrates the derivation of the American mortality cohort. The initial cohort inclusion criteria was defined as heads or spouses aged 18 to 64 (working-aged) at baseline who had at least three years of data prior to death, loss to follow-up or the study end in 2005. Cohort members were required to have at least three years of data so that baseline health and working histories could be established. A direct result of this is that in the first three years of the study there are no deaths and accordingly the first two years of the 37 In the past few years the CNEF datasets have also been developed for the Swiss Household Panel (SHP) and the Household, Income and Labour Dynamics in Australia (HILDA). 38 For some surveys the CNEF variables represent a significant value-added (and time saving) improvement over the measures available in the underlying datasets. For example in the PSID, yearly post-tax and transfer household income is derived by summing of broad range of household income sources, some of which need to be aggregated from monthly reports to the level of the year; state, federal and payroll taxes are estimated by using the NBER Taxsim model (Butricia and Burkhauser 1997). 50 time an individual was in the study was derfined as a baseline period with the third year as the first year of follow-up after which a person is at risk of dying. Entrance into the cohort was dynamic, to reflect the fact that individuals entered the underlying surveys at different times (e.g., the East German cohort) or met the cohort eligibility requirements after the initial baseline year 1984 (e.g., became a head or spouse after 1984) or both. The ‗Latino or Hispanic‘ sample in the PSID was dropped as it was only followed for five years, and thereafter mortality was no longer ascertained. Similarly the ‗Foreigner‘ sample in the GSOEP was also dropped; while this sample had been followed since 1984 there is an unusually low number of deaths compared to the ‗West German‘ sample that started at the same time (3.3% versus 8.5%). For the ‗Foreigner‘ sample it appears that deaths were incompletely ascertained as a higher proportion of individuals were lost to follow-up. This is likely because of repatriation back to their country of origin (see table C1 in the appendices for a description of the number of deaths by the different sample is the GSOEP). The biennial nature of the PSID after 1997 meant that it was not possible to collect complete labour market and health histories after 1997. While some information was available for the non- survey ‗off-years‘ these variables were not always collected;39 more vexing is that there is no information on current unemployment (the most reliable measure of unemployment) for the years 1998, 2000, 2002, and 2004. Accordingly, the baseline years were restricted to 1984 to 1995 (1995 was chosen so that a cohort member would have a minimum of three contiguous years at baseline), with years 1996 to 2005 contributing to follow-up and mortality where years and measures were available. While it is possible to follow individuals in the GSOEP cohort on a yearly basis after 1997, the same cohort restriction was applied to enforce comparability with the PSID cohort. Excluding individuals who would have otherwise met the cohort definition after 1995 led to only a small reduction in the number of deaths (25 deaths in the GSOEP and six in the PSID). The GSOEP 1984-2005 individual file contains 53,918 individuals, of whom 3,088 (5.7%) died. After excluding non-sample individuals and individuals who were never a head or spouse, the 39 For example, unemployment benefit payments are available for 1997 and 1998 from the 1999 survey and for 1999 and 2000 from the 2001, but in the 2003 and 2005 years unemployment insurance payments are aggregated with other public transfer payments for the off years. 51 potentially eligible cohort consisted of 35,050 individuals and 2,688 deaths. Excluding those aged 65 years or over at the first year of baseline decreased the cohort to 30,966 but halved the number of deaths to 1,357 (4.4%). Dropping the foreigner sample (Sample B) and samples with an intake after the last baseline year (Samples E, F and G) and excluding those without three or more years of complete data led to a final cohort of 10,866 heads or spouses and 879 (8.1%) deaths. The PSID 1968-2005 individual file contains 67,271 individuals, of whom 4,917 (7.3%) died. After excluding non sample individuals, the ‗Latino or Hispanic‘ sample, individuals no longer in the survey by 1984 and individuals who were never a head or spouse, the potentially eligible cohort was 14,874 individuals with 1,907 deaths (12.8%). Excluding those aged 65 years or over at the first year of baseline decreased the cohort to 13,605, but almost halved the number of deaths to 977 (7.2%). The additional cohort restrictions led to a final cohort of 9,786 and 876 (9.0%) deaths. 40 Two sub cohorts were also defined in order to account for the effects of health selection. The first sub cohort was restricted to individuals who were employed (defined as employed on the day of the survey) at both years before basline (t-1, and t-2). Individuals who were not working or unemployed in either year were excluded. This exclusion led to a sub cohort of 7,059 individuals and 395 deaths (5.6%) in the German cohort and 6,107 individual and 392 deaths (6.4%) in the American cohort. The second sub cohort was restricted to individuals who reported good or better health status at both baseline years. 41 Individuals who had poor or fair health in either year were excluded. The health exclusion led to a sub cohort of 8,797 individuals and 548 deaths (6.2%) in the German cohort and 7,724 individuals and 445 deaths (5.8%) in the American cohort. 40 The reason the final two cohort restriction led so many fewer individuals and deaths being dropped from the PSID compared to the GSOEP is in part attributable the PSID sample being cut after 1997, while the GSOEP added samples. 41 Heath is defined using heath satisfaction for the German cohort and self-reported health status for the American cohort. See section 3.4 for a description of these measures. 52 3.3.2: Self-reported Health Status Cohorts Many of the decisions around creating the self-reported health status cohorts mirrored the mortality cohorts. Cohort members were required to be heads or spouses aged 18 to 64 at baseline and had to contribute at least three consecutive years of follow-up to be included in the cohort. It was not possible to create cohorts across the three surveys that covered the same years or had a similar number of years of follow-up due to the changes in PSID data collection from annual to biennial in 1997 and the fact that the SLID followed up individuals for a maximum of six years. The change in the PSID data collection from annual to biennial is problematic for the three- country study in that the biennial period of data collection almost completely overlaps with SLID (cohort 1: 1996-2001 and cohort 2: 1999-2004). Furthermore, unlike the mortality analysis in which there is complete information on death across the entire study period, information on self- reported health status in the off-survey years is missing. The short panel time frame of the SLID and the biennial nature of the PSID for overlapping years makes it difficult to study concurrent unemployment and health dynamics across all three study countries. Sacker and colleagues‘ (2007) study of self-rated health trajectories in the United Kingdom and the United States using the PSID and BHPS is one of the few published studies that use longitudinal PSID data after 1997. Their study covers the time period of 1990 to 2001 or eight waves of the PSID. Sacker and colleagues pre-empt the missing years in their analysis by examining eight transitions in the PSID and nine in the BHPS. In this study, however, the missing data are more problematic as the main focus is on how two time-varying variables (unemployment and health status) are related to one another, rather than describing trajectories across time. This required creating study cohorts that span different time periods, 42 but allows the examination of the same dynamics or temporal relationship across years. Accordingly, the American cohort spans the years 1984-1997, the German cohort 1994-2005, and the Canadian cohort 1996-2005. Another difference between the three surveys is that the Canadian cohort only follows individuals for a maximum of six years although new cohorts are brought into the SLID every three years, while the Americans and 42 The Rodriguez (2001) study that used three-years panels of the BHPS, GSOEP and PSID to look at the relationship between unemployment and SRHS also used different study years. 53 Germans cohorts follow individuals until loss to follow-up or to the end of the study period. 43 This means that the Canadian cohort has more individuals than the American and German cohorts but fewer years of observations on those individuals. The GSOEP self-reported health status (SRHS) cohort starts in 1994, the first year of consistent SRHS data. Accordingly, the SRHS cohort is drawn from the later years of the survey (those with start dates after 1995) that were excluded in the mortality cohort. Sample B – the ‗Foreigner‘ sample that started in 1984 is also included, although sensitivity testing excluding the Sample B is also conducted to determine if the inclusion of this sample changes the results. The GSOEP 1984-2006 individual file 44 contains 57,758 individuals and 693,096 person years. Dropping individuals not in the sample or never a head or spouse between 1994 and 2004 decreased the cohort to 29,138 and 188,142 person years. Additional age, follow-up, and missing data exclusions lead to a final cohort of 19,029 individuals and 103,484 person years. The SLID cohort draws on panel one (1996-1998), panel two (1996-2001), panel three (1996- 2004) and panel four (2004-2005); panel five is not included as it starts in 2005 and only contributes one year of data. The SLID CNEF cohort contains 223,809 individuals and 760,396 person years. Dropping individuals from panel five, non sample cohabitants, 45 individuals not meeting the age restrictions and those who were never a head or spouse reduces the cohort by almost two thirds to 77,763 individuals and 334,609 person years. Restricting the cohort to individuals with three or more years of follow-up and those without missing data leads to a final cohort of 65,168 individuals and 217,530 person years of data. The PSID SRHS cohort is almost identical to the PSID mortality cohort in that the years for cohort eligibly overlap. The main difference between the two cohorts is that the person years after 1997 are not included in the SRHS analysis. The PSID SRHS cohort contains 67,271 individuals and 1,278,149 person years. Excluding the ‗Latino or Hispanic‘ sample, non-sample individuals, individuals lost to follow-up, individuals who enter the study after 1997 and those 43 As mentioned in section 3.2.1, the PSID stopped following a portion of the sample because of budgetary restrictions. 44 I used the more recent data release of the SRHS analysis, but as unemployment benefits variables changed in 2006 I still restrict the sample to data from 2005 or earlier. 45 Non-sample individuals are individuals in the household but not part of the sample frame; their information contributes to household and family measures, but otherwise they are not followed they leave the household. 54 never a head or spouse decreases the cohort to 12,779 individuals and 135,388 person years. Further restrictions on age, individuals with fewer than three years of follow-up and individuals with missing data leads to a final cohort of 9,545 individuals and 78,951 person years. Additional baseline restrictions on employment status and health status at baseline yielded an ‗employed only‘ at baseline cohort of 13,958, 46,507 and 6,857 respectively for the German, Canadian and American surveys, and a ‗good health‘ at baseline cohort of 16,603, 57,971, and 8,305 for the German, Canadian and American surveys, respectively. 3.4: Development of the Variables Used in the Studies This section develops the variables used in the following two chapters. The variables come from the CNEF file where there are existing comparable variables, although in a few cases (e.g., education) comparable variables specific to this study that more closely align with the study hypotheses are developed. Where comparable measures have not yet been developed (i.e., the labour force status variables) new cross-national equivalent variabled are created. The study variables are grouped into six categories – health (including dependent and control variables), labour forces status measures, education status, occupation, income measures, and demographic variables. 3.4.1: Health Variables There are four health variables used across the three studies. Death and self-reported health status are the respective dependent variables. Lagged self-reported health status, health satisfaction (GSOEP only) and disability status are used as control variables. 220.127.116.11: Death (G,P) Death is defined as all-cause mortality in any year between the last year of follow-up for an individual and the last-year of follow-up of the study. There were 879 and 876 deaths in the GSOEP and PSID cohorts, respectively (See Tables 3.1 and 3.2 for details). Most deaths occur in the year immediately following an individual‘s last year of follow-up (i.e., an individual can no longer be followed due to death), but some deaths occur after an individual has dropped out of the survey (157 or 17.8% of deaths in the GSOEP and 139 or 15.8 % of deaths in the PSID). All deaths are used to maximize the number of death events. 55 Mortality is ascertained at the time of survey for both the PSID and GSOEP and through re- contact efforts for those who have dropped out of the survey. Deaths in the PSID are periodically validated through the National Death Index and for most deaths, month and cause of death (ICD- 9) are also available in addition to year. For comparability with the GSOEP mortality is restricted to the year of death and all-cause mortality. 18.104.22.168: Self-reported Health Status (G, S, P) A five-category self-reported health status is available from 1984 onwards in the PSID, in 1992 and then 1994 onwards in the GSOEP, and from 1996 onwards in the SLID. This variable spans the categories ‗excellent‘, ‗very good‘, ‗good‘, ‗fair‘ and ‗poor‘ for the PSID and SLID and ‗excellent‘, ‗good‘, ‗satisfactory‘, ‗poor‘, and‘ bad‘ for the GSOEP. Five indicator variables corresponding to the five categories were created and used as a control for pre-existing health status in the mortality study for the PSID cohort. Excellent SRHS is the reference category in the analytic models. Depending on the model and labour forces status measure being examined, SRHS health status is lagged one year (t-1) or two years (t-2). In the self-reported health status study, SRHS is dichotomised as poor or fair versus good, very good or excellent (PF/GVGE) and poor, fair, or good versus very good or excellent (PFG/VGE) and the corresponding lagged version is used as a control. The rationale for this treatment of the SRHS is developed in detail in Chapter 5. 22.214.171.124: Satisfaction with Health (G) Self-reported health status is not available for the period of 1984 to 1991 or in 1993 in the GSOEP and an alternative measure of health – satisfaction with health – is used as the principal control for health status for the German cohort in the mortality study. Satisfaction with health is an 11-category variable derived from a question that asks individuals to rank how satisfied they are with their health. This variable ranges from ‗not satisfied at all‘ (0) to ‗completely satisfied‘ (10). For the years both variables are present the correlation between satisfaction with health and self-reported health status is high (ρ = 0.77); collapsing satisfaction with health into a five-category variable (see Table 3.6) led to a very small decline in the correlation (ρ = 0.74). Using health satisfaction as an 11-category variable compared to a five- 56 category variable does increase model fit, based on the AIC and BIC criterion in preliminary regression models on mortality. Differences in using a 5- or 11-category variable on the estimates of the labour force status variable are negligible and do not change the statistical significance or the interpretation of the results (i.e., the risk ratio of the unemployment variables differ by no more than 3% and there is no consistent pattern in the direction of the change in the risk ratio). For comparability purposes across surveys and to reduce the number of parameters in the statistical models the five-category variable is used. 126.96.36.199: Disability Status (CNEF – G,P,S) GSOEP In the GSOEP disability status is based on two questions that ask whether a person has a legally recognized disability and the degree of disability. The CNEF definition based on these questions is used thatdefines the disabled category as having a legally recognized disability of 30% or more. In sensitivity analysis, individuals with a disability of less than 30% had an odds ratio of close to one compared to no disability in a regression on mortality. PSID The PSID disability status measure is based on a question that asks whether a person has any physical or nervous condition that limits the type or amount of work they can do. SLID The SLID disability status measure is based on a series of questions that ask if the person has any difficulty doing any of the activities of daily living (e.g., hearing, seeing, walking communicating) or if the person has a physical or mental health condition that restricts their activities. 46 In the GSOEP, disability status is not asked in all years. Accordingly, for all three surveys, prior disability status is brought forward from previous years if the question was not asked in that year or if the question was not answered. 46 There are also differences in the derivation of this question prior to 1999 affecting the 1996 to 1998 measures in my cohort. The main difference is that after 1998 more questions were used to measure the activity limitations and disability construct. This tended to increases the number of individuals classified as disabled after 1998. 57 188.8.131.52: In Good Health at Baseline ‗Good health‘ at baseline was defined as reporting good or better self-reported health status in both baseline years for the PSID or a health satisfaction level of five or higher for both years for the GSOEP for the mortality analysis. For the SRHS analysis, in ‗good health‘ at baseline was defined only in the first year of baseline due to the shorter number of years of follow-up in the SLID. These variables were used to exclude individuals in poor health at baseline in some of the models. 3.4.2: Labour Force Status Variables 184.108.40.206: Deriving a Comparable Measure of Unemployment. Deriving a comparable definition of labour force status between these three countries is challenging as unemployment is defined differently within Germany compared to Canada and the United States. The standard definition of unemployment in North America and that used by the International Labor Office (ILO) is that an individual be without a job, looking for work and available to work. 47 While this also applies to Germany, the unemployed must also be registered with the local employment agency and meet certain job search and labour market activation measures. They may work up to 15 hours a week and earn a nominal sum per month (€165 in 2005) without any reduction in benefits; individuals who work a main job eligible for unemployment benefits and a secondary job may also keep earnings of up to €400 from the secondary job (what might be considered underemployment in Canada and the United States) once unemployed. Individuals on maternity or child rearing leave and those performing compulsory military or community service may also be classified as registered unemployed. In contrast, in Canada and the United States any paid work, irrespective of the number of hours or amount earned, leads to a classification of employed. Thus using the native definition of German unemployment would lead to a classification of unemployment for some individuals who would otherwise be classified as employed or out of the labour force in Canada and the United States. 47 There are also minor differences between the ILO, Canadian and American definition of who is considered unemployed. For example, full-time students seeking full-time work and available for work would be considered unemployed in the United States, but are not considered a part of the labour force in Canada. Conversely, a job search consisting only of reading newspaper advertisements would lead to a assignment of not in the labour force in the United States, but unemployed in Canada. When the Canadian unemployment rate is adjusted to the American definition the Canadian unemployment rate decreases by between a half and one percent during the period of 1976 and 1998 (Sorrentino 2000). 58 Moreover, some of the underemployed in Germany receive unemployment benefits and access to labour market activation programmes, which is a key institutional difference between Germany versus Canada and the United States and one that may modify the relationship between these types of working arrangements and health status. Accordingly, a measure of unemployment is developed that corresponds to the ILO and North American concept. The GSOEP has a rich set of questions relating to labour market and other activities, both at the time of the survey and retrospectively. Using these questions and applying a set of decision rules that give precedence to any paid work irrespective of registered unemployment status, and secondly to removing individuals not available for work (e.g., those in community service or on child rearing leave), a set of labour force status measures was derived that is comparable to the labour force status definition in the PSID and the SLID. The PSID has fewer measures on labour market and other activities so the converse – a PSID measure of labour force status that was comparable to the German definition – was not possible. Three comparable labour force status measures – current labour force status, labour force status in the year prior to the survey, and cumulative labour force status – were derived creating three mutually exclusive categories spanning employed, unemployed, and not working (out of the labour force). Current unemployment was also dichotomised into those who reported receiving any unemployment compensation benefits in the survey year and those who did not report receiving any unemployment compensation benefits in the survey year. 48 220.127.116.11: Current Labour Force Status Current labour force status is defined by three variables indicating being employed, unemployed (and laid off in the PSID) and not working at the time of the survey. GSOEP 49 Current labour status in the GSOEP is based on an amalgam of questions that cover employment status (working versus not working), registered unemployment, maternity leave or child rearing 48 Any unemployment compensation is also included with these variables in its own right. 49 The GSOEP provides a derived labour force status measure, but there is a coding error that classifies individuals who are registered unemployed but in marginal unemployment as out of the labour force. This variable was discarded and alternative labour force status measure was derived consistent with the decision rules outlined in the text. 59 leave, military or community service, and other secondary employment (odd job, second job or family job). An individual could report multiple affirmative answers (e.g., they could report ‗yes‘ to working, being registered unemployed, on maternity leave, and having a second job). Individuals were assigned to one of four mutually exclusive labour force categories – working full time, working part-time, unemployed, and not working – using the following hierarchy: - employed full-time which includes full-time employment and on the job vocational training and those on maternity leave or child rearing leave who indicate they are working full-time; - employed part-time which includes part-time employment, second, odd jobs or work in a family business and those on maternity leave or child rearing leave who indicate they are working part-time; - not working which includes those not working, the registered unemployed who are in training or on maternity or child-rearing, those in compulsory military or civilian service, those on maternity or child rearing leave and not working; and, - unemployed which includes the registered unemployed with no indication of any paid work, not on maternity leave, and not in military or community service. A Germany-specific definition of labour force status was also derived that gave precedence to registered unemployment in order to conduct sensitivity analysis on the labour force status definition. Full-time and part-time employment were collapsed for comparability with the PSID and SLID. PSID In the PSID current labour force status is based on one question indicating labour force status at the time of the survey. This question spans: working now, temporarily laid off, looking for work or unemployed, retired, permanently disabled, keeping house, student, other, and don‘t know. Those working are defined as employed and those laid off, looking for work or unemployed are defined as unemployed. All other responses are classified as not working. This question is asked up to four times in a survey year and respondents do not always give consistent answers. For the years 1984-1993 and 2003 onwards a derived variable is provided that resolves these discrepancies across the underlying questions. For the years 1994-2001 (the ‗early release‘ files), no derived employment status variables are provided. To enforce consistency across years the following hierarchy used for the PSID derived variables was also applied to the ‗early release‘ 60 variables: temporarily laid off, working now, looking for work/unemployed, retired, permanently disabled, keeping house, student, other, don‘t know. SLID In the SLID current labour status is based on a question indicating labour force status at the end of the reference year during the January interview (e.g., the respondent‘s labour force status on December 31 st , 2001 would be asked during the January 2002 interview). This question was not asked in 1999 for individuals aged 70 or older and for the years 1996 to 1998 this question referred to the entire year and not just the end of the year. Similar to the PSID, the question spans: working at a job or business or self employed, looking for work, going to school, keeping house, caring for other family members, retired, long-term illness or disability, doing volunteer work, no main activity and other. Those working at a job or business or self-employed are classified as employed. Those looking for work are classified as unemployed, and all others are classified as not working. 18.104.22.168: Labour Force Status in the Year Prior to the Survey Monthly labour force status was derived for each month of follow-up based on labour force status and activity questions pertaining to the year prior to the survey. These variables were then summed for each year and three variables indicating the number of months a person was employed, unemployed, or not working was created for each year. The number of months employed is the omitted reference category and the number of months unemployed and number of months not working are used as continuous variables ranging from zero to twelve. The parameter estimate of the months unemployment variable represents the effect of an additional month of unemployment on the risk of mortality or being in a worse SRHS state, controlling for the number of months of not working. GSOEP In the GSOEP monthly labour force status is based on a series of dichotomous but not mutually exclusive employment and activity questions similar to the current labour force status questions. These questions also changed and generally expanded across study years, necessitating modifications to the ranking over the years. A similar set of decision rules was applied to define 61 a mutually exclusive monthly labour force status for each month. Appendix table C2 outlines this hierarchy as it evolved across the survey years. PSID In the PSID the algorithm to assign monthly labour force status varies across years due to changes in the survey design and availability of variables. All questions are asked separately of heads and spouses. 1984-1992 Survey Years For the years 1984 to 1992 the derivation of monthly labour force status spanning employed, unemployed and not working is outlined in Table 3.7. ‗Missing‘ is assigned as not working for the consistency of generating a complete year history. While this may categorize very few working and unemployed persons as not working it will not have an effect on the working versus unemployed comparison. 1993-2001 Survey Years For the 1993 through 2001interview years, monthly labour force status is defined through a series of underlying questions that change depending on whether a person is currently employed or not working at the time of the survey. Individuals currently employed are asked two sets of questions on whether they were unemployed or out of the labour force for at least one week in a given month. Using these two questions a hierarchy similar to the one outlined in Table 3.8 is then applied. For individuals who are not currently working the derivation is more complex. Monthly labour force status for these individuals is based on a nested series of five questions. Individuals are first asked if they ever worked and when they last worked. Individuals reporting never working or last working prior to the previous year (e.g., before 1992 if the survey year is 1993) are then asked in which months did they look for work in the previous year. These individuals are considered to be out of the labour force except for the months they reported looking for work, which leads to a classification of unemployment for these months. Current non-working individuals who report working in the previous year are asked which months they were unemployed, working or out of the labour force. A hierarchy similar to the one outlined in Table 3.8 is then applied. Unemployment status is missing for unemployed spouses for February 1994 and 1995. For these two months their labour force status for January is carried over. 62 1998, 2000 between Survey Years The years 1998 and 2000 are off wave years and limited information is available. In 1999 and 2001, a series of questions on monthly unemployment and receipt of earning is asked for 1997 and 1999 (t-2), respectively. For these years a modified hierarchy is applied: individuals who report being ‗unemployed‘ are considered unemployed, then individuals who report earnings are considered working, and the rest are considered out of the labour force. There is an error in the unemployment variable for January 1997; for this month unemployment is reported at seven times the rate of adjacent months. Accordingly, this variable is not used and February 1997 labour force values are assigned to January 1997. Monthly labour force status for the t-2 years of 1997 and 1999 are only used in the cumulative labour forces status variable (see below). Monthly labour force status for the t-2 years of 2002 and 2004 is not available. SLID Monthly labour force status is a derived variable based on a series of activity questions by time period that determines whether an individual experienced a jobless spell, was available for work and looked for work at any time in a given month. Individuals in full-time studies are not considered unemployed even if they report being jobless and looking for work (Noreau, Hale, and Giles 1997). If an individual meets the definition of being unemployed in a given month they are considered unemployed for that month, even if they report working or being out of the labour force. If an individual reports no unemployment and worked at least some of the month, they are considered employed for that month; otherwise they are considered not working for that month. 22.214.171.124: Cumulative Labour Force Status (G,P)50 Cumulative labour force status is a variable that represents the percentage of time spent in a given labour force status accumulated across the period an individual is in the study. This 50 The cumulative labour force status measure was only used in the mortality study. Given that the SLID only followed individuals for a maximum of six years, while the GSOEP and SLID maximum follow-up was more than double that for the SRHS study, it was deemed that there was insufficient comparability for the cumulative labour status variable across the three surveys to proceed with the analysis. 63 measure is motivated by a study of the relationship between cumulative psychosocial and job characteristics and mortality in the PSID (Amick, McDonough, Chang, Roger, Pieper, and Duncan 2002), in which the psychosocial and physical aspects of job characteristics were attributed to a specific occupation through a job-exposure matrix and accumulated yearly. For this derivation, labour force status is accumulated monthly. Number of months in a labour force status is accumulated across years and then divided by the total number of months followed to get the proportion of observed time spent in each labour force status. For individuals with missing variables or years (e.g., 2002 and 2004 for the PSID), the proportion of time spent in each labour force status is adjusted so that it always sums to 100 percent. Table 3.8 provides a stylized example of the dynamics of the cumulative labour force status variable. Labour force status accumulates from baseline year t-2 onwards meaning that every individual has accumulated three years of labour force status prior to becoming at risk of dying. This measure is more sensitive to large changes in labour force dynamics earlier in follow-up compared to later in follow-up (e.g., 12 months of unemployment in years three of follow-up counts as 33% of lifetime spent in unemployment while 12 months of unemployment in year 10 would count as 10% of lifetime spent in unemployment). Because individuals are not observed for their entire adult life and only up to 22 years this has the potential to introduce bias in that there is the potential for more variation or change earlier in the follow-up period even though the individual may have already been in the labour force for some time. Most individuals who die, do not die in the early years of follow-up. The median number of years followed for those who die is 11 years in the GSOEP and 12 years in the PSID; the likelihood of dying also increases with years followed so large variations in percent of time unemployed in the early years would likely introduce a conservative bias (a bias toward the null) rather than magnify a relationship between percent of observed lifetime unemployed and mortality. Cumulative percent of lifetime employed is the omitted reference category and the percent of lifetime unemployed and percent of lifetime not working are used as continuous variables, ranging from zero to one hundred. 64 126.96.36.199: Current Unemployment and Unemployment Compensation An interaction variable between current unemployment and receipt of any unemployment compensation (see section 188.8.131.52 below) was also created. Current unemployment was dichotomized into the unemployed who reported receiving unemployment benefits and the unemployed who did not report receiving unemployment benefits. In sensitivity testing an unemployed variable that indicates whether a person had reported any months of unemployment in the year prior to survey was also created. This variable was then dichotomised by whether an individual had reported receiving unemployment benefits or whether they had not. Current labour force status and the receipt of unemployment benefits were harmonized so the variables referred to the same calendar year. 51 184.108.40.206: Working at Baseline Individuals had to report a current labour force status of employed for both years to be considered working at baseline, other labour force status combinations over the two years at baseline were assigned a value of not working at baseline. This variable is primarily used to exclude individuals who were not employed in both baseline years in some of the models. 3.4.3: Educational Status Variables on education status are derived based on a modified version of the Comparative Analysis of Social Mobility in Industrial Nations (CASMIN) classification of education. 52 The CASMIN educational classification distinguishes two different education dimensions, one based on hierarchy (length, quality and value of education) and the other based on skill orientation (vocational versus general) (Brauns, Scherer, and Steinmann 2003). The application of the CASMIN educational classification to the educational variables across these three surveys allows the creation of comparable education variables that distinguish between skill type and level of education within and between study countries. The GSOEP has a derived CASMIN educational variable, while a CASMIN equivalent variable can be derived in the PSID based on years of 51 In the GSOEP and PSID the income variables refer to the previous calendar year and not the year corresponding to the current survey year. Accordingly, income variables from the following survey year are brought back a year. 52 Initially an educational classification system based on the International Standard Classification of Education (ISCED) was considered. The CASMIN classification is preferred to the ISCED classification as the CASMIN classification makes the distinction between vocational and general training whereas the ISCED classification does not. 65 education and degree type and in the SLID based on terminal degree type and whether the terminal degree would be considered primarily a general-skilled degree (e.g., high school completion) or a specific-skilled degree (e.g., business or trade school diploma or certificate). The modified CASMIN classification developed by Kerckhoff and colleagues (2002) refines the original CASMIN classification as applied to the United States and from this a skill level (minimal, medium and high) and skill type (specific or general) classification is developed (Table 3.9) and applied across all three study cohorts. 53 There are marked differences in educational classification by study country. For Germany 66% of the cohort have specific-skilled qualifications compared to only 18% of the American cohort. Conversely 60% of the American cohort has a general skilled qualification compared to only 21.1% of the German cohort. The American and Canadian cohorts have a higher proportion of both the minimum skilled (or the inadequately educated) and of those with high general skill qualifications (university degrees). The Canadian cohort is similar to the American cohort except that there are a greater number of individuals with intermediate specific-skilled qualifications, although this may be due to differences in the questionnaire. 54 American and Canadian cohort members that have specific-skilled qualifications typically also have a general-skilled degree (high school or GED equivalency), while skill type for German cohort members tend to be either a general-skilled or specific-skilled degree. Due to the low number of individuals who have specific skill classifications (United States and Canada) or intermediate level general skilled qualification (Germany), education is instead classified by CASMIN level (minimal, medium, or high) in most of the comparative and education-stratified analyses 55 . In sensitivity analyses the skill categorization (minimal, specific, or general) is used for the German cohort, while for the Canadian cohort there are sufficient numbers to distinguish between the medium general skilled and the medium specific skilled. 53 The detailed application of the CASMIN classification as applied to the mortality study cohorts can be found in Appendix Table C3. 54 The SLID asks educational attainment every year, whereas the PSID only asks educational status periodically. 55 Table 3.9 and Appendix Table C3 denote how the cohort was classified into specific or general skills and by minimal, medium, or high educational level using the CASMIN rubric for the SRHS and mortality studies, respectively. 66 3.4.4: Occupation (G,P,S) It was challenging to create a comparable construct for occupation across the three surveys because of the heterogeneity in classification standards. Each of the three surveys uses different occupational classification schemes with the GSOEP using the International Standard Classification of Occupations 1988 (ISCO88), the PSID using the United States 1970 Census occupational classification system until 2001 and the 2000 Census classification thereafter, and the SLID using the Canadian 2000 National Occupational Classification system (NOCS). While occupational classification cross-walks have been developed between United Kingdom‘s Standard Occupational Classification system and an early version of United States‘ Standard Occupational Classification system, 56 there are no validated cross-walks between Statistics Canada‘s NOCS, the American Census classification systems and the ISCO88. Moreover, the lag in time between the American 1970 and 2000 Census classification systems presents challenges in harmonizing occupation across these years in the PSID. While the CNEF has a cross-national equivalent variable on occupational for the PSID and GSOEP, there is no cross-national equivalent variable for the SLID. Rather than attempt to code the SLID to a post hoc occupational classification, all three surveys were coded to the ISCO88 standard for the SRHS analysis. This was because the ISCO88 groups occupation is based on skill level and skill specialisation (Hoffman 1999), which is consistent with the conceptualization of skill level being a central moderating variable between unemployment and health. While it was perferable to code the mortality analysis to the ISCO88 standard, coding the PSID occupational categories across both the 1970 and 2000 Census standards proved prohibitive, so the CNEF equivalent variable was used for this analysis. The CNEF occupational variable codes the country-specific occupational classification (the 1988 – International Standard Classification of Occupations for the GSOEP and the 1970 United States Census classification for the PSID) into 100 occupational categories (e.g., chemist, office manager, janitor, farm hand, painter, machine fitter). From these a set of six occupational variables were created: no occupation (applied if an individual was out of the labour force at baseline); professional and technical 56 An extensive set of occupational cross-walks have been developed through Cambridge Social Interaction and Stratification initiative (see http://www.camsis.stir.ac.uk/index.html) that has sought to harmonize country-specific occupational scales to provide comparable occupational scales based on social stratification. At the time of this writing, cross-walks for the Canadian NOCS and US 1970 Census occupational classification system have not been included. 67 occupations; business and sales occupations; service occupations; agriculture and forestry and mining occupations; and, manufacturing occupations. The CNEF-based classification, however, is more akin to an industry classification with the grouping based upon production of goods or services rather than skills and tasks performed. 57 The ISCO88 is a high-level occupational classification spanning ten categories (Table 3.10, column 2), and from these a set of nine occupational categories were created: no occupation; managers; professionals and technicians; clerical, sales and services; skilled trades, plant and equipment machinery operators; agriculture, forestry and fishing; and, labourers (Table 3.10). Professionals and technicians and associate professionals were grouped together as it was not possible to distinguish between these categories across the three surveys (i.e., the high-level SLID and PSID occupational categories did not map well to professionals or technicians separately). Armed forces (a very small group in the PSID and GSOEP, and none in the SLID) were put in the same category as other security and protective services. Individuals not working at baseline do not have an occupation, but individuals with a pre-existing occupation who are out the labour force after baseline have their occupation carried forward. Professional and technical occupations is the reference category for the empirical models. 3.4.5: Income and Transfers58 Most income variables for the GSOEP, PSID and SLID are taken from the CNEF, but the PSID and SLID unemployment compensation are derived directly from the underlying surveys. All income variables across years are adjusted to current 2005 Euros (GSOEP), United States Dollars (PSID) or Canadian Dollars (SLID) using respective consumer price indices. 220.127.116.11: Post Tax and Transfer Household Income (CNEF – G,P,S) Post tax and transfer household income represents the sum of all income sources (labour, asset income, private and public transfers) minus reported taxes paid for all individuals in a household. 57 I attempted to create occupational cross-walks between both PSID occupational classification systems and the ISCO-88 standard using the finest grained level occupation available in the PSID (the third digit of the Census Occupational Classification system) but obtained different cross-tabulation across the two classifications when mapping them to the ISCO-88. In order to successfully code to the ISCO-88 across survey years I would need to code directly from the underlying occupational titles. 58 Most income and transfer variables are not used in the analytic models, but are used to describe income dynamics by labour and educational status in the descriptive tables. 68 In the GSOEP and SLID this is derived directly from the sum of the income and tax questions. The PSID variable is derived similarly except taxes paid is estimated using the National Bureau of Economic Research (NBER) TAXSIM algorithm that estimates the tax burden for each member in the household (Butricia and Burkhauser 1997). In the GSOEP all income variables are restricted to a positive range, while in the PSID and SLID negative values (corresponding to business or investment losses) are allowed. For comparability and estimation purposes all negative income values in the PSID and SLID are recoded to zero. To capture the non-linear effect of income on health the log of income is used in the analytic models. 18.104.22.168: Individual Labour Income (CNEF – G,P,S) Individual labour income represents the pre-tax wage and salary from all employment, including self-employment. 22.214.171.124: Unemployment Compensation (G,P,S) GSOEP Unemployment compensation is a derived variable taken from the GSOEP-CNEF file and is the sum of all unemployment-related transfers including unemployment assistance, unemployment benefits and subsistence allowance. These variables are collected at the monthly level and then summed to create a yearly total. PSID Unemployment compensation is not available in the PSID-CNEF file and was derived directly from the underlying yearly family files. Unemployment compensation is calculated based on 14 variables comprising the amount of unemployment benefits received, a time unit variable (bi- weekly, month, or year), and 12 single-month variables indicating whether unemployment benefits were received in a given month. The month variables are summed across the calendar year to give the number of months that unemployment benefits were received. Annual totals are derived by applying the following algorithm: - if time unit equals year, then unemployment compensation equals the reported amount; 69 - if time unit equals month, then unemployment compensation equals the reported amount times the number of months unemployed; - if time unit equals biweekly, then unemployment compensation equals the reported amount times 2.167 times the number of months unemployed; or, - if time unit equals weekly, then unemployment compensation equals the reported amount times 4.333 times the number of months unemployed. 59 For the years 1984 to 1993 and 2005, a derived variable based on the above algorithm has been provided in the respective family data files for heads and wives (this variable is only available from 1985 onwards for wives) corresponding to the tax years 1983 to 1991 and 2004 (Survey Research Centre 1998). For the years 1984 to 1992 these variables went through extensive consistency checks; implausible and missing values were corrected or assigned using hand coding procedures (Hill 1992), while for years 1993 and 2005 the variable was cleaned and missing values were assigned using computer coding and statistical imputation. For the years 1994 to 2003 derived and imputed variables are not available in the family data files. For these years a comparable unemployment benefit variable is created based on the underlying 28 variables corresponding to head‘s and spouse‘s unemployment experience. For the years 1984 to 1992 only the derived variables are available in the family data file, however for 1993 and 2005, both the derived variable and the underlying component variables are provided. For these two years the unemployment benefit variable was recalulated from the underlying 14 variables and compared this with the derived variable in the family data files to ensure that the calculation was done consistently across years. SLID Unemployment compensation in the SLID is taken directly from Line 119 (Employment Insurance and Other Benefits) of the respondent‘s income tax return for the previous year for those respondents who provided permission for a tax-filer linkage. For those individuals who did not provide permission, the amount of unemployment compensation received in the previous year is asked in the May income interview. Employment benefits can also refer to maternity 59 The biweekly and weekly multipliers of 2.167 and 4.333 are scaled up to the biweekly and weekly amounts to reflect the duration of an average month. 70 leave and seasonal benefits for workers in the fishing industry that would not be included as unemployment benefits in the GSOEP or PSID. 126.96.36.199: All Public Transfers (CNEF – G,P,S) All public transfers represent the sum all of public transfers including welfare, social assistance, unemployment compensation , workers compensation, food stamps (PSID only), child benefits, and maternity leave (GSOEP only). 3.4.6: Demographics 188.8.131.52: Gender (CNEF – G,P,S) Gender is a dichotomous variable with women as the reference category as men are hypothesized to have a higher risk of mortality. 184.108.40.206: Age (G,P,S) Age is derived by subtracting year of the survey from the year of birth. Age at survey is not used as the timing of the survey varies each year for the GSOEP and PSID and so an individual does not necessarily age one full calendar year between each survey. Age is the principal measure of time in the mortality analysis (see the methods section in Chapter 4 for more detail) so it was imperative that individuals age in discrete and uniform increments, even if a full calendar year does not elapse between each survey wave. Age is defined as a continuous variable and polynomials of age were also considered to capture any non-linear relationship between age and the health dependent variables. 220.127.116.11: Relationship to Head (CNEF – G,P,S) All cohort members in the three surveys are either a head or spouse. In the GSOEP, a head is self-identified and can be either a man or a woman irrespective of the gender of the other partner. In the PSID the head is always a man unless it a women-only household. Accordingly the relationship-to-head variable is not included in the the men-only models for the PSID cohort. In the SLID, the household head is defined as the individual with the highest greatest individual income but for comparability purposes with the PSID this has been revised to correspond to the PSID definition of head. Head is the reference category. 71 18.104.22.168: Martial Status (CNEF- G,P,S) Marital Status is a set of four variables indicating whether a person was married or living with a partner, single, divorced or separated, or widowed. Married is the reference category. 22.214.171.124: Household size (CNEF - G,P,S) Household size is a continuous variable indicating the number of people in the household. 126.96.36.199: Number of Children (CNEF - G,P,S) Number of children is a continuous variable indicating the number of individuals under the age of 18 in the household. 188.8.131.52: East German (G) East German is a variable indicating whether an individual was drawn from the 1991 sample (sample C) covering individuals from the former German Democratic Republic (GDR). 184.108.40.206: Immigrant (G) Immigrant is a variable indicating if an individual immigrated to Germany after 1948. In general immigrants tend to be from Eastern Europe (e.g., Poland, Romania, Russia), the countries comprising the former Republic of Yugoslavia and Turkey. Immigrant is included as a control variable in the German cohort as previous research has indicated that immigrants to Germany tend to have different labour market experiences and health dynamics than German-born individuals (Elkeles and Seifert 1996). 220.127.116.11: Race (P) Race is a set of three variables indicating whether an individual is white, black or of other ethnicity. Race is used to capture the health and labour market effects of discrimination and segregation in the United States. Race is not conceptualized as a biological construct, but rather as a marker of shared experiences and identities related to the social construction of historic racialized categories. Accordingly, race can be viewed as marker for radicalized identity (Veenstra, 2009). Race other than white or black are not distinguished due small sample sizes. All individuals are assigned a consistent racial indicator across time using the following ranking: black and African American, other (Asian, Pacific Islander, American Indian), or white. Race is 72 generally asked only of new heads (a former spouse can be a new head) and of new spouses and periodically of the whole sample. For new heads and spouses, ethnicity is asked up to four times in the survey. Any answer of ‗Black‘ or ‗African American‘ leads to an assignment of black or African American over other or white. Similarly any answer of ‗Other‘ leads to an assignment of other over white. For a small number of respondents race does change across years although this is usually from ‗White‘ to ‗Other‘, as the PSID expanded the number of race categories in the latter years of the survey. The same ranking algorithm is applied across years to create a consistent, time invariant race variable. In the SLID, there are a series of questions that asks about mother tongue, country of origin and visible minority status. Deriving a similar variable to the PSID leads to only a small number of individuals (less than 5%) being assigned an race other than white. Accordingly, race variables are not included in the SLID models. 18.104.22.168: Oversample Indicator (P) Oversample is a variable indicating if an individual was drawn from the survey of economic opportunity (SEO) sample of the PSID. The SEO sample is a sample of low-income, predominately black families. While income, race and other variables that would characterize this sample are included as covariates in the models, the oversample indicator is included to capture any residual fixed effects between the two samples not accounted for by other demographic and income variables. 22.214.171.124: Geography (CNEF – G,S,P) Geography was defined as residence in a state (Germany), census-division (United States) or province (Canada) in a given year. GSOEP Geography is defined as the level of the 16 Länder (States): eleven states from West Germany (Baden-Württenberg, Bayern, Berlin, Bremen, Hamburg, Hessen, Niedersachsen, Nordrhein- Westfalen, Rheinland-Saarland, Schleswig-Holstein) and five states from the East Germany (Brandenburg, Mecklenburg-Vorpommern, Sachsen, Sachsen-Anhalt, Thuringen). PSID 73 In the PSID, states are not uniformly represented in the sample with some small states (e.g., Delaware, North Dakota, Montana, and Vermont) having very few study subjects. Inclusion of state as the geography indicator variable led to estimation problems in some model specifications given the small number of observations for some states. 60 Accordingly, states are grouped into nine census divisions (New England, Middle Atlantic, East North Central, West North Central, South Atlantic, East South Central, West South Central, Mountain and Pacific) and a residual category indicating state unknown. The census division – state correspondence is: - New England: Connecticut, Maine, Massachusetts, New Hampshire, Rhode Island, and Vermont; - Middle Atlantic: New Jersey, New York and Pennsylvania; - East North Central: Illinois, Indiana, Michigan, Ohio, Wisconsin; - West North Central: Iowa, Kansas, Minnesota, Missouri, Nebraska, North Dakota, South Dakota; - South Atlantic: Delaware, District of Columbia, Florida, Georgia, Maryland, North Carolina, South Carolina, Virginia, West Virginia; - East South Central: Alabama, Kentucky, Mississippi, Tennessee; - West South Central: Arkansas, Louisiana, Oklahoma, Texas; - Mountain: Arizona, Colorado, Idaho, Montana, Nevada, New Mexico, Utah, Wyoming; and, - Pacific: California, Oregon, Washington, Alaska, Hawaii. SLID Geography is defined as the ten Canadian provinces (Newfoundland, Prince Edward Island, Nova Scotia, New Brunswick, Québec, Ontario, Manitoba, Saskatchewan, Alberta, and British Columbia). 60 Including a set of 51 state indicator variables (the 50 states plus District of Columbia and state unknown minus California as a reference state) led to some observations being dropped as the state indicator variables predicted the outcome perfectly especially in the stratified models. For example, in the state of Hawaii there were no high-skilled individuals who reported being in poor or fair self-reported health status. 74 126.96.36.199: Year Survey year (e.g., 1984 though 2005) was included as an indicator variable to control for time effects in the SRHS analysis. Survey year was not included in the mortality analysis as the inclusion of age, year and the exposure offset variable (see section 4.3.2 for a discussion of the offset variables) led to collinearity. 3.5: Assessment of Data Quality and Comparability of Study Cohorts This chapter has described the survey data sets, the development of the study cohorts and derivation of the study variables. To the extent possible comparable cohorts and harmonized variables have been developed drawing on the existing comparable variables in the cross- national equivalent file and derivations specific to this study. However there are differences in the cohorts and variables both within and across surveys that may affect the comparability, reliability and validity of the survey measures or otherwise introduce bias into this study. Some of the differences are intentional. The inclusion of race as a confounder and stratification variable in the PSID is imperative not only because of the longstanding health and social stratification by race in the United States, but also because the PSID deliberately oversampled blacks. Not accounting for race in the PSID would reduce the comparability of the results to countries where there is a different legacy of racial and ethnic segregation. A similar argument can also be made for including an East German variable in the GSOEP analysis. Some of the differences are unavoidable. The PSID SRHS cohort spans the decade prior to the GSOEP and SLID SRHS cohorts because of changes in the PSID survey data collection. Unless the relationship between unemployment and health changed within the United States from the 1980s to the 1990s using an earlier American sample should not weaken the comparisons with Germany and Canada. The review of unemployment and mortality from Chapter 2 would suggest that the relationship between unemployment and health has remained consistent across the 1980s and 1990s. Moreover, the institutional arrangements pertaining to unemployment and employment protection did not change in the United States across these decades (unlike Canada, which saw significant retrenchment in the benefit available to the unemployed). 75 The shorter follow-up period of the SLID also presents challenges to the comparability of the study cohorts. The differences in follow-up across the three cohorts has the potential to affect the health dynamics in the models in that the longer a person is the study the more likely a change in health status will be observed. Further a shorter follow-up places more importance on baseline and initial conditions compared to events that occur during follow-up. Thus the SLID cohort is inherently a less dynamic cohort than the longer GSOEP and PSID cohorts. Accordingly, in the analytic chapters an exposure offset of the number of years in the study to account for difference in follow-up is included. As with any longitudinal survey attrition, differential loss to follow-up, and ongoing representativeness are main concerns. Hill (1992) summarizes a number of assessments of representativeness and validity of the PSID during its early years and concludes that the PSID contains valid survey measures and does not have substantial non-response biases. The effect of attrition on the representativeness of the PSID has been well studied (Fitzgerald, Gottschalk, and Moffitt 1998; Lillard and Panis 1998; Zabel 1998). Fitzgerald, for example, reports 50% attrition in the original households from 1968 to 1989 and that attrition is more likely to occur among respondents of lower socio-economic status and those with unstable earnings and labour market histories. He finds that this does not significantly affect regression results for earnings, marital status, and welfare participation when compared with United States Current Population Survey. Poor health status was found to be a predictor of non-response in the BHPS and the ECHP, but health non-response did not substantively affect regression estimates of lagged health, income and education on self-reported health status across balanced or unbalanced models and a model corrected for non-response using inverse probability weights (Jones, Koolman, and Rice 2006). 61 Section 3.4.2 detailed the development of comparable labour force status variables, and while care was taken define a similar measure of unemployment across the three surveys there are undoubtedly still differences in who is considered unemployed and who is not across the three surveys. Even within a single country, how labour forces status was measured changed across years (e.g., appendix table C3 illustrates how the number of possible labour force statuses grew over time in the German cohort). 61 Inverse probability weighting corrects for non-response bias by weighting respondents by the inverse of the likelihood remaining a respondent such that individuals likely to drop but remain in the sample are given a higher weight to account for similar individuals who dropped out (Jones et al. 2007). 76 Retrospective measures of income and labour market activity have been found to be less accurate than current measures (Bound, Brown, and Mathiowetz 2000; Mathiowetz and Duncan 1988; Poterba and Summers 1995). A comparison of the SLID estimates of unemployment to estimates derived from Canada Labour Force Survey found that the SLID underreported the unemployment possibily due to the errors in recalling short spells of unemployment (Noreau, Hale, and Giles 1997). Jurges (2007) found in the GSOEP that 20% of current unemployment was not reflected in the retrospective unemployment measure collected the year following and that errors were more likely for individuals with weak labour force attachment and specifically women with children from West Germany, but not women from East Germany. Similarly, Jacobs (2002) studied the concordance between current and retrospective unemployment measures in the BHPS and found errors in retrospective recall rates in both men and women, with women being two to three more likely to not recall retrospective episodes of unemployment compared to men. Instead women tended to attribute the previous unemployment episode as being involved in family care. The agreement between current and retrospective unemployment within the three surveys using the SRHS cohorts was investigated to acertain whether there were difference in the recall errors across the three surveys. In the GSOEP, about 11% of current episodes of unemployment are not captured in next year‘s retrospective monthly measure of unemployment (10% for males and 12% for females), while in the PSID almost 50% of current unemployment episodes were not captured in the following retrospective monthly measure of unemployment (56% for women and 42%). In both surveys the lack of concordance is largely explained by whether an individual also reported receiving unemployment compensation. In the GSOEP the lack of concordance for those not receiving unemployment compensation was 42% and for those receiving unemployment compensation it was 2%. Similarly in the PSID, the lack of concordance for those not receiving unemployment compensation was 56% and for those receiving compensation it was 19%. There is more consistency among these measures in the SLID. About 23% of current episodes of unemployment in the SLID are not captured in the retrospective monthly measure of unemployment and there was no gender difference in the agreement rate. Concordance between current and retrospective measures of unemployment also did not vary by receipt of 77 unemployment (21% for those receiving unemployment compensation and 24% for those not receiving unemployment compensation). There was also a higher portion of individuals in the SLID who reported receiving unemployment benefits, but did not report ever being unemployed in the same survey year compared to the other surveys. This can be explained, in part, by the provision of employment benefits (i.e., maternity or paternity benefits and seasonal payments to fishers) through the same federal benefit system as unemployment benefits. The greater disagreement between current and retrospective unemployment measures in the PSID compared to the GSOEP may be explained by the higher overall level of unemployment compensation in Germany as invididuals who receive unemployed compensation are more likely to recall a previous episode of unemployment. Further, to be considered unemployed in Germany a worker is also required to register at the municipal employment office, which may also decrease recall bias. That the same recall pattern is not found in the SLID may be due to the SLID collecting labour market information in January for the prior year, while the receipt of unemployment compensation comes directly from tax filer records for many respondents. 3.6: Concluding Remarks The creation of comparable constructs and data lies at the heart of comparative research. This chapter has outlined the strengths of the constituent panel surveys used in this study, namely the ability to follow similar working-age cohorts with a consistent set of health and labour market variables over time. This represents a step forward from previous comparative health research which has relied on ecological or cross-sectional study designs. There are, however, limitations to the data and study design. The study cohorts were not designed to be comparable, ex ante, leading to some irresolvable differences in cohort design and study variables. Differences in survey design, variable construction and measurement error may also be reflective of institutional differences in their own right. While the data, cohort and measures used in this study allow for powerful analyses of how contextual and institutional factors can influence the unemployment and health relationship, care must also be taken to account for differences across the three studies that could bias the comparability of the results. 78 Tables Table 3.1: Derivation of GSOEP mortality cohort (1984-2005) Individuals (Person Years) Deaths (%) SOEP Cohort 1984-2005 (95% GSOEP public use sample) 53,918 3,088 (5.7%)1 - Drop individuals not eligible for an interview or who never had a successful interview -11,904 - 99 (0.8%) - Drop individuals who were never Head or Partner between 1984-20052 - 6,964 - 303 (4.3%) Individuals with at least one year of complete data as Head or Spouse (1984- 2005) 35,050 (304,804 PY) 2,686 (7.7%) -Drop individuals who were 65 or older at baseline (t-2) -4,084 (23,118 PY) - 1,329 (32.4%) -Drop person years in which individuals are not yet sample Heads or Spouses or are sample Heads or Spouses but are in an institution or otherwise not followable3 - 0 (19,492 PY) - 0 (NA) Heads and Spouses with at least one year of complete data (1984-2005) and between the ages of 18 and 64 at baseline 30,966 (248,675 PY) 1,357 (4.4%) - Drop Samples B, F & G4 (see Table C1 in appendices) - 14,066 (76,841 PY) 214 (1.5%) - Drop 1990 data from Sample C5 (East German cohort) - 153 (3,192 PY) 26 (17%) Heads and Spouses with at least one year of complete data (1984-2005) and between the ages of 18 and 64 at baseline (t-2) in retained sample cohorts 16,747 (168,642 PY) 1,117 (6.8%) - Drop individuals with less than three waves of follow-up -2,901 (4,101 PY) -213 (7.3%) Heads and Spouses with three or more years of complete data (1984-2005) and between the ages of 18 and 64 at baseline (t-2) in retained sample cohorts 13,846 (164,541 PY) 904 (6.5%) - Drop individuals whose baseline year (t-2) is after 1995 (this drops Sample E as well as later entrants from other retained samples) -2,980 (18,809 PY) -25 (0.8%) Heads and Spouses with three or more years of complete data (1984-1997) and between the ages of 18 and 64 at baseline (t-2) in retained sample cohorts 10,866 (145,732 PY) 879 (8.1%) Subsamples derived from baseline characteristics (these are not mutually exclusive) - Drop individuals who were not employed at baseline (t-1 or t-2) - 3,807 (50,555 PY) -484 (12.7%) Cohort members employed for both years at baseline (t-1 and t-2) 7,059 (95,177 PY) 395 (5.6%) - Drop individual who report poor or bad health satisfaction at baseline (t-1 or t-2) -2,069 (27,015 PY) -331 (16.0%) Cohort members with satisfactory health satisfaction or better for both years at baseline (t-1 and t-2) 8,797 (118,717 PY) 548 (6.2%) Notes: 1 The percentages in the Deaths column represent the percentage of deaths of the total number of individuals in a given row. 2 These individuals are mostly children. 3 In contrast to the PSID complete information is collected on all adult household members. The majority of these individuals are adult children who have yet to form their own households. This group is dropped for comparability with the PSID cohort on the basis that adults who never been a head or spouse likely have different employment experiences than adults who have been a head or spouse. 4 Sample B is dropped as deaths are incompletely ascertained in the foreigner sample due to individuals leaving Germany and returning to their country of origin once leaving the workforce. Samples E, F & G are dropped due to the short follow-up period. See Table C1 in the appendices for a more detailed breakdown. 5 The 1990 data from the East German cohort was dropped as prior year income and work history data was not collected. 79 Table 3.2: Derivation of American mortality cohort (1984-2005) Individuals (Person Years) Deaths (%) PSID Cohort 1968-2005 67,271 4,917 (7.3%)1 - Drop Latino/Hispanic sample (1990-1995)2 -10,607 - 189 (1.8%) - Drop non-sample individuals3 -20,725 - 817 (3.9%) - Drop individuals lost to follow-up prior to 19844 - 7,470 -1,743 (23.3%) - Drop individuals who were never Head or Spouse between 1984- 20055 -13,595 - 261 (1.9%) Individuals with at least one year of complete data as Head or Spouse (1984- 2005) 14,874 (191,931 PY) 1907 (12.8%) -Drop individuals who were 65 or older at baseline -1,269 (11,577 PY) - 930 (73.3%) -Drop person years in which individuals are not yet sample Heads or Spouses or are sample Heads or Spouses but are in an institution or otherwise not followable6 - 0 (45,384 PY) - 0 (NA) Heads and Spouses with at least one year of complete data (1984-2005) and between the ages of 18 and 64 at baseline 13,605 (134,970 PY) 977 (7.2%) Drop individuals with less than three waves of follow-up -1,963 (2,868 PY) -95 (4.8%) Heads and Spouses with three or more years of complete data (1984-2005) and between the ages of 18 and 64 at baseline 11,642 (132,102 PY) 882 (7.6%) - Drop individuals whose baseline year (t-2) is after 1995 -1,856 (7,760) -6 (0.3%) Heads and Spouses with three or more years of complete data (1984-1997) and between the ages of 18 and 64 at baseline (t-2) in retained sample cohorts 9786 (124,342 PY) 876 (9.0%) Subsamples derived from baseline characteristics (these are not mutually exclusive) - Drop individuals who were not employed at baseline (t-1 or t-2) - 3,679 (45,080 PY) -484 (13.2%) Cohort members employed for both years at baseline (t-1 and t-2) 6,107 (77,262 PY) 392 (6.4%) - Drop individual who report fair or poor health self-reported health status at baseline (t-1 or t-2) - 2,062 (29,945 PY) -431 (20.9%) Cohort members with good self-reported health status or better for both years at baseline (t-1 and t-2) 7724 (99466 PY) 445 (5.8%) Notes: 1 The percentages in the Deaths column represent the percentage of deaths of the total number of individuals in a give row. 2 Between the years of 1990 and 1995 the PSID was supplemented by a Latino/Hispanic cohort. Due to budgetary restrictions the follow-up of this cohort was discontinued after 1995 and deaths were not ascertained. 3 Non-sample individuals are individuals who moved into a sample household (usually through marriage). Non- sample individuals usually are the partner of a sample Head or Spouse, but could also be the parent of a followable sample child. These individuals are followed and complete information is collected on them while they are part of a sample household, but sampling weights are not calculated for them. They are not followed once they leave a sample household. 4 These are sample individuals who have been lost to follow-up or who have died prior to 1984. 5 This group is comprised almost entirely of children or child-age dependants of sample heads or spouses. A small number of individuals (940) are adult sample household member who were never a head or spouse between 1984 or 2005 (e.g., brother, sister, mother or father of Head or Spouse); these individuals account for most of the deaths in this group. 6 Typically these would be years prior to a sample child creating a split-off sample household and becoming a head or spouse. A small number of original sample Heads or Spouses have person years dropped due to being in an institution (prison or hospitals etc) for a given survey year. 80 Table 3.3: Derivation of the German SRHS cohort (1994-2005) Individuals (Person years) GSOEP Cohort 1984-2006 (95% GSOEP W public use sample)1 57,758 (693,096) - Drop individuals not in sample during 1994 to 2005 -23,434 (485,178) - Drop individuals who were never Head or Partner between 1994-2005 -5,186 (19,776) Individuals with at least one year of complete data as Head or Spouse (1994-2005) 29,138 (188,142 ) -Drop individuals who were 65 or older at baseline -3,978 (27,169) Drop individual never older than age 17 -25 (672) -Drop person years in which individuals are not yet sample Heads or Spouses or are sample Heads or Spouses but are in an institution or otherwise not followable -0 (2,960) Heads and Spouses with at least one year of complete data (1984-2005) and between the ages of 18 and 64 at baseline in retained sample cohorts 25,135 (157,341) - Drop individuals with less than three waves of follow-up -5,322 (7,251) Heads and Spouses with three or more years of complete data (1994-2005) and between the ages of 18 and 64 at baseline 19,813 (150,090) - Drop baseline and final year person years -0 (39,626) - Drop individuals and person years with missing data on covariates for the fully specified analytic models -738 (6,980) Analytic sample (1995-2004) 19,029 (103,484) Subsamples derived from baseline characteristics (these are not mutually exclusive) - Drop individuals who were not employed at baseline -5,071 (27,991) Cohort members employed at baseline 13,958 (75,493) - Drop individual who report poor or bad health status at baseline -2,426 (13,806) Cohort members with good self-reported health status or better at baseline 16,603 (89,678) 1. The GSOEP SRHS cohort is dawn from the 95% GSOEP public use sample including 2006 data; however as the 2006 data includes different unemployment compensation measures relating to policy changes in the unemployment insurance system in 2005, and data from this year is not used. 81 Table 3.4: Derivation of the Canadian SRHS cohort (1996-2005) Individuals (Person years) SLID Cohort 1996-2005 (CNEF cohort)1 223,890 (760,396) - Drop individuals from panel 5 (2005) - 34,895 (34,895) - Drop non-sample cohabitants - 41,049 (92,841) - Drop children (less than 18 yrs of age), adults aged 65 or older at baseline and those who were never a head or spouse -70,210 (299,309) Individuals with at least one year of complete data as Head or Spouse (1994-2005) and between the ages of 18 and 64 at baseline 77,763 (335,609) - Drop individuals with less than three waves of follow-up -7,768 (12,316) Heads and Spouses with three or more years of complete data (1984-2005) and between the ages of 18 and 64 at baseline 69,995 (323,293) - Drop baseline year observations - 0 (69,995) - Drop individuals with missing data on covariates for the fully specified analytic models3 - 4,827 (35,767) Analytic sample (1995-2005) 65,168 (217,530) Subsamples derived from baseline characteristics (these are not mutually exclusive) - Drop individuals who were not employed at baseline - 18,661 (61,206) Cohort members employed at baseline 46,507 (156,324) - Drop individual who report poor or fair health self-reported health status at baseline - 7,197 (22,565) Cohort members with good self-reported health status or better 57,971 (194,965) 1. The cross national equivalent SLID cohort is dawn from the underlying SLID panels which follow individuals for a maximum of 6 years. This study draws from panel one (1996-1998), panel two (1996-2001), panel three (1999- 2004) and panel four (2002-2005). 2. Baseline year observations are dropped as there is no lagged health measure for the first year of observation. 3. This includes individuals who do not have complete data in any year, as well has individuals who are missing data in only some years. 82 Table 3.5: Derivation of the American SRHS cohort (1984-1997) Individuals (Person years) PSID Cohort 1968-2005 67,271 (1,278,149) - Drop Latino/Hispanic sample (1990-1995)2 -10,607 (201,533) - Drop non-sample individuals3 -20,725 (393,775) - Drop individuals lost to follow-up prior to 19844 - 7,470 (142,172) -Drop individuals and person years after 1997 - 3,341 (311,795) - Drop individuals who were never Head or Spouse between 1984-19975 -12,349 (93,728) Individuals with at least one year of complete data as Head or Spouse (1984-1997) 12,779 (135,388) -Drop individuals who were 65 or older at baseline - 1,263 (10,662) -Drop person years in which individuals are not yet sample Heads or Spouses or are sample Heads or Spouses but are in an institution or otherwise not followable6 - 5 (19,548) Heads and Spouses with at least one year of complete data (1984-1997) and between the ages of 18 and 64 at baseline 11,511 ( 105,178) Drop individuals with less than three waves of follow-up - 1,754 (2,256) Heads and Spouses with three or more years of complete data (1984-1997) and between the ages of 18 and 64 at baseline 9,757 (102,922) - Drop baseline and final year person years - 0 (19,514) - Drop individuals with missing data on covariates for the fully specified analytic models3 - 212 (4,457) Analytic sample (1985-1996) 9,545 (78,951) Subsamples derived from baseline characteristics (these are not mutually exclusive) - Drop individuals who were not employed at baseline - 2,688 (19,797) Cohort members employed at baseline 6,857 ( 59,154) - Drop individual who report fair or poor health self-reported health status at baseline - 1,414 (12,041) Cohort members with good self-reported health status or better at baseline 8,305 (71,189) Table 3.6: Cross-tabulation between self-rated health status and health satisfaction for the years 1992, 1994-2005 for the German cohort Health Satisfaction (on a scale of 1 to 10) Current Self-Rated Health Status Excellent Good Satisfactory Poor Bad Total Excellent 9/10 4666 7562 722 59 40 13049 Good 8/7 1286 23874 12325 838 63 38386 Satisfactory 5/6 97 4477 16730 4285 155 25744 Poor 3/4 17 579 4135 6122 12 11465 Bad 0/1/2 19 143 309 2170 2276 4917 Total 6085 36635 34221 13474 3146 93561 83 Table 3.7: Assignment of monthly labour force status based on the monthly labour force status question for 1984-1992 62 Code Meaning LFS Status 1 Unemployed or temporarily laid off Unemployed 2 Out of the labour force, but not unemployed or temporarily laid off Not Working 3 Both unemployed and out of the labour force in the month Unemployed 7 Either unemployed or out of the labour force, but not clear which Not Working 9 Missing Not working 0 Neither unemployed, temporarily laid off or out of the labour force Employed Table 3.8. Stylized cumulative labour force status example depicting the transition from employed to not working in the mortality cohort. Years Followed Months Working Months Unemployed Months Not Working Cum. % Working Cum. % Unemployed Cum. % Not Working Baseline: T-2, T-1, T0 1 12 0 0 100% 0% 0% 2 12 0 0 100% 0% 0% 3 8 4 0 89% 11% 0% T+1 4 6 6 0 79% 21% 0% T+2 5 4 4 4 70% 23% 7% T+3 6 0 3 9 58% 24% 18% T+4 7 0 0 12 50% 20% 30% 62 Unemployment dominates working, as working is implied through the union of the null answer to both the unemployment and out of the labour force. This is different than the labour force hierarchy developed for the GSOEP data in which working dominates registered unemployment. It is necessary to have a different ranking algorithm for GSOEP due to different meaning of registered unemployment (i.e. individual can work while registered unemployed) compared to unemployment in the PSID. 84 Table 3.9: Highest degree of education based on a modified CASMIN classification at baseline for the SRHS cohort Skill Classification Germany United States and Canada United States Canada Neither – minimal Compulsory general elementary certificate (1a, 1b) 2,528 (12.9%) Less than high school (1a, 1b) 2,179 (22.5%) 15,413 (22.33%) Specific – medium Basic vocational qualification (1c) 5,785(29.6%) These educational qualifications do not exist in the United States or Canada NA NA Specific– medium Intermediate vocational qualification (2a) 5,016 (25.7%) Specific – medium Vocational Maturity Certificate (2c_voc) 1,195 (6.1%) Vocational degree or certificate (2c_voc, 3a_voc) (Not able to distinguish between 2c_voc and 3a_voc in the PSID or the SLID) 1,737 (17.9%) 19, 612 (30.7%) Specific -high Tertiary Education (3a_voc) 891 (4.6%) General – medium Intermediate general qualification or maturity certificate (2b, 2c_gen) 1,162 (5.9%) High school or GED - includes some college including CEGEP in Canada (2b, 2c_gen) 3,905 (40.3%) 19,893 (30.1%) General – high Tertiary Education – (3b, 3c) 2,965 (15.2%) Associate, bachelor, professional or graduate degree (3a_gen, 3b, 3c) 1,867 (19.3%) 11,157 (17.9%) 19,552 9,688 66,075 Notes: While classifications 1c and 2a do not exist in the United States (Kerckhoff, Ezell, and Brown 2002), it is likely that some the vocational training received in the United States would more comparable to basic vocational qualification (1c) and not at a higher level (2c_voc, 3a_voc). Most holders of vocational certificates in the PSID also have high school or GED completion. There may also be some additional technical school degrees in the PSID classified as 3a_gen or 3b. 85 Table 3.10: Creation of the occupational variable used in the SRHS study across the three surveys Occupational classification used in analysis ISCO-88 (GSOEP) 2000 NOC (SLID) US 1970 Census occupational classification (PSID) Managers Legislators, senior officials, managers (1) Senior managers (1), Other managers (2) Managers and administrators, except farm (201-245) Professionals and Technicians Professionals (2), Technicians and associate professionals (3) Professionals in business and finance (3), Natural and applied science (6), Professionals in health (7), Technicians in health (8), Social science, government service and religion (9), Teachers and professors (10), Art culture, recreation, and sport (11) Professional, technical and kindred workers (1-195) Clerical Clerks (4) Financial, secretarial and administrative (4), clerical workers (5) Clerical and kindred workers (301-395) Sales and Services Service workers and shop and market sales workers (5), Military (0) Wholesale, technical, insurance, real estate sales (12), Retail sales (13), Good and beverage sales (14), Protective services (15), Childcare and home support (16), Travel and accommodation (17) Sales workers (260-280), Service workers (901-984), Current and former members of the armed forces (580,600) Skilled Trades Craft and related workers (7) Contractors (18), Construction trades (19), Other trades (20) Craftsmen and kindred trades (401-575) Plant, equipment and machinery operators Plant and machine operators and assemblers (8) Transport and equipment operators (21), Machine operators and assemblers in manufacturing (24) Operatives except transport (601-695), Transport equipment operatives (701- 715) Agriculture, forestry and fishing Skilled agricultural and fishery workers (6), Occupations unique to primary industry (23) Farmers and farm mangers (24, 801, 802), Farm foremen (821) and forestry worker (25, 605), fisher/hunter (752) logger (761), Drillers (614), Blasters (603) Labourers Elementary occupations (9) Trades helpers, construction and transportation labourers (22), Labourers in processing manufacturing and utilities (25) Labourers, (740-785), Farm labourers (822-824) 86 Chapter 4: Unemployment and Mortality: A Study of Germany and the United States 4.1: Introduction The relationship between unemployment and mortality has been well studied in single country studies. Unemployment has been found to be associated with all cause mortality (Costa and Segnan 1987; Cubbin, LeClere, and Smith 2000; Iversen, Andersen, Andersen, Christoffersen, and Keiding 1987; Norstrom 1988), for both men (Bethune 1996; Lavis 1998; Nylen, Voss, and Floderus 2001) and women (Blomgren and Valkonen 2007; Saarela and Finnas 2005), for cause- specific outcomes (Johansson and Sundquist 1997; Kposowa 2001), and after controls for health selection into unemployment (Gerdtham and Johannesson 2003; Gognalons-Nicolet et al. 1999; Kiuila and Mieszkowski 2007; Rogers, Hummer, and Nam 2000). Unemployment is weakly associated with mortality in older workers (Hayward et al. 1989; Moser et al. 1987; Osler et al. 2003; Sorlie, Backlund, and Keller 1995; Stefansson 1991). Some studies have found that the relationship between unemployment and mortality is smaller during times of high unemployment and not in times of low unemployment (Martikainen and Valkonen 1996; Martikainen, Maki, and Jantti 2007); studies of plant closure, however, have found a consistent relationship (Eliason and Storrie 2007; Sullivan and Wachter 2007). While a relationship between unemployment and mortality has been found in both CME and LME countries (see section 2.5), differences in study and cohort design, measurement of unemployment, control of covariates, and model specification make it difficult to draw conclusions about whether the relationship between unemployment and health varies by country cluster. This chapter presents a comparative longitudinal cohort study of unemployment and mortality in Germany and the United States to determine whether this relationship varies between coordinated market and liberal market economies. In order to maximize comparability, comparable cohorts, measures, and statistical methodology have been used. 4.2: Research Objectives The hypotheses articulated in Chapter 2 lead to the following objectives: 87 - to examine the relationship between unemployment and mortality in working-age cohorts of Germans and Americans to determine whether and how this relationship differs by study country; - to examine how the relationship between unemployment and mortality changes after controlling for health selection and measure of unemployment; and, - to examine if the unemployment-mortality relationship is modified by educational status or gender and whether this also varies by study country. The specific hypotheses developed in Chapter 2 for the unemployment and mortality study are: 1. The association between unemployment and mortality will be weaker in Germany compared to the United States. 2. There will also be effect modification by educational status that will vary by study country. The relationship between unemployment and mortality will be weaker for the minimally skilled and medium skilled in Germany compared to their counterparts in the United States, with the minimally skilled in the United States being especially disadvantaged. The effect of unemployment for the high-general skilled in the United States will be weaker compared those in lower skill categories in the United States. There is no a priori expectation, however, that high-skilled workers in Germany will have a different unemployment-health relationship than those with lower skills. 3. Controlling for health selection will account for some but not all of the association between unemployment and health and a larger proportion of the association will be accounted by health selection into unemployment in Germany. 4. The direction of effect modification across genders is indeterminate, but the ranking across countries will be consistent by gender with the stronger associations being in the United States. 4.3: Methods 4.3.1: Study Cohort and Variables The derivation of the mortality cohort and variables were described in detail in Chapter 3. In brief, two dynamic cohorts were developed; eligible individuals were required to be between 18 and 64 at baseline and have a minimum of three waves of data. Baseline was defined as the first two waves of data, with the first year of follow-up being the third year, after which 88 individuals were at risk of dying. Individuals were followed until death or loss to follow-up or were censored in 2005. The German cohort included 10,886 individuals and 145,732 person years, yielding an average follow-up of 13.4 years. Listwise deletion of individuals with observations with missing information on variables led to a final cohort of 10,754 individuals and 117,123 person years that was used in the statistical models. The American cohort included 9,786 individuals followed for 124,342 person years, yielding an average follow-up of 12.7 yrs. Some individuals (2,108 or 22.5% of the sample) in the American cohort were censored in 1997 because of the reduction in the study sample. Listwise deletion of individuals with observations with missing information on variables led to a final cohort of 9,523 individuals and 98,721 person years. 63 Three sets of labour force status variables were examined: current unemployment, months unemployed in the year previous to the survey, and cumulative lifetime unemployment. The interaction between current unemployment and receipt of unemployment insurance could not be examined in the mortality models as individuals had to survive until the following year for their income variables to be observed. 64 Six groups of covariates were defined and sequentially placed in the model: age and gender; other demographics; household income; education status; occupation; and health status in the previous year. 4.3.2: Health Selection into Unemployment Three approaches were implemented to control for health selection into unemployment. Health selection is directly controlled for by including health status in the year prior to unemployment measures in all final models. Lagged health status, however, may be correlated with both prior and current unemployment and may not be a sufficient control for health selection. This is a form of the ‗initial conditions‘ problem in longitudinal study designs in which it is not possible to directly account for the effects of and temporal sequencing of prior unobservable life events on those observable during the study period (Jones, Rice, Basho d'Uva, and Balia 2007). To account for the challenge of ‗initial conditions‘, two sub cohorts were created that attempted to control 63 The large reduction in person-years for eligible cohort members with complete data was due to excluding the first two years of their study data in the statistical models. 64 Most German unemployed receive unemployment benefits, while most American unemployed do not (see Section 188.8.131.52) and one interpretation of the mortality analysis is that is a comparison between a cohort of unemployed that receive unemployment benefits and a cohort that does not. 89 for health selection prior to the study period through cohort construction rather than through statistical methods. The first sub cohort only includes individuals who are in good or better health at both baseline years and who are therefore unlikely to experience unemployment due to prior poor health. The second sub cohort only includes individuals currently employed at both baseline years and controls for health selection by removing those who have already experienced poor labour market outcomes that may be associated with poor health prior to the start of the study. 4.3.3: Statistical Model While many studies of unemployment and mortality have adopted a Cox proportional hazard approach (16 of the 40 studies reviewed in Chapter 2), this study uses a discrete-time survival approach (Rabe-Hesketh and Skrondal 2008; Singer and Willett 1993; Singer and Willett 2004). In the GSOEP, only the year of death is collected (i.e., the year a person died after the last year of follow-up); while in the PSID the day and month of death are collected on most deaths in addition to year. The metric for death in the GSOEP is discrete time while for the PSID it approaches continuous time. For comparability purposes deaths are coded in both surveys at the level of the year. This leads to interval censoring in which an event that occurs in continuous time (death) is captured in discrete-time intervals (Rabe-Hesketh and Skrondal 2008) and survival analysis techniques that assume continuous time, such as Cox proportional hazards, may not be appropriate (Singer and Willett 2004). 65 The discrete-time survival model can be formulated as: Where h(tij) is the hazard or probability of dying at time Ti that is conditional on having survived to Ti , the baseline hazard parameterization Dij, a set of fixed covariates Xi, and a set of time varying covariates Zij. While the hazard can be estimated using any binomial link (e.g. logit, probit or complementary log-log link), this study uses the complementary log-log (clog-log) link 65 The main reason Cox proportional hazard (PH) model is not acceptable is the abundance of failures (deaths) with the same date. In the case at hand there are multiple deaths (or ties) in every year. The PH model works on the rank ordering of failures and in its basic formulation failures with the same date or identical ranking create analytic difficulties. While there are methods to account for tied rankings in the PH model (e.g. the Breslow or Effron approximation), the abundance of ties in the survey data would lead to a less efficient (i.e., larger variance) modeling compared to the discrete time formulation. ijiijiiij ZXDjTjTth ,,,Pr)( (1) 90 that is the discrete-time equivalent of the Cox proportional hazard model. This leads to the convenient property that relative risks from these models can be directly compared to relative risks from other studies that use Cox proportional hazards (e.g., Lavis 1998; Ahs and Westerling 2006). The clog-log model, like the Cox model, also assumes a proportional hazard (i.e., a one unit change in a covariate in the model causes a proportional shift in the hazard across all values of the covariate). The clog-log model for a given individual i can be expressed as: ititit Xhh )1ln(ln)(cloglog (2) Where hit is the hazard from Equation 1 and XitB is the matrix product of the data matrix of the ith person at time t and the parameter vector. The hazard can be directly obtained by taking the inverse of the clog-log link. )exp(exp(1h (3) BXitit The transformation of which: )exp( )))exp(exp(1ln( )))exp()exp(exp(1ln( (4) j it jit it ijt B BX BBX h h provides the hazard ratio representing the relative risk of dying for a one unit change in xij that is independent of the values of all other covariates in the model. The conditional hazard specification enables us to account for both left censoring and delayed entry (i.e., individuals are observed at age 18 or older and not when they first become at risk of dying) and right censoring (i.e., individuals are lost to follow-up before dying or do not die before the end of the study). Delayed entry is accounted for by conditioning on age (i.e., age is how time is parameterized in the model) and thus the hazard at Ti becomes the probability of dying at agei conditional on having survived to that age (Rabe-Hesketh and Skrondal 2008). This enables the construction of a dynamic cohort in which cohort members are followed starting in different calendar years (e.g., 1984 for the West German cohort and 1991 for the East German cohort) and for different lengths of follow-up. One consequence of both delayed entry and right censoring is that cohort members are at risk of dying for different lengths of time. This may introduce bias into the model if time at risk is correlated with the variables in the model. 66 To 66For example, East Germans are followed for a maximum of 14 years and have seven fewer years at risk compared to West Germans who are followed for a maximum of 21 years. Holding everything else constant, there would be a lower overall risk of dying in the East German cohort compared to the West German cohort due to the difference in length of maximum follow-up. Introducing an exposure offset standardizes for this difference in time at risk. Indeed, 91 account for this potential bias the natural logarithm of time at risk (i.e., years followed) is introduced as an exposure offset to standardize the estimated hazard. The clog-log model is estimated using standard maximum likelihood estimation and can be extended to account for frailty (i.e., unobserved heterogeneity) or non-independence of observations at other levels of clustering (e.g., by region or year) through the inclusion of an appropriate random effect using adaptive Gauss-Hermite quadrature (StataCorp 2007). 184.108.40.206: Assessing Model Fit Model fit was assessed through the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) with smaller criterion statistics indicating a better fit to the data. The AIC and BIC are extensions to the log-likelihood statistic that penalize the log likelihood for the number of parameters in the model (AIC) and additionally the sample size (BIC). Unlike the likelihood ratio test or the deviance statistic, these fit criteria enable comparisons across non- nested models (e.g., different specifications on age, income or the labour force status variables) (Singer and Willett 2004). That said, these fit statistics were used as a guide only and in some cases models with higher (worse) fit statistics were preferred based on conceptual or comparability grounds. For example, including current self-reported health status in the model compared to lagged health status led to a better model fit (i.e., current health status was a much better predictor of mortality than was lagged health status). Health status measures that are contemporaneous with labour force status measures, however, are no longer controls for health selection into unemployment as current health status may be on the pathway between current labour force status and mortality. Accordingly the health status variables were always lagged one year prior to the labour force status variables. 67 220.127.116.11: Model-based Versus Design-based Approach The GSOEP and PSID have a non-random multi-stage stratified sampling frame with the PSID oversampling low-income and black households. Cross-sectional and longitudinal weights and information on the sampling frame (i.e., the strata and primary sampling unit of a respondent) have been created for both surveys in order to account for the survey design through design- the relative risk on the East German variable changes from being protective to representing an increased risk once the exposure offset is introduced. 67 For models that used current labour status, health status in the year prior to the survey was used. While models that used retrospective measures of labour force status (e.g., the number of months unemployed in the year prior to the survey) health status two years prior to the survey was used. 92 based estimation. In this analysis, however, a model-based approach is used that specifies the structural relationship between the dependent and independent variables and which accounts for the non-independence of observations through direct estimation. Deaton (1997) provides a concise summary of the two approaches and why one would choose a model-based approach over a design-based approach. In a fully-specified model (i.e., a model that has correct functional form and no missing variables) the structural relationship should be invariant to the sampling frame. Indeed, the inclusion of survey weights may lead to less efficient estimation (Reiter, Zanutto, and Hunter 2005). In this study, two dynamic cohorts are defined with entry possible in any year between 1984 and 1995. These cohorts do not correspond to a specific reference population and, as such, none of the supplied survey weights are applicable. 68 A model-based approach also allows for the consideration of a more flexible range of statistical and estimation techniques (e.g., accounting for multiple levels of correlation) than a design-based approach. Nevertheless, if the objective is to make inferences to the general population about the prevalence of an outcome within that population or to estimate the association between an outcome and exposure a design-based approach should be used. In sensitivity analysis the effect of the multi-stage stratified design of the surveys was accounted for with the inclusion of a random effect at the level of the primary sampling unit (PSU). These results are compared to the final models for the three labour status specifications in Tables D1 (GSOEP) and D2 (PSID) in Appendix D. There is some evidence of modest correlation at the level of the PSU in the German cohort, but not in the American cohort. For both cohorts and across all specifications there are no differences in the fixed-effect parameter estimates and the AIC and BIC statistics indicate that the models without the random effects are preferable. Accordingly results are presented only on the more parsimonious fixed-effect only models. 18.104.22.168: Alternative Specifications A number of alternative specifications were implemented to account for other potential departures from the statistical assumptions of the model that were not related to the survey design. Specifically, in separate models random and fixed effects for region and year were included to examine if the observed relationship between unemployment and mortality could be 68Both cross-sectional and longitudinal survey weights are designed to make the sample representative of a population at a given point in time. For example using the 1984 cross-sectional weight would make the sample representative to the 1984 population, while using the 1984 longitudinal weights creates a representative fixed cohort of individuals in 1984 that is followable over time by adjusting the 1984 cross-sectional weights for differential loss to follow-up. 93 explained by regional or temporal factors. In other models an individual-level random effect (i.e., what would be considered a frailty parameter in the survival analysis literature) was included to account for residual confounding (i.e., unobserved heterogeneity). None of the alternative specifications improved model fit; nor were there any changes in the relative risk between any of the labour force status variables and mortality. 4.4: Results 4.4.1: Descriptive Statistics 22.214.171.124: Mortality For the German cohort, 879 individuals or 8.1% of the cohort died and the average length of follow-up until death was 10.8 years. For the American cohort, 876 individuals or 9.0% of the cohort died and the average length of follow-up until death was 11.6 years. Figures 4.1 through 4.5 depict the survivor function for both cohorts and stratified by labour force status, gender, educational status, and baseline health status. The survivor functions for the both cohorts indicate an identical cumulative survival probability of about 86% (a risk of dying of 14%) by the end of follow-up. 69 Men have lower survival probabilities than women (83% versus 88%) across both cohorts. The survivor functions for the unemployed and employed at baseline are similar across cohorts (90%), while Germans who were not working have a survival probability of 83% versus 77% for non-working Americans. The high (92%) and medium skilled (89%) also have similar survival functions across cohorts, but there are differences among the minimum skilled (79% for the Germans and 74% for the Americans). Being in poor health at baseline leads to a survival probability of 70% for Americans and 85% for Germans by the end of follow-up. 126.96.36.199: Labour Market Status and Other Variables At baseline there are marked differences in the distribution of covariates by labour force status and by study cohort. Table 4.1 describes the baseline statistics of the German and American cohorts by current labour force status. The German cohort is older than the American cohort, particularly for the unemployed, who are on average eight years older in the German cohort compared to the American. There is no difference in the age of unemployed and employed in the 69 The reason that the probabilities from the survival function are greater than the simple probabilities is that the survivor function accounts for censoring. 94 German cohort, but the unemployed in the American cohort are younger than their employed counterparts. East Germans make up a disproportionate proportion (41%) of the unemployed in the German cohort, as do blacks in the American cohort (64%), reflecting, in part, that these groups are over sampled. Overall Germans are more likely to be married, divorced or separated but less likely to be single or widowed. Unemployed Americans are less likely to be married and more likely to be single. Notably, similar gender distributions are found across both cohorts, with men less likely to be unemployed or not working. Of those not working, about 80% are women in both cohorts. Household size is similar across cohorts, but Americans report more children in the household across all three labour force states; non-working Americans, in particular, have more children. A higher proportion the unemployed or those not working in the American cohort are minimum skilled, while a higher proportion of the unemployed or those not working in the German cohort are medium or high skilled. At baseline only individuals who are working have an occupation; working Germans are more likely to be in manufacturing occupations, while working Americans are more likely to be in professional and technical occupations or business and sales occupations. Unemployed and employed Germans report similar levels of health satisfaction at baseline, but unemployed Americans report lower self-reported health status compared to employed Americans. 70 Non-working Americans and Germans both report lower levels of baseline health status, with poor health being particularly prevalent in non-working Americans. The unemployed report two thirds of the household income of the employed in the German cohort, while the unemployed report half of the household income of the employed in the American cohort. This ratio is consistent the year previous and the year after the current unemployment episode. However, the fraction of current labour market income is higher (28%) for unemployed Americans compared to unemployed Germans (19%). In the year following average labour market income recovers, but is still only 28% for the German unemployed and 34% for the American unemployed. Unemployed Germans report higher relative levels of unemployment compensation and of total household public transfers compared to their employed 70 Recall that baseline health status is health satisfaction, rather than self-reported health status and these health measures could have different distributions. In particular, health satisfaction appears to be less associated with labour force status than self-reported health status. 95 counterparts. Unemployment compensation also makes up a greater proportion of public transfers for unemployed Germans than for unemployed Americans. Table 4.2 describes the relationship between months unemployed in the year previous to the survey by demographic measures, other unemployment measures and income and transfers across all survey years. 71 Overall, Americans report a slightly higher proportion of any months of unemployment (11%) compared to Germans (9%), but Americans were more likely to be unemployed for fewer months; 50% of unemployed Americans reported being unemployed for three or fewer months and 14% reported being unemployed for ten or more months. In contrast, 30% of unemployed Germans reported being unemployed for three or fewer months and 35% reported being unemployed for ten or more months. Unemployed Americans were much less likely to report receiving unemployment compensation. The proportion receiving unemployment compensation was never more than 36%, peaking for those reporting three or four months of unemployment and falling thereafter to 12% for those reporting 12 months of unemployment. The proportion of Germans receiving unemployment compensation was 76% for those reporting one month of unemployment; rising to about 90% coverage at eight months and declining to 83% for those unemployed the entire year. Unemployed Germans also reported higher mean benefit levels across all month profiles. Unemployment and total public transfers represented a greater proportion of household income. There was a greater decline in household and individual labour income for Americans compared to their employed counterparts than for Germans. 4.4.2: Current Unemployment Tables 4.3 and 4.4 show the full results for the German and American cohort respectively, with the cumulative addition of groups of covariates. The relative risk (RR) of dying for unemployed Germans compared to employed Germans is 2.1 (95% CI: 1.5-3.1) in the age and gender model. The progressive inclusion of other demographic variables attenuates the relative risk to 2.0 (95% CI: 1.4-2.9). With the inclusion of the socio-economic status variables (household income, education and occupation) the relative risk drops to 1.7 (95% CI: 1.2-2.4). Once health status in the previous year is added the relative risk (1.4 95% CI%: 1.0-2.0) is no longer statistically 71 Measures have been harmonized across years to ensure that reflect the same calendar year. 96 significant at the 95% confidence level, although the unemployed still have a 40% increased chance of dying. The relative risk of dying for unemployed Americans compared to employed Americans is 3.7 (95% CI: 2.6-5.2). In the age and gender model, the progressive inclusion of other demographic variables attenuates the relative risk to 2.9 (95% CI: 2.0-4.1). With the inclusion of the socio- economic status variables the relative risk drops to 2.5 (95% CI: 1.7-3.5). With the inclusion of lagged health status the relative risk is 2.4 (95%CI: 1.7-3.4). Unemployed Americans have a higher relative risk of dying compared to unemployed Germans (2.4 versus 1.4 in the final lagged health model) representing 1.5 times larger risk (Table 4.5). While the attenuation of the relative risk is similar with the inclusion of other covariates for unemployed Americans and Germans, the inclusion of lagged health status has a greater attenuating effect for the German unemployed. 188.8.131.52: Results for the Other Variables72 For the German cohort, each additional year of age increases the risk of dying by 5% (RR 1.05); men have a 2.5 relative risk compared to women; being East German is associated with a 1.3 relative risk compared to West Germans, but there is no increased risk for immigrants, or for being a spouse. Single people (RR 1.5) and those divorced or separated (RR 1.4) have increased risks compared to married individuals, but those widowed do not. Household size is associated with an increased relative risk of 1.2 for each additional person, but each additional child has a protective effect (RR 0.7). The log of household income is protective (RR 0.83). The minimum skilled have a relative risk of 1.4 compared to the high skilled, but there is no association for the medium skilled. Having no occupation is associated with a relative risk of 1.6 compared to management and professional occupations, but there are no associations for the other occupational categories. Lagged health satisfaction and disability exhibit strong associations with mortality. While there was no association for good health satisfaction compared to excellent health satisfaction, being of satisfactory, poor or bad health is associated with increasing relative risks of 1.5, 1.8, and 3.8; the relative risk for being disabled is 1.4. For the American cohort, each additional year of age increases the risk of dying by 5% (RR 1.05); men have a relative risk of 1.7 compared to women; being black is associated with a 72 I only report the results on the other covariates in the current labour force status model as the specification and the results of the other variables were the same across the three different labour force status models. 97 relative risk of 1.2 compared to being white, while there was no association for those of other ethnicities. There is no association for spouses compared to heads. Single people have a relative risk of 1.6 compared to those married, but there is no association for those divorced or separated or those widowed. Neither household size nor number of children yields an association. The log of household income is protective (RR 0.93), but there is no association by educational or occupational status. Lagged self-reported health status exhibits a strong and increasing negative gradient with very good, good, fair, and poor health being associated with 1.5, 2.0, 3.8 and 7.2 relative risks compared to excellent health, respectively, while there is no association for disability status. 4.4.3: Months Unemployed and Cumulative Unemployment Tables 4.6 and 4.7 show the results for months unemployed and cumulative lifetime unemployment across both cohorts. Each additional month of unemployment for unemployed Germans is associated with an increased risk of dying of 1.07 (95% CI: 1.03-1.10) and 1.05 (95% CI: 1.01-1.09) for the full- and health- adjusted models. For unemployed Americans the risk of additional months of unemployment is identical (1.09 95% CI: 1.05-1.14) in both the full- and health-adjusted models. Evaluated at the average number of months unemployed (6.8 months for the German unemployed and 4.3 months for the American unemployed), yields an average relative risk of 1.4 for the German cohort and 1.5 for the American cohort. Each additional percent of follow-up spent unemployed (lifetime unemployment) yields a 1% increased risk of mortality (RR 1.007 95%: CI: 1.001-1.012) for the German cohort and a 2% increased risk of mortality (RR 1.016 95% CI: 1:008-1.024) for the Americans. 4.4.4: Gender Stratified Results Tables 4.8 through 4.10 present the gender stratified results across both cohorts. In the German cohort there is effect modification by gender across all labour force status measures. Men have consistently higher and statistically significant associations between current unemployment and mortality (RR: 1.6 95% CI: 1.0-2.4) and for the other labour force status measures. No associations are found for women for current unemployment (RR: 1.0 95% CI: 0.5-2.1) or for the other labour status measures. In the American cohort, men and women have statistically significant risks for current unemployment, with women (RR: 2.6 95% CI: 1.5-4.5) having a slightly higher risk than men (2.4 95% CI: 1.4-3.5); however, for months unemployed (RR: 1.11 98 95% CI: 1.06-1.18) and for cumulative lifetime unemployment (RR: 1.023 95% CI: 1.016-1.03) men have higher and statistically significant associations, while there is no association with mortality for women. 4.4.5: Education Stratified Results Tables 4.11 through 4.13 present the education stratified results across both cohorts. There is marked effect modification by educational status in both the German and American cohorts. The relative risk of dying is highest for unemployed high-skilled Germans and minimum-skilled Americans and lowest for unemployed medium-skilled Germans and high-skilled Americans. Compared to the full cohort, unemployed minimum-skilled Germans have a higher, but not statistically significant relative risk (1.6 95% CI: 0.7-3.7), while the relationship for the unemployed medium skilled is close to one and not statistically significant (RR: 1.1 95% CI: 0.7- 1.7). In contrast, unemployed high-skilled Germans have a larger and statistically significant relative risk (RR: 3.0 95% CI: 1.3-7.0). For unemployed minimum-skilled (RR: 2.6 95% CI: 1.4- 4.7) and medium-skilled (RR: 2.4 95% CI: 1.5-3.8) Americans the association is slightly higher than the relative risk in the full cohort for current unemployment, but there is no association for the unemployed high skilled (RR 1.0 95% CI: 0.2-4.3). For the other labour force status measures, unemployed high-skilled Germans continue to have a higher relative risk of mortality, while there are no associations for the minimum and medium skilled. The unemployed medium- skilled Americans have the highest relative risks for months unemployed and cumulative unemployment, while there are no associations for these measures for the unemployed minimum and high skilled. 4.4.6: Exclusions Tables 4.14 through 4.16 present the results based on excluding those in poor health at baseline and those unemployed or not working at baseline across both cohorts. The relative risk for current unemployment was higher in both the good health sub cohort (G: 1.7 95% CI: 1.1-2.6; P: 3.0 95% CI: 2.0-4.4) and the working cohort (G: 1.7 95% CI: 1.1-2.7; P: 3.4 95% CI: 2.2-5.4) compared to the full cohort for both study countries (G: 1.4; P: 2.4). No differences in the relative risk of months unemployed and cumulative lifetime unemployment were found in comparing the sub-cohort results with the full cohort results, except that mortality risk of cumulative lifetime unemployment was higher in the American working sub 99 cohort (RR 1.027 95% CI: 1.01-1.04) compared to the full cohort (RR 1.016). 4.4.7: Country-specific Analyses The German-specific analysis that stratified the cohort by whether an individual was from East or West Germany found large effect modification (Table 4.17). The relative risk of mortality for unemployed East Germans was 2.1 (95% CI: 1.2-3.6) for current unemployment and 1.08 (95% CI: 1.02-1.16) for months unemployed compared to employed East Germans, while there was no association for unemployed West Germans who had a relative risk of 0.9 (95% CI: 0.5-1.6) for current unemployment and 1.03 (1.0-1.1) for months unemployed. No association was found in either group for cumulative unemployment. In contrast, stratifying the analysis by black and white or other did not yield evidence of effect modification (Table 4.18). The relative risk of mortality for unemployed blacks and white or other was 2.5 (95% CI: 1.5-4.0) and 2.3 (95% CI: 1.3-2.9) for current unemployment; 1.09 (95% CI: 1.03-1.15) and 1.11 (95% CI: 1.04-1.18) for months unemployed; and for cumulative unemployment it was 1.01 (95% CI: 1.00-1.02) and 1.02 (95% CI: 1.01-1.03) compared to employed blacks and whites or other. 4.5: Discussion 4.5.1: Assessment of Chapter Hypotheses This study found an increased risk of dying for current unemployment for both Germans and Americans, but in almost all cases the risk was much higher for the American unemployed compared to the German unemployed. There is a statistically significant and higher risk of dying for men, the high skilled and East Germans in Germany compared to other unemployed groups, while for the American unemployed there is a consistent relative risk of dying among all groups except for the high skilled (Figure 4.6). Men in both Germany and the United States have elevated risks of dying for months unemployed and cumulative unemployment, but the risk in the United States is about twice as high; there is no association for women for these measures in either cohort. The higher risk of dying for the unemployed in the United States compared to Germany supports the hypothesis that the institutional environment, including higher levels of unemployment and 100 employment protection, mediates the unemployment-mortality relationship. This finding held across all labour force specifications and for the sub-cohort exclusions, suggesting that the ranking of unemployment-mortality risks by country are robust to different measures of unemployment, health selection into unemployment, and labour force composition. Nonetheless, it is worthwhile to consider whether other within-country factors, including differences in the measurement of unemployment, cohort definition, survey design or other unmeasured country confounders, could explain the differences in the risk of dying for the German and American unemployed. In other words, are there factors that could to lead the findings in Germany being biased downward and the findings in the United States being biased upwards such that if these factors were controlled for results between these two countries could converge? This study used an internationally standardized measure of unemployment across the two surveys, but there are individuals in Germany who meet a German-specific definition of unemployment (described in Chapter 3) who were excluded from the unemployed in this study and it may be these individuals had poorer health outcomes than those who were considered unemployed according to the international definition. To test for this, the German-specific definition of unemployment rather than the international measure was used in secondary GSOEP-only models. The results from these models were similar to the results in the main models, indicating that the relationship in Germany is not dependent on how unemployment was defined. Research has shown that individuals in poor health are more likely to remain unemployed (Stewart 2001). 73 In Germany these individuals have an incentive to remain unemployed due to the continued receipt of unemployment benefits, while in the United States these individuals are less likely to be eligible for benefits and thus do not have this incentive. Based on differences in incentives relating to health selection out of unemployment, the results in Germany are more likely to be biased upward and the results in the United States more likely to be to be biased downward, which would magnify the differences in risk between these two countries. The results from this study are also consistent with the one other study using the GSOEP, which found no relationship between unemployment and mortality in Germany (Frijters, Haisken- DeNew, and Shields 2005a). Nine unemployment and mortality studies use American data (see 73 This refers to health selection out of unemployment which was discussed in Chapter 2 section 2.4.3. 101 Section 2.5.3) and while it is challenging to draw direct comparisons across these studies to give a modal or median estimate of the risk of dying associated with unemployment due to differences in study design and methods, unemployment measures, and length of follow-up, the results on current unemployment fall within the range for risk ratios found in the American studies that use current unemployment as a measure. One other study found a higher risk ratio of over three (Lavis 1998), with most studies finding a risk ratio between 1.5 to 2.2 (Cubbin, LeClere, and Smith 2000) (Kiuila and Mieszkowski 2007; Sorlie, Backlund, and Keller 1995; Sorlie and Rogot 1990), and a few finding no association at all (Hayward, Grady, Hardy, and Sommers 1989; Rogers, Hummer, and Nam 2000). While other factors that may introduced a differential bias into the unemployment-mortality relationship cannot be ruled out, 74 the consistency of the results in this study from those of other studies and the fact that known biases are likely to magnify the differences in the relative risk across countries support the claim that the differences in risks are not artifactual, but reflect real differences in the unemployment-mortality relationship across these two countries. The education stratified results are also consistent with the hypotheses developed in Chapter 2, which specify that there should be a stronger unemployment-mortality gradient by skill level in the United States, but that the modal medium- (and vocationally-) skilled worker in Germany would be the most protected from the negative health consequences of unemployment. In the American cohort there is no relationship between unemployment and mortality for the high skilled across any measure of unemployment. It appears that individuals with a high level of education are best suited to take advantage of the flexible labour markets within LMEs. 75 The high skilled are more likely to receive unemployment benefits when unemployed than those of lower skill levels. Further these individuals may also have other resources (e.g., savings, familial resources, and social or business contacts from educational or professional organisations) to draw upon that would buffer the effect on unemployment on health. The drop in household income for the unemployed high skilled in America was smaller than for those of lower skill levels. The median household income for the unemployed high skilled was 64% of the employed high skilled, while for the unemployed medium and minimum skilled it was 48% and 45%. 74 For example differences in attrition among the unemployed across the surveys could also introduce bias into the results. But for the bias to be differential and increase the relative risk in Germany and decrease the relative risk in the United States, the unhealthy unemployed would have to be more likely to drop out in Germany and the healthy unemployed would have drop out in the United States. 75 Almost all the high skilled in the United States would have a general skilled education. See section 3.4.3. 102 Moreover, the household income of unemployed high skilled was similar to that of the employed medium skilled and higher than that of the employed minimum skilled. Both the unemployed minimum and medium skilled have an elevated risk of dying and, across all three measures of unemployment, the medium skilled have the highest risks in the United States. In Germany, the medium skilled have the lowest risk of dying across all three unemployment measures. This is the strongest evidence that institutional environment can affect the relationship between unemployment and health as institutional protection is targeted towards medium- (and vocationally-) skilled worker in Germany. Both medium-skilled groups are the largest group of workers and also have the largest number of unemployed in both cohorts (although the unemployment rate is highest among the minimum skilled in both countries) and this comparison does not suffer from small sample size in the number of unemployed and the number of deaths. Further, the contrast between the two medium-skilled groups is striking with respect to receipt of unemployment compensation and household income that may be mediators of the unemployment-mortality relationship. The unemployed medium skilled in Germany have a median household income of 70% of their employed counterparts and 75% of them report receiving unemployment compensation, while the unemployed medium skilled in the United States have a median household income of 48% of their employed counterparts and only 19% report receiving unemployment compensation. 76 The elevated relative risk for the unemployed high skilled in Germany requires some interpretation. While there was no prior expectation that the unemployed high skilled would fare better than the unemployed medium skilled in Germany, it was not hypothesized that they would fare worse. There are two possible explanations for this finding. First, the low number of unemployed and deaths among high-skilled Germans leads to results that may be sensitive to only a few events. There are 109 deaths for the high skilled overall (only 4% of the high skilled died compared to 9% of the minimum or medium skilled). Accordingly, this difference may be driven by a low baseline hazard for the employed high skilled, even if there are only a few deaths among the unemployed high skilled (only 7% or seven died). The wide confidence intervals for this relationship also support this interpretation. Secondly, the unemployed high skilled in Germany may be relatively worse off compared to employed high skilled as the institutional 76 This is likely an understatement of the proportion who receive unemployment compensation due to the reporting errors in the unemployment compensation variable that was discussed in section 3.5. 103 supports are targeted towards medium-skilled workers. Inspection of the data, however, reveals that the unemployed high skilled who died are almost all (six of seven) from East Germany, with no elevated risk of mortality in the unemployed high skilled in West Germany. This is consistent with the country-specific results that indicated that the elevated mortality risk was only found in the East German cohort. Overall, the gender differences between countries are also consistent with the study hypotheses. American women have higher risks of dying for current unemployment, while no relationship is found for any of the measures of unemployment for German women. German men have statistically significant risks of dying for all measures of unemployment, but this risk is about two thirds to one half the relative risks for American men, depending on the measure. There are differences between men and women for both cohorts. American men and women have similar risks for current unemployment, but these risks diverged for the other labour force status measures. For German men and women, the risks differed across all measures. For women, the weaker relationship between unemployment and mortality may be due socio-economic gradients being expressed more through morbidity than mortality (Wingard 1984). Support for this interpretation will be found in the SRHS study if a robust relationship for women is found. The divergence among the unemployment-mortality relationship across the three labour force status measures for American women suggest that some of the effect modification may be due to differences in recall bias and in interpretation of the unemployment construct for women compared to men (Jacobs 2002; Jurges 2007). This is also consistent with the gender differences in recall bias discussed in section 3.5 in which there was a higher lack of concordance between retrospective and current measure of unemployment for women. Accordingly, caution needs to be applied when interpreting the retrospective measures for women. The strong and similar relative risk for current unemployment for American women suggests that unemployment is a health risk for both men and women in the United States. In Germany, the fact that no relationship was found for women for any of the labour force status measures may also be due to recall bias (Jurges 2007), but could also be related to the gender segmentation in institutional and social support arrangements in Germany (Estevez-Abe 2005). Indeed, in Esping-Andersen‘s Three Worlds typology one of the distinguishing features of the corporatist or Christian Democratic welfare regime is the emphasis in these regimes on social and state support 104 reinforcing traditional familial roles (Esping-Andersen 1999). Health status prior to unemployment was found to explain more of the relationship in Germany compared to the United States. The inclusion of lagged or baseline health status in the PSID did not attenuate the risk of dying, indicating that prior health status is not a confounder in the unemployment-health relationship in the American cohort. In contrast, prior health status is a confounder for this association in the German cohort, attenuating the risk for all measures of unemployment, sometimes to statistical insignificance. This supports the hypothesis that health selection is more important in Germany because of the protective institutional effects, but that social causation explains more of the relationship in the United States. The results are also robust to the control for health selection in that similar or higher risks are observed in the good health and working sub cohorts. Notably, the relative risks increased in the sub cohorts that excluded either those not employed or those in poor health at baseline rather than decreasing. Removing those in poor health at baseline from both the unemployed and employed controls may affect either groups‘ underlying hazard. If the baseline hazard for the employed controls drops but the risk difference between the two groups stays the same, then the relative risk will necessarily increase. As such, a direct comparison of the risks across the exclusion sub cohorts is not advisable as the composition of the control groups has changed. The result of a robust relationship between the employed and unemployed who are healthy at baseline, however, indicates that health selection into unemployed does account for the observed relationship between unemployment and mortality. That a strong relationship between unemployment and mortality is also found in the working-only cohort at baseline provides additional support for this interpretation. Further, this suggests that the association between unemployment and mortality is also present in individuals with strong labour force attachment. This argument that health selection has been sufficiently accounted for hinges on the validity of our baseline health status measures. Two points support the argument that they are sufficient controls. First, the health status measures are the strongest predictors of mortality in the model – poor health predicts death well; second, baseline health and labour force exclusions were based on two years of data and as such those in good health were persistently in good health. 77 77 Results from SRHS analysis in Chapter 5 indicate that SRHS is highly correlated across years. 105 4.5.2: Within Country Differences The American results are not sensitive to stratification by race; unemployed blacks and whites have similar relative risks compared to their employed counterparts. This is not to say that race does not play a role in the unemployment and health relationship as blacks are much more likely to experience unemployment than whites (e.g., the unemployment rate for blacks in 2005 was 9.5%, while it was 4.4% for whites). 78 Indeed the proportion of mortality attributable to unemployment (i.e., adjusted population attributable fraction (Rothman, Greenland, and Lash 2008) is higher for blacks than for whites given the higher prevalence of unemployment for blacks. Societal context – the institutional, economic, and socio-cultural environment – can matter in two ways; both in determining who is unemployed and in how unemployment affects health. In the United States, blacks are more likely to become unemployed than whites, but once unemployed both black and whites are at an increased relative risk of dying. This suggests that country and institutional patterns relating to unemployment may, in part, be codetermined by the legacy of racism and segregation in the United States as blacks are more likely to be unemployed. In contrast, in Germany, East Germans have both an increased risk of being unemployed and an increased relative risk of dying compared to West Germans. It is also the high-skilled East Germans that drive the relationship between unemployment and health in the high-skilled stratification. The results from Germany suggest that for West German workers, who have spent their entire working life embedded within the CME institutional environment, the institutional supports are effective. For unemployed East Germans, who come from a different institutional environment (a planned economy), the institutional supports are not as effective. 4.5.3: Unemployment and the Accumulation of Disadvantage While a strong and robust association between unemployment and mortality was found in the United States and for some groups in Germany, this study does not definitely establish whether this relationship is causal. Unemployment may also be a marker for other mechanisms and for the accumulation of socio-economic disadvantage that may affect health. For example, workers in hazardous jobs may be more likely to face involuntary job loss (Robinson 1986). Unemployed workers are also more likely to come from groups already vulnerable to negative health 78 See ftp://ftp.bls.gov/pub/special.requests/lf/aa2005/pdf/cpsaat24.pdf. 106 outcomes (i.e., unemployment is concentrated among the low-waged, the minimally-skilled, East Germans in Germany and blacks in the United States). Disentangling the confluence of these determinants of health is challenging, but insight can be gained from the comparative study design by moving beyond the comparison of relative risks across countries and comparing the average predicted risk of dying for specific unemployed groups across countries. Table 4.19 and Figure 4.7 depict the predicted hazard of dying evaluated at the mean of other covariates across the three models and stratified by educational status. 79 Figure 4.7 shows that the average risk of dying across all ages is lowest for employed Germans, while unemployed and non-working Americans have the highest risk. When stratified by skill level (Table 4.19), the differences in risks are magnified; unemployed minimum-skilled American are about seven times more likely to die than employed high-skilled Germans and employed or unemployed high-skilled Americans. In contrast, the risk of dying for non-working Germans (the worst off German group) is four times than those with the lowest risk. There is also a doubling of risk between the unemployed medium-skilled Germans and Americans. Taken together, these results support the idea that distributional and institutional factors contribute to the flattening of the socioeconomic- health gradient (Hertzman 2001; Siddiqi and Hertzman 2007). Further unemployed minimum and medium skilled Americans may be less likely have had access to health care insurance while employed and be more likely to lose it once unemployed compared to the high skilled. Access to health insurance and health care may be a key institutional feature that explains steeper socio- economic gradients in mortality in the United States compared to Germany and other countries (Kunitz and Pesis-Katz 2005). While it may not be possible to distinguish between determinants that are truly upstream and those that lie along the pathway to health, what is clear is that among individuals with multiple health vulnerabilities or disadvantages (in this case low education status and unemployment) the effects of these risks on health are modifiable. 4.5.4: Strengths and Limitations The strengths in this study are the focus on creating comparable cohorts across study countries and the emphasis on creating similar labour market and educational variables. This study used a full range of covariates spanning demographic, socioeconomic status, and health status variables to control for potential confounding. It also examined two alternative measures of unemployment 79 The predicted probabilities also reflect the effect of the other covariates on the risk of death; they represent the clustering of risk among groups of which the effect of unemployment and education would be only part. 107 in addition to current unemployment and a number of approaches were also taken to control for health selection. There are number of limitations to this study. First, in spite of the efforts to create comparable measures across the study, some measures across countries were different. In particular the health status controls were different across countries. In the German cohort, the health status controls were measured through health satisfaction and disability defined as having a registered disability, while in the American cohort, health controls were measured through self-reported health status and disability defined as self-reported activity restrictions. Second, differences in attrition and measurement error across the studies could have introduced differential bias into the study. Third, there may be other variables that might confound the relationship between unemployment and mortality that were not controlled for in the models (i.e. residual confounding). For example, it was not possible to control for health-related behaviours such as drinking, smoking or physical activity that may have differed across the unemployed groups. It was also difficult to make the distinction between voluntary and involuntary unemployment. 4.6: Conclusion The findings from this study support the idea that context matters to the health of the unemployed. In Germany, a CME with high levels of employment and unemployment protection, the unemployment-mortality association is only found for East Germans. For West German workers, who have spent their entire career within the CME institutional environment, there is no association. In the United States there is no unemployment-mortality association for the high skilled who are best positioned to take advantage of the flexible labour market found in liberal market economies. But for the remainder of workers – the minimum and medium skilled – unemployment comes with an increased risk of death. In particular, those at the bottom of labour market and educational hierarchy – the minimum skilled – are much more likely to die, reflecting the accumulation of health disadvantage within this group in the Unites States. The VOC framework is predicated on the idea that there are two macroeconomic equilibria that lead to similar levels of aggregate national wealth and economic growth. This study provides evidence that these equilibria may also have profound distributional consequences when it comes to workers‘ health. The findings from and questions raised in this study point towards a continued research 108 agenda. Does the mediating effect of the institutional environment hold in other countries that have strong social support for the unemployed? The literature review in Chapter 2 found that there was an unemployment-mortality relationship in other CME countries (Ahs and Westerling 2006; Gerdtham and Johannesson 2003) (Eliason and Storrie 2007) and it cannot be ruled out that the findings from this study are peculiar to Germany and the United States. Further exploration needs to focus on the specific mechanisms that may buffer the unemployment- mortality association. This study focused on overall difference in the levels of unemployment and employment protection, but future research should also focus on the role of other government cash transfers and on post-unemployment labour market trajectories. 109 Figures and Tables Figure 4.1: Survivor function for the German and American cohorts by years followed Figure 4.2: Survivor function for the German and American cohorts by years followed stratified by labour force status at baseline (t-2) 110 Figure 4.3: Survivor function for the German and American cohorts by years followed stratified by gender Figure 4.4: Survivor function for the German and American cohorts by years followed stratified by educational status 111 Figure 4.5: Survivor function for the German and American cohorts by years followed stratified by health status at baseline (t-1, t-2) Figure 4.6: Summary of the relative risks of dying for current unemployed for the German and American cohorts 112 Figure 4.7: Predicted hazard of dying by current labour force status and age for the German and American cohorts (adjusted for all covariates including t-1 health status) 113 Table 4.1: Descriptive statistics at baseline stratified by current labour force status and study country Working (72.1%) Unemployed (5.3%) Not Working (22.5%)) Working (70.3%) Unemployed (8.5%) Not Working (21.4%) Age 37.3 (37.0-37.6) 38.1 (37.1-39.1) 44.21 (43.7-44.7) 34.13 (33.8- 34.4) 29.9 (29.1-30.8) 40.04 (39.5-40.6) Male .56 (.55-.58) .43 (.4-.47) .20 (.18-.22) .52 (.51-.53) .46 (.42-.49) .19 (.16- .21) East German .30 (.29-.31) .41 (.37-.45) .19 (.17-.20) Immigrant .06 (.06-.07) .18 (.16-.2) .09 (.08-.10) White .61 (.6-.62) .32 (.28-.35) .53 (.51- .55) Black .34 (.33-.35) .64 (.61-.67) .41 (.39- .43) Other .05 (.04-.06) .04 (.03-.06) .06 (.06- .07) Married .70 (.69-.71) .64 (.60-.67) .77 (.76-.79) .59 (.58-.6) .32 (.29-.36) .64 (.61- .66) Single .22 (.21-.23) .22 (.19-.25) .12 (.10-.13) .28 (.27-.29) .48 (.45-.51) .18 (.16-.2) Div or Sep .07 (.06-.07) .13 (.11-.15) .05 (.04-.06) .02 (.01-.02) .02 (.01-.03) .06 (.05- .06) Widowed .01 (.01-.02) .02 (.01-.03) .06 (.05-.07) .08 (.08-.09) .10 (.08-.12) .07 (.06- .08) Household size(#) 2.87 (2.84-2.9) 2.75 (2.66-2.85) 2.91 (2.87-2.96) 2.73 (2.7-2.77) 2.66 (2.55-2.76) 3.18 (3.11- 3.24) Children(#) .73 (.71-.75) .71 (.63-.79) .75 (.71-.78) .83 (.80-.85) .98 (.90-1.06) 1.12 (1.07- 1.17) Minimum skill .10 (.09-.10) .19 (.16-.21) .28 (.26-.29) .16 (.15-.17) .36 (.33-.38) .39 (.38- .41) Medium skill .67 (.66-.68) .67 (.63-.71) .59 (.58-.61) .60 (.59-.61) .57 (.54-.6) .51 (.49- .53) High skill .22 (.21-.23) .13 (.1-.17) .12 (.11-.14) .24 (.23-.24) .07 (.04-.09) .10 (.08- .11) No occupation .05 (.05-.06) 1.00 (.98-1.02) 1.00 (.99-1.01) .02 (.01-.02) 1.00 (.99-1.01) 1.00 (1.00- 1.00) Professional and Technical .16 (.15-.17) 0.00 (-.03-.03) 0.00 (-.01-.01) .19 (.18-.19) 0.00 (-.02-.02) 0.00 (-.01-.01) Bus/Sales occ .30 (.29-.31) 0.00 (-.03-.03) 0.00 (-.02-.02) .34 (.33-.35) 0.00 (-.03-.03) 0.00 (-.02-.02) Serives occ .15 (.14-.16) 0.00 (-.02-.02) 0.00 (-.01-.01) .14 (.14-.15) 0.00 (-.02-.02) 0.00 (-.01-.01) Agr/For/Min occ .03 (.03-.03) 0.00 (-.01-.01) 0 .00(-.01-.01) .04 (.04-.05) 0.00 (-.01-.01) 0.00 (-.01-.01) Manufacturing occ .30 (.29-.31) 0.00 (-.03-.03) 0.00 (-.02-.02) .26 (.25-.27) 0.00 (-.03-.03) 0.00 (-.02-.02) Health excellent .31 (.3-.32) .28 (.24-.31) .27 (.25-.28) .29 (.28-.3) .25 (.22-.28) .19 (.17- .21) Health good .38 (.37-.39) .28 (.24-.32) .3 (.28-.32) .35 (.34-.36) .28 (.25-.31) .23 (.21- .25) Health satisfied .21 (.2-.22) .26 (.22-.29) .25 (.24-.27) .27 (.26-.28) .29 (.26-.32) .27 (.25- .29) Health fair .07 (.06-.07) .10 (.08-.12) .09 (.08-.1) .08 (.07-.09) .14 (.12-.16) .21 (.19- .22) Health poor .03 (.03-.04) .08 (.06-.1) .09 (.08-.09) .01 (.01-.02) .04 (.02-.05) .11 (.1-.12) Disabled .04 (.03-.04) .03 (.02-.05) .12 (.11-.13) .08 (.07-.09) .15 (.12-.17) .33 (.32- .34) Health Occupation GSOEP (N=10866) PSID (N=9786) Demographics Education 114 Table 4.1: Descriptive statistics at baseline stratified by current labour force status and study country (continued) Notes: Dollars $ and Euros € are in 2005 values. Working Unemployed Not Working Working Unemployed Not Working household income(t-1) 30655 20839 24855 38539 20116 31017 (30238-31072) (19243-22434) (24159-25551) (37898-39181) (18262-21970) (29847-32188) household income(t0) 31614 21683 24852 40653 20281 31651 (31255-31974) (20349-23017) (24209-25495) (39892-41414) (18079-22482) (30260-33042) household income(t+1) 32936 22830. 25820 41759 21284 31631 (32566-33306) (21471-24189) (25159-26480) (41094-42423) (19357-23211) (30418-32844) individual labour income(t-1) 24924 7617 2018 29896 9372 2977 (24473-25375) (5893-9342) (1266-2770) (29248-30545) (7498-11246) (1794-4160) individual labour income(t0) 24344 4555. 1283 32382. 8607. 2340 (23949-24739) (3089-6020) (577-1989) (31735-33029) (6735-10478) (1157-3522) individual labour income(t+1) 24503. 7093 2433 32709 11125 3303 (24159-24853) (5809-8376) (1809-3057) (32046-33372) (9202-13048) (2093-4514) unemployment compensation (t-1) 207 2843 282 209 753 123 (171-243) (2705-2981) (222-342) (175-243) (662-844) (49-196) unemployment compensation (t0) 206 3233. 379 116 649 47 (173-239) (3107-3359) (309-430) (96-137) (591-708) (10-84.38) unemployment compensation (t+1) 470 2210 294.81 166.29 262.66 24.47 (426-514) (2048-2372) (216.24-373.39) (143-189) (196-329) (-18-67.) household public transfers (t-1) 1199 5288 2026 737 3258 2851 (1124-1273) (5003-5573) (1902-2150) (659.75-814.1) (3034-3481) (2710-2991) household public transfers (t0) 1544 5934 2301 561 3435 2920 (1474-1615) (5674-6195) (2176-2427) (486-632) (3210-3651) (2783-3056) household public transfers (t+1) 2079 5188 2018 657 2684 2661 (1996-2162) (4886-5491) (1871-2165) (584-730) (2473-2896) (2528-2794) Transfers GSOEP (N=10866) PSID (N=9786) Income 115 Table 4.2: Unemployment, unemployment compensation, income and public transfers by number of months unemployed by study cohort American cohort (PSID) German cohort (GSOEP) Age 47.4 39.5 40.3 41.3 40.7 40.8 41.6 41.7 42.4 42.4 43.4 44.4 47.7 Male 46.9 53.9 51.5 50.9 47.1 43.7 47.2 41.1 40.5 42.8 41.1 41.3 42.9 Unemp at survey 0.8 18.8 34.4 37.9 43.6 43.8 39.0 48.5 48.2 49.6 56.3 57.9 86 Lifetime unemp 2.3 14.5 18.3 20.5 25.4 27.5 29.6 34.6 35.5 36.5 41.4 42.2 54.3 Unemp comp % 0.9 76.1 82.6 85.8 88.0 88.8 89.4 89.0 93.8 91.3 88.8 92.0 82.9 Unemp comp € 55.0 1652.3 2280.9 2687.4 3446 3950.7 4627.4 4654.3 5910.6 5927.3 6077.9 6650.8 6543.7 Public transfers € 1755.3 4212.1 4800.6 5140.6 6203.1 6684 7452.3 7658 8917.8 8741 8951.5 9871.4 9658.9 Ind lbr inc (t-1) € 19590.5 13986.6 12855.1 13632.3 11328.4 11159.8 12494.4 9623.4 10517.9 11302.8 9708.8 10200.6 4826.5 Ind lbr inc (t0) € 19910.7 15648 12790.5 12135.7 9349.2 8204.7 7635.4 5586.5 4412.1 3869.8 2727.4 2612.8 223.6 Hhld inc (t-1) € 33599.2 27859.2 27305.1 27752.5 27217.4 26808.8 27151 27039.2 26881.4 27717.6 26481.7 26795.8 23235.1 Hhld inc (t0) € 33723.6 28380.6 27326.4 27259.5 26914.8 26522.4 26240.6 27258.2 25836.8 25045.4 24943.8 25558 21567.5 Uecomp/Hhld inc 0.2 6.8 9.6 11.8 15.0 17.7 20.6 22.3 26.9 28.5 29.5 31.5 38.2 Pub Trans/Hhld inc 6.2 17.0 20.2 21.9 26.0 29.7 33.0 34.6 39.8 41.8 43.0 46.7 55.9 N (Person Years) 128696 1162 1217 1289 927 707 837 574 533 635 428 327 3425 % (Person Years) 94.1 0.8 0.9 0.9 0.7 0.5 0.6 0.4 0.4 0.5 0.3 0.2 2.4 Notes: Number of months unemployed is based on the number of months unemployed reported for the year prior to the survey year. Unemployment at survey has been brought forward so that is for the same year as the months unemployed measure. Dollars $ and Euros € are in 2005 values. Months unemployed zero one two three four five six seven eight nine ten eleven twelve Age 43.4 34.2 34.1 35.2 35.4 35.2 35.7 35.8 34.8 35.2 35.5 33.9 37.0 Male 42.7 49.5 48.3 47.2 48.0 50.7 47.2 40.9 42.3 39.7 40.4 39.1 43.4 Unemp at survey 3.6 13.3 18.1 23.0 27.7 32.9 41.1 45.8 49.0 54.2 51.1 48.5 56.0 Lifetime unemp 2.9 9.5 13.3 17.5 21.2 24.5 27.6 27.6 33.4 36.1 40.7 43.9 49.1 Unemp comp % 1.5 23.2 30.3 36.1 36.1 35.9 35.0 32.2 31.6 25.1 21.8 23.7 12.0 Unemp comp $ 41.5 313.2 601.9 962.9 1263.4 1594.6 1707.5 1666.5 1669.7 1566.4 1315.5 1529.8 826.0 Public transfers $ 914.3 1557.3 1852.6 2342.1 2752.8 3013.6 3594.5 3484.4 3895.8 4102.2 4208.4 4491.5 4918.0 Ind lbr inc (t-1) $ 27464.5 21309.0 20510.0 18614.8 17086.9 17685.0 16218.9 14622.0 14328.7 12011.3 10293.4 11318.5 6531.0 Ind lbr inc (t0) $ 28601.7 21957.2 20297.9 17786.3 15194.2 14191.2 13347.4 10074.8 8201.7 6315.0 4696.8 4232.0 1226.3 Hhld inc (t-1) $ 44762.9 32302.5 30889.5 30613.5 27308.2 29094.7 28341.2 28260.2 28983.5 25082.5 23112.7 21415.6 18987.2 Hhld inc (t0) $ 46472.9 32464.1 31168.0 29996.8 27715.5 26856.9 26569.0 27295.2 25175.9 22509.7 19379.1 18122.5 16064.5 Uecomp/Hhld inc 0.2 1.2 1.9 3.4 4.4 5.6 6.4 6.3 6.6 7.7 6.6 7.8 5.6 Pub Trans/Hhld inc 5.6 9.8 10.7 12.6 14.0 15.8 18.4 18.3 23.4 27.3 30.0 34.9 45.5 N (Person Years) 117757 3190 2491 1840 1238 933 905 670 575 506 381 294 1322 % (Person Years) 89.1 2.4 1.9 1.4 0.9 0.7 0.5 0.4 0.4 0.4 0.3 0.2 1.0 116 Table 4.3: Relative risk of dying by labour force status at the time of the survey for the German cohort, 1986-2004 Age & Sex Demo- graphics Household Income Education Occupation Health Unemployed 2.145*** 2.002*** 1.795** 1.734** 1.676** 1.417 [1.501,3.066] [1.395,2.871] [1.248,2.580] [1.205,2.495] [1.163,2.413] [0.984,2.042] Not Working 3.007*** 3.153*** 2.894*** 2.841*** 2.452*** 2.005*** [2.443,3.703] [2.555,3.891] [2.341,3.579] [2.295,3.517] [1.964,3.062] [1.605,2.505] Age 1.057*** 1.051*** 1.052*** 1.051*** 1.047*** 1.045*** [1.049,1.065] [1.042,1.060] [1.043,1.061] [1.042,1.060] [1.037,1.056] [1.035,1.054] Male 2.121*** 2.074*** 2.182*** 2.289*** 2.599*** 2.466*** [1.843,2.441] [1.726,2.493] [1.812,2.627] [1.894,2.768] [2.124,3.181] [2.017,3.015] East German 1.205* 1.176 1.253* 1.221* 1.280** [1.017,1.428] [0.992,1.395] [1.053,1.492] [1.026,1.454] [1.071,1.529] Immigrant 0.856 0.830 0.819 0.778 0.814 [0.624,1.173] [0.606,1.138] [0.597,1.123] [0.566,1.068] [0.593,1.117] Spouse 0.906 0.950 0.919 0.887 0.941 [0.737,1.113] [0.772,1.169] [0.746,1.132] [0.717,1.096] [0.763,1.162] Single 1.395 1.376 1.390 1.358 1.487* [0.985,1.975] [0.971,1.949] [0.982,1.969] [0.956,1.928] [1.045,2.115] Div or Sep 1.500** 1.408* 1.411* 1.425* 1.359* [1.134,1.982] [1.062,1.868] [1.063,1.872] [1.072,1.894] [1.021,1.808] Widowed 1.189 1.171 1.128 1.093 1.120 [0.927,1.525] [0.911,1.505] [0.876,1.452] [0.847,1.410] [0.867,1.445] Household size(#) 1.103 1.195** 1.180** 1.171** 1.171** [0.990,1.228] [1.070,1.334] [1.056,1.319] [1.047,1.310] [1.046,1.311] Children(#) 0.725*** 0.686*** 0.695*** 0.683*** 0.715*** [0.601,0.874] [0.569,0.828] [0.576,0.839] [0.566,0.824] [0.593,0.863] Hhld income (t-1,1og) 0.779*** 0.794*** 0.801*** 0.825*** [0.721,0.841] [0.730,0.864] [0.733,0.874] [0.752,0.906] Min_skill 1.540*** 1.464** 1.365* [1.196,1.982] [1.113,1.927] [1.036,1.800] Med_skill 1.257* 1.230 1.186 [1.015,1.556] [0.976,1.551] [0.939,1.499] No occupation 1.735*** 1.582** [1.277,2.355] [1.163,2.152] Bus/sales occ 1.078 1.079 [0.793,1.466] [0.793,1.469] Services occ 1.025 0.991 [0.724,1.451] [0.699,1.406] Agr/For/Min occ 1.142 1.103 [0.694,1.881] [0.669,1.819] Manufacturing occ 1.027 0.995 117 [0.748,1.409] [0.724,1.367] Hlth sat good (t-1) 0.919 [0.676,1.251] Hlth sat satisfied (t-1) 1.512** [1.123,2.035] Hlth sat poor (t-1) 1.895*** [1.381,2.601] Hlth sat bad (t-1) 3.808*** [2.772,5.231] Disabled (t-1) 1.426*** [1.217,1.671] Observations 117123 117123 117123 117123 117123 117123 AIC 9956.3 9942.0 9917.1 9909.3 9882.2 9658.0 BIC 10004.7 10067.7 10052.5 10064.1 10085.3 9909.4 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 1. Complementary log-log models. 2. Includes the log of years followed as an exposure offset. 3. Employed, Female, West German, German born, Head, Married, High skill, Professional/Management occupations, Excellent health satisfaction, and Not disabled are the respective reference categories for the categorical variables. 118 Table 4.4: Relative risk of dying by labour force status at the time of the survey for the American cohort, 1986-2004 Age & Sex Demo- graphics Household Income Education Occupation Health Died after final yr Unemployed 3.661*** 2.888*** 2.636*** 2.565*** 2.454*** 2.353*** [2.599,5.158] [2.042,4.085] [1.859,3.739] [1.808,3.639] [1.727,3.486] [1.656,3.343] Not Working 4.239*** 4.086*** 3.762*** 3.618*** 3.325*** 2.428*** [3.486,5.153] [3.353,4.978] [3.080,4.595] [2.959,4.425] [2.694,4.104] [1.956,3.014] Age 1.054*** 1.053*** 1.054*** 1.054*** 1.054*** 1.052*** [1.048,1.061] [1.045,1.061] [1.046,1.062] [1.046,1.062] [1.046,1.062] [1.044,1.060] Male 1.860*** 1.782*** 1.834*** 1.839*** 1.752*** 1.722*** [1.620,2.134] [1.450,2.189] [1.495,2.249] [1.499,2.257] [1.416,2.169] [1.397,2.125] Black 1.695*** 1.543*** 1.448*** 1.423*** 1.203* [1.457,1.970] [1.320,1.803] [1.229,1.705] [1.206,1.678] [1.021,1.419] Other 0.973 0.930 0.910 0.881 0.864 [0.693,1.367] [0.662,1.308] [0.646,1.281] [0.625,1.241] [0.613,1.218] Spouse 0.971* 0.976 0.975 0.973* 0.981 [0.946,0.998] [0.950,1.002] [0.950,1.002] [0.947,1.000] [0.955,1.007] Single 1.643*** 1.531** 1.568** 1.563** 1.596** [1.234,2.188] [1.148,2.041] [1.175,2.092] [1.170,2.087] [1.201,2.122] Div or Sep 1.198 1.148 1.127 1.096 1.049 [0.935,1.537] [0.896,1.470] [0.880,1.444] [0.855,1.406] [0.819,1.343] Widowed 1.401** 1.323* 1.325* 1.348* 1.239 [1.090,1.802] [1.028,1.703] [1.029,1.706] [1.046,1.736] [0.965,1.591] Household size(#) 0.981 1.014 1.003 1.001 0.985 [0.893,1.078] [0.923,1.115] [0.912,1.103] [0.910,1.101] [0.895,1.084] Children(#) 0.857* 0.839* 0.849* 0.848* 0.881 [0.741,0.991] [0.726,0.970] [0.734,0.982] [0.733,0.981] [0.762,1.019] Hhld income (t-1,1og) 0.872*** 0.886*** 0.891*** 0.925* [0.826,0.920] [0.837,0.938] [0.840,0.945] [0.864,0.990] Educ - Minimum skill 1.516*** 1.314 0.916 [1.183,1.943] [0.998,1.732] [0.691,1.214] Educ - Medium skill 1.406** 1.283* 1.109 [1.119,1.766] [1.002,1.642] [0.864,1.422] No occupation 1.318 1.115 [0.995,1.745] [0.840,1.479] Bus/sales occ 0.895 0.924 [0.677,1.185] [0.699,1.223] Services occ 0.992 1.003 [0.725,1.357] [0.733,1.372] Agr/For/Min occ 1.096 1.177 [0.726,1.656] [0.778,1.781] Manufacturing occ 1.258 1.213 119 [0.939,1.684] [0.906,1.625] Very good SRHS (t-1) 1.468* [1.034,2.084] Good SRHS (t-1) 2.028*** [1.447,2.841] Fair SRHS (t-1) 3.810*** [2.681,5.414] Poor SRHS (t-1) 7.249*** [4.976,10.56 0] Disabled (t-1) 1.103 [0.926,1.315] Observations 99175 99175 99175 99175 99175 99175 AIC 9308.3 9193.1 9175.2 9167.2 9161.4 8907.9 BIC 9355.8 9316.6 9308.2 9319.3 9361.0 9155.0 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 1. Complementary log-log models. 2. Includes the log of years followed as an exposure offset. 3. Employed, Female, White/Other race, Head, Married, High skill, Professional/Management occupations, Excellent self-reported health status, and not disabled are the respective reference categories for the categorical variables. 120 Table 4.5: Relative risk of dying by labour force status at the time of the survey, adjusted for potential confounders, German and American cohorts, 1986-2004 Demog GSOEP Demog PSID All GSOEP All PSID Health GSOEP Health PSID Unemployed 2.002*** 2.888*** 1.676** 2.454*** 1.417 2.353*** [1.395,2.871] [2.042,4.085] [1.163,2.413] [1.727,3.486] [0.984,2.042] [1.656,3.343] Not Working 3.153*** 4.086*** 2.452*** 3.325*** 2.005*** 2.428*** [2.555,3.891] [3.353,4.978] [1.964,3.062] [2.694,4.104] [1.605,2.505] [1.956,3.014] Observations 117123 99175 117123 99175 117123 99175 AIC 9942.0 9193.1 9882.2 9161.4 9658.0 8907.9 BIC 10067.7 9316.6 10085.3 9361.0 9909.4 9155.0 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 The ‗Demog‘ model includes all demographic variables (e.g., age, sex, race, marital status, household characteristics. The ‗All‘ model includes all variables except for lagged health status and the ‗Health‘ model adds lagged health status. Table 4.6: Relative risk of dying by labour force status in the year prior to the survey, adjusted for potential confounders, German and American cohorts, 1986-2004 Demog GSOEP Demog PSID All GSOEP All PSID Health GSOEP Health PSID Died after final yr Number of months unemployed 1.084*** 1.113*** 1.066*** 1.094*** 1.049** 1.090*** [1.047,1.122] [1.069,1.158] [1.029,1.103] [1.050,1.139] [1.013,1.086] [1.047,1.135] Number of months not working 1.117*** 1.104*** 1.092*** 1.084*** 1.073*** 1.061*** [1.096,1.138] [1.087,1.121] [1.070,1.114] [1.066,1.102] [1.052,1.095] [1.043,1.079] Observations 115649 99129 115649 99129 115649 99129 AIC 9736.1 9222.3 9683.2 9187.6 9536.7 9020.2 BIC 9861.6 9345.9 9886.0 9387.2 9787.8 9267.3 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 Table 4.7: Relative risk of dying by cumulative labour force status, adjusted for potential confounders, German and American cohorts, 1986-2004 Demog GSOEP Demog PSID All GSOEP All PSID Health GSOEP Health PSID Died after final yr % of yrs followed unemployed 1.014*** 1.020*** 1.009*** 1.016*** 1.007* 1.016*** [1.009,1.019] [1.013,1.028] [1.004,1.015] [1.008,1.024] [1.001,1.012] [1.008,1.024] % of yrs followed not working 1.014*** 1.013*** 1.010*** 1.012*** 1.007*** 1.007*** [1.011,1.016] [1.011,1.015] [1.007,1.013] [1.009,1.015] [1.004,1.011] [1.004,1.010] Observations 116877 99129 116877 99129 116877 99129 AIC 9902.1 9244.7 9877.3 9225.3 9717.2 9047.2 BIC 10027.8 9368.2 10080.3 9424.9 9968.6 9294.4 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 121 Table 4.8: Relative risk of dying by labour force status at time of survey stratified by sex, adjusted for potential confounders, German and American cohorts, 1986-2004 Males All GSOEP Males Health GSOEP Males All PSID Males Health PSID Females All GSOEP Females Health GSOEP Females All PSID Females Health PSID Unemployed 1.853** 1.555* 2.363*** 2.248*** 1.192 1.041 2.590*** 2.570** [1.208,2.843] [1.013,2.387] [1.504,3.714] [1.429,3.536] [0.581,2.446] [0.507,2.137] [1.475,4.547] [1.465,4.509] Not Working 2.857*** 2.220*** 3.156*** 2.280*** 1.892*** 1.668** 3.599*** 2.745*** [2.142,3.812] [1.662,2.965] [2.377,4.190] [1.704,3.050] [1.295,2.765] [1.143,2.434] [2.583,5.015] [1.957,3.851] Observations 54741 54741 42945 42945 62382 62382 56230 56230 AIC 5672.8 5569.8 4716.1 4588.7 4223.3 4108.0 4468.3 4347.1 BIC 5851.0 5792.6 4880.8 4796.7 4404.1 4334.0 4647.1 4570.5 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 The ‗All‘ model includes all variables except for lagged health status and the ‗Health‘ model adds lagged health status. Table 4.9: Relative risk of dying by labour force status in the year prior to the survey stratified by sex, adjusted for potential confounders, German and American cohorts, 1986-2004 Males All GSOEP Males Health GSOEP Males All PSID Males Health PSID Females All GSOEP Females Health GSOEP Females All PSID Females Health PSID Months unemployed 1.078*** 1.059** 1.120*** 1.108*** 1.028 1.017 1.044 1.051 [1.034,1.123] [1.015,1.104] [1.067,1.175] [1.056,1.163] [0.962,1.098] [0.952,1.086] [0.967,1.128] [0.974,1.135] Month not working 1.114*** 1.090*** 1.080*** 1.051*** 1.057** 1.046** 1.087*** 1.072*** [1.085,1.144] [1.062,1.120] [1.056,1.104] [1.027,1.075] [1.022,1.095] [1.011,1.083] [1.059,1.114] [1.045,1.099] Observations 54115 54115 42914 42914 61534 61534 56215 56215 AIC 5489.5 5420.5 4732.4 4642.4 4204.0 4134.1 4474.9 4403.3 BIC 5667.5 5643.0 4897.1 4850.4 4384.6 4359.8 4653.6 4626.8 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 122 Table 4.10: Relative risk of dying by cumulative labour force status stratified by sex, adjusted for potential confounders, German and American cohorts, 1986-2004 Males All GSOEP Males Health GSOEP Males All PSID Males Health PSID Females All GSOEP Females Health GSOEP Females All PSID Females Health PSID % of yrs followed 1.012*** 1.009* 1.024*** 1.023*** 1.000 0.998 0.998 1.001 unemployed [1.005,1.019] [1.002,1.016] [1.015,1.033] [1.014,1.032] [0.990,1.011] [0.988,1.009] [0.982,1.014] [0.985,1.017] % of yrs followed not 1.013*** 1.010*** 1.012*** 1.005* 1.003 1.002 1.011*** 1.008*** working [1.009,1.017] [1.006,1.015] [1.007,1.016] [1.001,1.010] [0.997,1.009] [0.996,1.008] [1.006,1.016] [1.004,1.013] Observations 54606 54606 42914 42914 62271 62271 56215 56215 AIC 5625.5 5549.3 4743.2 4647.3 4260.4 4183.2 4496.1 4420.5 BIC 5803.7 5772.0 4907.9 4855.3 4441.2 4409.1 4674.9 4644.0 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 The ‗All‘ model includes all variables except for lagged health status and the ‗Health‘ model adds lagged health status. Table 4.11: Relative risk of dying, by labour force status at time of survey stratified by educational skill level, adjusted for potential confounders, German and American cohorts, 1986-2004, health (t-1) model Full GSOEP Min skill GSOEP Med Skill GSOEP High Skill GSOEP Full PSID Min skill PSID Med Skill PSID High Skill PSID Unemployed 1.417 1.629 1.086 2.983* 2.353*** 2.554** 2.373*** 1.014 [0.984,2.042] [0.718,3.692] [0.678,1.739] [1.272,6.993] [1.656,3.343] [1.395,4.677] [1.487,3.786] [0.240,4.283] Not Working 2.005*** 2.354** 1.691*** 2.800*** 2.428*** 2.406*** 2.494*** 1.698 [1.605,2.505] [1.394,3.974] [1.282,2.229] [1.564,5.010] [1.956,3.014] [1.635,3.540] [1.850,3.363] [0.949,3.039] Observations 117123 17359 74928 24836 99175 20546 57291 21338 AIC 9658.0 2364.1 6000.8 1341.1 8907.9 3430.6 4350.4 1150.5 BIC 9909.4 2542.6 6213.0 1527.9 9155.0 3613.0 4556.4 1333.8 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 The ‗Full‘ model refers to the complete unstratified cohort and the ‗Min Skill‘, ‗Med Skill‘ and ‗High Skill‘ models refer to the respective education stratified cohorts. 123 Table 4.12: Relative risk of dying by labour status in the year prior to the survey stratified by educational skill level, adjusted for potential confounders, German and American cohorts, 1986-2004, health (t-1) model Full GSOEP Min skill GSOEP Med Skill GSOEP High Skill GSOEP Full PSID Min skill PSID Med Skill PSID High Skill PSID Died after final yr Months unemployed 1.049** 1.014 1.034 1.152*** 1.090*** 1.047 1.121*** 0.916 [1.013,1.086] [0.933,1.102] [0.990,1.080] [1.061,1.250] [1.047,1.135] [0.971,1.130] [1.067,1.177] [0.701,1.197] Months not working 1.073*** 1.056* 1.062*** 1.140*** 1.061*** 1.058*** 1.060*** 1.042 [1.052,1.095] [1.009,1.106] [1.036,1.089] [1.080,1.203] [1.043,1.079] [1.029,1.088] [1.035,1.085] [0.993,1.094] Observations 115649 17137 73980 24532 99129 20527 57263 21339 AIC 9536.7 2349.8 5932.9 1295.0 9020.2 3452.5 4413.5 1160.8 BIC 9787.8 2528.0 6144.8 1481.5 9267.3 3634.8 4619.5 1344.0 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 The ‗All‘ model includes all variables except for lagged health status and the ‗Health‘ model adds lagged health status. The ‗Full‘ model refers to the complete unstratified cohort and the ‗Min Skill‘, ‗Med Skill‘ and ‗High Skill‘ models refer to the respective education stratified cohorts. Table 4.13: Relative risk of dying by cumulative labour force status stratified by educational skill level, adjusted for potential confounders, German and American cohorts, 1986-2004, health (t-1) model Full GSOEP Min skill GSOEP Med Skill GSOEP High Skill GSOEP Full PSID Min skill PSID Med Skill PSID High Skill PSID Died after final yr % of yrs followed 1.007* 1.006 1.004 1.014 1.016*** 1.010 1.019*** 1.012 unemployed [1.001,1.012] [0.996,1.016] [0.996,1.011] [0.998,1.031] [1.008,1.024] [0.998,1.022] [1.008,1.030] [0.976,1.049] % of yrs followed not 1.007*** 1.002 1.009*** 1.007 1.007*** 1.007** 1.006* 1.005 working [1.004,1.011] [0.995,1.009] [1.005,1.013] [0.999,1.016] [1.004,1.010] [1.002,1.012] [1.001,1.011] [0.995,1.015] Observations 116877 17342 74749 24786 99129 20527 57263 21339 AIC 9717.2 2418.9 5993.2 1349.2 9047.2 3460.6 4432.3 1163.2 BIC 9968.6 2597.4 6205.3 1535.9 9294.4 3643.0 4638.3 1346.4 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 124 Table 4.14: Relative risk of dying by labour force status at time of survey with exclusions for baseline health and labour force status, adjusted for potential confounders, German and American cohorts, 1986-2004, health (t-1) model Full GSOEP Health Good GSOEP Working GSOEP Full PSID Health Good PSID Working PSID Died after final yr Unemployed 1.417 1.658* 1.718* 2.353*** 2.987*** 3.414*** [0.984,2.042] [1.064,2.583] [1.083,2.728] [1.656,3.343] [2.014,4.430] [2.152,5.416] Not Working 2.005*** 2.406*** 2.129*** 2.428*** 2.794*** 3.557*** [1.605,2.505] [1.842,3.142] [1.597,2.838] [1.956,3.014] [2.141,3.645] [2.704,4.679] Observations 117123 95618 76540 99175 79683 63669 AIC 9658.0 6463.8 4803.6 8907.9 5117.8 4367.3 BIC 9909.4 6662.6 4997.7 9155.0 5312.8 4548.6 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 Table 4.15: Relative risk of dying by labour status in the year prior to the survey with exclusions for baseline health and labour force status, adjusted for potential confounders, German and American cohorts, 1986-2004, health (t-1) model Full GSOEP Health Good GSOEP Working GSOEP Full PSID Health Good PSID Working PSID Died after final yr Number of months unemployed 1.049** 1.055* 1.048* 1.090*** 1.086** 1.099** [1.013,1.086] [1.012,1.099] [1.001,1.097] [1.047,1.135] [1.032,1.142] [1.034,1.169] Number of month not working 1.073*** 1.069*** 1.058*** 1.061*** 1.049*** 1.058*** [1.052,1.095] [1.044,1.096] [1.030,1.086] [1.043,1.079] [1.026,1.072] [1.035,1.082] Observations 115649 94413 75612 99129 79674 63667 AIC 9536.7 6243.6 4649.3 9020.2 5100.1 4339.1 BIC 9787.8 6489.5 4889.4 9267.3 5341.5 4565.7 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 Table 4.16: Relative risk of dying by cumulative labour force status stratified with exclusions for baseline health and labour force status, adjusted for potential confounders, German and American cohorts, 1986-2004, health (t-1) model Full GSOEP Health Good GSOEP Working GSOEP Full PSID Health Good PSID Working PSID Died after final yr % of yrs followed unemployed 1.007* 1.006 1.007 1.016*** 1.019*** 1.027*** [1.001,1.012] [0.999,1.014] [0.995,1.018] [1.008,1.024] [1.009,1.029] [1.011,1.044] % of yrs followed not working 1.007*** 1.007** 1.009*** 1.007*** 1.006** 1.007** [1.004,1.011] [1.003,1.011] [1.004,1.014] [1.004,1.010] [1.002,1.011] [1.002,1.012] Observations 116877 95425 76377 99129 79674 63667 AIC 9717.2 6343.5 4722.8 9047.2 5105.5 4351.9 BIC 9968.6 6589.6 4963.2 9294.4 5346.9 4578.4 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 125 Table 4.17: Relative risk of dying by all three labour force status variables stratified by East and West Germany, adjusted for potential confounders, German cohort, 1986-2004, health (t-1) model East German West German East German West German East German West German Died after final yr Unemployed 2.079** 0.888 [1.209,3.574] [0.500,1.576] Not Working 2.051** 2.041*** [1.217,3.457] [1.594,2.612] Months unemployed 1.082** 1.027 [1.023,1.145] [0.980,1.077] Months not working 1.104*** 1.068*** [1.053,1.157] [1.045,1.092] % of yrs followed 1.007 1.005 unemployed [0.998,1.017] [0.998,1.013] % of yrs followed not 1.005 1.009*** working [0.998,1.012] [1.005,1.013] Observations 27920 89412 27601 88252 27780 89306 AIC 2098.6 7692.0 2063.8 7599.8 2081.7 7763.5 BIC 2296.3 7917.6 2261.2 7825.1 2279.2 7989.1 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 Table 4.18: Relative risk of dying by all three labour force status variables stratified by race, adjusted for potential confounders, American cohort, 1986-2004, health (t-1) model Black White/Other Black White/Other Black White/Other Unemployed 2.456*** 2.261** [1.521,3.964] [1.315,3.887] Not Working 3.084*** 1.996*** [2.152,4.422] [1.519,2.623] Months unemployed 1.085** 1.108*** [1.027,1.146] [1.043,1.178] Months not working 1.092*** 1.038*** [1.061,1.123] [1.016,1.060] % of yrs followed 1.013* 1.019** unemployed [1.003,1.023] [1.006,1.032] % of yrs followed not 1.009** 1.005* working [1.003,1.014] [1.001,1.009] Observations 34155 65020 34122 65007 34122 65007 AIC 3685.9 5208.8 3731.8 5266.2 3758.8 5273.4 BIC 3888.5 5426.8 3934.3 5484.2 3961.3 5491.4 Exponentiated coefficients; 95% confidence intervals in brackets * p < 0.05, ** p < 0.01, *** p < 0.001 126 Table 4.19: Predicted hazard of dying in a given year by current labour force status, skill level and age, based on the skill stratified estimates Country Labour Force Status Skill Level Age 20 Age 30 Age 40 Age 50 Age 60 Age 70 Ratio at Age 40 Germany Employed High Skill 0.0005 0.0008 0.0012 0.0017 0.0026 0.0039 1.00 United States Employed High Skill 0.0004 0.0007 0.0012 0.0019 0.0031 0.0050 1.00 United States Unemployed High Skill 0.0004 0.0007 0.0012 0.0019 0.0031 0.0050 1.01 Germany Employed Medium Skill 0.0006 0.0010 0.0016 0.0026 0.0043 0.0069 1.38 United States Employed Medium Skill 0.0005 0.0009 0.0016 0.0029 0.0053 0.0097 1.39 Germany Unemployed Medium Skill 0.0007 0.0011 0.0018 0.0028 0.0046 0.0075 1.50 Germany Employed Minimum Skill 0.0008 0.0012 0.0020 0.0032 0.0052 0.0084 1.70 United States Not Working High Skill 0.0010 0.0014 0.0020 0.0029 0.0042 0.0060 1.71 Germany Not Working High Skill 0.0010 0.0017 0.0027 0.0044 0.0072 0.0117 2.34 United States Employed Minimum Skill 0.0013 0.0020 0.0032 0.0049 0.0075 0.0115 2.70 Germany Unemployed Minimum Skill 0.0016 0.0023 0.0033 0.0047 0.0068 0.0098 2.79 Germany Not Working High Skill 0.0014 0.0022 0.0033 0.0049 0.0073 0.0108 2.80 Germany Unemployed High Skill 0.0015 0.0023 0.0035 0.0052 0.0077 0.0115 2.98 United States Unemployed Medium Skill 0.0012 0.0021 0.0038 0.0070 0.0127 0.0231 3.29 United States Not Working Medium Skill 0.0012 0.0022 0.0040 0.0073 0.0133 0.0243 3.46 Germany Not Working Minimum Skill 0.0023 0.0033 0.0047 0.0068 0.0098 0.0141 4.03 United States Not Working Minimum Skill 0.0032 0.0049 0.0076 0.0117 0.0180 0.0278 6.50 United States Unemployed Minimum Skill 0.0034 0.0052 0.0081 0.0124 0.0191 0.0295 6.90 Notes: The hazard of dying is evaluated at the mean of the other covariates in each of the stratified models (e.g. the hazard of dying for the minimum skilled is based on the mean of the covariates of the minimum skilled in each survey and so forth). 127 Chapter 5: Unemployment and Self-rated Health: A Study of Canada, Germany and the United States 5.1: Introduction This chapter builds on the mortality analysis presented in Chapter 4. Canada, an LME country with higher levels of unemployment and employment protection than the United States, is introduced as a middle case between Germany and Canada. Other research has established that socio-economic gradients are flatter in Canada compared to the United States (see section 2.2.2), although no research has conducted a direct comparison of unemployment-related health inequalities between these two countries. This study enables a more detailed exploration of whether the overall institutional context matters to the health of the unemployed, and whether within LMEs whether there can be effective mediation of the unemployment-health relationship. Self-reported health status (SRHS) is the dependent variable in this study. While SRHS is strongly associated with mortality (see section 5.3.2), it enables the consideration of the effect of unemployment on morbidity. Self-reported health status is also a more ‗powerful‘ variable as there is more variation in it across years compared to mortality. In the mortality study only about eight percent of the cohorts died. With the greater variation in SRHS it possible to directly examine whether there are differences in the health of the unemployed who receive unemployment compensation and those that do not. Self-reported health status may also be a better measure of the test of the psychosocial effects of unemployment on health because it may capture aspects of morbidity related to psychological and function. 5.2: Research Objectives The hypotheses articulated in Chapter 2 lead to the following research objectives: - To examine the relationship between unemployment and self-reported health status in three working-age cohorts of Germans, Canadian and Americans, and whether and how this relationship changes by study country; - To examine if the receipt of unemployment benefits modifies the relationship between unemployment and self-reported health status; and, - To examine if these relationships are modified by skill level or gender. 128 The specific hypothesis for the unemployment and self-reported health status study are similar to the mortality study, but include hypotheses related to the receipt of unemployment benefits: 1. The association between unemployment and SRHS will be weaker in Germany compared to the United States. It is unclear where the strength of the association in Canada will rank among the three countries, given that Canada has higher levels of unemployment and employment protection than the United States, but lower levels of long-term unemployed than Germany. 2. The receipt of unemployment compensation will mediate the effect of unemployment on health within countries. The higher prevalence of the long term unemployed in Germany compared to the LME countries, however, may confound this comparison. 3. There will continue to be effect modification by educational status that will vary by study country. The relationship between unemployment and SRHS will be weaker for the minimally skilled and medium skilled in Germany compared to their counterparts in the United States and Canada, with the minimally skilled in the United States being especially disadvantaged. Effect modification by skill level will be similar in Canada and the United States given the similarities in these countries educational systems. 4. The relationship between unemployment and SRHS will be stronger for women than the in unemployment and mortality study and there will continue to be stronger associations in Canada and the United States compared to Germany. 5.3: Methods 5.3.1: Self-reported Health Status 184.108.40.206: Validity of Self-reported Health Status A number of reviews have found that SRHS is predictive of mortality and exhibits a robust gradient (i.e., poor health is more predictive than fair, and fair more than good) (Benyamini and Idler 1999; DeSalvo et al. 2006; Idler and Benyamini 1997). The reviews by Idler and Benyamini established a strong link between SRHS and mortality, while more recently, DeSalvo and colleagues conducted a systemic review of all published studies between 1996 and 2003 that examined the association between general self-rated status and mortality. Based on a pooled analysis from the 22 cohort studies that met their eligibility criteria, they report relative risks of 1.25 for good SRHS, 1.39 for fair SRHS, and 1.92 for poor SRHS when compared to excellent SRHS. 129 The relationship between SRHS and other health conditions is not as well established. Benjamins and colleagues (2004), report that SRHS is predictive of disease-specific mortality, but not for accidents, after linking the United States National Health Interview Survey to the National Death Registry, and Idler and colleagues (2004), drawing on the United States National Health and Nutrition Examination Survey (NHANES), report that poor SRHS is predictive of mortality in the presence of a previously diagnosed chronic condition, but not without. In contrast, Burstrom and Fredlund (2001), using the Swedish Survey of Living Conditions, report that SRHS was predictive of mortality regardless of the presence of a chronic condition. Other research has shown that the relationship between self-reported health status and mortality and other measures of health may vary by socioeconomic status (Burstrom and Fredlund 2001; Dowd and Zajacova 2007; Quesnel-Vallee 2007; Singh-Manoux et al. 2007) or by employment status (Kerkhofs and Lindeboom 1995). Quesnel-Valle (2007) argues that the differences in this relationship by socio-economic may be due to the underlying social safety net in that the differences seem to show a steeper gradient in the relationship between SRHS and mortality in the United States compared to other countries in her review. Burstrom and Fredlund (2001) also report a stronger relationship between self-reported health and mortality in individuals of high occupational standing compared to those of lower occupational standing. They attribute the differences in the relationship between occupational groupings to differences in the baseline risk of death between the categories (i.e., the relative risk changes among SES categories due to changes in the baseline risk, even though the absolute risk difference is similar across SES categories). The mortality analysis conducted in Chapter 4 can be expanded to directly address the robustness of SRHS as a measure of mortality and whether this relationship varies by socioeconomic status in the PSID. When current SRHS is used as a regressor, the relative risk of poor self-reported health status on mortality is 9.7, for fair it is 3.3, for good it is 1.7 and, for very good it is 1.0. 80 Consistent with Burstrom‘s findings when the SRHS-mortality relationship is stratified by educational status the strength of this relationship is strongest in the high skilled (the relative risk for poor is 13.9, for fair is 4.29, for good is 2.4 and for very good 1.5) and weakest in the minimally skilled (the relative risk for poor is 4.0, for fair is 2.7, for good is 1.7 and for very 80 Recall that I do not use current self-reported health status as a control variable in Chapter 4 as I want to enforce the temporal ordering between health status and labour force status in those models, however in this case current self-reported health status is the best measure. 130 good 1.3). In addition to Burnstrom‘s interpretation of why these risks may vary, this may also be due to differences in how SRHS is perceived across the spectrum of socio-economic status (Kerkhofs and Lindeboom 1995). These findings indicate that some caution should be applied in interpreting the differences in odds ratios across SES strata. 220.127.116.11: Modelling Self-reported Health Status Self-reported health status has most often been dichotomised as an outcome variable into poor or fair versus good, very good, or excellent (based upon a five category variable). The dichotomising of self-reported health status into a binary variable has the potential to ignore health dynamics within the collapsed categories (i.e., movement between excellent to very good or good and between fair and poor). Indeed when year-to-year transitions in SRHS are examined in the study cohorts, the transitions between excellent, very good and good are the most common. The ordered logit model, an extension of the standard logit model, enables the estimation of the probability of an individual being in a particular self-reported health state conditional on a set of observed characteristics and an additional set of threshold parameters which demarcate the logit probability density function into the underlying ordinal categories. From this, the odds ratio for a one unit change in any variable can be obtained. The proportional odds assumption of the logit model also extends to the ordered logit model in that the odds ratio is not dependent on threshold parameters or covariate values in the model. The ordered logit model assumes that the odds ratio will be invariant across the ordinal categories. In other words, the odds ratio of being in a worse SRHS category will be the same irrespective of how SRHS is dichotomised (e.g., the odds ratio of being in poor health versus fair, good, very good or excellent health will be the same as the odds ratio of being in poor or fair health versus good, very good or excellent health and so forth). To test the proportional odds assumption, the ordered logit results for the dynamic health model was compared with four sets of logit models that dichotomised self-reported health status across the four possible dichotomisations for each of the study cohorts (See appendix tables E1 to E3 for the results of these models). 131 There was a declining gradient in the estimated odds ratio as the dichotomisation moves up the ordinal scale, with the odds ratio of current unemployment compared to current employment ranging from between 1.7 (SLID), 2.4 (GSOEP) and 2.8 (PSID) for the poor versus fair or higher categorization to an odds ratio of 1.0 (PSID), 1.1 (GSOEP) and 1.1 (SLID) for the very good and lower versus excellent categorization. Of the three surveys, the SLID has the most similar odds ratio for current unemployment across the four logit models, but it is clear that for current unemployment the proportional odds assumption does not hold. Indeed, the ordinal logit odds ratio on unemployment is most similar to the poor, fair or good versus very good or excellent dichotomisation for the GSOEP and PSID and to the poor, fair, good or very good versus excellent for the SLID, suggesting that the ordinal logit specification would underestimate the odds of falling into the worst self-reported health statuses. Not all covariates in the model violated the proportional odds assumption, in particular, the log of household income and skill level tended to exhibit proportionality across the ordinal logit and four logit specifications. While there are a number of extensions of the ordinal logit model that allow for the relaxation of the proportional odds assumptions, 81 this study focuses on the two middle dichotomisations of poor or fair versus good, very good (PF/GVGE) or excellent and poor, fair or good versus very good and excellent (PFG/VGE) in logit models. This is done for two reasons. First, the PF/GVGE and PFG/VGE self-reported health status specifications span the part of the SRHS distribution where most of the health dynamics take place. Second, the logit estimation framework enables one to attend to other statistical issues in a comparable and parsimonious manner across the three surveys (e.g., an individual-level random effect to account for within subject correlation). 5.3.2: Statistical Methods 18.104.22.168: Description of Data and Study Cohort This chapter examines three labour status specifications: current unemployment, number of months unemployed in the year prior to the survey, and the interaction between current 81 For example, extensions have been developed that relax the proportional odds assumption by modeling the threshold parameters as a function of the covariates or additional variables (Jones, Rice, Basho d'Uva, and Balia 2007; Rabe-Hesketh and Skrondal 2008). Other possibilities include the generalized ordered model (Williams 2006) that is similar to performing multiple logit models as I did above, but with constraints to reduce the number of parameters estimated and the stereotype logistic model (Long and Freese 2006) that is related to the multinomial logit model. While all these models are feasible with my data, they introduce an additional set of constraints or parameterization that may vary across the three datasets. Accordingly to maximize comparability and parsimony across the three datasets, I adopt the two logit dichotomizations describe above of self-reported health status. 132 unemployment and the receipt of unemployment benefits. 82 As the Canadian data only has a maximum of six years a cumulative labour force status measure is not developed for the SLID and this measure is not used in the SRHS analysis. For the unemployment and unemployment benefit interaction models, the receipt of unemployment benefits was also included as a separate indicator variable as this variable represented the receipt of unemployment benefits at any time during the survey year. Accordingly, the reference group is those employed at the time of the survey and who reported no unemployment benefits in that year. To test for differences among the two groups of unemployed and the employed reference group three separate Wald tests were conducted: 1. Unemployed, no benefits = Employed, no benefits H1: βunemployed, no benefits = 0 2. Unemployed, benefits+ unemployment benefits = Employed, benefits H2: βunemployed, benefits + βunemployment benefits = 0 3. Unemployed, benefits + unemployment benefits = Unemployment, no benefits. H3: βunemployed, benefits + βunemployment benefits = βunemployed, no benefits The third test is a direct test of whether the unemployed in receipt of benefit are more likely to report better self-reported health status than the unemployed not in receipt of benefits. 22.214.171.124: Balanced Versus Unbalanced Cohort A balanced design means that individuals are present for every wave of follow-up, while an unbalanced design implies that individuals do not need to be present for every survey wave, can drop out of the study and can have gaps across survey waves. This study adopts an unbalanced design, which enables individuals to enter and leave the study cohort in different years, but with the restriction that individuals have a minimum of three contiguous years of survey data at baseline and for every subsequent year of follow-up two contiguous years of data. The restriction ensures that individuals have a minimum number of years to be present across all labour force status and model specifications (i.e., for individuals to be included in models that look at number of months unemployed in the dynamic health specification three years of data are required for the GSOEP and PSID – t-1 for lagged health status, t0 for covariates, and t+1 for number of months unemployed in the previous year). 82 Current unemployment and unemployment benefits are harmonized so that they refer to the same survey year. 133 In practise, given the sample construction of the three cohorts, the SLID cohort is closer to a balanced design, while the PSID and the GSOEP are more unbalanced. The SLID cohort has an average of 4.6 years of follow-up and 42% of individuals are present for a maximum of 6 years; by construction no individuals are present for the entire study period. The PSID cohort has an average of 10.8 years of follow-up and 40% are present for the entire study period and the GSOEP cohort has an average of 7.7 years of follow-up and 28% are present or the entire study period. Listwise deletion was used in the statistical models to ensure that all models, irrespective of measures used, had the same individuals and person years within study cohorts. For the German cohort, 738 (3.7%) members of the eligible cohort were excluded due to missing data yielding an analytic cohort of 19,029 individuals and 103,484 person years. For the Canadian cohort, 4,827 (6.9%) members of the eligible cohort were excluded due to missing data yielding an analytic cohort of 65,168 individuals and 217,530 person years, For the American cohort, 212 (2.2%) %) members of the eligible cohort were excluded due to missing data yielding an analytic cohort of 9545 individuals and 78,951 person years. 83 126.96.36.199: Estimation Strategy This study adopts the methodology outlined by Jones and colleagues (2007) to examine the dynamics of health in a longitudinal and panel data context. Two model specifications are examined – a static health model and a dynamic health model84 – and random effects logit estimation is used to estimate the odds of a transition into poor health, conditional on a set of fixed effects, lagged health status variables (the dynamic health model only). In sensitivity testing, the survey design is accounted for with the inclusion of a second random effect at the level of the primary sampling unit in the PSID and GSOEP (see appendix tables E4 and E5). 85 The static health model can be expressed as: 83 See Tables 3.3 to 3.5 for a more detailed development of the three cohorts. 84 I do not use a statistical specification to account for the problem of initial conditions (i.e., that exposures and heath prior to the observation period confound the observed relationship) (Wooldridge 2005), rather I argue that the exclusions of cohort members not working at baseline or not in good health at baseline is an acceptable test of whether prior health and working history confounds the observed relationship. 85 These models are very computationally intense given the large sample size and number of parameters. I was able to estimate these models for the PSID and SLID using a multi-core processor and STATA 10 MP. Even so, it took between four days to a week to estimate each model. I was not able to estimate the model with two random effects for the SLID cohort, given the processing limitations of the computers that housed the SLID data at the UBC Research Data Centre. 134 The model examines the relationship between the level of self-reported health status (yit) and labour force status (xit) conditional on a set of individual-level covariates (zit), and region and year fixed effects (wit ), a random effect (μi) to account for correlation within individuals across waves and an idiosyncratic error term ( . The model can be extended to included lagged health status, which can account for state dependence between health at time t0 and lagged health at t-1. Where δ is the lagged health (yit-1) coefficient, which can be viewed as the persistence of being in a particular health state between years and the effect of the fixed effects can be viewed as the likelihood of transitioning into the lower health state. The final model accounts for health selection through lagged health status, and unobservable heterogeneity (unexplained variance at the individual level) is dealt with through the random effect specification. These models are further supplemented, as was done in the mortality analysis, by excluding individuals whose baseline characteristics may indicate health selection or poor labour market outcomes prior to the start of follow-up. Random effects for region and year could also be parameterized, but for computational simplicity and given the large sample size relative to the number of fixed effect parameters to be estimated, the fixed effect formulation for region and year was used rather than the random effect formulation. 5.4: Results 5.4.1: Descriptive Statistics 188.8.131.52: Self-reported Health Status There are noticeable differences in the distribution of self-reported health status by labour force status and by country at baseline (Figure 5.1). Germans report the lowest levels of excellent self- reported health status and highest levels of fair or poor self-reported health status compared to Canadians or Americans across both the employed and unemployed categories. The unemployed report lower self-reported health status compared to the employed in all countries. The proportion of unemployed and employed reporting poor or fair health was 23% and 13% in Germany, 11% and 6% in Canada and 17% and 9% in the United States. While the risk 135 difference ranged from 5% in Canada to 10% in Germany, the relative risks were similar (1.8 for Germany and Canada and 1.9 for the United States). 86 184.108.40.206: Independent Variables There are significant differences in the distribution of covariates by labour force status and by study cohort that may confound the unemployment and health relationship. Table 5.1 describes the baseline descriptive statistics of the German, Canadian, and American cohorts by current labour force status. The German and Canadian cohorts are older than the American cohort, particularly for the unemployed who are on average 13 and 10 years older in the German and Canadian cohorts, respectively, compared to the American cohort. Men make up a greater proportion of the unemployed (58%) in Canada compared to Germany and the United States (45%), while the proportion of the employed who are men is similar. Unemployed Americans are more likely to be single than their counterparts in Germany or Canada, while the distribution of marital status is similar. Blacks are the majority of the unemployed in the American cohort (65%) and almost half of the unemployed are East Germans (45%) in the German cohort. Household size is similar across cohorts, but Canadians and Americans report more children than Germans. A larger proportion of the unemployed and those not working in the American and Canadian cohorts are minimum skilled (35%) compared to the German cohort (19%). The high skilled comprise a larger proportion of the employed in Germany (29%) and the Unites States (24%) compared to Canada (19%), whereas they are equally represented in the unemployed in Germany (11%) and Canada (10%), and less so in the United States (7%). Employed Germans are more likely to be of the professional and technical occupations than Canadians and Americans, while services and sales occupations are more represented in Canada and the United States. 87 The unemployed report about 60% of the household income of the employed in the German and Canadian cohorts, while the unemployed report half of the household income of the employed in the American cohort. The proportion of current unemployed who report receiving any 86 The difference in the underlying distribution points to the inherent subjectivity of the SRHS measure which may vary by cultural context and implies that comparing average levels of SRHS across countries may not be appropriate. 87 The occupational classification used here appears to more accurately capture the variation in the occupational structure among the three countries than in the mortality study. 136 unemployment compensation is 73% in Germany, 52% in Canada 88 , and 17% in United States. Average public transfers as a proportion of total household income was similar for the unemployed in Canada (31%) and Germany (33%), but smaller in the United States (17%). Unemployment compensation was a larger component of public transfers in Germany compared to Canada. Table 5.2 describes the relationship between months unemployed in the year previous to the survey and to demographic, other unemployment measures, and income and transfers across all survey years. The proportion of individuals with any unemployment is similar across cohorts, but slightly lower in Canada (9.3%), and Germany (10.3%) compared to the United States (12%). 89 The pattern of more months of unemployment in Germany, but a higher level of coverage and more generous public transfers compared to the United States was observed in the mortality study. Here Canada emerges as a middle case (Figures 5.2-5.5). The unemployed in Canada are less likely to report short term unemployment (zero to three months), but more likely to report being unemployed for the entire year than Americans, while the converse is true with Germans. In particular, the German unemployed (31%) are much more likely than the Canadian (15%) or American (9%) unemployed to be unemployed for the entire year (Figure 5.2). Overall the average number of months unemployed is 7.1, 5.2 and 4.5 months for the German, Canadian, and American cohorts, respectively. Unemployed Canadians are more likely to report receiving unemployment compensation across all unemployment-month profiles than Americans, with the proportion receiving benefits peaking at over 60% between months three to eight. In contrast, unemployment benefit coverage averages over 80% across all months unemployed in the German cohort, and never reaches 40% in the American cohort. Notably, Canadians (20%) and Americans (13%) unemployed for the entire year both report low likelihoods of receiving benefits (Figure 5.3). 90 Canadians and Germans also report similar declines in average household income compared to the continuously employed with this proportion being between 70% and 80% for one to eleven months, but declining to 54% for Canadians unemployed the entire year (Figure 5.4). For Americans there is 88 Canada‘s employment insurance system also provides maternity and seasonal fishers benefits, so this proportion is likely overstated. 89 These percentages are higher for the American and German SRHS cohorts compared to the cohort from the mortality analysis. The SRHS German cohort spans the period when Germany had a level of higher level of unemployment and excludes the 1980s when it had a lower period. Similarly the American cohort excludes the late 1990s during which the United States had low levels of unemployment. 90 These coverage rates are similar to those presented from population-based labour force surveys in Chapter 2 (see figure 2.1). 137 a larger immediate drop in household income and then a gradual decline such that the household income of those unemployed for the entire year is only 42% of the household income of the continuously employed. 91 As the duration of unemployment increases, unemployment benefits represent an increasing share of household income for the unemployed in Germany rising to 40% for those unemployed for 12 months. Unemployment benefits are never more than 20% or 10% household income in Canada and the United States. 5.4.2: Multivariate Results Multivariate results were estimated for both the static and dynamic health models and across the two formulations of SRHS. Overall the results for the labour force status variables are similar for the static and dynamic health models within each country (Tables 5.3 and 5.4). 92 The goodness- of-fit statistics – the AIC and BIC – indicate that the dynamic health model is the superior specification and accounts for much of the unexplained variance at the individual level. Similarly, the goodness-of-fit statistics indicate that the empirical specification (i.e., the choice of independent variables and their functional form) better explains variation in the PF/GVGE dichotomization of SRHS compared to the PFG/VGE dichotomization. Accordingly, the results are presented for dynamic health models and the PF/GVGE dichotomization of SRHS. Results from the other specifications are only presented where they differ from the dynamic health or PF/GVGE dichotomization. Complete results on the PFG/VGE dichotomization, however, can be found in the appendix to this chapter (Tables E6 to E14). 5.4.3: Full Cohort 220.127.116.11: Current Labour Force Status Current unemployment is associated with higher odds of fair or poor SRHS in all three cohorts in both the static and dynamic health models (Table 5.3). The odds ratio of fair or poor SRHS is 1.9 (95% CI: 1.7-2.1) for the static model and 1.7 (95% CI: 1.5-1.9) for the dynamic health model in the German cohort, 1.9 (95% CI: 1.7-2.2) and 1.5 (95% CI: 1.4-1.9) in the Canadian cohort, and 1.8 (95% CI: 1.6-2.0) and 1.7 (95% CI: 1.5-1.9) in the American cohort. 91 The results are similar if median measures of household income are used. 92 This is not the case for all the variables in the model. In particular there is large attenuation in the household income and education variables across the static and dynamic health specifications. Unlike the labour force status variables, these variables have a greater stability across years (education status can be considered a fixed effect even though I do not restrict it to be the same across waves) and are more highly correlated with prior health status than the labour force status variables. 138 When the good, fair or poor SRHS specification is examined there is also an association between unemployment and worse SRHS across both the static and dynamic health models in all three countries (Table 5.4). The odds ratio of good, fair or poor SRHS is 1.3 (95% CI: 1.3-1.4) for the static model and 1.2 (95% CI: 1.1-1.4) for the health dynamic model in the German cohort, 1.4 (95% CI:1.3-1.5) and 1.3 (95% CI:1.2-1.4) in the Canadian cohort, and 1.2 (95%CI:1.1-1.4) and 1.2 (95% CI:1.1-1.3) in the American cohort. 18.104.22.168: Results for the Other Variables The results of the other covariates from the final dynamic health model are summarized for SRHS dichotomized as PF/GVGE (Table 5.3) as follows: In all three countries age is negatively related to worse SRHS, while age squared is positively related, yielding a convex relationship in the odds by age. Being a man is protective (OR 0.87) in the German cohort, while there was no relationship by gender in the Canadian and American cohorts. Black (OR 2.9) and other (OR 1.8) ethnicities are associated with greater odds of reporting worse SRHS in the American cohort. Being single (OR 1.2) in the Canadian cohort or being divorced or separated in the Canadian (OR 1.2) or German (OR 1.3) cohorts is associated with a higher odds compared to being married. There is no association by marital status in the American cohort. Being a spouse is protective in the Canadian (OR 0.8) and American (OR 0.8) cohort, but not in the German cohort. Household size is negatively associated with poor SRHS (GSOEP (G): OR 1.03; SLID (S): OR 1.04; PSID (P): OR 1.05), while number of children is protective (G: OR 0.9; S: OR 0.8; P: OR 0.9). The log of household income is associated with worse health status in all three countries (G: OR 0.8; S: OR 0.9; P: OR 0.8). There is a consistent gradient in the odds of reporting worse SRHS by educational status, with a steeper gradient in the American cohort. The odds ratio for the minimum skilled are 1.6, 2.1 and 7.2 for the German, Canadian and American
The Open Collections website will be unavailable July 27 from 2100-2200 PST ahead of planned usability and performance enhancements on July 28. More information here.
UBC Theses and Dissertations
Unemployment and health in context and comparison : a study of Canada, Germany, and the United States… McLeod, Christopher Bruce 2009
Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.
- 24-ubc_2010_spring_mcleod_christopher.pdf [ 3.58MB ]
- JSON: 24-1.0070886.json
- JSON-LD: 24-1.0070886-ld.json
- RDF/XML (Pretty): 24-1.0070886-rdf.xml
- RDF/JSON: 24-1.0070886-rdf.json
- Turtle: 24-1.0070886-turtle.txt
- N-Triples: 24-1.0070886-rdf-ntriples.txt
- Original Record: 24-1.0070886-source.json
- Full Text